Lead story / research

MIT researchers use AI to repair a model, lifting its math-test score to 48.1%

MIT CSAIL researchers say OpenAI's GPT-6 Astra directly edited another AI model's internal parameters, raising its MATH-500 accuracy from 7.9% to 48.1%. Their November 2 report attributes most of the gain to stopping repetitive output—not adding mathematical knowledge. The target, Qwen3.5-0.8B, could finish more answers using abilities it already had.

Astra changed 131,072 weights, about 0.016% of the model, suppressing a numerical direction associated with repetition. Finding that small patch took a substantial search: the strongest attempt ran for 20.3 hours, including 4.7 hours waiting through a service outage, and allowed up to 50 evaluation requests. It was not a single instruction followed by an instant fix.

The benefits extended beyond the headline benchmark. Under the researchers' evaluation conditions, GSM8K math accuracy rose from 13.8% to 67.1%, while instruction-following performance on IFEval increased from 21.0% to 62.3%. But Astra produced no established mathematical improvement in Gemma 3 270M, which lacked Qwen's severe repetition problem. That contrast supports the narrower explanation: removing an obstruction can improve measured performance without adding reasoning abilities.

Reliability and comparative value remain open questions. The work does not establish that direct weight editing consistently beats conventional training on speed, cost or effectiveness, nor does it demonstrate recursive self-improvement—a cycle in which AI improvements help generate further improvements. The researchers have released experiment logs and editing artifacts and say they are developing a benchmark and supporting framework.

launch

OpenAI rebuilds its Git backend to handle more parallel coding

OpenAI says its new Pando backend now serves all development in its shared code repository, after Codex-driven parallel work exposed bottlenecks. It reports nearly fivefold lower median push latency and nearly a petabyte of Git data served daily. Pando lets unrelated branches update independently and reuses download bundles. Git workflows stay intact, but reading several branches does not give one synchronized snapshot.

security risk

North Korean hackers use AI to select crypto targets, investigators say

Investigators say North Korea's TraderTraitor uses AI to identify cryptocurrency businesses capable of yielding large payouts. That assessment comes from techniques in recent hacks; the report does not detail the AI systems involved. The FBI attributes a recent $6.3 billion Binance theft to the unit. ISACA's Chris Dimitriadis argues that responsibility for a theft on this scale belongs in the boardroom, not only with security teams.

launch

Anthropic lets Claude agents stop uploading full conversations each turn

Anthropic's message threads public beta temporarily stores conversation history, so follow-up API requests send only new messages and a prior response ID. The company reports faster response starts for Claude Code requests containing 500,000 to 1 million tokens. Claude still processes the full history, and token charges do not change. Storage is best-effort, so developers must retain their own records and replay the history if a thread fails.

Also worth knowing

•  Alphabet shareholder sues Pichai over alleged unlawful AI training

•  OpenAI introduces a benchmark for AI help with shopping, budgets and childcare

•  Apple is reportedly developing a home robot that could cost $1,000 or more

•  OpenAI will more than quadruple rewards for critical security flaws

•  Anthropic warns faster AI research could outpace its safeguards

•  Musk restricts Grok Bot usage, citing unsustainable costs

•  OpenAI launches a program to cut campus power use when grids are strained

Daily tool drop

6 AI tools worth knowing today

The Internet Had a Point

Apple Store’s hidden extinguisher

Reader check-in

Help shape tomorrow's briefing

One click tells us what to keep, improve, or tighten.