A hearing is scheduled for May 19

Claude skews high-income; Meta AI skews low-income

this means that existing estimates overstate the returns to software R&D, and makes the software intelligence explosion seem much less likely.

But I think we have enough evidence to think that software progress might really be several times a year, and to make a best guess contextualized with a lot of uncertainty.

gpt-oss-20b does substantially better than GPT-3 on MMLU, despite using the same amount of training compute.

This means that almost all existing estimates of software progress were misleading.

these estimates rely on an overly conservative estimate of software progress of 3× per year

Synthetic data can help push beyond this — a good example that Millidge raises is the Phi series of models.

If doubling cumulative research effort also doubles compute efficiency, then the returns to R&D are 1. If it quadruples, then the returns are 2.

Almost all the evidence points to very fast software progress: each year, the training compute needed to get to the same capability declines several times — possibly even ten times or more.

AI software progress is about reducing the training compute you need to get to the same level of capability, through better algorithms or data.

context management plus engineering improvements may well push the task horizon to weeks or even months.

if a model cannot learn new things while performing a task, it will struggle when the task horizon grows very long.

"The set of efforts aimed at breaking past the feasible horizon of current techniques."

new techniques may initially underperform existing ones but eventually surpass them — a pattern we've seen repeatedly, most recently in the wave of agentic coding progress

We can treat the task horizon that an LLM can reliably handle as a north-star metric for model progress, analogous to transistor density in Moore's Law

The key reason for the confusion is that people think in terms of methods that each contribute a discrete piece to the system — pretraining, SFT, RL.

I'd view continual learning more as an "arrow" than a "line" — it's the collective effort to push the task horizon that an LLM can reliably handle.

I've been thinking about this stuff for decades, and I had not broached the topic of platonic patterns until until this year. And that's because I think it is now actionable.

When there are imperfections, we rely on users and our active community to tell us how the software is not working correctly, so we can fix it. The way we do that, and have done for 15 years now, is via bug reports. Discussion is great, but detailed bug reports are better for letting developers know what’s wrong.

Alerts are actionable, not informational: We believe that an alert should provide concise and accurate security advice. For an unsafe account, that means resetting your password.

it is worth opening a merge request with the minimal viable change instead of opening an issue encouraging open feedback on the problem without proposing any specific change directly.

The nature of MRs facilitate discussions around a proposed solution to a problem that is actionable. An MR is actionable, while an issue will take longer to take action on.