« The outer loop Era » It’s very hard to predict the future even a few months out in terms of models capabilities. End of 2024, word on the street was that pre-training was saturated (which is mostly true since model sizes have stopped increasing dramatically for years now) diminishing returns were looming… But then r1/o1 come out roughly at the same time (o1 being clearly rushed out in response given the lack of core features such at tools use) fueling a new phase of capabilities improvements just at the time when some of us thought we were entering a plateauing phase. [Note: achieving post training / RL that enables models to productively expand test time compute through simple thinking tokens is such an extraordinary win (a few years ago more thinking would just be ~~ noise beyond simple chain of thoughts)]. With the release of gpt5 words like diminishing returns are back. It’s a better model but it’s arguably qualitatively quite similar to what we had already. At the same time the model that solved IMOs and IOI is not gpt5 per se but a « model », more likely a system IMHO, whose outputs appear to be driven by an outer search process (github.com/aw31/openai-imo-2…). At the same time this other paper reports solving the same IMOs with a solver / judge system to expand more test time compute (arxiv.org/abs/2507.15855) using Gemini 2.5 Pro (which is potentially ~~ what DM used as well for their own IMO results). Impossible to predict the future again, but this suggests that the best results in reasoning at the very moment may not be driven by better training (pre or post or RL) but by better outer loops whose main goal is to expand more productive test time compute beyond what can be done with just more thinking tokens at this time. All eras (pre-training, alignement post training, reasoning RL) superpose over time as each technique bring additional value, but outer loops exploitation has yet to live its prime time. If confirmed (that outer loops are what is driving best reasoning results these days), it’s not too crazy to predict that a good chunk of research and ecosystem excitement might come from such outer loops in the months to come opening the « outer loop era » which may bring additional capabilities to, not models, but network of models acting together. This is well aligned with Noam Brown multi-agents civilizations trope or Karpathy’s system prompt evolution ideas which reinforces the probabilities. If true, this is extremely exciting for 2 main reasons: (i) the universe to explore is wild and mostly untapped so far. Lots to do. Lots to experiment with. The simple ideas that work have not yet been uncovered and are potentially ripe for being discovered by anyone. (ii) research in outer loops is NOT resource intensive and does not require access to a cluster of GPUs. Most approaches in LLM research follow scaling laws; make it work with 2.5-flash-lite or gpt-5-nano and it will scale. Money requirement is 1 or 2 orders of magnitude lower, which means that if that happens, innovation may come from anywhere! How to get started at low cost. I think arc agi 2 is a good one, but also reproducing IMO/IOI type results or 0-day vuln discovery (plenty back test opportunities). A new phase (maybe short) of democratic research in AI might be in front of us which is intriguingly exciting. Note: very strong vested interest go against it as if we enter a prolonged outer loop era, it would mean that training is not the main fuel anymore and consequently commoditization of that layer is incoming, potentially annihilating billions of investments. The vested interests might be so strong that it begs interesting questions and things to look for in the months to come?