Habr AI Author Proposes Deterministic AI on MacBook Air Over GPU Race

An opinion article was published on Habr AI, in which the author proposes an alternative to the race for increasingly expensive GPU clusters. Instead of scaling probabilistic models, he advocates for a deterministic AI core on Rust, which, according to him, already runs locally on MacBook Air M2 with 8 GB of memory.

Why the Author Disputes the Market

The text begins with a direct attack on the current strategy of market leaders — OpenAI, Google, and Meta. The author calls it "brute force": fundamental limitations of LLMs are not solved mathematically, but masked by an ever-increasing number of GPUs, megawatts, and investments. According to his logic, the industry compensates for architectural problems with computational power rather than seeking a more rigorous decision-making model.

The post on Habr is marked as "Opinion," so this is more of a manifesto and polemic with the market than an academic work. The article lists three reasons why the author considers such an approach to be a dead end. The first is energy consumption: training and running large models require increasingly more electricity and hardware.

The second is the stochastic nature of LLMs: the system does not derive an answer by strict rules, but guesses the next token. The third is hallucinations, which the author considers a built-in property of the probabilistic architecture, not a bug that can be completely fixed with additional filters.

"The future of artificial intelligence is not a question of money.

It is a question of the right mathematics."

What He Proposes

Instead of a classical neural network scheme, the author describes a deterministic core with O(1) logic. The key idea is that the decision-making time and its validation remain constant, and the system works not with a long probabilistic context, but with intents represented as mathematical formulas. This, by design, should eliminate the principle of "guessing" and make the model's behavior predictable.

Hence the claim to more reliable AI for scenarios where the answer needs not just to be generated, but verified. The text also provides specific theses about the project. The author writes that the current version of the core v0.

26.0 has already been tested in comparison with cloud models, and the solution itself is deployed on an ordinary MacBook Air M2. The main stack is Rust, without Python and intermediate interpreters: high-level intents, according to him, are translated directly into executable commands.

Special emphasis is placed on the fact that the author bets not on the general "magic" of the model, but on the controllability of the architecture.

• Deterministic O(1) logic for inference and validation
• Local execution on MacBook Air M2 with 8 GB of RAM
• Focus on Rust and DMA instead of heavy abstraction layers
• Claimed performance of up to 5.4 million CPU operations
• Comparison with cloud AI systems in core v0.26.0 tests

How Protection is Organized

A separate section of the article is devoted to self-learning and security. The author claims that the core has a built-in "constitution of universal human values" — not as a set of text prohibitions, but as a system of mathematical axioms within the logic. The article directly emphasizes that this level of constraints is harder to bypass with ordinary prompt engineering, because unwanted actions should be filtered out at the validation stage itself.

The key thesis is simple: AI should be not only smart but also verifiable at the architecture level itself. Here also passes the main ideological turn of the text: the author contrasts local, controlled, and mathematically rigid AI with cloud models that scale along with costs and risks. He also writes that the technology is already protected by a BSL license, is being prepared for patenting, and technical details and roadmap are published in the project's repositories.

All of this is presented as groundwork not for "chatter," but for autonomous systems where an error is critical. The author's final thought sounds even broader: the future of such AI should start not in a data center, but on the user's device.

What This Means

The article on Habr is important not as proof of the new architecture's victory, but as a signal of growing demand for more compact, predictable, and local AI. Even if the author's theses still need independent verification, the vector itself is clear: the market is tired of the idea that any AI problem can only be solved with one more GPU cluster. Against the backdrop of rising compute costs and interest in edge scenarios, such texts are already setting an alternative agenda.