Section

Models & Research

Latest Models & Research News

Anthropic Researchers Tested Their Interpretability Tools and Found No Uplift at All

Sparse autoencoders, natural-language autoencoders and activation oracles were measured against a baseline that just reads the transcript. None of them beat it.

11h ago

An alphaXiv preview card showing the paper title 'Would this change your answer? Evaluating Explanations of LLM Behavior In The Wild with Counterfactual Experiments' beside the first page of the paper

AI Safety

Typos and Paraphrases Add Up: Two Researchers Steered Frontier Models With Invisible Cues

Individually meaningless prompt choices combine almost additively. Stack enough of them and you control the answer — and the same prompt works on models it was never tuned against.

11h ago

An alphaXiv preview card showing the paper title 'Model Hypnosis: Strong control of AI via additive subliminal effects' by Enric Boix-Adsera and Benedict Tessler, beside the first page of the paper

AI Models

DeepSeek's New Prices Took Effect at 16:00 UTC. Cache Hits Cost Twelve Times More.

DeepSeek twice warned of a "significant" increase without naming a number or a date. Both have now arrived — and the largest rise is not the one being quoted.

1d ago

AI Agents

A Huawei-Backed Open Project Shipped a Seven-Agent Office Swarm — to HarmonyOS PC First

WorkSwarm puts the human inside the agent team rather than in front of it. The performance claims are the vendor's own, and there is no new model here.

2d ago

AI Agents

Tencent Wired DeepSeek's Agent Framework into QQ. Install Takes Three Steps.

Two days after DeepSeek released Harness as a developer preview, China's largest chat platform made it a plugin — npm install, launch, scan a QR code.

2d ago

Google AI

Google Open-Sourced a Compiler That Runs AI Models on Data It Cannot Read

HEIR converts pre-trained models to run inference directly on homomorphically encrypted inputs. Four reference applications ship with it, and none of them are a hosted product.

2d ago

AI Models

Alibaba Put a 27B Model Under Apache 2.0 and Says It Beats Its Own Bigger One

Qwen3.8-27B is dense, not mixture-of-experts, carries a 262,144-token native context, and arrives with every benchmark self-reported.

2d ago

The Hugging Face model-card graphic for Qwen/Qwen3.8-27B.

AI Models

Zhipu Says It Took Terminal Bench From 4.6 to 28.3 Without a New Base Model

GLM-5.3 sits on the same 753B foundation as GLM-5.2. Every claimed gain comes from post-training — and the weights are not out for two more weeks.

4d ago

Zhipu's own bar chart headed LLM Performance Evaluation, comparing GLM-5.3 with GLM-5.2, Kimi K3, Fable 5 and GPT-5.6 Sol across six benchmarks including Terminal Bench 3.0 and DeepSWE.

Google AI

Google Shipped Gemini 3.7 Flash Three Weeks After 3.6 — at Half the Price

Enterprise automation scores nearly doubled and long-horizon software engineering went from 49.0% to 65.3%. Introductory pricing is $0.75 per million input tokens until New Year.

4d ago

AI Models

DeepSeek Updated Its Flagship and Warned Prices Are Going Up "Significantly"

V4-Pro-0813 adds the Responses API and tool calls. The more consequential line is a note on the pricing page about an imminent increase.

5d ago

Open Source AI

A 40-Minute Window on PyPI Reached 434,000 Build Pipelines

New research reconstructs the March LiteLLM compromise: 2,500+ organisations potentially exposed through a file that ran whether or not you imported the library.

5d ago

An illustration of a glowing blue chain against a dark circuit background, with one link breaking apart and red warning triangles around the break.

Google AI

DeepMind Put Sign-Language Translation on a Phone. It Took 100,000 Hours of Video.

SL2T ships in Gboard and Live Transcribe on the Pixel 11 — the first sign-language model in a consumer product. It launches with one language pair.

5d ago

A dark interface panel showing a signing person rendered as coloured pose-tracking lines over the body and hands, beside a translation box reading The Galapagos Islands have many different animals.

AI Models

xAI Shipped Grok 4.6. It Is the Same Foundation Model, Trained Harder.

An Intelligence Index of 61 and a 500K context, built by post-training Grok 4.5 rather than pretraining anything new. On coding it still trails GPT-5.6.

5d ago

The words Grok 4.6 in white type over a dark grey field with a soft swirl of light curving behind them.

AI Safety

North Korean Operators Installed Their Own LLMs. That Is the Point.

Ollama, GPT4All and Msty found inside Kimsuky infrastructure, with a retrieval index pointed at stolen documents. Nothing reaches a vendor's abuse team.

6d ago

The North Korean flag overlaid across a laptop keyboard, its red star centred over the keys.

AI Models

Meta Put a 30B Model Under Apache 2.0. There Was No Blog Post.

Muse Glimmer runs on one consumer GPU and carries a genuinely permissive licence. The release went out through Hugging Face and a Zuckerberg essay.

6d ago

Mark Zuckerberg photographed in a crowd at a public event, wearing a black T-shirt and gold chain.

AI Agents

Sixty Claude Subagents Improved a Riemann Bound. It Is Not a Proof.

An unreleased model raised a lower bound on zeta zeros from 41.6% to 67.2% over about a day and a half. Anthropic says the technique will not lead to a proof.

6d ago

Anthropic's research illustration: a field of small scattered marks on cream paper beside a drawing compass resting on graph paper.

Open Source AI

Torvalds Called a 400-Commit rc7 the New Normal. He Blamed AI Review Tools.

Release candidates are supposed to shrink as a cycle ends. This one did not, and the kernel's maintainer says the bugs are being surfaced by machines.

Aug 10, 2026

A gentoo penguin standing on a beach with its flippers spread, photographed against the surf — the mascot image The Register ran with the release.

Google AI

DeepMind's Cyclone Model Gained a Day of Warning. Intensity Is Still the Hard Part.

Published in Nature and built with the National Hurricane Center, WeatherNext Cyclones cuts three-day track error to roughly 100 km. The weights are on GitHub.

Aug 9, 2026

A flat illustration of a browser window over a world map, ringed by weather, clock and cursor icons

AI Safety

Only 26% of AI Security Patches Were Clean. Another 20% Fixed the Bug and Broke the App.

1Password's research lab graded 6,080 model-written patches against six recent CVEs. More than a third of the successful ones were rated fragile.

Aug 9, 2026

AI Models

Alibaba Wants a Cut of What You Earn From Qwen. The Weights Stay Free.

The next open-weight release is set to carry a revenue-share licence for large commercial users, following Moonshot's Kimi K3 terms. Alibaba's own percentage is not set.

Aug 9, 2026