Astar: a model that redesigns other models from their commit log
Two stories this morning about who gets to decide what an AI system does next: a training pipeline that took the idea step away from Alibaba's engineers, and a lab chief telling the world's two presidents what to agree on.
A model with 8 billion parameters now proposes the next upgrade to Alibaba's ad-recommendation system more reliably than the human engineers who used to do it, according to a paper from Alibaba and Zhejiang University posted to arXiv. The system, Astar, is trained on the one dataset industrial AI teams already have and never use: their own commit history, where every entry pairs a proposed change with the loss curve and business metric it produced. It has been deployed on the recall model behind Lazada's advertising system, where it guided 20 consecutive iterations over two weeks without human intervention, lifting offline Hitrate@200 by 23.6% and, in an online A/B test, GMV by 4.86% and advertising revenue by 1.82%. On single-proposal success rate measured against real execution, the 8B model scored 0.6786 against 0.3229 for human experts and 0.3071 for the strongest general-purpose large language model; even the 0.6B variant beat both, at 0.5435. The paper is precise about what got automated: implementation, training and evaluation were already automated, and proposing effective evolution directions was the last step that still relied on senior experts.
The interesting part is not the headline number but the mechanism. General-purpose models gave generic advice because the expertise lives in a specific system's data distribution and training dynamics, not in any paper, so the team trained a reward model that predicts whether a proposed change will help — AUC 0.8487, against 0.6142 for human experts — which turns a multi-day validation cycle into a roughly one-second pre-screen. That is what makes the throughput claim real rather than rhetorical: the bottleneck in model iteration was never writing the code, it was paying for the experiment. The case studies show what the proposals look like. In one, the model found that a shared feed-forward network was forcing one parameter set to fit both coarse and fine codebook levels, and proposed a mixture-of-experts replacement that stays FLOPs-neutral because each token still activates exactly one expert. In another, it diagnosed that Muon's Newton-Schulz orthogonalization was amplifying noise directions, and proposed a spectral denoising step using a random-matrix-theory threshold rather than the usual fallbacks of tuning the learning rate or reverting to AdamW. Both were accepted, and both are the kind of judgment call that was previously the job description of a senior algorithm engineer.
The honest caveat is scale: this is one production system and a self-reported comparison against a small number of human experts, and the paper's efficiency claim — ideas proposed 10 to 100 times faster — measures a loop a large lab can run continuously and most teams cannot. It is also worth separating from this week's pacing debate. Astar does not improve itself; it improves a recommender, and a human still owns the roadmap it optimizes. What it establishes is narrower and more useful than the singularity framing: the proposal step in machine-learning engineering is now trainable, given enough logged history.
Sam Altman has a price for a US-China AI accord: the Nobel Peace Prize. In an interview with Fortune editor-in-chief Alyson Shontell, the OpenAI CEO said both Presidents Trump and Xi Jinping would win the prize if they agreed on shared standards and testing for frontier AI development, and that the deal need not be complicated. "Even if just the US and China could agree on some shared standards and testing for development of this technology, I think that'd be a wonderful accomplishment that the two of them can deliver," Altman said. "I don't think this is hard. This is like a one-page document."
The substance is in what he declined to endorse. Shontell asked whether a straight ban on recursive self-improvement, until the technology could be made collectively safe, would suffice. Altman backed the spirit and flagged the mechanics: "It's very hard to say what a ban on RSI means," he said, adding that such a ban probably would not be enough on its own. That is a notable position from the head of the lab whose own pacing statements this week were widely read as a response to a rumour about exactly that capability — and it is a lab chief asking for a verification regime rather than a prohibition. Shared standards imply shared testing, which implies inspectors, which is the part no draft has yet costed.
We covered the diplomatic track when the first dedicated US-China AI safety talks of this administration were scheduled — US, China schedule first dedicated AI safety talks of Trump's second term — with Treasury Secretary Scott Bessent leading the US side ahead of the September 24 Trump-Xi meeting in Washington. Altman's interview lands eleven days before that meeting, and the meeting is the venue he pointed to. A frontier-lab CEO publicly pitching summit deliverables is a lobbying position, not neutral commentary; the useful read is that OpenAI now wants a treaty it can predict.
If a small model can out-propose your senior engineers at redesigning a production system, is the job that's left the one that decides what the system optimizes for? Tell us in the comments.
