This is an all-in-one crosspost of a scenario I originally published in three parts on my blog, No Set Gauge. Links to the originals:

Thanks to Luke Drago, Duncan McClements, Theo Horsley, and Bilal Chughtai for comments.

2025-2027

Below is part 1 of an extended scenario describing how the future might go if current trends in AI continue. The scenario is deliberately extremely specific: it’s definite rather than indefinite, and makes concrete guesses instead of settling for banal generalities or abstract descriptions of trends.

The return of reinforcement learning

From 2019 to 2023, the main driver of AI was using more compute and data for pretraining. This was combined with some important "unhobblings":

Post-training (supervised fine-tuning and reinforcement learning for instruction-following) helped the LLMs be usable without difficult prompting.
Starting in 2024, Anthropic showed that judgement and taste in data curation—and the evaluation metrics that guide data curation—could give you a "magic sauce" effect in perceived LLM quality.

Most real-world LLM uses, of course, involved generating a sequence of tokens to try to achieve some task. So there were a lot of untapped gains from doing reinforcement learning (RL) for performance on concrete domains, rather than just RL for the models following instructions and being "safe"—i.e. a combination of avoiding PR hazards, and preparing for misuse mitigations on actually capable models down the line.

OpenAI fires the starting gun in 2024 with the release of o1, which was based on RL on chains-of-thought (COT), i.e. the model is trained to reason step-by-step towards correct answers, i.e. "test-time compute" in the horror-filled annals of machine learning jargon. In late 2025 they release “GPT o5” (“GPT” to normal people, and “o5” to those keeping track of the version number), a model which can take text, image, audio, video, computer screen state, real-life footage, whatever, process and understand it (choosing itself whether it should do chain-of-thought reasoning before answering or not), and output text, image, audio, video, computer actions.

Whereas the labs had spent almost four years racing down the scaling graph on pretraining compute, they had not yet done so for COT RL, ...

Fernando Abiti

Fernando Abiti