We scan new podcasts and send you the top 5 insights daily.
Modern reinforcement-learned reasoning models do not follow strict human-like propositional deduction. Instead, they produce a broad token soup enumerating dozens of prospective solutions—frequently repeating words like 'wait'—before skimming back over that linearized search space to select an outcome.
AI models struggle to plan at different levels of abstraction simultaneously. They can't easily move from a high-level goal to a detailed task and then back up to adjust the high-level plan if the detail is blocked, a key aspect of human reasoning.
Reinforcement learning incentivizes AIs to find the right answer, not just mimic human text. This leads to them developing their own internal "dialect" for reasoning—a chain of thought that is effective but increasingly incomprehensible and alien to human observers.
Contrary to fears that reinforcement learning would push models' internal reasoning (chain-of-thought) into an unexplainable shorthand, OpenAI has not seen significant evidence of this "neural ease." Models still predominantly use plain English for their internal monologue, a pleasantly surprising empirical finding that preserves a crucial method for safety research and interpretability.
Despite full access to a model's internal reasoning, its decision-making remains opaque. Models explore and backtrack through many ideas using a "linearized tree search," and the critical point where a final decision is made is often unclear, making simple reading of the CoT insufficient for effective supervision.
Lila observed its AI models achieving high-reward outcomes despite generating pathological or nonsensical 'chain of thought' reasoning. This suggests the human-legible text is often a post-hoc justification, not a transparent window into the model's true computational process happening in latent space.
The argument that LLMs are just "stochastic parrots" is outdated. Current frontier models are trained via Reinforcement Learning, where the signal is not "did you predict the right token?" but "did you get the right answer?" This is based on complex, often qualitative criteria, pushing models beyond simple statistical correlation.
Classifying a model as "reasoning" based on a chain-of-thought step is no longer useful. With massive differences in token efficiency, a so-called "reasoning" model can be faster and cheaper than a "non-reasoning" one for a given task. The focus is shifting to a continuous spectrum of capability versus overall cost.
AI models learn to reason like mathematicians—including backtracking and exploring dead ends—even though their training data (textbooks, papers) often presents clean, final proofs that hide the messy discovery process. This suggests the models are developing a general-purpose reasoning capability, not just mimicking final outputs.
Allocating substantial compute budgets to reinforcement learning rather than pure pre-training trains models to prune unproductive search paths. This produces models that reach correct conclusions in a fraction of the inference time, cutting token latency and infrastructure costs for production workloads.
The benefit of discrete reasoning (like generating tokens or tool calls) over a continuous 'neuralese' is error correction, analogous to why digital computing beat analog. A slightly wrong token can be 'rounded' to the correct one, preventing the compounding errors that would plague a purely continuous process.