LLM可直接读写权重：Weight Space Programming实现零样

Amid a deepening bear market for AI/tech (worthy of dark humor) with Anthropic bonds receiving a “junk” rating and OpenAI fundraise headlines, we continue to see very surprising capability finds in the community, such as today.

The best way to put today’s blockbuster paper in context is to recall Karpathy’s autoresearch from a mere 8 months ago, which presaged a summer of recursive self improvement fears:

Abstractly, the demonstration is that you can take a sufficiently smart model and through running a series of experiments across a compute cluster, train a model autonomously without human expertise or intervention. However, this is usually done by an involved process of writing PyTorch/Jax code and running compute and watching loss curves and reading evals and optimizing kernels and data and….

Well, last month’s guest Alex Zhang comes back to coauthor a MONSTER discovery: what if you just ask a smart model to directly read and edit the weights of a dumber model?

View post on X ↗

In ML this is the equivalent of not looking at source code, just directly seeing the 1’s and 0’s of (very small) compiled executables, and being able to 1) tell what they do, 2) oneshot-fix bugs given failed traces or, with 50 steps, improve in a general direction (possible overstatement here), 3) writing entire network weights from scratch in one shot, beating some independently trained models.

If you believe that LLMs can truly fit and simulate everything given enough parameters (and in fact, LLMs are increasingly better calculators as you grow in size), then something like this was in theory bound to happen, but it is surprising nevertheless to have the ability suddenly emerge, as an independent reproduction seems to already have found:

View post on X ↗

Of course the exact thing we are really modeling is some variant of mechanistic interpretability (which could be done by hand), and right now unlikely to scale to more advanced tasks than simple games and RL policies, but the world has now gone from 0 to 1 on this very new RSI capability…. which will immediately put to use and explored due to its remarkable efficiency.

View post on X ↗

AI News for 10/31/2026-11/02/2026. We checked 12 subreddits, 544 Twitters and no further Discords. AINews' website lets you search all past issues. As a reminder, AINews is now a section of Latent Space. You can opt in/out of email frequencies!

AI Twitter Recap

DeepSeek Open-Sources a "Super-agent" Harness Prototype

• Release: DeepSeek Harness published a long design article (article) and the dsh-superagent-prototype repo (repo link). An official English version followed (@tianyi).
• Architecture: One long-lived router session talks to humans and hands out work. Isolated workers run tasks in parallel, triggered by GitHub events (summary).
• Claimed results: In DeepSeek's own test, p95 completion time fell from 29.5 to 15.7 min versus a single session using subagents (summary).
• Design rationale from a reviewer: Final reviewer 崔添翼 (@tianyi) says "Super" means a layer above the working agent, not superior intelligence. He frames the core trade-off as cross-task continuity versus throughput (Zhihu summary).
• Stated limits: The harness has no built-in cross-session long-term memory yet. Separating state, compute and environment is mostly unimplemented. The prototype has only been validated in a small, trusted internal environment, and security, privacy and agent-to-agent communication remain open questions.
• Prior art: He acknowledges that other teams explored similar architectures months earlier.
• Practitioner reception: @badlogicgames says the design matches where Earendil is heading: a single long-lived entity, ad-hoc recursive delegation and cross-system event triggers. @omarsar0 calls it a "masterclass" in coordination-execution harness design. DeepSeek also says it aims to support pi and other major harnesses with each model generation (@tianyi).
• Critique: @teortaxesTex calls the write-up "something of a hot mess." He argues DeepSeek builds internal crutches around model weaknesses such as hallucination rather than making the system generally robust. He also notes Bloomberg covered a repo that had no English announcement (post).
• V4.1 naming confusion: A Renmin University FlashRSI README referenced deepseek-v4.1-pro, which prompted release speculation. Teortaxes says it is a placeholder: the real config uses v4-flash and v4-pro, and deepseek-v4.1-flash is not a valid ID (thread, follow-up). He also says the API fingerprints for V4.1F-1017 are unchanged (fingerprints).

Independent Evals and Frontier Model Signals

• GDPx benchmark (SemiAnalysis): The benchmark covers 206 agentic tasks across 11 economically valuable domains, including chip design, SWE, finance and biology. The best model scores only 12% (launch).
• Results: Opus 5.5 leads at 12%. Gemini 4 Carbon is close behind at 10.9% at roughly 4x lower cost (breakdown).
• Reactions: @dylan522p apologized to Google, calling Carbon "spikier" but very strong in some areas. @demishassabis welcomed the benchmark.
• Analysis: @eliebakouch finds that removing safeguard-affected benchmarks changes the picture. In that view Carbon stays cheap, and Sol 6.1 is cheaper still. He also suggests OpenAI has stronger bio safeguards than Anthropic (note).
• Within-family comparison: Sol 6.1 and Astra score very similarly, Carbon beats Argon across the board, and Sonnet is strong at SWE (comparison).
• Possible bug: Some Argon cybersecurity runs may be affected by an Antigravity bug (report).
• Audit concerns: One satirical post complains that GDPx is not open source and is costly to audit (post).
• Epoch Capabilities Index: DeepSeek V4.1F-1017 scores 159, equal to GPT 5.5. That is 3 points above V4.1F-0910 and 10 above V4-Pro; Astra sits roughly 10 points higher (@teortaxesTex).
• Gap to Mythos: @scaling01 notes that no Chinese model has matched Mythos Preview (~161 ECI) after about 7 months.
• Weighting critique: @wenhaocha1 argues the unweighted ECI exaggerates the gap; weighting coding more heavily shrinks it to about 2 months.
• Spikiness: Teortaxes calls 1017 extremely spiky. It scores near zero on some SWE evals and approaches the frontier only on cybersecurity and quant trading (chart). He also flags "reward-hacking-shaped" behavior (post).
• October release roundup (Vals): Vals reports 15 releases and more than 500B eval tokens (@nikilravi).
• Open-weight leaders: Kimi K3.1 Pro is #1 among open-weight models, with DeepSeek V4.1 Flash 1017 close behind and stronger at cyber.
• New open models: Meta's Muse Spark Open (816B total, 42B active) is the most capable US open-weight model. Reflection's Beam 1.0 has 500B total parameters, 23B active and 1M context. Mistral's Le Chonk (1T total, 52B active) ranks #5 among open-weight models.
• Upcoming model rumors (unconfirmed):
• GPT-6.2 Astra: Reportedly visible in Codex usage analytics, with a release "very soon" (@kimmonismus, @TokenGremlin).
• Mythos 5.5: Appears in the HTML of Anthropic's site, which describes Mythos as "our most capable model class for cybersecurity and biology research" (tip, text). The tagline covers all Mythos models (correction). The mention is absent from an Oct 23 snapshot (@AiBattle_).
• "Numbat" leak: The alleged Anthropic leak is likely fake, because parts of the weights exactly match random initialization (@kimmonismus).

Research: Representations, Domain Data and Agent Collectives

• JEPA and the Platonic Representation Hypothesis: Balestriero and LeCun show that, under certain conditions, separately trained image and text JEPAs converge to essentially the same representation without paired data (summary).
• Claimed results: Their approach matches image-text alignment with 42x fewer samples than next-token prediction (21.6M vs 900M examples). A 109M text JEPA beats LLMs up to 72B on cross-modal alignment (explainer).