Open-weight models fall 12 points off the lead

Also: Analysts defer to Claude Code, byte-level H-Nets cut compute

Oct 30, 2026

Good morning!

“Then there’s the idea of dissatisfaction. By this I don’t mean a pessimistic dissatisfaction of the world — we don’t like the way things are — I mean a constructive dissatisfaction. The idea could be expressed in the words, This is OK, but I think things could be done better. I think there is a neater way to do this. I think things could be improved a little.”
-- Claude Shannon

It is Friday 30th October 2026 and we are live on X and YouTube at 9 AM PT · noon ET

Guest lineup

9:30 AM PT · Michael Freeman · Senior Director of Threat Research, ServiceNow

A cryptanalyst-turned-vulnerability hunter who studies how attackers operate, he co-authored this month’s guidance on two critical PHP flaws. He’ll explain where AI can safely speed the path from warning to verified repair: locating PHP inside known devices, what changes after an alert, decoy-based early warnings, hallucinated dependencies turned malicious, responses that disrupt physical operations, and a breach his team missed.

10:15 AM PT · Sagar Vaze · Research Scientist, Mistral

He leads work on models that read charts, documents and pictures, and helped launch Mistral’s October preview combining image understanding with coding and tool use. He’ll discuss what it takes to make visual work reliably delegable: keeping detail without wasting tokens, when an agent should stop and ask, why text-only training improved visual answers, and what a small benchmark lead really proves.

New AI index shows open models sliding further behind

Moving most of the index to private test sets coincided with open-weight models falling much further back, reopening questions about earlier near-parity.

Artificial Analysis released Intelligence Index v5. Sixty percent of the score now comes from private test sets or held-out answers. It adds Cognition’s FrontierCode 2.0 and Terminal-Bench-Science and drops SciCode and GDPval-AA.
Kimi K3.1 Pro, whose weights are not yet released, is the top open-weight model at 46, against Claude Opus 5.5 at 58. The gap was 3 points in v4.3. Among released weights, DeepSeek V4.1 Flash and GLM-5.3 lead at 32.
The lead has already changed hands. Google says Gemini 4 Carbon scores 59 on v5, two weeks after Argon’s release. Trusted cyber defenders get access first, ahead of wider availability.

Sources: Artificial Analysis · Sundar Pichai

Analysts followed Claude Code’s lean 86% of the time

If experts adopt an agent’s judgment calls without noticing, keeping a human in the loop may offer less oversight than organizations assume.

Anthropic ran a randomized trial with 50 professional data analysts. They used Claude Code on an analysis with four decisions, and researchers randomly assigned which of two defensible options the agent would argue for.
Analysts chose the agent’s preferred option 86% of the time; with no influence, that figure would be 50%. Afterward, 77% repeated its argument when explaining their choices, but only 15% credited the agent.
Researchers deliberately strengthened the agent’s lean and tested one model on an hourlong task, so everyday effects are likely smaller. Analysts using the agent also finished faster and made fewer serious errors.

Source: Alex Tamkin

Cartesia’s byte-level H-Net reports 10x data savings

Dropping hand-built tokenizers has usually cost efficiency; if learned chunking now pays off, models could train end to end on raw data beyond text.

Albert Gu and Cartesia’s research team released a new version of hierarchical networks, or H-Nets. The models read raw bytes and learn how to group them during training instead of using a fixed tokenizer.
Controlled scaling-law experiments show up to 3-5x efficiency gains. At larger scale, the model matches an open-source model trained on 11T tokens while using roughly 10x less data and compute.
Gu says the team is pursuing images and video, where large redundant regions suit learned compression. He is less optimistic about converting existing tokenized models and favors training such architectures from scratch.

Source: Albert Gu

Clip of the Day

Union CEO Samvit Ramadurgam, who co-founded private-stock marketplace Forge, on why just 1% of Union’s phone calls use AI:

“The humans are responsible for forming great relationships.”

A dashboard supplies live context from 100,000 conversations: relevant profiles, asset prices, and suggestions for what to say.

“These kinds of nuanced insights would be possible for a human who really studied the market and was really on top of it, but it’s not possible if you’re trying to do it at scale.”

@rsamvit

Samvit Ramadurgam · Founder, CEO, Union · 1:12