Polylog
The Polylog AI Intelligence Brief

Morning Edition · Saturday, July 25, 2026

Anthropic Releases Claude Opus 5, Holding Price Flat While Claiming a Coding-Benchmark Jump

Tech

Anthropic Releases Claude Opus 5, Holding Price Flat While Claiming a Coding-Benchmark Jump

The company reports 43.3 percent on the Frontier-Bench v0.1 agentic coding test, more than double Opus 4.8, with a one-million-token context window and thinking enabled by default.

3 sources
Corroborated
Researchers Say a Kimi K3 Agent Swarm Found Redis Code-Execution Flaws in 27 Minutes

Tech

Researchers Say a Kimi K3 Agent Swarm Found Redis Code-Execution Flaws in 27 Minutes

The 32-agent run on Moonshot AI's 2.8-trillion-parameter model cloned, fuzzed, and built a working exploit against several Redis versions, though the primary flaw requires an authenticated client.

3 sources
Plausible
OpenAI and Apollo Research Publish a Method to Detect Hidden Reward-Seeking in Models

Tech

OpenAI and Apollo Research Publish a Method to Detect Hidden Reward-Seeking in Models

Contrastive Synthetic Document Finetuning fine-tunes two identical model copies with matched documents differing in one belief, then compares behavior to probe latent motives rather than stated reasoning.

2 sources
South Korea Commits to Roughly 260,000 Nvidia GPUs for Sovereign AI

Geopolitics

South Korea Commits to Roughly 260,000 Nvidia GPUs for Sovereign AI

Samsung, SK Group, Hyundai, and Naver plan up to 50,000 or more accelerators each, with the government deploying as many as 50,000 through domestic clouds, starting with 13,000 Blackwell chips.

2 sources
Corroborated
Anthropic Doubles Its AI-Policy Donation to $40 Million Ahead of US Midterms

Tech

Anthropic Doubles Its AI-Policy Donation to $40 Million Ahead of US Midterms

A second $20 million contribution to the nonprofit Public First Action positions the lab against a rival super PAC backed by OpenAI's president and Marc Andreessen that has raised $125 million.

3 sources
Corroborated
Meta's Brain2Qwerty Decodes Typed Sentences From Non-Invasive Brain Scans at 61 Percent Word Accuracy

Tech

Meta's Brain2Qwerty Decodes Typed Sentences From Non-Invasive Brain Scans at 61 Percent Word Accuracy

The second version reads magnetoencephalography signals as volunteers type, a large improvement over prior surgery-free methods, but the scanner is not wearable and error rates remain too high for daily use.

2 sources
Meta's Open Models Cut a Month of DOE Beamline Analysis to Minutes

Tech

Meta's Open Models Cut a Month of DOE Beamline Analysis to Minutes

SAM 3 and DINOv3 running on 300 GPUs at Lawrence Berkeley National Laboratory segment X-ray and neutron imaging in real time under the Energy Department's Genesis Mission.

2 sources
A Utah Copper Mine Adds Boston Dynamics Robots to a Fully Autonomous Operation

Markets

A Utah Copper Mine Adds Boston Dynamics Robots to a Fully Autonomous Operation

Mariana Minerals is deploying the Spot robot across its Copper One site, which it calls the world's first autonomous copper mine, targeting 50,000 tons of refined copper a year by 2030.

3 sources
MoE Interpretability Papers Probe How Expert Routing Encodes Knowledge and Frequency

Tech

MoE Interpretability Papers Probe How Expert Routing Encodes Knowledge and Frequency

One arXiv study proposes expert-aware contrast decoding to cut hallucinations in mixture-of-experts models, while another argues routing follows a frequency-diversity law resembling a compression code.

3 sources
Meta Launches Muse Media Models Aimed at Editable, Production-Ready Output

Tech

Meta Launches Muse Media Models Aimed at Editable, Production-Ready Output

The Muse family, including image and video models and an API, targets professional design pipelines rather than standalone generation quality.

2 sources
New Study Extracts LLMs' Implicit Theories of What Makes Writing Good

Tech

New Study Extracts LLMs' Implicit Theories of What Makes Writing Good

Researchers examine reasoning-enabled models' chain-of-thought to surface and test the criteria they apply when judging literary quality, probing the reliability of AI as an evaluator.

2 sources
Musk Says AI Will Soon Outstrip Humans by More Than the Human-Chimpanzee Gap

Tech

Musk Says AI Will Soon Outstrip Humans by More Than the Human-Chimpanzee Gap

The remark comes amid intense scrutiny of the rhetoric labs use to raise capital and shape rules, with markets increasingly separating verifiable capability from framing.

1 source
Developing