AI will advance rapidly but will not reach superintelligence

Summary

Nathan Lambert argues that efficiency in training and inference infrastructure will increase at a near-exponential rate over the next few years. However, he notes that these gains will not translate into general superintelligence or a qualitative leap in model capability. Lambert reframes the concept of 'RSI' as 'parallelized, AI-assisted language modeling.' He emphasizes that the bottleneck is shifting from engineering to idea generation, rather than from engineering to research.

Why it matters

The article offers a balanced counterweight to both the 'AGI is around the corner' and 'AI has plateaued' camps. It suggests that the most transformative changes in the near term will be economic rather than cognitive, as the cost per unit of intelligence is collapsing. It also highlights a practical blind spot: the RL environment data market is generating revenue streams between $100 million and $1 billion, even though practitioners openly label the outputs as low quality. This indicates that much of the current 'progress' is built on a fragile foundation. For practitioners and investors, the distinction between efficiency gains and capability leaps determines whether to bet on infrastructure investments or frontier model breakthroughs.

Highlights

• Lambert predicts that pre-training architecture and data selection research could be largely automated within 2-3 years. He notes that this process will be driven by verifiable metrics such as tokens per GPU and cost per response.
• A Jevons paradox effect is expected: cheaper and faster inference will expand agentic demand rather than saturate it. Value will shift from raw model performance to agent orchestration (e.g., Meta's Muse).
• The RL environment data market is booming, with multiple companies generating revenues exceeding $100 million and $1 billion. However, practitioners widely label the average output as low quality, pointing to a significant and addressable quality gap.

Critical take

Lambert's claim that infrastructure acceleration will not change the 'nature' of models relies on the assumption that scaling inference-time compute and automating architecture search are categorically different from the self-improving feedback loops that could produce genuine leaps. Lambert hints at this boundary but never tests it. Additionally, he presents the timeline for automating pre-training research within 2-3 years as 'reasonable,' yet ignores how quickly this automation can feed back into the human-centric research process he claims will remain dominant.