Lambert Expects Rapid AI Progress, but Not a Path to Superintelligence

Nathan Lambert, a researcher known for his work on post-training and open models and the author of the Interconnects newsletter, put his position in the essay's title: he expects rapid AI progress, but not movement toward general superintelligence. The trigger was conversations with leading researchers in the industry. Many of them are convinced that within a few years AI will be better than they are at their own jobs. Lambert found this surprising, although at first he could not explain his own distrust.

Only the introduction was available in the version that reached the feed. What follows is therefore a breakdown of the thesis and its context, not of the author's full argument.

Two Forecasts That Are Easy to Confuse

The statement "AI will be better than me at my job" and the statement "we are building general superintelligence" sound similar, but they are tested differently. The first is a forecast about a specific profession: researcher, engineer, analyst. It can be measured on benchmarks, on closed tasks, and by how fast agents close tickets. The second is a forecast about a system that surpasses humans in every domain and expands its own capabilities.

Lambert separates these expectations. He considers rapid improvement of models on narrow and mid-level tasks likely. By his logic, it does not follow that the trajectory leads to a single, universal intelligence.

Why Insiders' Words Should Be Read With a Correction

People at top labs see models before everyone else and watch quality grow every day. That is a valuable source, but it comes with distortions:

• a researcher judges progress on tasks they are automating themselves, so the sample is biased;
• work inside a lab consists of well-formalized tasks with fast feedback, and those are the areas where AI does best;
• "better than me at my job" often means speed and volume, not posing problems and choosing a direction;
• the "few years" timeframe in these conversations stays vague, with no criteria for checking it.

This does not refute the forecast, but it shows where it may fail to generalize. Work that is convenient to automate in a lab is not the same as the work of a doctor, a lawyer, or a product manager.

What the Text Itself Tells Us

• Author: Nathan Lambert, Interconnects newsletter.
• Publication date: 9 October 2026.
• Main thesis: rapid progress is expected; a path to general superintelligence is not.
• Source of doubt: disagreement with the forecasts of top industry researchers.
• Horizon cited by the author's interlocutors: a few years.
• The available excerpt gives no detailed arguments or numerical estimates.

How to Use This Debate in Practice

For teams planning AI adoption, the difference between the two forecasts has direct consequences. If Lambert is right, it makes sense to plan for gradual improvement of tools: agents for code, analysis, and support. Betting that a single model will replace entire functions looks riskier.

In practice, there are three things worth doing. First, measure the effect on your own tasks, not on public leaderboards. Second, budget for frequent model changes: the pace of improvement is high in either scenario. Third, when reading statements from lab executives and researchers, ask which specific set of tasks they have in mind. It is worth reading Lambert's full argument in the original: that is where you can see what underpins his confidence that rapid progress and superintelligence are different trajectories.