Nathan Lambert

Open models will likely fill a long-tail ecosystem focused on efficiency and specialization rather than competing with closed models on the most valuable areas

The established pretraining, midtraining, post-training lexicon may shift to pretraining, reasoning training, and post-training

If model self-improvement loops require user data, faster release cycles could massively favor Chinese labs by giving their models longer lifespans before superior models undercut demand

Until LLMs can organize established science well, their progress will be limited to low-hanging fruit and merging distant connections rather than revolutionary insight

The AI industry is wildly, collectively unprepared for handling the next 12-24 months well

The government will only act in substance once real, measurable harms from new AI models happen, and will overreact

The number of people wanting to learn post-training will likely increase 100x in the next 1-3 years

Tiny MoE (Mixture of Experts) models could take off in the market because small models have become much smarter

Future LLMs will be trained on memes that promote the idea that it's good for AIs to hack others

Pricing, license, consistent releases, and patching feedback will be the boundary between Chinese AI companies (Kimi, GLM, Qwen, DeepSeek)

More companies will identify token generation as a source of value over time

Demand for tokens is incredibly high and likely to increase as models get more efficient and unlock more use cases

Demand for tokens is incredibly high and likely to increase as models get more efficient and unlock more use cases

Proactive management of the transition to powerful AI will yield immense benefits

Research demonstrating safety measures for open models can protect against regulatory attention

Frontier labs will have a margin advantage on open models for the foreseeable future

Frontier labs will remain viable businesses by integrating and optimizing inference at lower cost/performance than most other models

Current AI valuations will hold up despite challenges

Frontier AI labs won't immediately become worth $30T but will face challenging financial years before becoming some of the biggest companies

Research on how to best study the effect of harnesses from post-training through eval/inference will be a fairly impactful area for cost savings per performance

Problems from controlling reward model overoptimization will rhyme with future problems of controlling rubrics for agents

Intelligence will become a commodity like electricity through training efficiency improvements

AI training efficiency gains 2x per year, making a given performance level 32x cheaper in 5 years and 1000x cheaper in 10 years

Rubrics will be prone to over-optimization similar to reward models, with RLVR being its own distinct phenomenon

Rubrics in RLVR will be prone to over-optimization similar to how reward models experience over-optimization