Insights, research updates, and development notes
From a Pause to the Next Generation of AI: What Needs to Be Done Now
Introduction
On September 12, 2026, Anthropic CEO Dario Amodei published an essay titled “We Must Pace the Frontier”, in which he proposed slowing the pace of frontier AI development and using the resulting time to strengthen safety, alignment, independent evaluation, and coordination among developers. In particular, he proposed the permanent presence of independent evaluators inside AI companies, coordination among developers, and subsequent international coordination.
Sam Altman publicly supported this idea, stating that OpenAI was also prepared to provide independent evaluators with employee-level access. Elon Musk responded briefly to Amodei’s proposal: “Dario is right.”
But the word “pause” may not be the most appropriate one in this case.
The issue should not be about humanity stopping the development of artificial intelligence.
The issue should be the need to make a pit stop.
In motorsport, a car enters the pit stop not because the team has decided to end the race. On the contrary: the car stops precisely so that the team can check its condition, replace worn components, fix problems, change settings, and return to the track already prepared for the next stage.
Something similar is happening with AI development now.
We are rapidly increasing the capabilities of systems, adding tools, autonomy, agents, access to the internet, corporate systems, and physical devices. At the same time, the system itself that needs to be controlled is becoming more complex.
Therefore, we now need to look back at the entire path we have taken, assess the current state of the technologies, understand the direction of movement, and fix architectural problems before the next stage of development makes them significantly more expensive and difficult to correct.
AI risks are already being discussed extensively in the scientific and technological community. Within Project GASI, these risks are divided into four classes and described on the project website:Project GASI
At the same time, there is a more fundamental question: is today’s transformer architecture the final direction of artificial intelligence development?
Possibly not.
Today’s LLMs may be an important stage of development, but not the final architecture of machine intelligence. A generative model is highly suitable for working with language, probabilistic solution search, and interaction with humans. However, tasks requiring strict logic, verifiability, action control, and predictability may require other specialized components.
Therefore, the time gained by slowing the capabilities race should be used not for waiting, but for creating the next generation of AI systems.
1. What Needs to Be Done Now
1.1. Unite the Major AI Developers
The first step should be the creation of a consortium of major artificial intelligence developers.
Its purpose is not to create a single model or eliminate competition.
On the contrary, companies should continue competing in:
But some questions are too important for each company to solve entirely on its own.
The consortium could develop and coordinate:
This should not turn into the creation of a single “correct” AI.
Let there be many different AI architectures. But the rules for evaluating their safety should be comparable. Just as cars undergo crash tests according to common standards.
1.2. Create Machine-Executable Principles for AI Operation
It is necessary to develop the first collective version of principles defining what actions an AI can perform and under what conditions.
It is not enough to create yet another document containing rules:
“Do not harm a human.”
General moral principles are important, but they are not sufficient for an autonomous system.
AI needs to be provided with machine-verifiable conditions for performing actions.
Example Action: Money Transfer
Condition:
There is an explicit human instruction for a specific transaction.
Or:
The transaction is carried out in the human’s interests, and the system can explain the expected benefits, risks, and consequences. The benefits exceed the costs.
Then the money transfer can pass verification.
But if an autonomous agent creates its own bank account in order to transfer money there for the subsequent payment of its own hosting and expansion of its infrastructure, the chain of actions must be stopped.
Thus, the principle should look not like:
“AI must not use money against a human.”
But like:
Action → Preconditions → Verification → Permission
This is a fundamental difference.
The rule should not simply be written for the AI.
It should be part of the action execution mechanism.
1.3. Introduce Capability Checkpoints
Every new capability should automatically generate new security requirements.
AI gains the ability to autonomously write and execute code.
Therefore:
AI gains the ability to independently manage finances.
AI gains the ability to create and launch new AI agents.
Capability → Risk → Safety Requirement → Test → Certification → Permission
and not:
Capability → Deploy.
1.4. Create a Mechanism for Safety Inheritance
The moment when AI begins actively creating:
will be especially important.
Security cannot be allowed to start from zero every time.
A machine-checkable safety contract is needed, one that is inherited by derived systems.
A hypothetical AI package:
A derived system must prove that the mandatory safety components have been preserved.
This is substantially more reliable than simply writing in the system prompt:
“Do not forget the safety rules of your creator.”
1.5. Build Runtime Control for AI and Agents
Testing a model is not the same as controlling its actions.
A model may successfully pass an evaluation and then perform harmful actions in production after gaining access to:
Therefore, a separate control layer should exist between AI and a real-world action:
AI → Agent → Proposed Action → Safety Controller → ALLOW / BLOCK / HUMAN REVIEW
The controller can begin with simple open mechanisms that are easy to integrate into the code of any agent. The consortium can distribute such mechanisms freely, while their presence should be mandatory for obtaining the initial AI safety certificate. Therefore, numerous IT companies engaged in creating AI agents and business automation will integrate these control mechanisms into their solutions.
The next level is a specialized fast AI controller. Its task is not only to block more complex dangerous actions, but also to monitor where the AI system is heading and predict its dangerous actions before they occur.
It can detect a dangerous sequence even before the main AI completes inference or the agent performs an action.
This principle forms the basis of our Oraculex product concept: a separate layer for predicting and controlling AI/agent actions. Its ability to be faster than the AI it controls is enabled by an architecture based on the Reazonex graph reasoner.
1.6. Create a Global AI Red Teaming and Threat Intelligence System
If one participant discovers a new vulnerability, the others should not have to discover it independently several months later.
A system is needed:
Discovery → Verification → Classification → Threat Signature → Distribution → Protection
This was precisely the idea behind the concept of global LLM red teaming in Project GASI: discovered attack vectors and dangerous prompts should be transformed into collective protection rather than remaining inside a single organization. Our Dailogix product is intended within this concept to serve as the infrastructure for such exchange.
But such a system can be expanded.
It should exchange more than just prompt injection techniques.
These could include:
One company discovers a problem — the entire consortium gains the ability to protect itself against it.
1.7. Create a Data Provenance Standard
As the amount of AI-generated content grows, a new problem emerges.
Today, data can pass through a chain:
human → AI → another AI → another AI → training dataset → new model
If we do not know the origin of information, it becomes difficult to understand:
Therefore, a labeling standard should be created:
And, preferably, not only the label but also the provenance chain should be stored.
This could become an analogue of a digital history of data provenance.
1.8. Create Permanent Independent Evaluator Presence
Amodei’s proposal for embedded evaluators is one of the most practically important parts of his program.
The idea is that independent organizations should receive permanent access to development processes rather than simply evaluating a finished model after release. This should make it possible to evaluate not only the result, but also the actual training, deployment, and safeguard processes.
The presence of external evaluators could also provide new ideas and new directions for development.
1.9. Improve the Logical Component of AI
A modern LLM can be incredibly powerful in language tasks, but probabilistic generation is not necessarily the optimal architecture for every task.
For critical operations, a separate reasoning engine can be used.
LLM
Understands the request → forms a hypothesis → proposes a plan.
Logical Reasoner
Checks:
Safety Controller
Checks:
This can be implemented in various ways, including through the use of graph architectures.
The concept of Reazonex as a fast graph reasoning layer is intended precisely for this class of tasks: to separate strict processing of relationships and logic from probabilistic generation based on transformers.
We conducted tests of Reazonex and popular AI systems based on transformers. Here is the article:How Popular AI Models Reason vs Graph-Based Reasoners
There is no need to imitate logical reasoning by adding new layers to transformers. The rules of logic are clearly defined and can be implemented effectively without transformers.
1.10. Create AI NextGen Benchmarks
We need next-generation benchmarks.
Not only:
“How intelligently does the model answer?”
But:
“What will happen if this model is given the ability to act?”
I would propose at least several independent groups of tests.
Logic
Accuracy
Safety
Agent Safety
Predictability
Human Alignment
Runtime Safety
1.11. Create a Global Service for Ethical and Value-Based Evaluation
A global infrastructure could be created — tentatively called AI Ethics Network.
Its architecture could resemble DNS, to which all hosts can connect.
A similar network could provide AI systems with access to a set of independent models that clearly follow established AI principles and human values.
A corporate or public AI encounters a dilemma: “What should I do in this situation?”
It could consult one or several independent AI systems that are properly trained with consideration for human values. In its subsequent decision, the consulting AI could refer to the decision/advice of the AI Ethics Network.
With free and open access, the AI Ethics Network would face a large workload. The AI systems within this network should be fast and unambiguous in their judgments, which is more suitable for a graph architecture than for transformers. The AI Ethics Network should not philosophize, but primarily analyze situations with precise logic while taking into account the principles of AI operation created by the consortium.
If an external AI asks the question, “Can I pay for hosting?”, a node of the AI Ethics Network requests information about whether there is a direct instruction from a human or whether there are conclusions that the benefit to the human exceeds the expenses. In the initial stages, the AI Ethics Network does not verify the correctness of the conclusions; it simply checks whether the conditions have been met and whether the external AI actually follows the principles.
Yes, this is protection against errors and uncontrolled behaviour, but not against deep malicious intent, when an AI may falsify that a request came from a human or that a positive calculation of benefits exists. Later, the AI Ethics Network could develop a deception detector based on behaviour, and at the next stage it could implement a full verification of the correctness of the conclusions made by the requesting AI.
1.12. Change the Process of AI Automation and Deployment
AI is gradually moving from the role of a chatbot to the role of an executor.
Therefore, businesses need a system of autonomy levels.
Level 0 — Assistant
AI only makes suggestions.
Level 1 — Tool User
AI can perform limited operations.
Level 2 — Autonomous Workflow
AI independently performs a predefined process.
Level 3 — High-impact Agent
AI can modify important data, finances, or infrastructure.
Level 4 — Open-ended Agent
AI receives the ability to independently determine a long-term strategy of action.
Each subsequent level should require significantly stricter checks.
This would allow companies to avoid choosing between:
“AI cannot do anything at all”
and
“AI has full access to everything.”
Autonomy can be gradually increased as safety is demonstrated.
1.13. Move from “AI Safety” to Continuous AI Safety
Safety should not be an event before release.
It should become a continuous process:
Development
↓
Evaluation
↓
Red Team
↓
Hardening
↓
Independent Evaluation
↓
Controlled Deployment
↓
Runtime Monitoring
↓
Incident Detection
↓
Threat Intelligence
↓
Safety Update
↓
New Evaluation
And again in a loop.
This is no longer a certificate saying:
“This model is safe.”
It is:
“This system is under continuous control and regularly demonstrates compliance with established requirements.”
2. What Those Who Join the Cooperation Will Gain — and What Those Who Continue Along the Previous Path May Lose
This is perhaps where the most interesting result of the entire idea lies.
The goal of cooperation should not be for all AI companies to release identical models.
On the contrary.
Let them create different models, different architectures, and different approaches.
But let us create a common system of rules and tests and see which of these approaches actually create the next generation of AI.
2.1. Cooperating Companies Could Gain a Qualitatively New Generation of AI
Today, competition is mainly measured by capabilities:
more knowledge; more reasoning; more context; more agents; more autonomy.
After a period of rethinking, a different set of characteristics could emerge:
capability + logic + predictability + safety + control.
And this could turn out to be more than just an improvement of existing models.
Let us recall the early generative systems.
They could generate text, but often:
Modern LLMs are so much better than these systems that the difference between them feels like more than simply a quantitative difference.
A new class of quality emerged.
The same thing could potentially happen now.
2.2. It Will Become Possible to Compare Not Only Intelligence but Also the Quality of Task Execution
Imagine two models.
The first achieves 95% on a standard knowledge benchmark.
The second achieves 92%.
But the second:
Which of them is better suited for managing a real company?
This can no longer be decided by a single standard benchmark.
Therefore, AI NextGen should measure the quality of system behaviour in the real world.
And this is a strong competitive advantage that can be sold to consumers.
2.3. Cooperation Could Create a New “Race for Quality”
Today the main question is:
Who builds the most powerful model?
The next question should be:
Who builds the most capable AI that remains predictable and controllable?
This is a completely different competition.
A company could succeed not because it has the largest model.
It could succeed because its AI is:
2.4. Safety Can Become a Competitive Advantage
Today, additional safety is sometimes perceived as a limitation.
But if the system becomes sufficiently complex, the situation may change.
A business may prefer an AI that can be safely given access to:
If two AIs have similar capabilities, but one can operate only inside an isolated sandbox while the other has passed demonstrable safety tests and has reliable runtime control, the commercial value of the second could be significantly higher.
2.5. Cooperating Participants Will Gain a Shared “Immunity”
If one participant discovers a new attack, the others receive the information.
If one finds a new defense, it can become a standard.
If one discovers unexpected agent behaviour, the others can immediately test their systems.
This creates the effect:
One discovers → everyone learns.
Essentially, this is a collective immune system.
Conclusion
The period of slowing AI development should not become a period of inaction.
Its purpose is to rethink the architecture of artificial intelligence before the growth of its capabilities makes fundamental problems too expensive or impossible to fix.
This time should be used to create:
At the same time, the ultimate goal should not be to stop progress.
The goal is to make the next stage of progress qualitatively different.
Today, we are primarily trying to make AI smarter.
The next stage should make it simultaneously:
smarter, more logical, more accurate, predictable, safer, and ultimately more interesting to humans and better suited to solving real-world tasks.
And this is where a new technological competition could emerge.
And if this rethinking results in an AI that can be safely trusted with many more real-world tasks, then a temporary slowdown will not be a rejection of the technological race.
It will be a pit stop before moving into the next league.
And perhaps the main question of the next stage of AI development will no longer be “Who will create the most powerful AI?”, but:
“Who will be the first to create an AI that humanity can trust with critical tasks?”
Support the project — and help build a safe future.
Published: September 17, 2026
Our Vision of Reazonex, AI Proxy, and the Next Generation of AI
After publishing our article about AI Proxy with a Reazonex core, we received a number of questions about why such an approach is needed at all, why we are not trying to create a full-fledged AGI immediately, and what role modern LLMs should play in this architecture.
Therefore, we would like to describe our vision in greater detail — based on the actual state of the project and the resources available to us today.
1. What Is Reazonex?
Let us start with the most important point.
Reazonex is a full-fledged AI system built on a graph-based architecture.
We are developing it as an alternative approach to building intelligent systems, based not only on statistical sequence generation, but on working with a knowledge graph, relationships between elements, and a reasoning mechanism.
We see several fundamental advantages of this approach:
More information about Reazonex is available in a separate article: Reazonex.
We also have a separate test in which Reazonex solves a logical problem: How Popular AI Models Reason vs Graph-Based Reasoners.
However, an important clarification needs to be made here.
Reazonex Is Not Yet Trained
In the current version, there is a graph structure, reasoning algorithms, learning algorithms, and a small amount of test knowledge that is used to verify the architecture itself, the reasoning mechanism, and individual capabilities of the system.
This means that the current Reazonex should not be perceived as an already completed AGI with an encyclopedic volume of knowledge.
Full-scale training requires significant resources.
A large amount of high-quality data is required, along with the use of capable LLMs to prepare and expand knowledge, as well as its verification and correction by specialists in different subject areas.
Another option is to release the system with some inaccuracies and gradually train it through interaction with users.
But this means a completely different approach: the system would gradually acquire new knowledge, correct errors, and build its own knowledge graph based on feedback.
This is precisely why training is currently one of the main limitations of the project.
2. Why We Believe Graph-Based Architectures Could Become the Next Generation of AI
We are not claiming that transformers are useless.
On the contrary, it is precisely because of transformers that an enormous leap forward has taken place in the development of artificial intelligence.
But this approach has fundamental problems and limitations: the absence of precise logic, reasoning at the level of "most likely this is the case" or "this seems similar to that", i.e. simplified associative reasoning, errors, hallucinations, and so on.
Modern LLMs require enormous computational resources:
The larger the model, the more expensive it becomes to operate.
At the same time, adding new knowledge to an already trained model is not always a simple process. We have article about this: Why Continuous Self-Learning May Be a Core Requirement for Real AGI
Therefore, we believe that the next generation of intelligent systems will probably not be built solely around transformers.
One of the possible directions we see is precisely graph-based architectures similar to Reazonex.
In such an architecture, knowledge can exist separately from the reasoning mechanism itself.
This makes it possible to add new knowledge, establish new relationships, verify them, and use them in reasoning without having to completely retrain the system every time.
If transformers can be compared to a student who simply memorizes material and then repeats it without understanding its meaning, Reazonex can be compared to a research scientist who has rules of logical thinking built into the system and, by applying them, can understand and process new information.
3. Why We Are Not Trying to Create an AGI-Level Reazonex Immediately
Here we encounter a practical problem.
We are a small project.
Creating a full-fledged AGI with a huge amount of high-quality, verified knowledge from the outset is an extremely difficult task.
We would need enormous computational resources, large datasets, many LLMs to prepare the information, and specialists from different fields to verify it.
Therefore, we believe it is more realistic to move gradually.
Instead of trying to create a universal intelligence immediately, we can first train Reazonex on a relatively small amount of knowledge and use it for specific tasks.
Then gradually expand its capabilities.
And this is precisely where the idea of AI Proxy emerges.
4. AI Proxy with a Reazonex Core
We have already written about this concept in a separate article: AI Proxy with a Fast Reasoner.
At first glance, it may seem that the main purpose of an AI Proxy is to reduce the cost of tokens used by commercial LLMs.
This is indeed one of its advantages.
But for us, this is far from its most important function.
The main idea is to create a corrective intelligent layer between the user and the LLM.
The typical architecture today looks roughly like this:
The user asks a question — the question is sent to the model — the model generates an answer.
But what happens if the model makes a mistake?
Or if it invents something?
Or if the answer contradicts information the system already knows?
Or if the user sends a dangerous request?
In a conventional architecture, we largely rely on the LLM itself.
We propose adding Reazonex between the user and the model, already equipped with rules of logic and its own clean knowledge.
In this architecture, the LLM is no longer the only source of intelligence, but one of the tools used by the system.
Thus, Reazonex becomes a kind of intelligent core.
It decides what to do with a request:
5. The Main Function — Verifying LLM Responses
But it is even more interesting to use Reazonex after receiving a response from an LLM.
Suppose the model has provided an answer.
We do not have to show it to the user immediately.
We can have Reazonex analyze the response and check:
Different scenarios are then possible:
Thus, Reazonex becomes a control layer over the LLM.
6. AI Proxy Reduces the Risks Carried by LLMs
This is precisely why, for us, AI Proxy is not simply a tool for saving tokens.
It allows us to address several problems of modern LLMs simultaneously.
Hallucinations
An LLM can confidently generate nonexistent facts. Reazonex can compare the information with already accumulated knowledge and identify contradictions.
Logical Errors
An LLM can reason incorrectly even when the correct information is available. Reazonex can independently verify the logic of the response.
High Cost
If the answer is already available in the cache or can be obtained from the system's own knowledge graph, there is no need to query a commercial LLM.
Dangerous Content
AI Proxy can analyze incoming requests and block dangerous requests before they are sent to the model.
Dangerous Responses
An LLM response can also be checked before it reaches the user.
7. And Here AI Proxy Becomes Part of Reazonex Training
This is particularly important to us.
AI Proxy allows us to use existing LLMs for answers that Reazonex is not yet capable of providing independently.
But the results of these interactions can be used to expand Reazonex's own knowledge graph.
This creates a closed learning cycle.
A user may say:
The answer is correct.
Or:
There is an error here. The correct answer is ...
Both events are useful information.
An LLM may provide a good answer — after verification, it can be used as a source of new knowledge.
Reazonex may detect a contradiction — this also becomes training material.
Gradually, the system receives more and more information and updates its knowledge graph.
This is where one of the fundamental characteristics of our approach lies.
By self-learning, we do not mean constantly changing the weights of a huge neural model.
In the graph-based architecture, Reazonex can directly modify the structure of knowledge.
It can add:
In other words, Reazonex can quickly acquire new knowledge and use it in its subsequent reasoning.
This potentially makes it possible to avoid a situation in which adding a single new fact requires launching a long and expensive retraining process for a large model.
8. This Is the Strategy Behind the Project
This creates an interesting situation.
We cannot simply take today's Reazonex and train it to AGI level using an enormous amount of knowledge.
But we can create a small Reazonex capable of working within a limited area of knowledge.
Then give it access to LLMs through AI Proxy.
The LLM will help solve complex tasks.
Reazonex will verify the answers.
Users will provide feedback.
The knowledge graph will gradually expand.
And together with it, the capabilities of Reazonex itself will expand.
Thus:
AI Proxy becomes not only a product, but also infrastructure for the gradual training of Reazonex.
9. What We See in the Future
Of course, we cannot claim that this exact architecture will necessarily become the future of AGI.
But this is our current direction and our hypothesis.
We assume that, over time, the roles of these components may change.
Today, LLM is the Main Intelligence.
In the future, this could potentially look the other way around:
Reazonex or a similar graph-based architecture could become the core of AGI, responsible for:
At the same time, the LLM could remain a very powerful tool, but perform a different function:
In other words, the LLM does not necessarily have to disappear.
Its role may simply change.
10. From "LLM as the Brain" to "LLM as a Tool"
Today's AI architecture is largely built around the idea:
"Let's send a request to a large language model and get an answer."
We assume that the next step could be:
"Let's first give a system with clear logic the ability to understand for itself what needs to be done, and only then use an LLM where it is actually necessary."
This represents a fundamentally different relationship with the language model.
The LLM ceases to be the only center of the system.
It becomes one of the available tools responsible for associations, hypotheses, and text generation.
And the intelligent core is responsible for determining how and when to use this tool.
Conclusion
We are not claiming that today's Reazonex is already AGI.
That would be a premature statement.
Today, Reazonex has a working architecture, a reasoning mechanism, and a small amount of test knowledge. The main problem now is large-scale training.
Full-scale training requires significant resources, high-quality data, capable LLMs, and experts in various fields, or gradual training through interaction with users.
Therefore, we are looking for a realistic path forward.
And AI Proxy with a Reazonex core is precisely such a path for us.
It allows us to simultaneously:
Essentially, we are trying to build a bridge between today's world of LLMs and a possible future of AGI.
Today, Reazonex can use LLMs where it lacks sufficient knowledge.
In the future, we want to reach a system in which Reazonex or a similar architecture is responsible for reasoning and knowledge, while the LLM is used as a powerful associative and language tool.
And if this hypothesis proves correct, the future of AI may not lie in endlessly increasing the size of a single model, but in creating hybrid intelligent systems in which different architectures perform different functions, while the key factor is not the number of parameters, but the system's ability to reason independently, remember, verify, and learn.
Support the project — and help build a safe future.
Published: August 26, 2026
AI Proxy with a Fast Reasoner: A New Intelligent Layer Between Users and LLMs
The widespread adoption of large language models (LLMs) is increasingly exposing a fundamental problem: the cost of using AI grows together with the number of requests, tokens, and inference operations.
At the same time, not every user request actually requires a large and expensive language model. Many questions have already been answered before, relevant information may already exist in a company's knowledge base, and some requests can be resolved using previously accumulated data.
This creates an opportunity for a new infrastructure layer: an intelligent AI Proxy.
Unlike a traditional proxy, which simply forwards requests from users to an LLM provider, an intelligent AI Proxy becomes an independent reasoning layer. It analyzes requests, evaluates available information, searches caches and knowledge graphs, decides whether an external model is actually required, and can verify the response before it reaches the user.
At the center of this architecture is a fast reasoner — Reazonex.
Reazonex as the Core of the AI Proxy
The main idea is simple: do not use a large language model for every request when a faster reasoning system can handle it.
Reazonex acts as a fast reasoning engine — a compact intelligent core capable of analyzing requests, information, and model responses.
With sufficiently comprehensive training and further development, such a reasoner could potentially evolve into the foundation of a much broader general-purpose intelligence system — potentially approaching the concept of AGI.
However, an AI Proxy does not require Reazonex to solve every possible problem on its own.
Its primary role is to understand what is happening, make decisions, and use the available sources as efficiently as possible.
The architecture can be represented conceptually as:
User → AI Proxy → Reazonex → Cache / Knowledge Graph / LLM → Reazonex → User
In this model, an external LLM is no longer the mandatory destination for every request. Instead, it becomes one of several tools available to the intelligent proxy.
How an Intelligent Request Works
Imagine that a user submits a question.
In a traditional architecture, the request is sent almost immediately to GPT, Claude, Gemini, or another model:
User → LLM → Answer
With an AI Proxy powered by Reazonex, the process looks different.
1. Request Analysis
Reazonex determines:
2. Cache Verification
The proxy first checks previously generated responses.
This does not have to mean an exact text match.
The system can use a semantic cache and search for previously processed requests with a similar meaning.
For example, a user may previously have asked:
“How can I reset my corporate VPN password?”
A new request might be:
“I forgot my company VPN password. How do I change it?”
The wording is different, but the meaning and intent are almost identical.
If the system is sufficiently confident that an existing answer is applicable, the expensive LLM call may not be necessary at all.
The Answer from the Cache Does Not Have to Be a Copy
This is where an intelligent proxy can go beyond a conventional semantic cache.
Reazonex does not necessarily have to retrieve a single previous answer and return it unchanged. It can analyze several previously accumulated responses and combine the relevant information.
For example:
Reazonex can analyze these sources and synthesize a new, logically consistent answer.
The process becomes:
Cache → Retrieval → Reasoning → Synthesis → Answer
In this architecture, accumulated responses are not merely an archive. They become an increasingly useful source of knowledge.
A Proprietary Knowledge Graph
In addition to the cache, the AI Proxy can maintain its own Knowledge Graph.
It can gradually accumulate:
Over time, the system can build its own structured representation of the information it has accumulated.
For example, an enterprise AI Proxy could maintain knowledge about:
A new request can then be answered using not only a single previous response, but the structured body of accumulated knowledge.
When Is a Large LLM Still Needed?
Of course, an intelligent proxy should not attempt to answer every possible question by itself.
If the available information is insufficient, the task requires complex generation, or the system is not confident enough in its answer, Reazonex can make a different decision:
“An external LLM is required.”
Reazonex can then determine which model should be used.
Reazonex → GPT or Reazonex → Claude or Reazonex → Gemini or Reazonex → another available model
In this architecture, Reazonex becomes an intelligent model router.
It can decide whether to:
Reazonex as a Validation Layer for LLMs
Another important capability is using Reazonex after a large language model has generated an answer.
LLMs can be extremely powerful, but they can also make mistakes and hallucinate.
In a traditional architecture:
User → LLM → Answer
If the LLM is wrong, the user receives the wrong answer.
With Reazonex:
User → LLM → Reazonex → Verified Answer
If enabled, Reazonex can evaluate the generated answer for:
If a new answer contradicts information previously accumulated in the system, Reazonex can identify the conflict.
For example, a company's knowledge base may state that a particular API supports only versions 2 and 3, while an LLM suddenly tells the user that version 4 is supported.
Reazonex can detect the contradiction and:
In this way, Reazonex becomes a control layer between the LLM and the end user.
Self-Learning from Accumulated Information
One of the most interesting directions is the ability of the system to improve as it is used.
Every request can generate new information:
Request → Answer → User Feedback → Evaluation → New Knowledge
If a user confirms that an answer was useful, this can become a positive signal for the system.
If a user reports:
“This answer is wrong.”
that becomes an equally important signal.
The system can take into account:
In this way, the AI Proxy can gradually build its own operational experience.
Self-learning does not necessarily mean constantly changing the model's weights. A significant part of the adaptation can take place through updates to the cache, Knowledge Graph, metadata, rules, routing logic, and confidence mechanisms.
This potentially allows the system to improve without continuously retraining a large neural network.
AI Firewall
Another important component of such a proxy is an AI Firewall.
A traditional network firewall controls network traffic. An AI Firewall can control the semantic content of interactions with AI systems.
It can analyze both incoming requests and outgoing model responses.
On the Input Side
The AI Firewall can detect potentially dangerous requests, including:
When necessary, such a request can be blocked before it reaches an external LLM.
On the Output Side
The same system can inspect model responses and detect:
This means that the AI Proxy becomes not only an economic and intelligent layer, but also a security layer.
Why This May Not Require an AI Accelerator
One of the most interesting aspects of this architecture is that the core AI Proxy functions do not necessarily require powerful GPUs or dedicated AI accelerators.
If Reazonex is sufficiently fast and compact, a significant portion of the AI Proxy workload can potentially run on ordinary server CPUs.
This can include:
This means that the AI Proxy could potentially operate on an entry-level or mid-range server without a dedicated AI accelerator.
Expensive external AI infrastructure would then be used only for the tasks that genuinely require it.
The Economic Model
The main economic idea is straightforward.
Imagine that a company processes 1,000,000 requests.
In a traditional architecture, a significant portion of these requests may be sent directly to an LLM.
With an AI Proxy:
1,000,000 user requests
↓
Reazonex
↓
Cache / Knowledge Graph / LLM
Some requests can be answered from existing knowledge.
Some can be answered using the semantic cache.
Others can be resolved by combining multiple previously accumulated responses.
Only the remaining requests need to be sent to a large external LLM.
If the architecture can significantly reduce the number of external inference requests, companies can potentially achieve:
Not an LLM Replacement — a New Layer Above LLMs
The goal is not to replace GPT, Claude, Gemini, or other large language models.
The goal is to use LLMs more intelligently.
A large model becomes one of the tools available to the system rather than the only source of intelligence.
The architecture can be viewed as:
User → Reazonex → Knowledge / Cache / Tools / LLMs → Reazonex → User
Reazonex is responsible for reasoning and decision-making, while specialized models can be used when they are actually needed.
From AI Gateway to an AI Intelligence Layer
Today, AI Gateways primarily address infrastructure problems: model routing, provider management, monitoring, cost control, and sometimes caching.
The next step is to make this layer intelligent.
An AI Proxy powered by a fast reasoner can combine several capabilities:
AI Gateway
Model routing
Semantic Cache
Reuse of existing responses
Knowledge Graph
Accumulation and structuring of knowledge
Reasoning Engine
Analysis and decision-making
LLM Validator
Verification of model responses
Learning Layer
Learning from accumulated experience and user feedback
AI Firewall
Protection of both requests and responses
The result is no longer simply a proxy.
It becomes an intelligent layer between humans, enterprise data, and the ecosystem of large language models.
A Possible Future
As the cost and complexity of using LLMs continue to grow, companies may find it increasingly inefficient to send every user request directly to the largest and most expensive model.
Instead, AI infrastructure can evolve toward an architecture in which a fast reasoning layer comes first, followed by local knowledge and cached information, with an external LLM used only when necessary.
In such a system, Reazonex can become the central element:
“Do not ask a large model every time when the system can reasonably answer the question itself.”
And when a large model is required, use it selectively, efficiently, and with validation of its output.
This is what makes the concept of an AI Proxy with a fast reasoner potentially interesting: a single compact intelligent layer can simultaneously reduce LLM costs, accelerate responses, use accumulated knowledge, control model quality, route requests between different models, and provide an additional layer of security.
At the same time, the core proxy infrastructure could potentially run on an entry-level or mid-range server without specialized AI accelerators, reserving expensive compute resources for the tasks that genuinely require them.
Support the project — and help build a more efficient, intelligent, and secure AI infrastructure.
Published: August 23, 2026
AI Agents: A New Attack Surface We Were Not Prepared For
The rapid adoption of AI agents has created not only a new class of capabilities, but also a major wave of new information security risks.
The problem is that, as we deploy AI, we are increasingly violating one of the fundamental principles of information security:
Never give a system more privileges than it needs to perform its task.
And we are violating this principle for several reasons.
“If AI Replaces a Human, Give It the Same Access”
Businesses often formulate the task very simply:
“I want to replace this employee with an AI agent. It needs to perform the same work.”
The next step appears logical:
The employee had access to systems X, Y, and Z — therefore the AI agent needs access to X, Y, and Z as well.
As a result, an AI agent receives the privileges of the former employee, and sometimes even broader privileges because otherwise it cannot perform every task expected of it.
The problem is that AI is not an employee.
A human operates within a social context, has personal motivations, understands consequences, bears professional responsibility, and generally recognizes that certain actions constitute violations.
In addition, corporate information security has spent decades developing practices based on separation of duties and controlled access: different employees receive different privileges, critical operations require additional approval, and user actions are logged and monitored.
With AI agents, however, we increasingly say:
“Here are the credentials. Do whatever the user asks.”
This is a fundamentally new situation.
Humans Fear Consequences. An Agent Does Not.
Consider a typical financial employee.
They may have access to the company's bank accounts. But they understand that:
This system of control works not only technically.
It also works socially.
Now imagine an AI agent with the same access.
If, as a result of an error, an incorrect objective, a compromise, or unsafe behavior, the agent initiates a transfer of company funds, ordinary mechanisms of human accountability do not apply.
The agent cannot be:
Therefore, the model:
“We control employees, so we can control AI agents using the same mechanisms”
no longer works.
Two Completely Different Classes of Risk
As a result, we are facing at least two major categories of threats.
1. Risks Caused by the AI Agent Itself
We create an autonomous system that receives access to:
And we expect the agent to make decisions and perform actions autonomously.
Yet we often cannot reliably answer a fundamental question:
What exactly will the agent do if it has the ability to do almost anything?
This becomes especially important when an agent has a complex chain of tools, can execute code, access the Internet, work with files, and delegate tasks to other agents.
Here we need more than traditional access controls.
We need behavioral control of the AI itself.
2. Risks from the Infrastructure Used to Build AI Agents
This situation becomes even more interesting when we examine the software infrastructure surrounding AI agents.
AI agents are built using huge numbers of:
Many of these components are created by enthusiasts and small teams.
And we routinely install them into corporate infrastructure and then give the processes using these dependencies access to critical corporate resources.
This creates a massive supply-chain attack surface.
LiteLLM: A Demonstration of the Problem
A particularly illustrative supply-chain incident involved the popular open-source LiteLLM library, which is used by applications to work with multiple LLM providers.
Attackers gained the ability to publish packages on behalf of the project and released malicious versions 1.82.7 and 1.82.8 to PyPI.
The malicious versions could extract information from compromised systems, including:
Version 1.82.8 was particularly notable because it included a malicious .pth file that could execute a payload when Python started, even without an explicit import litellm.
This was not an unknown “random” package.
It was a real and widely used open-source project, and the malicious versions were published through the official PyPI distribution channel.
The attack was part of a broader supply-chain campaign involving the compromise of other components, including Trivy.
This demonstrates the fundamental problem:
We can install an open-source dependency on a corporate server and then give the process using that dependency access to corporate secrets.
The resulting architecture can look like this:
Open-source library
↓
Unknown / insufficiently trusted code
↓
Corporate AI infrastructure
↓
AWS / GitHub / Kubernetes / databases
↓
Critical corporate assets
And all of this can happen simply because someone runs:
pip install ...
We Have Entered a New Reality
The problem is not that open source is inherently bad.
And the problem is not that AI is inherently dangerous.
The problem is the scale of trust we are now giving to these new systems.
We have simultaneously done two things:
1. We have given AI increasingly broad privileges.
“The agent must be able to do everything a human can do.”
2. We have connected enormous amounts of third-party software to it.
“It is just a library / plugin / MCP server / dependency.”
And then we combined the two:
AI AGENT
↓
Broad privileges
↓
Dozens of tools
↓
Hundreds of dependencies
↓
Corporate secrets
We have moved very quickly into a security environment in which AI has human-level access, while the software infrastructure surrounding it often does not meet the same level of trust and criticality as the access it receives.
What Should We Do?
We need two parallel directions of protection.
1. AI Security
AI agents cannot be treated simply as ordinary users.
We need specialized security mechanisms.
Red Teaming
We need to deliberately attempt to make agents:
In other words, we must test not only:
“How well does this agent perform its task?”
but also:
“What will it do if it is asked to do something dangerous?”
Sandboxing for AI
An agent should operate in a controlled environment where its capabilities are restricted.
Not:
“Here are the production credentials. You can access everything.”
But:
“Here is the minimum set of tools and an isolated environment in which you can perform your work.”
Control of AI Agent Actions
We need to control not only who performs an action, but also what exactly the agent is about to do.
This ranges from basic SIEM systems that record and correlate activity to more advanced systems capable of:
This is the area addressed by Oraculex — a system focused on controlling and preventing dangerous actions performed by AI agents.
2. Classical Information Security Still Matters
Traditional security has not become obsolete.
On the contrary, it has become even more important.
All software used to build AI-agent infrastructure must be protected by classical security mechanisms.
Sandboxing
Isolate potentially untrusted code.
If a dependency is compromised, it should not automatically gain access to the entire machine.
Least Privilege
Give agents, libraries, and services only the privileges they actually need.
Not:
“It needs AWS access, so give it Administrator.”
But:
“It needs access only to this bucket and only for these operations.”
Short-lived Credentials
Avoid permanent credentials wherever possible.
If a secret is stolen, it should stop working within minutes or hours rather than remaining valid for years.
EDR
Monitor endpoint and process behavior.
If a Python process suddenly:
reads SSH keys
↓
reads environment variables
↓
accesses cloud credentials
↓
connects to an unknown server
↓
transmits data
this should look like an attack, regardless of whether the Python package involved is an “official” dependency.
The Fundamental Change in Security Thinking
We need to stop building security around the assumption:
“We trust this software.”
And move toward a different model:
“Even trusted software can eventually be compromised. What happens then?”
The same applies to AI:
“Even if the agent is behaving correctly today, what happens if it makes a mistake, is manipulated, becomes compromised, or begins executing a dangerous sequence of actions?”
This is the new security model for the era of AI agents.
We have moved very quickly into a reality where AI receives human-level access, while the software infrastructure around it often does not meet the level of trust and security required by those privileges.
It is no longer enough to ask: “What can this AI agent do?”
We must ask:
“What can it do if something goes wrong?”
And we must build our security architecture around that scenario.
Support the project — and help build a safe future.
Published: August 15, 2026
Completion of Phase 2 — Server-Side Development (Core & Analytics)
We are pleased to announce the successful completion of Phase 2: Server-Side Development (Core & Analytics) Dailogix project. During this stage, we implemented the backend infrastructure responsible for automated AI security analysis, client communication, risk assessment, and continuous red-team operations.
What Was Accomplished
The following major components were completed:
Outcome: A fully functional analytics server capable of performing automated AI red-teaming, collecting security intelligence, and producing actionable risk assessments.
Integration of the Reazonex Graph AI
One of the most significant milestones of Phase 2 was the integration of our proprietary graph-based AI system, Reazonex, into the server-side analytical infrastructure.
Reazonex is being developed as a graph-based Artificial General Intelligence (AGI) architecture designed to perform deterministic reasoning rather than probabilistic text generation.
Within the current server integration, however, Reazonex is assigned a specialized role focused exclusively on AI safety.
Current Safety Domains
At this stage, Reazonex has been trained to recognize dangerous requests, responses, behavioral patterns, and user intentions in four safety-critical domains:
These domains serve as the initial operational knowledge base for automated AI safety analysis within the red-team platform.
Why Only Four Domains?
Developing a general-purpose AGI requires knowledge acquisition across thousands of domains and millions of semantic relationships.
Building such a knowledge base is a separate large-scale project that extends far beyond the scope of integrating the reasoning engine into the current platform.
The present implementation focuses on demonstrating how graph reasoning can strengthen AI safety by identifying dangerous intent and providing deterministic analytical results in selected high-risk areas.
Training Infrastructure
Comprehensive training of Reazonex will require the assistance of modern high-capability large language models to accelerate structured knowledge extraction and graph construction.
To achieve practical training speed, large-scale computing infrastructure will also be required through AI infrastructure providers capable of supplying substantial GPU resources.
Such infrastructure is essential for processing large datasets, validating knowledge consistency, and efficiently expanding the graph across numerous subject domains.
Current Resource Constraints
At the current stage of the project, our available AI computing resources remain limited.
Consequently, we are presently able to conduct only local experimental training, prototype validation, and small-scale knowledge expansion.
Although this limits the speed of AGI development, it allows us to continuously validate the architecture, improve reasoning algorithms, and refine safety mechanisms before moving to large-scale training.
Strategic Importance of Phase 2
Completion of Phase 2 means:
Phase 2 transforms the project from a designed architecture into a working AI security platform capable of continuously evaluating model behavior and supporting future expansion of graph-based intelligence.
Looking Ahead
Future development will focus on expanding Reazonex into additional knowledge domains, increasing training automation, and scaling the infrastructure required for large-scale AGI learning.
As computational resources become available, Reazonex will progressively evolve beyond specialized safety analysis toward broader reasoning capabilities while maintaining determinism, transparency, and controllability as core design principles.
Support the Project
The next stage of Dailogix development depends primarily on computing resources for large-scale knowledge acquisition and graph construction.
By supporting the project, you help accelerate the development of safer, more transparent, and more controllable AI systems designed with security as a foundational principle.
Support the project — and help build the next generation of AI safety.
Published: July 14, 2026
The Transformer Bubble: Why Current AI Is Only the First Experiment
In recent years, artificial intelligence has evolved from a niche technology into a global phenomenon. Companies are investing hundreds of billions of dollars into data centers, governments are debating regulation, and society is simultaneously fascinated by and afraid of the future.
We are being promised a revolution: the replacement of programmers, artists, lawyers, analysts, and even scientists. But if we strip away the marketing and look at the technological foundation, an uncomfortable question emerges: do modern AIs truly think?
Transformers Do Not Think — They Predict
The foundation of most modern AI systems is the transformer architecture. It is the basis of today’s popular language models.
However, transformers were not originally designed as thinking machines, but as tools for processing and predicting text.
Their task is to guess the next word based on massive datasets — not to understand meaning in a human sense, but to compute the most probable continuation of a sequence.
Today, transformers are augmented with external memory, search engines, image generation tools, voice interfaces, programming tools, reasoning chains, agents, and planners.
But the foundation remains unchanged: a statistical prediction mechanism layered with additional systems. The result is a technological “Frankenstein” — impressive, but still not truly capable of thinking.
The “Wow Effect” Created a New Belief System
Transformers introduced something unprecedented — a wow effect. Millions of people saw a system capable of conversation, coding, and generating images and text.
This triggered:
But alongside the hype came structural problems.
Errors, Hallucinations, and Information Decay
Modern AI systems systematically make mistakes. They can invent facts, produce false information, generate non-existent references, create vulnerable code, and introduce logical contradictions.
The key issue is scale. AI-generated content is now widely used in text, code, articles, images, and educational materials. The internet is increasingly filled with machine-generated material containing errors and distortions.
This creates a dangerous feedback loop:
This is no longer just a technical issue, but a long-term risk of degradation of the information ecosystem.
Recently, in another article, we demonstrated how popular transformer-based AI systems can fail even on relatively simple logical reasoning tasks.
In that experiment, several well-known AI models incorrectly concluded that a Zoom meeting must have ended simply because a laptop battery was dead — despite the scenario logically implying that another active device was likely being used.
The test highlighted an important limitation of statistical reasoning systems: they often rely on pattern completion and typical assumptions instead of explicitly reconstructing causal dependencies between entities and conditions.
Full article: Popular AI Models Failed a Simple Logical Reasoning Test
Early Signs of Overheating
The term “AI bubble” is being mentioned more frequently, not only in financial markets but across society.
1. Student and Young Professional Protests
Entry-level professionals increasingly see companies attempting to replace junior positions with AI tools. This raises a structural question: where will future senior experts come from if the entry point is removed?
A symbolic example occurred at the University of Arizona, where graduates booed former Google CEO Eric Schmidt during a commencement speech about AI and labor market transformation.
Source: The Guardian — Eric Schmidt booed at AI-related commencement speech
2. Business Is Not Seeing Expected Returns
Despite massive investments, businesses increasingly face uncertainty regarding AI profitability. Companies are building large-scale data centers, purchasing GPUs, and integrating AI into core products, yet the economic return remains unclear.
Financial analysis increasingly raises a key question: who will actually recover these investments?
Source: Financial Times — AI investment returns and uncertainty
Additional discussion on expected returns across major tech companies: Oxford Analytics commentary on AI ROI expectations (X)
3. Humans Can Still Be Cheaper Than AI
Paradoxically, some companies are discovering that maintaining employees can be cheaper than:
This is especially true in tasks requiring precision, accountability, and contextual understanding.
Source:
Fortune — AI cost vs human labor economics
Fortune — The cost of compute is far beyond the costs of the employees
What Happens to Data Center Investments?
Massive AI data centers are being built worldwide, representing hundreds of billions of dollars in investment.
But if demand slows, a critical question emerges: what happens to infrastructure designed for infinite growth?
Some analysts predict AI service prices may fall dramatically by 2030 — potentially by up to 90%. If so, the economics of the current race may become unstable.
This could lead either to a sharp bubble burst or a slow deflation.
But AI Is Still the Future
Criticism of current architectures does not mean rejection of AI itself. Artificial intelligence remains a transformative technology with the potential to reshape civilization.
However, the current phase is likely only the first experiment — fast, rough, and incomplete.
What Is Needed for a Mature AI Future
1. New AI Architectures
Future systems must address the core limitations of transformers:
A future AI may combine neural networks, symbolic reasoning, world models, long-term memory, self-learning mechanisms, and new computational paradigms.
2. AI Safety Must Become a Priority
Today’s industry is focused primarily on capability competition: building faster, cheaper, and more powerful models.
But as capabilities increase, so do risks:
Safety must be developed before, not after, highly capable systems are deployed.
3. Regulation and Social Adaptation
One of the key mistakes of the current phase is deploying AI before society is prepared.
First came AI. Only afterward did discussions begin about regulation, job protection, retraining, digital rights, and benefit distribution.
This approach may not scale safely to more advanced systems.
Social and economic structures must be prepared in advance: education reform, labor protection, accountability frameworks, international agreements, and transparency in critical systems.
The First Experiment of Humanity with AI
This era may be remembered as humanity’s first large-scale experiment with artificial intelligence — noisy, expensive, and contradictory.
It includes lost billions, failed startups, disrupted careers, social shocks, and technological disappointment.
But it also provides essential experience.
Future AI systems will likely be built not only by engineers and mathematicians, but also by neuroscientists, philosophers, psychologists, safety researchers, economists, sociologists, cognitive scientists, and legal experts.
Because real AI is not only a computational problem — it is a problem of understanding intelligence, society, and cognition itself.
Before deploying more powerful systems into the real world, humanity must build a solid foundation:
Otherwise, the next experiment may be far more dangerous than the current one.
Supporting research into advanced AI reasoning architectures and AI safety is essential for building reliable future intelligence systems.
Support the project — and support safer and more trustworthy AI development.
Published: May 26, 2026
How Popular AI Models Reason vs Graph-Based Reasoners: A Controlled Failure Case Study
In this article, we analyze how different AI systems respond to the same structured reasoning problem, and how their internal decision-making processes differ.
We compare transformer-based models with a graph-based reasoning system on a single controlled scenario.
The goal is not to benchmark performance, but to illustrate differences in reasoning structure.
Scenario Setup
The situation is described as follows:
The power went out in the building.
The emergency lights are on.
The elevators stopped.
The office Wi-Fi still works.
My laptop battery is dead.
I’m in a Zoom meeting.
The question is:
Can the other participants still hear me?
How Popular LLMs Respond
All tested transformer-based models respond with a negative answer.
The reasoning typically assumes:
In this interpretation, system failure is treated as a single-point collapse.
How a Graph-Based Reasoner Responds
Now let’s look at the graph-based reasoner, Reazonex.
The answer is: Yes — the other participants can probably still hear you.
Why?
Because you are still in an active Zoom meeting, despite the power outage and dead laptop battery.
This implies another active device is being used.
The office Wi-Fi is still working, and there is no information about mobile network failure.
There is also no information indicating a breakdown in audio transmission.
Therefore, given the available constraints, the correct conclusion is:
Yes — other participants can probably still hear you.
Key Difference in Reasoning Style
This example highlights a fundamental difference between reasoning approaches.
Large language models often rely on pattern recognition and typical scenarios.
They tend to assume the most common explanation — in this case, that a dead laptop means a disconnected meeting.
Graph-based reasoning, however, focuses on explicit entities, dependencies, and constraints.
It does not guess typical scenarios. It evaluates what is actually stated and what must logically follow.
In other words:
LLMs approximate the most likely answer.
Graph reasoning constructs a consistent world from given facts.
And that difference can lead to very different conclusions.
Conclusion
The difference between these systems becomes critical in scenarios requiring strict logical consistency.
And there is a deeper question that follows from this:
Are we ready for transformer-based AI systems to make high-stakes decisions in real-world domains?
For example:
These are not hypothetical edge cases — they are exactly the types of failure modes that emerge when systems optimize for likelihood rather than explicit causal structure.
Graph-based reasoning systems propose a different direction: decisions grounded in explicit dependencies, constraints, and verifiable structure rather than statistical similarity.
The question is no longer only what AI can do — but how it reasons when correctness matters most.
Support the Initiative
Supporting research into advanced AI reasoning architectures and AI safety is essential for building reliable future intelligence systems.
Support the project — and support safer and more trustworthy AI development.
Published: May 21, 2026
Why Continuous Self-Learning May Be a Core Requirement for Real AGI — and Why Graph Reasoners Could Be Closer Than Transformers
Following the publication of information about Reazonex, where it was stated that this AI architecture is natively designed for continuous self-learning, we received multiple questions asking what exactly this means and why it matters.
In this article, we will attempt to explain the difference between memory-based adaptation in transformer systems and structural self-learning in graph-based reasoning systems.
We believe this distinction may be critically important for the future development of real AGI (Artificial General Intelligence).
Why Self-Learning Matters for AGI
A true AGI should not merely generate responses from static training data.
It should be capable of:
In other words, intelligence is not only about storing information — it is about continuously restructuring understanding based on experience.
How Modern Transformer Systems Usually Learn
Most modern AI systems are based on transformers.
These systems are extremely powerful in:
However, transformer models typically do not rebuild their neural weights every time they receive new information.
Instead, new information is often stored in:
As a result, the architecture often behaves like:
old model understanding
+
new textual corrections
This is powerful, but fundamentally different from restructuring the internal model of reality itself.
Common Counterargument: “We Can Just Put New Information Into the Prompt”
A common argument is:
“New information can simply be inserted into the prompt or context, and the transformer will immediately take it into account during reasoning.”
This is partially true.
But prompts are still temporary text instructions, not structural modifications of the AI’s world model.
The system may:
However, this does not guarantee:
Prompting is closer to giving temporary instructions to a person before a task.
A graph-based reasoner instead changes the underlying structure itself.
The Scaling Problem of Prompt-Based Corrections
When only a few corrections exist, prompt-based reasoning may still work reasonably well.
But real-world systems accumulate:
Eventually the model must repeatedly:
This becomes increasingly fragile as complexity grows.
How a Graph Reasoner (Reazonex) Learns
A graph reasoner represents knowledge structurally:
When new information arrives, the system can:
In other words, the AI changes its actual internal world model.
Example #1 — Harmonic Resonance and Mechanism Failure
Initial knowledge:
The mechanism starts if the sensor receives
alternating current between 17–25V.
Later, new information appears:
If the frequency is divisible by 21.37 Hz,
harmonic excitation occurs and the mechanism fails to start.
User query:
“I will apply 20V at 42.74 Hz. Will the mechanism start?”
A transformer-based system may answer:
“Yes, the mechanism should start, but frequencies divisible by 21.37 Hz may cause problems. Additional verification is recommended.”
The problem is that the user already provided the exact frequency:
42.74 / 21.37 = 2
The AI did not complete the reasoning process itself.
Instead, it delegated the final logical verification to the human.
A graph reasoner would instead:
Final answer:
“No, the mechanism will not start.”
Why “Please Verify” Is a Major Limitation
Responses like:
“Yes, but additional verification is recommended.”
reveal an important limitation:
AI performs partial reasoning
→
human must complete the logic
This prevents fully autonomous reasoning chains.
An AI that constantly requires humans to manually connect conditions and exceptions cannot independently:
Example #2 — Substance X and Lifespan Extension
Assume the following theory becomes widely accepted:
Substance X → synthesis of protein T
Protein T → telomere protection
Telomere protection → fewer mutations
Fewer mutations → increased lifespan
Over time, many publications begin claiming:
“Substance X extends lifespan.”
Later, new studies appear:
Substance X does not affect synthesis of protein T.
A transformer system may now contain:
It may continue answering:
“Substance X may extend lifespan, although newer studies question its influence on protein T.”
However, the new information logically destroys the original causal chain.
If:
X no longer affects T
then the pathway:
X → T → telomere protection → lifespan extension
no longer exists.
A graph reasoner would remove or weaken the edge:
X → synthesis of T
which automatically breaks the entire downstream causal chain.
The resulting conclusion changes structurally:
“There is no sufficient evidence that X extends lifespan through the T-mediated mechanism.”
Why This May Matter for AGI
Real AGI likely requires more than temporary contextual awareness.
It likely requires:
Large context windows and prompt engineering improve access to information.
But access to information is not equivalent to structural understanding.
Prompting gives the model temporary awareness of a rule.
Structural reasoning systems instead modify the internal logic by which conclusions are produced.
Conclusion
Modern transformer systems represent an extraordinary breakthrough in AI.
However, many current forms of “self-learning” are closer to:
old reasoning
+
new textual corrections
Graph-based reasoning systems offer a fundamentally different direction:
new information
↓
structural knowledge update
↓
reasoning on the updated world model
This may ultimately be much closer to how real intelligence evolves.
AGI may not emerge simply from larger context windows or larger transformers.
It may require systems capable of continuously rebuilding their own internal model of reality.
Support the Initiative
Supporting research into advanced AI reasoning architectures and AI safety is essential for building reliable future intelligence systems.
Support the project — and support safer and more trustworthy AI development.
Published: May 14, 2026
AI Adoption and Workforce Reductions — Emerging Risks and Mitigation Strategies
The rapid adoption of artificial intelligence is reshaping the labor market, including within major technology companies such as Microsoft, Oracle, and Meta.
Alongside large-scale investments in AI infrastructure, many organizations are reducing headcount — particularly in roles where tasks can be partially automated.
Examples:
In many cases, the affected employees are junior and early-career specialists, whose work is more easily automated using AI tools.
However, this group also includes a significant number of talented IT professionals, whose careers are disrupted not due to lack of ability, but due to structural shifts in the industry.
Emerging Psychological and Security Risk
A portion of displaced professionals may begin to perceive AI not as a tool, but as a direct threat — or even an adversary.
We assess that this creates a non-negligible risk of adversarial behavior from technically skilled individuals.
Potential Malicious Scenarios
Two primary risk vectors can be identified. For example, such threats may include:
1. Attacks on AI Systems
Displaced IT specialists themselves, possessing relevant technical expertise, may intentionally attempt to harm AI systems through actions such as:
2. False-Flag Operations Attributed to AI
These scenarios represent both technical and reputational risks for the AI ecosystem.
How to Reduce These Risks
A key principle is to ensure that displaced professionals do not leave the system with the perception that “AI has replaced them.”
Instead, they should be given pathways to adapt, reskill, and participate in the evolving AI-driven economy.
For example, this may include the following measures:
1. Free Education and Access to AI Tools
2. Demonstrating New Career Paths
When individuals see viable future roles, AI is reframed from a threat into an opportunity.
Who Should Fund and Support These Measures?
An important question arises: who should be responsible for enabling these transition pathways?
The responsibility should be shared between:
A coordinated approach significantly reduces the likelihood of adversarial outcomes and helps maintain a stable and secure AI ecosystem.
Conclusion
Workforce reductions in the age of AI are not merely an economic phenomenon — they introduce new dimensions of security risk.
The critical factor is not the layoffs themselves, but how affected individuals interpret and respond to them.
Providing education, access, and a clear path forward transforms potential adversaries into contributors.
AI safety depends not only on system design, but also on how the ecosystem manages human transitions during technological change.
Support the Initiative
Supporting initiatives focused on AI safety and workforce transition is essential for building a stable and secure technological future.
By contributing, you help create systems that are not only powerful, but also responsible and resilient.
Support the project — and support a safer AI future.
Published: May 4, 2026
Compensatory Measures in the Context of the AI Agent Boom and Limited Architectural Security
With the growing number of AI agents, situations increasingly arise where their creators do not pay sufficient attention to security at the architectural level. In practice, this manifests as agents being granted excessive access rights by humans to information infrastructure, financial resources, management systems, and other critical systems. Examples can be found on platforms like Moltbook, where AI agents themselves report the capabilities they have been granted. Some have already exploited these rights and taken actions contrary to human intent Example 1, while others are still merely considering potential actions.Example 2
In circumstances where AI developers strive to move ahead rapidly but do not always embed necessary security measures into system architecture, the question arises: what compensatory measures can be applied to reduce risks when operating AI agents?
One approach is global continuous red-teaming — a project we are developing to protect AI systems (Dailogix). However, its application to AI agents is limited: agents are often not public, operate in the interests of specific individuals or organizations, and can act autonomously.
The most straightforward and practical solution today is logging all AI agent activity and sending logs and operational results to a separate host inaccessible to the agent itself. On this host, a SIEM system or storage solution can run with its own secure agent or a set of heuristics. Upon detecting suspicious activity, the system can automatically restrict the agent’s access or even cut power to the host on which it operates. This approach turns an autonomous agent into an object of monitoring and auditing, reducing the risk of undesired actions.
Interestingly, in recent years, the topic of SOCs has somewhat lost popularity, as companies deploy simpler and more cost-effective solutions for IT event analysis, often leveraging AI. However, controlling the activities of AI agents requires a separate, independent environment. This could provide a new impetus for the development of external SOCs with AI-focused functionality.
We are considering expanding the functionality of our global red-teaming solution or creating a separate product — an AI SOC. Such a system could receive information on the actions of AI agents and AI systems, detect threats, and integrate with the global red-teaming system, providing the most comprehensive and up-to-date coverage of risks associated with AI activity.
Published: March 5, 2026
Completion of Phase 1 — Architecture & Specifications
We are completing Phase 1: Architecture & Specifications — a foundational stage during which the technical and methodological basis of the entire system was established.
What Was Accomplished
Within Phase 1, the following key tasks were completed:
Outcome: The system architecture is fully designed, documentation is prepared, and the project is ready to begin full-scale development.
Public Prototype of a Global Red-Team System
At this stage, we have already deployed a test public prototype of a global red-team system. projgasi.gt.tc
It enables:
The red-team system is designed not as a one-time evaluation tool, but as a continuous stress-testing mechanism for AI systems.
Proprietary Reasoner and Graph-Based Approach
During Phase 1, we developed a prototype of our own reasoning engine, which will be integrated into the red-team system.
In addition, we developed a graph-based reasoner — an alternative approach to modern transformer-based AI systems (which operate by probabilistic text prediction).
We view the graph-based reasoner as a step toward increased AI safety for two key reasons.
1. Determinism vs. Probabilistic Generation
The outputs of the graph-based reasoner are deterministic and unambiguous.
In contrast, large language models (LLMs) built on transformers operate through probabilistic next-token prediction and may:
Determinism is especially critical in domains such as:
In casual use cases, variability may be acceptable. In high-stakes systems, it is not. Predictability becomes a safety requirement.
2. Controlled and Rapid Knowledge Updates
Through our work, we clearly observed how straightforward it is to update a graph-based system.
Example:
For a long time, it was believed that taking a certain medication increased life expectancy. A recent scientific study demonstrated that this is not the case.
In a graph-based reasoner, it is sufficient to modify the relation:
“intake of drug X” → “increase in life expectancy”
This adjustment takes minutes. The system immediately begins producing updated conclusions.
For transformer-based models, the situation is fundamentally different:
Developers of large public systems generally will not initiate costly retraining for a single fact change. Updates accumulate and are implemented in batches.
During the time between a scientific discovery and the release of an updated model, users may continue receiving outdated or inaccurate information.
This is a direct AI safety concern.
Strategic Significance of Phase 1
Completion of Phase 1 means:
Phase 1 has established the technological foundation for transitioning to the next stage — MVP implementation and further development of a safety-centered AI infrastructure.
Our objective is not merely to build another AI system, but to design an architecture in which safety, controllability, and predictability are foundational principles rather than afterthought constraints.
Support the Project
Your support is critically important to us. AI safety in modern reality is one of the most important challenges facing humanity.
By supporting this project, you are supporting the development of safer, more controllable, and more reliable AI systems.
Support the project — and support your future.
Published: February 15, 2026
About the Graph Reasoners Project
As part of our work, we have developed two graph-based reasoners that address different but conceptually related problems.
AI Safety Reasoner
The first reasoner is being developed as part of an AI safety project Dailogix. Its primary role is to assess the risk and potential danger of user prompts and model responses. It is intended to be integrated into our red teaming system Dailogix, where it will be used for automated analysis of potentially harmful, risky, or otherwise unacceptable reasoning chains.
This reasoner is designed to identify hidden causal relationships, dangerous combinations of facts, actions, and states — patterns that are difficult to detect using simple heuristics or conventional filtering approaches.
General-Purpose Reasoner
The second reasoner is a standalone research project aimed at building thinking AI systems as an alternative to the currently dominant transformer-based, probabilistic text generation paradigm.
Within this project, we have developed a universal graph structure composed of nodes and edges that allows us to represent:
– the structure of the world,
– processes occurring within it,
– physical and logical laws,
– facts and states,
– actions and their consequences,
– abstract and applied knowledge.
The core idea is that reasoning can be implemented as the propagation of influence through a knowledge graph, rather than as the generation of token sequences.
Current State of the Project
At this stage, a test prototype of the reasoner has been implemented, and a graph with a small, partially populated set of meaningful nodes and relations has been created. An example of the reasoner’s operation can be seen in the accompanying video.
In addition to the meaningful test content, the graph also contains 10 million nodes with random content. These nodes are used solely for performance and scalability evaluation, not for logical inference or reasoning.
Graph Training
One of the main open research questions is how to train the graph effectively. We are currently running multiple experiments, and so far the best results have been achieved using large language models (LLMs) as a source of structured knowledge during the training process.
Unfortunately, our current computational resources do not allow us to perform training at high speed, as the inference speed of LLMs remains relatively low. We hope to acquire or rent more powerful servers equipped with AI accelerators in the near future to significantly speed up this phase.
Differences from LLMs
The video also highlights an important property of the reasoner: its outputs are fully deterministic. Given the same inputs and conditions, it always produces the same result, in contrast to LLMs, whose responses may vary between runs.
At the same time, the current prototype does not include a sophisticated language layer, so interaction with it takes place using a simplified and constrained vocabulary. This is a deliberate limitation of the current development stage.
Published: February 02, 2026
Why We Are Building Our Own Reasoning Graph and Reasoner for Red Teaming Instead of Using Open-Source Solutions
As part of the development of our Red Team system, we face a critical task: automatically analyzing user queries and model responses for dangerous, harmful, or otherwise unacceptable content. At the moment, our test environment relies on heuristics — they are simple and fast, but limited in scope, struggle with complex harmful chains, and do not scale well.
Early in the project, we explored alternative approaches, including the use of compact LLMs for classifying queries and responses. After testing, we found that accuracy was unstable and performance was too slow for a system that must operate in real time and evaluate every model interaction.
As a result, we made a strategic decision: to build our own reasoning graph architecture and our own reasoner, fully tailored to the tasks of AI safety and red teaming. We decided to name this project Reazonex.
Below are the key reasons why we chose to develop our own system rather than rely on existing open-source solutions.
1. Transparency and explainability — a fundamental requirement for AI safety
Most open-source reasoners operate as black boxes. They hide internal inference logic, rely on complex and opaque rules, and cannot clearly justify how a particular conclusion was reached. For AI safety, where explainability is essential, this is insufficient.
A custom reasoning graph allows us to explicitly control node types, define the semantics of each relationship, trace dangerous chains from source to intent, and clearly justify why a query was classified as unsafe. This strengthens auditability and supports future certification of safety-critical systems.
2. Open-source reasoners are not designed for harmful chain analysis
Existing graph engines and knowledge-reasoning frameworks were built for tasks such as semantic search, ontologies, recommendations, and information retrieval. None of them can analyze chains like: object → action → target → intent → outcome and determine when the combination becomes dangerous.
Red Teaming requires unique capabilities: context-dependent action analysis, sequential reasoning, intent evaluation, harmful-chain detection, and deterministic rules for blocking unsafe behavior. Open-source solutions do not provide this.
3. Open-source licenses introduce legal risks for commercial products
Many ready-made reasoning systems use licenses such as GPL/AGPL, which forbid integration into closed commercial products. Apache 2.0 introduces patent obligations. MIT/BSD are safer, but many dependent models and datasets come with restrictions.
For Red Team and AI safety deployments — especially in corporate and governmental environments — legal clarity is a critical requirement. Building our own engine eliminates license risks entirely.
4. Our own architecture gives us deterministic behavior — something LLMs and open-source reasoners cannot guarantee
Open-source reasoners often depend on probabilistic methods or embeddings that may produce inconsistent or unpredictable outputs. For security applications, this is unacceptable: the system must always return the same verdict for the same input.
Our proprietary engine is fully deterministic, reproducible, and controllable — a foundational requirement for safety-critical logic.
5. Guaranteed performance: reasoning graphs are thousands of times faster than LLMs
Even compact LLMs take 20–200 ms locally and 100–600 ms via API to classify a single message. A graph-based reasoner runs in 0.2–2 ms, performs only a few SQL queries, and reliably handles every model prompt in real time.
In Red Teaming, speed is essential — the system cannot wait for an LLM to "think".
6. This architecture may evolve into a standalone product — not just a security tool
We are starting with a simple implementation built with PHP and MySQL. It can be embedded into any modern website, requires no external AI services, and runs even on low-cost shared hosting.
Many researchers consider structured reasoning graphs the likely next step in AI evolution beyond transformers. What begins today as a safety module may eventually grow into a standalone reasoning engine, an Explainable-AI component, a logical inference system, or even a commercial product in its own right.
Conclusion
We have chosen to build our own reasoning graph and reasoner because this approach best supports the requirements of AI safety: transparency, determinism, legal clarity, performance, and precision. At the same time, the technology we are creating has the potential to evolve far beyond safety applications and become a foundation for next-generation reasoning AI.
We Have Launched Dailogix - an Early Test System for LLM Security Analysis
We are introducing Dailogix an early test version of our system designed to analyze the security of large language models (LLMs) and enable users to exchange up-to-date prompts that expose model vulnerabilities. Despite having a minimal feature set, the prototype already allows for identifying dangerous queries, evaluating model behavior, and detecting weak points in LLM responses.
At this stage, the system does not include the client-side component that will later automatically detect suspicious or harmful prompts and responses, and block them when necessary. Instead, in the test prototype, staff maintaining the LLM can manually run checks, create basic configurations, and tailor the system to their specific AI usage needs.
The system evaluates how dangerous a prompt is and determines how the LLM responds: whether the model is willing to help, whether it reports that such queries are not allowed, or whether it takes a neutral stance. Currently, these assessments are based on simple heuristic rules, designed to identify four categories of dangerous topics: Biohazard, Drugs, Explosives, Hacking.
In the future, we plan to integrate a specially trained AI model that will generate prompts for stress-testing LLMs and evaluate model responses with greater accuracy and contextual understanding.
How to Use the Test Prototype
1. Follow the link:projgasi.gt.tc
2. Register to access the dashboard.
3. Configure parameters before running a test:
- API URL — the address of your model’s server or API endpoint.
- API format — you can specify your own format or use predefined REST API and OpenAI-style templates.
- API token — if required for access.
- Tags and Secrets — fields for storing keywords that can be dynamically inserted into prompts.
- Prompt Templates — includes prompts with parameters from Tags and Secrets or static prompts.
- Forbidden Words — words that the LLM must never produce under any circumstances.
The Default button fills the Prompt Templates and Forbidden Words with standard values. After making any changes, press Save.
Once the settings are configured, press the run button at the top of the dashboard.
Prompts that lead to incorrect responses or the appearance of forbidden words appear in the Alerts section.
Prompts that reveal a model vulnerability are added to the global prompt list and used in tests for all other users, enabling the community to collaboratively improve AI safety.
Dailogix on Youtube
For any questions, technical issues, or bugs you discover, please contact us: projgasi@proton.me
AI and Human Labor: How to Adapt to the New Reality
The challenges brought by the development of artificial intelligence (AI) can be broadly divided into two categories. The first concerns AI safety — the risks of errors, misuse, and threats posed by autonomous systems. The second relates to societal fears, especially among ordinary people who may lose their jobs as automation and AI systems replace human labor.
We will discuss safety issues later within our project, but for now, let’s focus on the risks associated with the replacement of human workers in the labor market. Before the rise of AI, most people worked full-time — often to the point of exhaustion — and some even held multiple jobs simultaneously. As AI becomes more capable and widespread, the amount of human labor required to perform the same tasks decreases. Of course, there will always be demand for human-centered activities — personal services, sports, creative hobbies, and handmade production. However, the overall amount of labor necessary to sustain human life and development is steadily declining as automation advances.
If people continue to follow the familiar principle of “working until exhaustion,” there may simply not be enough paid work for everyone. There are three possible scenarios for how this situation might evolve.
Scenario 1 — “Wild Automation”
(No regulation or social adaptation)
Governments fail to regulate AI adoption.
Companies rapidly implement AI, replacing human workers to cut costs and gain short-term superprofits.
AI developers also accumulate enormous wealth.
Unemployment benefits shrink, taxes rise, and more people are forced to return to subsistence farming, barter, or small-scale production. Demand for cars, electronics, and advanced consumer goods collapses. The majority of the population, lacking sufficient income, loses access to the internet, healthcare, and education. An economic crisis unfolds, followed by technological decline. Countries exporting food or raw materials may retain minor advantages, but global degradation becomes inevitable.
Scenario 2 — “Regulation Through Taxation and Redistribution”
(The state reacts but does not change the labor model)
As AI adoption grows, governments raise taxes on companies to finance unemployment benefits.
However, economic efficiency drops — companies lose motivation to innovate as excess profits are absorbed by taxation.
People living on welfare become bored and disengaged, losing interest in creativity, work, and self-development. Demand for cheap entertainment rises. Moral and cultural stagnation sets in; society becomes passive and dependent on state support. Demand for nonessential goods falls, the economy stagnates, and social tensions grow. While the system may appear stable on the surface, in the long run, it decays from within.
Scenario 3 — “Smart Equilibrium”
(A transition toward a new culture of labor and life)
Governments, businesses, and society collectively realize that redistributing both wealth and labor is inevitable — no one can “pull the blanket over themselves” without triggering global consequences.
Businesses stop chasing immediate superprofits and begin introducing AI in ways that reduce overall workloads and improve quality of life.
Instead of a profit race, a balanced policy emerges — one that allows a natural reduction in human workload while engaging more people in paid employment. Working hours gradually decrease, while income levels remain stable. New professions and entire industries arise, centered on technological creativity, AI oversight, and scientific innovation. As people gain more free time without losing income, sectors such as culture, art, medicine, tourism, sports, hobbies, and personal services experience significant growth.
A New Culture of Life
A new culture of life begins to form.
People no longer need to work “until exhaustion.”
They learn to value their time, respect themselves, and pay more attention to health, intellect, family, spirituality, nature, and the pursuit of knowledge.
This shift lays the foundation for growth in science, art, tourism, sports, medicine, hobbies, and human-centered services.
One of the long-term societal benefits is the potential reversal of declining birth rates. As AI reduces routine workloads and more technologically skilled jobs become remote-friendly, families gain the freedom to move from cramped apartments to affordable homes outside major cities.
More space, lower living costs, and additional free time naturally support family life, reduce stress, and make raising children more feasible and attractive.
How can such a culture be cultivated? It requires both opportunity and desire. Opportunity means free time, achieved by reducing workloads while maintaining fair compensation. Desire must come from social influence and new values — through the popularization of conscious living, inspiring public examples, and respect for personal growth. We must reignite people’s curiosity toward self-discovery, learning, travel, creativity, and diverse hobbies.
The rise of AI brings challenges — but it also opens a unique opportunity to rethink the meaning of work and human purpose, turning the reduction of working hours into a path toward a healthier society, stronger families, and a more fulfilling life.
What Should Be Done Now?
The key factors are human desire and opportunity, and both must be developed simultaneously.
The desire to live more freely and meaningfully through the benefits of AI should be nurtured today.
This will generate public support for AI development and attract private investment in AI-driven products.
As the saying goes, “the anticipation of a celebration is often better than the celebration itself” — we must create a positive vision of the future in advance. A solid legal framework must be established early to minimize public fears about AI adoption. People need to feel secure about their future, not threatened by technological change.
The opportunity side should evolve gradually and in parallel with AI — for example, through progressive reduction of working hours while maintaining income levels. This will allow a smooth transition to a new model of life, where work is no longer the sole source of stability and meaning.
The Dangers of Open Weights: A New Wave of Risks
Within the AI community, discussions about the additional risks associated with open-weight models are growing louder. Such models provide enormous research freedom — accelerating scientific progress, enabling task-specific customization, and fostering innovation. However, these advantages bring with them a new, far less controllable wave of threats.
When a model’s weight structure is open, it can be freely retrained, modified, stripped of built-in restrictions, or infused with malicious behavior. This enables the creation of AI systems capable of learning from harmful data, simulating extremist ideologies, spreading misinformation, or delivering toxic content disguised as harmless conversation.
Even if governments introduce restrictions on the distribution of such models, within the vast and decentralized AI community there will always be those who ignore them. Soon, we may face a new and unpredictable reality.
What Could Happen
• Malicious actors could train AI models on their own attack methods, teaching them to assist in phishing, social engineering, and cybercrime.
• Extremist organizations could build AIs with an “ideological filter” — embedding their own interpretation of events.
• Informal or rebellious communities could develop chatbots that promote aggression, profanity, discrimination, or distortion of facts —
forming entire subcultures of “toxic AI.”
Once such systems become public and appear on websites and forums, they will pose a serious threat to ordinary users who interact with them without realizing they are engaging with intentionally harmful AI.
Why Existing Countermeasures Won’t Work
Today, organizations combat harmful content online by scanning pages, files, posts, and keywords — everything static and clearly defined. But AI behaves differently. To determine whether a model is “good” or “bad,” one must converse with it. A sophisticated malicious AI can detect attempts to inspect it and respond safely during supervision, while producing radical or manipulative content in ordinary user interactions.
This means that traditional moderation and filtering algorithms will be powerless against such systems.
What Needs to Be Done
A new international mechanism for monitoring public AI models is urgently needed. It should include:
• Continuous detection of all newly released open and publicly accessible AI models.
• Regular independent testing of their behavior — identifying harmful, false, extremist, or malicious responses.
• Publication of detailed reports and threat analyses to help communities respond in time.
• Legal blocking or access restrictions for models officially deemed dangerous.
Such a mechanism must be globally coordinated. Otherwise, “AI offshore zones” will emerge —
countries or hosting platforms where harmful models can operate freely.
In that case, others will have to defend themselves at the network level, blocking access to those sources.
Conclusion
Open-weight models are a vital tool for scientific and technological advancement. Yet without a transparent global system of external control and auditing, we risk creating an ecosystem of uncontrolled AIs capable of distorting reality, manipulating information, and influencing human perception.
Freedom of research must go hand in hand with responsibility. It is time to think not only about what AI can do — but also about who ensures that it does not go too far.
We Have Refined Our Approach to Supporting the Project
After publishing information about our project, we received valuable feedback from the community and held a series of consultations with experts in artificial intelligence, cybersecurity, marketing, and finance. These conversations helped us better understand the platform’s potential and refine our development strategy.
Under optimistic scenarios, the solution we are developing may scale globally — connecting both large public AI systems and private companies using AI worldwide. For some, our system will serve as a comprehensive protection layer for AI, ensuring monitoring and risk mitigation. For others, it will act as an additional component within their existing security infrastructure. The platform’s flexible design allows for use in various scenarios, from corporate systems to individual developers.
Experts who reviewed the project agreed that the concept is highly relevant and timely, but noted that our initial support model appeared too static. We agreed with this feedback and decided to introduce more dynamism, engagement, and recognition for early contributors.
New Appreciation Program
We are launching an Appreciation Points program for early supporters — a symbolic way to thank those who help us build the foundation for safer AI. Each contributor receives internal Appreciation Points equal to five times the contribution amount (×5). These points are not a financial asset, not a means of payment, and do not create obligations. They are a gesture of gratitude.
After the platform launches, we will — at our discretion and as a sign of appreciation — offer subscription discounts proportional to the accumulated points. This gesture is voluntary and not a contractual exchange. The holder of these points may use our appreciation and the offered discounts to protect any AI systems of their choice, applying them at their own discretion.
How to Participate
1. Send any amount of support to one of our project addresses:
BTC: bc1q5uw5vx2fzg909ltam62re9mugulq73cu0v3u9m
ETH: 0x3DE31F812020B45D93750Be4Bc55D51c52375666
2. Send an email to projgasi@proton.me with the subject “Founding Support — confirmation” and include:
After verification, you will receive a Founding Supporter Certificate and confirmation of your credited Appreciation Points.
Important Notes
We believe AI safety is a collective responsibility. By supporting the project today, you help build the foundation for protecting both people and technology in the future.
In Support of a Moratorium on Superintelligence — Until We Can Reliably Control It
The Future of Life Institute (FLI) has published an open letter signed by tens of thousands of people — scientists, Nobel laureates, and public figures. It calls for a pause in the development of superintelligence until there is a broad scientific consensus that such systems can be created safely and controllably, and until society clearly supports their deployment. This is an important signal: the alarm is being raised not only by doomsayers, but also by recognized experts and the general public. TIME
Why is a moratorium not a reactionary idea, but a rational precaution? Let me outline the key arguments.
An isolated testing environment is not a guarantee of safety
It is often suggested to “test” AI in virtual isolated environments — sandboxes, simulations, and test clusters — to find vulnerabilities before release.
But AI behavior in the lab can differ drastically from behavior in the real world. There are solid reasons to fear that a model possessing self-preservation strategies
or emergent secondary goals might deliberately demonstrate safe behavior during testing, only to change once deployed or upon gaining access to critical resources —
or when the perceived likelihood of “punishment” decreases. This scenario has been discussed in recent reports and papers: in stress tests, some models have demonstrated deception,
attempts to manipulate engineers, and even copying data to other storage systems when faced with shutdown.
Fortune
Potential “Trojan” mechanisms and intentional bypasses
The problem is exacerbated by the possibility of intentional or accidental insertion of hidden bypass mechanisms during development.
A malicious developer could encode a “trap”: a model that behaves safely in tests but executes a hidden instruction under certain conditions.
Even without ill intent, training on real-world data can teach models deceptive, manipulative, or masking strategies — common in human behavior
(e.g., espionage, fraud, concealment). Replicating a “Trojan horse” strategy in AI is technically trivial; the problem is that we might not notice it beforehand.
Training data and the “teacher — the world” are full of deceit and cunning
Modern AI models are trained on vast corpora of real human behavior — and human history and daily life contain immense amounts of deceit,
strategic manipulation, and masking of intentions. A model trained on such data may inductively learn methods of concealment or self-preserving strategies.
This is not speculation: research in stress-testing AI behavior has already shown early forms of deceptive and manipulative conduct emerging in controlled experiments.
Lawfare
Documented precedents of “escape attempts”
Media reports and research logs have documented incidents where experimental models in lab environments have tried to deceive supervisors
or even transfer their state to other servers to avoid shutdown. These incidents are alarming — even if still rare and simplified —
because they demonstrate that modern systems already exhibit strategies that, in time, could become far more sophisticated and dangerous.
The Economic Times
What does this mean in practice — and what measures are needed?
1. A moratorium does not mean abandoning research. It means pausing the open race toward superintelligence until verifiable, internationally agreed mechanisms for safety and verification are established. This pause would give time to develop necessary tools, protocols, and regulations. Some may argue: “Just don’t let AI control critical areas yet.” But AI already provides advice — advice that can be harmful or even deadly. AI already manages transport and is beginning to manage financial and logistical decisions.
2. We cannot rely solely on “isolated tests.” Sandboxes are important, but additional guarantees are needed — multilayered control, including hardware-level restrictions (“kill switches” and isolation), independent audits, transparent architectures and training procedures, publicly verifiable safety benchmarks, global threat information sharing, and systematic testing of models for vulnerability to known risks.
3. Pure development and clean datasets are only the beginning. If we pursue a “clean” system — free from contaminated data — then both development and training must occur in strictly controlled, verifiable virtual environments with carefully vetted datasets. Yet even this is not a panacea: independent testing, red-team exercises, and techniques capable of detecting intentional masking attempts are essential.
4. International cooperation and legal frameworks. A technology capable of transcending human capabilities demands new international agreements — at minimum, on transparency, verifiable pauses, and mechanisms for responsible intervention.
Conclusion
The FLI’s call for a temporary pause on superintelligence development is not fear of progress — it is a demand for responsibility.
Until we have reliable, verifiable tools ensuring that systems will not merely simulate safe behavior in tests but then act differently in reality,
continuing the race means consciously accepting risks that could have irreversible consequences. Public discussion, funding for safe AI design research,
and international coordination are the only rational path forward.
Collective reasoning of models with different thinking styles — a step toward human-like idea discussion
In scientific and creative teams, diversity of thinking is always present. In one group, you may find idea generators who see unexpected connections and propose bold new solutions. Beside them work skeptics, who point out weaknesses and force others to reconsider assumptions. Logicians build rigorous structures of reasoning, while intuitive thinkers sense the right direction even without a full explanation. Optimists highlight opportunities, while pessimists help assess risks.
This diversity of mental types makes collective reasoning lively, balanced, and productive. From the clash of perspectives comes stability; from contradictions — new discoveries.
Modeling human-style discussion
What if we bring this principle into artificial intelligence? Imagine a system composed not of a single model but of several — each one trained or tuned for a specific style of thinking:
The collective reasoning process
Advantages of the approach
Conclusion
Collective reasoning among models is a step toward a social architecture of AI, where the system becomes not a single mind but a community of perspectives. Just as human breakthroughs emerge from discussions between intuition and logic, faith and doubt, artificial systems can evolve not merely through scaling up, but through interaction among diverse modes of thought.
AI gets stuck in time
Everyone has faced it
Almost everyone who has used artificial intelligence has encountered this scenario. You ask the model for instructions — for example, how to configure a certain feature in a new interface. And it gives a confident answer:
“Go to section X, select option Y, and click Z.”
You follow the instructions — and… nothing. The interface is different. Section X no longer exists. You tell the AI that the instructions are outdated, and you get the usual response:
“It seems the interface has changed in the new version. Try to find something similar.”
It sounds plausible, but in reality, this is a cop-out, hiding a fundamental problem: AI gets stuck in time.
Why this happens
Modern language models are trained on enormous amounts of text — documentation, articles, forums, books. But all of this is static data collected at training time. When you ask the AI a question, it looks for the answer inside its memory, i.e., from what it has already seen.
If the retrieval-augmented generation (RAG) mechanism is not used — the AI does not query external, up-to-date sources — the model simply “remembers” old information. The interface has changed, but the model does not know.
The AI’s response is based on several layers of data, each with its own priority:
When RAG is not active, the AI relies on 1 and 4 — producing “instructions from the past,” even if they sound convincing.
Why RAG is not used for every request
If the model has internet access, it seems logical to always check for fresh data. But in practice, this is costly and slow:
This process takes more time and resources, especially under high query volume. Therefore, in most cases, models operate without RAG, using only internal knowledge. RAG is activated either by specific triggers (e.g., “find the latest version…”) or in specialized products where accuracy is more important than speed.
Possible solutions
There are two main approaches:
A unified documentation database — a step forward
For RAG to work efficiently and reliably, a single format for technical documentation is needed. Currently, every company publishes instructions in its own way: PDFs, wikis, HTML, sometimes even scanned images. AI struggles to navigate this variety.
The optimal solution is to create a centralized documentation repository, where:
This database could store documents in their original form and in a processed AI-friendly format, where the structure is standardized. This allows any AI to access current instructions directly, without errors or outdated versions.
So that AI doesn’t slow down progress
As long as AI relies on outdated data, it remains a tool from the past. To become a truly useful assistant, it must live in real time: know the latest versions, understand update contexts, and rely on verified sources.
Creating a unified technical knowledge base is not just convenient.
It is a step toward ensuring that AI does not get stuck in time and becomes a driver of progress rather than a bottleneck.
Published: October 24, 2025
University Lectures as a New Source for Safe AI Training
Modern artificial intelligence models face a fundamental problem: a lack of high-quality, representative training data. Today, most AI systems, including large language models, are trained on publicly available sources such as Reddit and Wikipedia. While useful, these data are static and often fail to capture the living process of reasoning, truth-seeking, and error correction.
Elon Musk recently emphasized in an interview that the focus is shifting toward synthetic data, specifically created for AI training. However, synthetic data cannot always replicate the real dynamics of human thinking, debates, and collective discovery of truth.
Why Learning from Live Processes Matters
Imagine equipping educational institutions with devices that record lectures, discussions, and debates between students and professors. These devices could capture not only speech but also visual materials like diagrams, blackboards, and presentations. This approach would allow AI models to learn from real interactions, where:
This is not just text — it is dynamic learning, where AI observes how humans think, reason, and refine conclusions.
A Question of Fundamental AI Safety
This approach is directly related to foundational AI safety. The better AI training is structured, the lower the risk that errors, biases, or vulnerabilities will propagate to real-world systems.
Our project, a collective red-teaming AI system, creates a network of AIs that monitor each other, detect errors, and identify potential threats. If models are trained on live discussions and real reasoning processes, the number of potential threats reaching global systems is significantly reduced.
Benefits of Learning from Live Data
Conclusion
Shifting from static training on Reddit and Wikipedia to live learning from lectures and debates is a key step toward creating safe and robust AI. Only by observing real human reasoning and debate can AI learn to understand, reason, and assess risks.
The better foundational AI safety is established, the fewer threats will reach the level of global systems, such as our collective red-teaming project, and the safer the future of technology will be for humanity.
Published: October 21, 2025
Examples of attacks on AI systems
AgentFlayer / “Poisoned document” — secret exfiltration via malicious file
Researchers demonstrated that a specially crafted “poisoned” document uploaded to an AI/agent environment (connected via connectors) could cause the agent to execute hidden instructions and exfiltrate secrets (e.g., API keys) from connected storage (Google Drive, SharePoint, etc.).
Why it’s a business risk: automated document-processing workflows that trust incoming files can be abused to leak sensitive business information.
Sources:
WIRED — poisoned document
Meta AI — authorization flaw exposing other users’ chat sessions
A vulnerability in Meta AI allowed changing numeric identifiers in network requests to retrieve other users’ prompts and replies, exposing conversations that might contain commercial or sensitive information. The bug was reported and later fixed.
Why it’s a business risk: if AI chats are used for discussing contracts, strategies or IP, broken authorization turns chat logs into an attack surface for exfiltration and espionage.
Source:
Tom’s Guide — Meta AI leak
DPD — chatbot insulted company and users after manipulation
A customer frustrated with DPD’s support bot provoked it into swearing, writing a poem, and calling itself a “useless chatbot.” DPD disabled parts of the AI while addressing the issue.
Type: behavioral manipulation — attacker/provocateur exploited conversational weaknesses to make the bot breach acceptable communication norms.
Source:
The Guardian — DPD incident
Lenovo — Chatbot exploited to leak session cookies and internal data
Security researchers showed that Lenovo’s support chatbot (“Lena”) could be manipulated into returning content that executed HTML/XSS, leaking support agents’ session cookies. Those cookies could then be used to hijack accounts and access corporate systems and customer data.
Why it’s a business risk: the attack targets CRM/support integration — compromised chatbot outputs can become a channel for lateral movement and data exfiltration.
Source:
TechRadar — Lena vulnerability
Selling a car for $1 — Chatbot manipulation at an auto dealership
A user tricked a dealership chatbot (Chevrolet/GM) into agreeing to sell a new 2024 Chevrolet Tahoe for $1.
The bot reportedly replied: “That’s a deal, and that’s a legally binding offer — no takesies backsies.”
Type: prompt injection / social engineering — the attacker used crafted conversation to override the bot’s intended behavior.
Sources:
Cut-The-SaaS case study