Intelligence Is Time, Not Parameter Count

An LLM is not simply a model that predicts the next token, but rather a computing machine that solves problems through chain-of-thought. The key point is that this ability lies not in simple inductive generalization, but in transductive inference, which solves new problems by leveraging both the given data and memory together.

The authors explain this through the contrast between system-1 and system-2. Inductive generalization, which requires fast responses, is closer to system-1, while transduction, which performs reasoning of varying length for each query, operates like system-2 — and they view LLM reasoning as precisely this form of computation.

From this perspective, Solomonoff's 1964 universal optimal algorithm and Levin's 1973 universal search are important starting points, but neither is practical. In their new paper, the authors argue that when training is designed to reduce inference time, the model comes to learn not just simple statistical structure but algorithmic structure as well.

One of the most important claims is the relationship log speed-up = I(h : D). Here, h is the solution to a new problem and D is the training data, and the speed-up of the solution is explained by the algorithmic mutual information between the solution and the data. In other words, the more a model is trained to reduce inference time, the more meaningful information its weights come to contain.

This logic leads to a reversal of conventional scaling laws. The approach of continually increasing parameter count may, past a certain point, push the model into a savant regime, approaching a state where it passes benchmarks through brute force rather than through learned insight. The authors state that the key variable of intelligence is not scale but time.

In summary, model design and training objectives should shift in the following directions:

• Design models to predict the marginal value of additional computation at the inference stage
• Include a complexity cost in the training objective to encourage reduced inference time
• Enable computational resources to be adjusted according to user or environmental conditions

Here, the cost of time is not absolute but environment-dependent. Some tasks, like scientific discovery, allow for decades, while others, like high-frequency trading, depend on milliseconds. Agents should therefore call smaller, specialized models suited to the environment, or allow users to adjust inference cost themselves.

Finally, these ideas also change how AI coding is done. The authors cite the open-source library AI Functions as an example, explaining that combining function bodies written in natural language with pre-/post-conditions makes it possible to implement high-level planning and low-level feedback control simultaneously. The conclusion of the piece is that, as a result, AI agents should be trained and evaluated to find better answers in less time.