The AI talent war reaches fever pitch as both Transformer co-author Noam Shazeer and Nobel laureate John Jumper leave Google DeepMind for rivals. A new benchmark reveals even Claude Fable 5 only solves 3% of realistic knowledge work tasks. OpenAI acquires Astral (uv, ruff) to supercharge Codex's Python toolchain.
1. Google DeepMind Loses Two Giants in 48 Hours
In an unprecedented talent exodus, Google DeepMind has lost two of its most prominent researchers within two days. On Thursday, Noam Shazeer — co-author of the landmark "Attention Is All You Need" paper and Gemini co-lead — announced he was leaving Google to join OpenAI as Lead for Architecture Research. Google had spent an estimated $2.7 billion on a licensing deal to bring Shazeer back from Character.AI in 2024. Just 24 hours later, John Jumper — the 2024 Nobel Prize in Chemistry winner who led the AlphaFold team — announced he was departing for Anthropic after nearly nine years at DeepMind.
Both exits shake the foundations of Google DeepMind's talent moat. All eight original Transformer paper authors have now left Google, with two at OpenAI. Jumper's move to Anthropic signals that the company's biology/chemistry research ambitions are a strategic priority. The combined signal suggests a structural shift: frontier AI talent is flowing away from Google toward smaller, more aggressive competitors — Anthropic at a $965B valuation and OpenAI preparing for an $852B IPO.
— VentureBeat · The Decoder · aidailypost.com · Korea JoongAng Daily
🔗 VentureBeat: Shazeer Leaves Google for OpenAI
🔗 aidailypost: Jumper Departs DeepMind for Anthropic
🔗 Korea JoongAng Daily: Anthropic Seoul Office
2. AA-Briefcase: Frontier AI Passes Only 3% on Real Knowledge Work
Artificial Analysis released the AA-Briefcase benchmark, a sobering reality check on AI's capabilities for long-horizon, multi-week knowledge work tasks. The benchmark tests models on complex projects requiring research, planning, cross-referencing documents, and synthesizing findings. Even the best-performing model — Claude Fable 5 — fully solves only 3% of tasks. Cost per task ranges from $0.04 to over $31.
The results expose a critical gap between benchmark performance and production value. While models continue to dominate coding benchmarks (GPT-5.6, Opus 4.8), real knowledge work remains fundamentally unsolved. China's GLM-5.2 ranked third, surprising many. The benchmark was built by experienced professionals rather than researchers, making it more representative of actual enterprise workflows. For AI agents to deliver on the promise of autonomous knowledge work, this 97% failure rate must drop dramatically.
— The Decoder · Artificial Analysis · aidailypost.com
🔗 The Decoder: New benchmark exposes AI struggles
🔗 Artificial Analysis: Announcing AA-Briefcase
3. OpenAI Acquires Astral (uv/ruff) to Supercharge Codex
OpenAI announced the acquisition of Astral, the company behind the ultra-fast Python tools uv (package installer/resolver) and ruff (linter/formatter). The acquisition integrates Astral's team and technology directly into OpenAI's Codex platform, which now has 5M+ weekly active users. Financial terms were not disclosed.
The strategic logic is clear: Codex's primary use case is Python and TypeScript code generation. Making Python developer tooling faster, more reliable, and deeply integrated with AI-assisted coding workflows creates a moat against competitors like Claude Code and Cursor. However, the developer community has raised concerns about the open-source future of uv and ruff, both of which are widely used outside Codex. OpenAI stated it would "continue supporting the open-source community," but specific licensing commitments remain unconfirmed. This acquisition follows Codex's recent expansion into enterprise workspaces, Goal Mode GA, and domain-specific plugins.
— The New Stack · ITHome · CSDN
🔗 The New Stack: OpenAI acquires Astral
🔗 ITHome: OpenAI transaction details
4. Anthropic Fable 5 Ban: The Full Backstory Emerges
New reporting reveals the intricate chain of events that led to the US government ordering Anthropic to globally disable its Claude Fable 5 and Mythos 5 models on June 12. The triggering event was not a direct US-China AI security concern but rather SK Telecom — South Korea's largest carrier and a $100M Anthropic investor — being identified by the White House as "suspected of having ties to China."
Dario Amodei, Anthropic's CEO, was given two options by David Sacks (Co-Chair of PCAST): fix the jailbreak vulnerability or voluntarily de-deploy. He refused both. This refusal escalated the situation from a Korea-specific restriction to a full global ban. The administration now demands "zero jailbreaks" before relaunch — a standard security experts call technically impossible for any frontier model. Meanwhile, Anthropic opened its Seoul office on June 17-18, signing an MOU with Korea's Ministry of Science and ICT, and its MD of International expressed confidence models would return "within days."
— WIRED · The New Stack · Korea JoongAng Daily · LLMBase
🔗 LLMBase: SK Telecom trigger analysis
🔗 The New Stack: US Gov orders Anthropic to pull models
🔗 Korea JoongAng Daily: Seoul office opening
5. OpenAI: Small Doses of "Beneficial Trait" Training Make Models Broadly Safer
OpenAI researchers published findings showing that reinforcement learning on desired behavioral traits — such as truthfulness and corrigibility — makes AI models safer across multiple unrelated domains. The research demonstrated that a modest amount of beneficial trait training improved model safety scores on 44 out of 53 benchmarks, with no degradation in task performance.
The implications are significant for AI safety research. Rather than requiring separate safety mechanisms for each risk category, the research suggests that cultivating "meta-traits" creates broad-spectrum safety improvements. Training on health data also improved deception resistance, a surprising cross-domain benefit. This approach could provide a practical pathway for AI labs to implement safety before deploying frontier models — especially relevant given the ongoing Fable 5 regulatory standoff.
— The Decoder · aidailypost.com · LM Market Cap
🔗 The Decoder: OpenAI beneficial trait training
🔗 aidailypost: OpenAI safety breakthrough
6. Nvidia Robots Train Themselves Through AI Coding Agents
Researchers from Nvidia, Carnegie Mellon University, and UC Berkeley demonstrated a breakthrough in physical AI: robots that learn dexterous grasping by using AI coding agents to generate their own training programs. A fleet of eight robots achieved hit rates of up to 99% in real-world grasping tasks through this self-supervised approach.
The technique represents a fundamental shift in robotic learning. Instead of humans writing training curricula or manually collecting demonstration data, the robots leverage LLM coding agents to analyze task requirements, write training code, and iteratively improve their own performance. This creates a closed loop where the robot generates its own improvement. The method is particularly relevant for warehouse logistics, manufacturing, and domestic robotics — scenarios where programming every edge case manually is infeasible. The research suggests that the convergence of LLM agents and embodied AI may accelerate robotics learning far faster than traditional reinforcement learning approaches.
— The Decoder · VentureBeat
🔗 The Decoder: Nvidia robot self-training
7. Google Returns to Smart Speakers: Gemini AI Assistant Built In
Google has released its first smart speaker in six years (since Nest Audio in 2020), featuring built-in Gemini AI as its core assistant. The device represents Google's renewed push into the voice AI hardware market, competing directly with Amazon's Alexa (powered by Nova models) and Apple's HomePod (now running Gemini-powered Siri via iOS 27).
The Gemini integration enables natural conversation capabilities, contextual awareness across queries, and Google search grounding — an area where the company holds a structural advantage over competitors. The timing coincides with an increasingly heated consumer AI hardware cycle: Amazon and Apple have both refreshed their smart speaker lines with AI upgrades, and the $99 price point from the previous generation Google Nest Audio suggests a competitively priced entry. For Google, the smart speaker is not just a hardware play — it's a data and distribution channel for Gemini.
— LLMBase · AI News · Kersai
🔗 LLMBase: Google smart speaker return
🔗 Kersai: June 2026 AI news roundup
Top comments (0)