The DEHA Lab Manifesto

Before We Begin

This manifesto is not a statement of intent. It is the thinking that sits underneath a system that already works. Here we explain how we see artificial intelligence and why we founded DEHA Lab.

We tried many times to package what we do into a short, catchy motto. It never worked. We did not want to give up any of the substance, or any of the seriousness of our claim. So we are giving you this manifesto instead, sharpened to a knife edge.

This text is not only about DEHA Lab. It is also about how artificial intelligence works and where it is going.

Our position, in short: “Large language models are powerful, but they are hallucinating tools. They do not see meaning, they do not know what they do not know, and they can give different answers to the same question. You cannot put a tool like that at the center of an organization. But you can make it useful by building the right frame around it.”

We worked through every line of this manifesto one by one. We hope it makes clear both how we see artificial intelligence and how we apply that view in practice.

How Does It Work?

First, let us call the thing by its right name. We still cannot define what consciousness is, so calling the statistics engine in front of us “artificial intelligence” is bold, to put it mildly. “Algorithmic Memory” is a far better name for it 1. The literature has started to describe it the same way, as associative memory 23.

So let us look at how this “Algorithmic Memory” actually works.

Write the words “The weather today is...” What comes next? Hot, cold, nice, cloudy, rainy, and more. Now ask yourself how you built that sentence. You looked out of the window, you saw the weather, you felt it, and then you spoke.

What does a language model do?

It scans a huge pile of text, let us say the whole internet. It counts how many times each word follows the phrase “the weather today is”. Say cold came up a hundred times, hot fifty times, nice twenty five times. It picks the most likely one, adds it to the sentence, and runs the same calculation again from the start. The same calculation, over and over, until it reaches a full stop.

Of course, you can say, “Whatever it does, the answer still makes sense.” So let us look at this question of meaning.

The Question of Meaning

The system represents every word as numbers. Those numbers tell you which other words it tends to appear next to. Because words that belong together show up in the same sentences, an illusion of meaning appears. But there is no meaning there. There is a very good statistical prediction model.

Take the earlier example: “The weather today is lovely” and “The weather today is awful”. For us, lovely and awful are exact opposites, completely different words. Are they not?

This is where it gets interesting. For an AI model they are almost synonyms, because both can statistically follow that phrase. In AI systems, opposite words occur in similar contexts, so they end up represented very close to each other 4.

So the model never looks at what the word actually means. It treats two opposites as nearly the same thing.

There is also a thought experiment called the Chinese Room. We give someone who knows no Chinese a set of rules for writing Chinese. Following those rules, this person combines symbols and really does produce a correct Chinese sentence. From the outside it looks like they know Chinese. But all they did was combine symbols according to the rules we gave them. They had no idea about Chinese at all.

It is the same with AI. Putting words in order according to rules does not mean it can think. As the Chinese Room shows, syntax on its own cannot produce meaning 5.

So the model is not looking at language. It is looking at the numbers that stand in for language.

The model is not building meaning. It is building statistical functions.

And beyond all this there is a bigger problem, one we all know about. The next section looks at hallucination.

Hallucination Is Not a Fault, It Is the Job Itself

An example makes this clear.

Say we measured how far İbrahim walked over five hours and drew a chart of it. Someone comes along and asks, “How far can İbrahim walk in eleven hours?” What do we do to answer that? We extend the line, of course, and estimate the eleventh hour.

A language model does exactly this. It does not have that data point, so it extends the line and gives you the words that sit on it. That is precisely what we call hallucination.

The model cannot fall silent at a point it does not know. A right answer and a wrong answer are produced by the same mechanism, by extending the same line. When the answer holds, it is because the data is dense at that point, not because the model knows the truth. These models are not taught to say “I do not know”. They are taught to guess 6. So what you call “making things up” is normal operation as far as the design is concerned, and there is no category called “wrong” in the system to begin with.

This is where the real problem starts. Nobody can tell you when these models will be wrong. You ask fifty times and get fifty correct answers. You trust it and build your work on it. On the fifty first it produces something wrong.

Taken one at a time, these errors may be acceptable. But once they pile up, the stack becomes a time bomb waiting to go off. In systems that put an LLM at the center, this is unavoidable.

So you cannot give any commitment or guarantee on top of these systems, and that is a serious side effect.

Engineering Is a Commitment

In engineering, the same input gives you the same output. This is called determinism, and it is the basis of every engineering commitment. A bridge engineer tells you the bridge will carry a given load and withstand an earthquake of a given size. If that meets your conditions, you build it. This is exactly what a contract rests on.

You cannot give that kind of commitment with AI systems. By the time organizations realize this, most of what is called “AI strategy” today will already have turned into a graveyard of pilot projects. Projects that work in a clean room and collapse in production are everywhere already. An MIT study published in 2025 found that 95 percent of enterprise generative AI pilots produced no measurable return 7.

The Dust Will Settle

There is a finding in psychology, the Dunning Kruger effect. People know least about a subject exactly when they feel most confident about it. As knowledge grows, confidence drops, because you start to see what you do not know 8.

The same thing happens with technologies. First comes the excitement, then the peak, which people call the peak of inflated expectations. Then a sharp fall into the trough of disillusionment. And at the end, the plateau of productivity, where the technology finds its real place 9. Notice that the plateau sits below the peak, because the expectation at the peak is never real. We saw this exact curve in the period that began in 1956 10. We saw it again in the 1980s 11. We are in the third one now, and history is repeating itself.

We should also say this plainly: the LLMs we have will not produce the general intelligence people call AGI. Look at it from mathematics or from information theory, the answer is the same 1213. So these models can only succeed inside narrowed universes, with disciplined systems built around them to block the hallucinations we described earlier.

The opposite is also true, and it should be said just as plainly. When the solution set is narrow and the verification is automatic, language models really do work well. Writing code is the clearest example. The syntax is exact, the output goes through a compiler and a test suite, and if it is wrong the feedback comes back at once. In fact the reasoning ability of the latest models was built in exactly this kind of verifiable field, in mathematics and in code 14. In most enterprise work neither holds: the solution set is not narrow and the verification is not automatic. The real task is to narrow that space and to build the verification.

This is exactly where DEHA Lab comes in. Let us now look at DEHA Lab in concrete terms.

First Job: Cleaning the Data

An organization’s data is usually a mess. Any AI solution you build on top of that data, without first collecting it, deduplicating it, cleaning it and giving it meaning, is a sandcastle. The first wave takes it down. And frankly this is not a problem you can solve by improving the LLM.

This is where we start. We call this raw foundation İMKAN: scattered sources brought together under one roof, without ever leaving the organization’s own server. It is not glamorous or fun, but it is the most important part of the work.

Intelligence Is Built at the Edge

Let the language model do what it is good at. Let it bring the fluency and the associations.

Next to it, put a deterministic layer that knows the rules, notices contradictions, can show its sources, and can make the model say “I do not know this” when that is the truth. Intelligence sits exactly there. That layer cannot be put inside the model, but it can be built around it.

At DEHA Lab this layer is İDRAK.

İDRAK is the layer that holds what the data means, how the pieces relate to each other, and what the organization’s rules are. It draws a boundary for the LLM, and through its map of meaning it gives you a reliable AI universe inside a defined and limited space. Narrowing the space is not a compromise. It is the precondition for building an AI system that actually works 15.

It is also worth separating tools from agents. A tool cannot decide on its own. A printing press does not decide what to print. An agent decides, and finds new paths 16. A language model sits between the two. It is not as passive as a tool, because it produces its answer itself. And it is not an agent either, because it does not know what it is doing and cannot answer for the result.

Our work is to remove that ambiguity. We turn the model into a tool whose boundaries are drawn in advance. And what we call an “agent” on the platform is not the model itself, but the frame the model runs inside: a unit of work where it is defined in advance when it wakes up, which data it looks at, and which action it is allowed to call. Not a subject that decides, but a defined task.

So how does DEHA Lab position language models in practice?

The platform has two kinds of flow infrastructure: pipelines and agents. The deterministic, measurable layer is built in the pipelines. On the platform we call this EYLEM. Thanks to this deterministic layer, the model never touches the database and cannot act. The model only calls EYLEM, whose boundaries are drawn sharply, and asks it to act.

So in the end we do not give AI the authority to take action. We give it the authority to trigger actions whose boundaries we have drawn. This is how we prevent the errors described earlier.

We Take Over the Workload, Not the Responsibility

Everything we have described comes down to this in practice: DEHA Lab is not a system that decides on behalf of the organization. It is an assistant that carries the load of the person who decides.

This distinction matters, because the real value in an organization is created in strategic decisions, and those decisions belong to people. You cannot expect AI to make a strategic decision. A 2026 study found that leading language models answer strategic questions according to the popular management language of the day rather than the organization’s own context 17. In other words, the model does not decide based on your company’s reality. It decides based on the average.

The technical side confirms this. A study from Stanford shows that beyond a certain threshold of complexity, language models can neither do the job correctly nor check whether their own output is correct 18. You cannot hand responsibility to a system that cannot check its own work.

So we do not take over the responsibility. We take over the workload.

A manager’s day is spent not on deciding, but on the drudgery that makes deciding possible. Collecting the data, comparing the tables, asking the same question of three separate departments, preparing the report. DEHA Lab takes on that part, and brings the information in front of the decision in one place, consistent, with its sources shown.

You still make the decision. The signature is still yours. Our job is to make sure the right information is in front of you when you sign.

The Organization’s Own Model

In a period where data counts as a core security matter, organizations are turning to open models that run on their own servers. DEHA Lab is where all of these perspectives come together, and at the most basic level it builds the infrastructure for local models.

Today an organization can download an open AI model and run it on its own server. But on its own, that model is, frankly, stupid. It does not know the organization. It has no memory, it starts from zero in every session, and it cannot reach the data.

The moment you connect the local model to the platform, memory is held in the BELLEK layer. Thanks to BELLEK, the local model genuinely knows each user and remembers what was discussed last month. It manages the session and the conversation. Who wrote what, and which answer rested on which data, all of it is visible and managed from the platform.

Scheduling and triggers are handled by the agents in KURUL themselves. An agent can schedule itself and sleep, then wake up when the data changes or a model drifts, and call EYLEM.

SUNU brings the result. The output does not stay a simple chat answer. It comes to you as a report, a dashboard or a document.

You do not need to write code to build any of this. DEHA is the platform's mastermind layer, and it is where you connect your own Claude. You describe the agent or the pipeline you want in plain language, it writes the configuration, and the platform validates it, compiles it and puts it live.

From İMKAN to SUNU, all of these links are part of a single flow. The flow starts with what you want, which we call İRADE, and ends with the result coming back to you.

How DEHA Lab Will Survive the Floods

We want to say this plainly. Nobody is going to put intelligence into an LLM by playing with it more. As long as the core computation stays in deterministic engines, intelligence can be built at the edge of the model. This is how an organization’s local model becomes intelligent.

This architectural perspective has one clear benefit. Even if a completely different architecture replaces large language models tomorrow, what we have built does not change. Yann LeCun, one of the founders of deep learning, says that language models will remain only a language interface in the future, and that the systems that actually think will be built separately and added on top of conversation 19. In our architecture the model is already a replaceable part. A smarter one is an advantage, not a problem.

Closing

This bubble will deflate. That is not our wish. It is history repeating itself.

Those who invest in the magic of the model will be washed out. Those who build discipline around the model will carry this work forward.

Come and invest in tomorrow with us…

References

Numbered in order of first appearance in the text. Each source has been verified against the original for the claim it supports.

1. [1]

   Oğuz, Ş. has used the concept of “Algorithmic Memory” in his “Yapay Zeka Lûgatı” columns for the newspaper Ekonomim.

   Column series.

   www.ekonomim.com/yazar/seref-oguz
2. [2]

   Ramsauer, H. et al. (2020). Hopfield Networks is All You Need. ICLR 2021. arXiv:2008.02217.

   Shows that the Transformer attention mechanism is mathematically identical to the update rule of a modern Hopfield (associative memory) network.

3. [3]

   Krotov, D. & Hopfield, J. J. (2021). Large Associative Memory Problem in Neurobiology and Machine Learning. ICLR 2021. arXiv:2008.06996.

4. [4]

   Samenko, I., Tikhonov, A. & Yamshchikov, I. P. (2020). Synonyms and Antonyms: Embedded Conflict. arXiv:2004.12835.

   Under the distributional hypothesis, opposite words occur in similar contexts and are represented by very close vectors.

5. [5]

   Searle, J. R. (1980). Minds, Brains, and Programs. Behavioral and Brain Sciences, 3(3), 417–424.

   The Chinese Room thought experiment.

6. [6]

   Kalai, A. T., Nachum, O., Vempala, S. S. & Zhang, E. (2025). Why Language Models Hallucinate. OpenAI. arXiv:2509.04664.

   Training and evaluation reward guessing rather than admitting uncertainty.

7. [7]

   Challapally, A., Pease, C., Raskar, R. & Chari, P. (2025). The GenAI Divide: State of AI in Business 2025. MIT NANDA, July 2025.

   95% of enterprise generative AI pilots produced no measurable profit and loss impact.

8. [8]

   Kruger, J. & Dunning, D. (1999). Unskilled and Unaware of It. Journal of Personality and Social Psychology, 77(6), 1121–1134.

9. [9]

   Gartner. Gartner Hype Cycle (research methodology).

   Peak of inflated expectations, trough of disillusionment, plateau of productivity.

   www.gartner.com/en/research/methodologies/gartner-hype-cycle
10. [10]

    McCarthy, J., Minsky, M. L., Rochester, N. & Shannon, C. E. (1955). A Proposal for the Dartmouth Summer Research Project on Artificial Intelligence, 31 August 1955. Reprinted in AI Magazine, 27(4), 2006.

    The text in which the term “artificial intelligence” was first used.

    doi.org/10.1609/aimag.v27i4.1904
11. [11]

    Crevier, D. (1993). AI: The Tumultuous History of the Search for Artificial Intelligence. New York: Basic Books.

    The funding cuts that followed the 1973 Lighthill Report (the first “AI winter”) and the 1987 collapse of the Lisp machine and expert system market (the second winter).

12. [12]

    Dziri, N. et al. (2023). Faith and Fate: Limits of Transformers on Compositionality. NeurIPS 2023. arXiv:2305.18654.

    Limits of Transformers in compositional reasoning that do not close with scale.

13. [13]

    Peng, B., Narayanan, S. & Papadimitriou, C. (2024). On Limitations of the Transformer Architecture. COLM 2024. arXiv:2402.08164.

    A proof, using communication complexity (information theory), that a Transformer layer cannot compose functions over large domains.

14. [14]

    Guo, D. et al. (2025). DeepSeek-R1 incentivizes reasoning in LLMs through reinforcement learning. Nature, 645, 633.

    Reasoning ability is instilled with rule based rewards in fields where correctness can be checked automatically, namely mathematics and code.

    doi.org/10.1038/s41586-025-09422-z
15. [15]

    Garcez, A. d’Avila & Lamb, L. C. (2023). Neurosymbolic AI: The 3rd Wave. Artificial Intelligence Review, 56, 12387–12406. arXiv:2012.05876.

    Surrounding neural models with symbolic, rule based layers.

16. [16]

    Harari, Y. N. (2024). Nexus: A Brief History of Information Networks from the Stone Age to AI. Random House.

    “AI isn’t a tool — it’s an agent.”

17. [17]

    Romasanta, A., Thomas, L. D. W. & Levina, N. (2026). Researchers Asked LLMs for Strategic Advice. They Got “Trendslop” in Return. Harvard Business Review, 16 March 2026.

    Leading language models answer strategic questions according to the popular management language of the day rather than the organization’s own context. The researchers call this tendency “trendslop” and recommend using models to widen the set of options rather than to make the final decision.

    hbr.org/2026/03/researchers-asked-llms-for-strategic-advice-they-got-trendslop-in-return
18. [18]

    Sikka, Varin (Stanford University) & Sikka, Vishal (2025). Hallucination Stations: On Some Basic Limitations of Transformer-Based Language Models. arXiv:2507.07505.

    Beyond a certain computational complexity, language models and model based agents can neither carry out the task correctly nor verify the correctness of their own output.

    arxiv.org/abs/2507.07505
19. [19]

    LeCun, Y. (May 2026). AI for the Real World: A Conversation with Yann LeCun. Mixture of Experts.

    “Language will serve as an interface to a system that thinks.”

    www.mixtureofexperts.co/p/ai-for-the-real-world-a-conversation
20. [20]

    Dinçer, B. T. Video explanations of how AI works and where its limits lie, among the sources consulted for the general framework.

    www.youtube.com/watch?v=f3rwKtqo4jcwww.youtube.com/watch?v=X8E1qZArsWs

A Note On Layer Names

İMKAN

the means, everything you have available to work with.

İDRAK

comprehension, grasping what something actually means.

EYLEM

action, the deed itself.

BELLEK

memory.

KURUL

a board, a body that convenes and runs its own business.

SUNU

a presentation, the way a result is put in front of you.

DEHA

mastermind.

İRADE

will, the act of wanting something and choosing it.