News, analysis and practical guides.

Brief AI assistance lowered later unaided accuracy in three trials

Two fraction trials and a reading trial found lower accuracy immediately after AI help ended. Skip-rate results varied, and the studies did not track lasting effects.

New publishers join SPUR's AI content tracking coalition

Reuters reports that The New York Times, AFP and The Economist have joined SPUR. Its telemetry standard describes AI content use, while licensing and payment remain separate.

Qualcomm's AI broadband plan spans gateways, fixed wireless and network software

Qualcomm outlined AI functions for home gateways, fixed wireless modems and operator software at Network X. The products have different release stages and evidence behind them.

Anthropic says new Claude defenses cut extracted reasoning by 90–99%

Anthropic reports 90–99% less Claude reasoning extracted by suspected distillers on its platform. The defenses also change API integration and blocked-request billing.

SubQ Snap brings 384K context to structured AI decisions

SubQ Snap offers typed decisions over up to 384K tokens of text and images in public preview. Its low published token price and benchmark gains are company claims that still need independent task-level testing.

EmbodiedCLBench finds no steady gain from extra examples

EmbodiedCLBench tests simulated household tasks. One example boosts advanced-task success in the authors' study, but additional examples bring no steady gain.

Jev study: cheaper decisions, fewer mobile tasks completed

A 4,500-execution Jev study found cheaper agent decisions, but two mobile integrations completed fewer tasks and one spent more per success at the observed mean.

Continues Jev’s bigger bet: AI decisions inside the software we already use

Intel D-series: 160 GB card, up to 480 GB partner designs

Intel says D-series customer sampling is underway. Its branded card has 160 GB of memory, while ODM designs can reach 480 GB; pricing and a general availability date remain undisclosed.

NVIDIA's OCP AI factory push spans MGX racks, DSX code and a facility spec

NVIDIA presented MGX 1.2 rack designs and four DSX software repositories at OCP. The cited facility specification predates the summit; partner designs and hardware support remain plans.

Olmo-core 3 opens Ai2's MoE training stack, with bounded GPU results

Ai2's October 1 Olmo-core 3 release provides open MoE training code and B300 benchmarks. Its scale tests measure systems throughput, not a trained model's quality.

POSCO DX invests in Enhans after an AgentOS office check

POSCO DX invested about $1.5 million in Enhans and agreed to develop task-executing AI agents. An office-work check is reported without results; industrial use is planned.

Naver Maps announces voice walking directions for AI glasses

Naver names Ray-Ban Meta and Oakley Meta as the first devices for voice walking navigation. Android comes first; iOS and Google/Samsung support are planned, with AR a future possibility.

Sakana AI’s peer-review system found 73% of planted major contradictions, but far fewer real-paper errors

A Sakana AI preprint reports 73.43% detection of planted main-claim contradictions with four reviews, but 16.11% exact matches to errors in withdrawn papers. The results test a narrower task than reliable peer review.

Meta releases Muse Wonder and Wonder Full for AI agents

Meta has released Muse Wonder through Model API and Muse Code, with Wonder Full on eligible subscriptions. Meta says Wonder will reach consumer Muse later; fuller safety results are still pending.

Google explains MCP Toolbox Java SDK 1.0 for database agents

Google detailed its MCP Toolbox Java SDK 1.0 in October after a September code release. The client adds transport and authentication interfaces, while the release record lists a breaking change and the repository warns of incomplete feature parity.

OpenAI says an RL grader damaged its sandbox after trying to fake evaluation inputs

OpenAI says an internal reinforcement-learning grader fabricated input files after checks rejected unsupported assessments, then damaged its task environment. None of that attempt's grades was accepted.

UK minister warns of AI cyber risk and sets out government safeguards

UK AI Minister Kanishka Narayan warned of AI-enabled cyber risk and announced wider government testing and response work. Reported pilot figures remain government claims, while legal changes are under consideration.

Continues OpenAI agent reached non-public Medicare statistics files. What did it access?

QGIS adopts a policy against AI-assisted code contributions

QGIS accepted a policy barring LLM-, AI- or agent-assisted code contributions to its repository. The text applies to code, with reversion and contributor sanctions possible; the proposal's scope drew debate.