The Informed Layer · Newsscan
Daily briefing

GPT-6 goes broad; agents catch what humans miss

← All issues

Anthropic / Claude ecosystem

Claude Code surfaces bugs developers overlooked for months

A real-world writeup from an ESPHome project shows Claude Code identifying subtle, long-lived bugs that human developers had missed for months. The piece positions the CLI agent as a genuine second pair of eyes on mature codebases rather than a boilerplate generator.

Terence Tao denies claims that Claude solved the Millennium Prize Problems

After Anthropic reported that Claude agents formalised Fermat's Last Theorem in Lean in 11 days, viral claims escalated to assert Claude had cracked the Millennium Prize Problems. Terence Tao has publicly denied ever making such a statement, correcting the record on how far the AI proof work actually goes.

Frontier model providers

Google DeepMind introduces advanced drug discovery AI model

DeepMind has introduced a new AI model for accelerating drug discovery by predicting properties of chemical compounds. It extends DeepMind's AI-for-science lineup beyond weather, biology and materials.

Grok Imagine Video 1.5 agent now available

xAI's Grok Imagine has released its Video 1.5 agent, built on Image 2.0, aimed at improving continuity and storytelling across AI-generated video scenes. The agent focuses on multi-scene narrative output rather than single clips.

DeepSeek releases V4 Vision, an open-source multimodal model under MIT licence

DeepSeek has released V4 Vision, described as its first open-source multimodal model, under a permissive MIT licence. If confirmed, it gives the open-weights ecosystem a frontier-lab vision-language option.

GPT-6 Astra goes live in ChatGPT, Microsoft 365 and GitHub Copilot

OpenAI's GPT-6 Astra is now live across ChatGPT, Microsoft 365 and GitHub Copilot, extending the flagship model's reach into mainstream productivity and developer tools. This broadens availability beyond the initial launch and Azure AI Foundry limited access.

GPT-6 Astra agent builds its own 'simulation within a simulation'

A widely shared demo shows a GPT-6 Astra agent creating a world inside Unreal Engine, populating it with Astra-powered agents, and then building a second simulation inside that world. The experiment highlights open-ended agentic tool use and autonomy.

AI developer tooling & infrastructure

AI Agents Listing launches as first directory indexing agents, MCP servers and skills together

A new directory claims to be the first to index and cross-link AI agents, MCP servers and agent skills in one place. It aims to provide unified discovery across the three layers of the agent ecosystem.

Artificial Analysis v4.2 doubles private test weight to 40%

The v4.2 Intelligence Index raises the weighting of private test sets from 20% to 40% and details new evaluations and leaderboard shifts. The change reduces the influence of potentially contaminated public benchmarks on model rankings.

Google open-sources Mantis, an agentic vulnerability-scanning harness that cuts false positives

Google released Mantis, an open-source harness that uses agentic techniques and sandboxed reproducers to validate AI-generated vulnerability findings. The goal is to filter out false positives before they reach security teams.

Cloud & platform providers

No significant new developments.

AI policy, regulation & governance

EU tech chief says AI guardrails are coming to the US anyway

EU technology chief Henna Virkkunen argued that US AI guardrails are effectively inevitable, as state-level laws converge on EU-style preemptive regulation regardless of federal inaction. Her comments frame a transatlantic regulatory convergence thesis.

Labor plans digital duty of care with algorithm opt-out as Taylor voices deep scepticism

Australia's Labor government will introduce digital duty of care laws requiring social media platforms to let users opt out of recommender algorithms, backed by a $100m-plus enforcement regime. Opposition frontbencher Angus Taylor says he is 'deeply sceptical' of the crackdown.

Industry & market moves

Mistral AI raises €3 billion mega-round

Mistral AI has closed a €3 billion Series C+ round, the largest European AI funding event of the week. The raise reinforces the sovereign-AI capital narrative around Europe's leading frontier lab.

Australia's Nanoveu acquires Singapore chipmaker Embedded A.I. Systems

Australian-listed Nanoveu is acquiring Singapore-based ultra-low-power AI SoC maker Embedded A.I. Systems (EMASS). The deal aims to strengthen Nanoveu's AI-enabled semiconductor portfolio and integrate EMASS technology into EyeFly3D for robotics, medical and unmanned-vehicle markets.

AI product & feature launches

Google launches India-specific AI project for farm productivity and climate resilience

Google is launching its first India-specific agricultural AI initiative, combining satellite-derived agricultural understanding (AnthroKrishi) with DeepMind-backed monitoring tools. The project targets higher farm productivity and climate resilience for Indian agriculture.

Five enterprise vendors converge on the same three-layer agent infrastructure stack

Broadcom/VMware, Citrix, CrowdStrike, ServiceNow and Genesys have each shipped a governance-first, three-layer agent infrastructure stack embedded in their existing platforms — without coordinating. The pattern suggests de facto standardisation of enterprise agent architecture.

Research with immediate practical relevance

UC Berkeley releases CUA-Lite, an open platform for computer-use agents

Berkeley researchers released CUA-Lite (Lite.OSWorld), unifying agents, environments, traces and RL training/evaluation under a single schema and action space. Its VM-free sandboxes reportedly match OSWorld fidelity.

TUM's MIRMI deploys embodied laboratory intelligence to decode cancer cells

Munich's MIRMI-AIM programme showcases autonomous robotic co-exploration of cancer-cell machinery, using embodied laboratory intelligence to map cellular processes. The work aims to open new therapy avenues through AI-driven experiments.

ARC Prize: OpenAI's AGI score hinged on the harness, not just the model

Analysis of GPT-6 Astra's ARC-AGI-3 results shows widely different scores (62.7% vs 99.9%) depending on which evaluation harness was used. ARC Prize cautions that it is not itself claiming AGI, deflating the 'AGI number' narrative around Astra.