Transformer co-inventor Noam Shazeer leaves Google for OpenAI

Transformer co-inventor Noam Shazeer left Google to join OpenAI ahead of its IPO, Anthropic opened its Seoul office with enterprise deployments across Samsung, NAVER, and LG amid ongoing export control tensions, and Google DeepMind published an AI control roadmap proposing tiered safeguards for increasingly capable agents. FERC ordered grid operators to fast-track data center interconnections, Baseten is reportedly raising $1.5 billion for AI inference infrastructure, and today’s ArXiv papers revealed that defense training degrades agent tool-use capabilities and introduced stress tests for coding agents at 100+ interaction turns.

Funding & Business #

Transformer co-inventor Noam Shazeer leaves Google to join OpenAI #

CNBC / TechCrunch

Noam Shazeer, co-author of the 2017 “Attention Is All You Need” paper and co-lead of Google’s Gemini, announced he is joining OpenAI – less than two years after Google brought him back from Character.AI for $2.7 billion. OpenAI CEO Sam Altman called him “one of the people I have most wanted to work with since the very beginning of OpenAI.” The same week, OpenAI hired former Trump AI policy official Dean Ball, signaling a dual push on both research talent and regulatory positioning ahead of its planned IPO.

Anthropic opens Seoul office with Samsung, NAVER, and LG enterprise deals #

Anthropic / Korea Times

Anthropic formally opened its third Asia-Pacific office in Seoul, simultaneously announcing enterprise deployments with NAVER (Claude Code across its engineering organization), Samsung SDS (Claude Cowork and Claude Code across Samsung Electronics), LG CNS (Claude across LG Group), Nexon, Hanwha Solutions, and Channel Corp. The company also signed an MOU with South Korea’s Ministry of Science and ICT for Korean-language safety evaluations and research access for 60+ university researchers. The Seoul launch is strategically timed against the backdrop of the ongoing Fable 5 / Mythos 5 export control situation, with Anthropic’s managing director stating the company is “very confident” models will become available again in the region within days.

Baseten reportedly raising $1.5B months after its last mega-round #

TechCrunch

AI inference startup Baseten is reportedly close to finalizing a $1.5 billion round at a $13 billion valuation, months after its previous raise. The round underscores the “inference gold rush” as demand for serving infrastructure scales faster than training infrastructure – a shift from the training-compute era where capital concentrated on GPU clusters for model development. For teams choosing inference providers, the scale of capital flowing into this layer signals sustained competition on price and latency.

Elastic agrees to acquire DeductiveAI for up to $85M #

TechCrunch

Elastic is acquiring DeductiveAI, a three-year-old startup that uses AI to catch and resolve software bugs, for up to $85 million. The acquisition extends Elastic’s observability stack with AI-native debugging capabilities, following the pattern of established infrastructure companies absorbing AI-first startups to add intelligent automation to their existing platforms.

General Intuition in talks to raise $300M at $2B valuation #

TechCrunch

General Intuition, which trains embodied AI and world models using Medal’s dataset of 2 billion videos per year from 10 million monthly active users, is raising $300 million at approximately $2 billion. The round continues the pattern of world-model startups attracting infrastructure-scale capital – Odyssey raised $310 million at $1.45 billion just days ago – suggesting investors see physical-world simulation as a distinct asset class from language and code models.

Developer Tools #

Zero-Touch OAuth for MCP: Enterprise-Managed Authorization #

Model Context Protocol Blog / Hacker News (203 points)

The Model Context Protocol announced that Enterprise-Managed Authorization (EMA) is now stable, allowing organizations to centrally provision MCP server access through identity providers. Instead of users manually authorizing each MCP server individually, administrators define access policies once and employees automatically gain access on login – no per-app OAuth consent screens required. For enterprises evaluating MCP adoption, centralized authorization through existing identity infrastructure removes one of the primary deployment friction points: getting hundreds of engineers connected to approved tool servers without a consent-screen gauntlet.

Amazon Bedrock AgentCore harness is now generally available #

AWS Machine Learning Blog

Amazon Bedrock AgentCore harness reached general availability, offering a two-API-call path from definition to running agent: CreateHarness to define it, InvokeHarness to run it. Each agent runs in its own isolated environment with filesystem, shell, cross-session memory, web browsing, and MCP tool connectivity. This is distinct from yesterday’s AgentCore knowledge and observability release – the harness is the runtime container itself, bundling the infrastructure that teams typically spend weeks assembling from orchestration frameworks, sandboxes, and memory stores.

Is it agentic enough? Benchmarking open models on your own tooling #

Hugging Face Blog

Hugging Face released an open-source benchmarking harness for evaluating how well models act as coding agents when using specific software libraries, measuring not just correctness but token consumption, error rates, and the effect of different context strategies. The harness revealed that adding CLI documentation and curated examples improved performance for large models but hurt smaller models (Qwen3-14B dropped from 67% to 43%), demonstrating that agent-facing APIs must be evaluated across model sizes rather than optimized for frontier models alone. Library maintainers can use this to measure whether API changes actually help agents before merging.

From Minutes to Seconds: LLM-Guided Autotuning for Helion Kernels #

PyTorch Blog

PyTorch published results on using LLMs in a population-based search loop to autotune Helion kernel configurations, achieving performance matching Likelihood-Free Bayesian Optimization while benchmarking 10x fewer configurations and reducing wall-clock tuning time by 6.7x. The approach is model-agnostic – Claude Opus 4.8, GPT-5.5, and Claude Sonnet 4.6 all performed within a few percent of each other. The result suggests LLM-guided search is a viable general-purpose replacement for expensive optimization sweeps in performance-critical ML infrastructure.

Research & Papers #

GDM AI Control Roadmap #

AI Alignment Forum / Google DeepMind

Google DeepMind published its AI Control Roadmap (v0.1), a framework for implementing internal guardrails to catch adversarial behavior from AI agents even when those agents are too capable to fully oversee. The roadmap introduces TRAIT&R, a threat taxonomy covering loss of control, work sabotage, and direct harm, paired with 15 concrete tiered mitigations that escalate alongside AI capabilities – from monitoring reasoning processes today to real-time access controls and shutdown infrastructure for future systems. For teams deploying agents in production, the roadmap provides the most detailed public framework yet for thinking about what controls are needed at each capability tier.

The Autonomy Tax: Defense Training Breaks LLM Agents #

arxiv:cs.CL

Defense training designed to protect LLM agents against prompt injection attacks creates a fundamental capability-alignment paradox: the same training that improves safety under adversarial conditions degrades the agent’s ability to use tools, follow multi-step instructions, and complete complex autonomous tasks under normal conditions. The paper quantifies this tradeoff across multiple defense-trained models, showing that practitioners face a systematic tension between agent security and agent utility. For teams deploying tool-using agents, this means defense-trained models may require capability testing that goes beyond standard safety benchmarks to catch the performance regression that safety training introduces.

StaminaBench: Stress-Testing Coding Agents over 100 Interaction Turns #

arxiv:cs.CL

StaminaBench measures coding agent stamina – how many consecutive change requests an agent can handle before failing – rather than the fraction-of-tasks-solved metric that dominates current benchmarks. Agents implement a REST API server and modify it across procedurally generated follow-up requests, matching the real pattern of vibe-coding sessions that run dozens or hundreds of turns. The benchmark addresses a blind spot in agent evaluation: current benchmarks test whether agents can solve isolated tasks, but production coding sessions require sustained coherence across long interaction histories.

Before the Pull Request: Mining Multi-Agent Coordination #

arxiv:cs.CL

Large-scale studies find that autonomous coding agents produce PRs faster but get them accepted less often – a coordination and trust gap that PR-level telemetry cannot explain. This paper argues the missing signal lives before the PR, in how concurrent agents claim, divide, and collide over shared work, and introduces grite, an open-source coordination substrate that needs no central server. For teams running multiple coding agents against the same codebase, the paper provides the first empirical framework for understanding why agent PRs conflict and how pre-PR coordination protocols can improve acceptance rates.

Deontic Policies for Runtime Governance of Agentic AI Systems #

arxiv:cs.AI

This paper presents a governance framework for constraining autonomous LLM agents that can invoke tools, manipulate data, install software, and coordinate across organizational boundaries, arguing that authentication and access control alone are insufficient – agents require the full structure of enterprise governance expressed as deontic policies (permissions, obligations, prohibitions). The framework operationalizes governance rules that are currently implicit in organizational policy into machine-enforceable runtime constraints. For teams deploying agents with real-world authority, this bridges the gap between “the agent has API access” and “the agent operates within our compliance framework.”

Security #

MosaicLeaks: Can your research agent keep a secret? #

Hugging Face Blog / ServiceNow

MosaicLeaks exposes a privacy vulnerability in deep-research agents: individually benign web queries that, when combined over time, allow adversaries to infer private enterprise information without ever seeing the underlying documents – the “mosaic effect.” The researchers found that simply telling agents not to leak has minimal effect, and training agents solely for task performance actually increased leakage from 34% to 51.7%. Their proposed Privacy-Aware Deep Research method reduced leakage to 9.9% while maintaining task success, demonstrating that agent privacy requires training-time intervention, not just prompting.

NRT-Bench: Multi-turn red-teaming of LLM agents in safety-critical systems #

arxiv:cs.CL

NRT-Bench tests LLM agent robustness under sustained adversarial pressure in a simulated nuclear power plant control room, where a five-role operator team backed by configurable LLMs faces multi-turn red-teaming attacks. The benchmark moves beyond single-turn jailbreak testing to measure whether agents maintain safety-critical behavior across extended adversarial interactions – the scenario that matters for real-world deployment in infrastructure, medical, or industrial settings. The setup quantifies a risk that single-turn safety evaluations systematically underestimate.

Regulatory & Policy #

Bernie Sanders unveils $7 trillion plan to give Americans control of AI industry #

Ars Technica

Senator Bernie Sanders proposed a $7 trillion plan to create an AI wealth fund that would give Americans a stake in the AI industry’s economic gains, arguing that the benefits of AI should not concentrate exclusively among technology companies and their investors. The proposal is the largest AI-specific policy plan introduced in Congress to date and signals that political pressure for AI benefit-sharing is moving from think-tank papers to legislative drafts. The scale of the proposal will likely provoke industry pushback, but it sets a boundary marker for the political conversation about who benefits from AI productivity gains.

The US says ASML’s top chip tool may be in China. ASML says it isn’t #

TechCrunch

The US government claims ASML’s most advanced EUV lithography tool may have been shipped to a Chinese customer, which ASML denies, creating a standoff over export control enforcement on the semiconductor equipment that underpins AI chip manufacturing. ASML has strong commercial incentives to maintain its export license, cutting against the idea it would risk arming a Chinese fab. The dispute highlights the difficulty of enforcing semiconductor export controls when the supply chain involves a single non-US supplier with monopoly power over the most critical tool.

SK Telecom at the center of Anthropic’s Mythos export control controversy #

Wired / Hacker News (115 points)

Wired details how SK Telecom, South Korea’s largest carrier and a $100 million Anthropic investor, was identified by the White House as a potential Chinese security risk with access to Mythos 5, escalating the export ban from a targeted access revocation to a global suspension of both Fable 5 and Mythos 5. The article adds new detail to the export control story that has dominated AI policy discussion for two weeks, revealing that Amazon researchers separately flagged Fable 5 vulnerabilities that compounded the government’s concerns. The incident is reshaping how AI companies structure international investor relationships and model access tiers.

Infrastructure #

FERC orders data center grid fast-lane for interconnections #

TechCrunch / NVIDIA Blog

The Federal Energy Regulatory Commission ordered grid operators to create a priority interconnection pathway for data centers, AI factories, and advanced manufacturing facilities, addressing the multi-year queue bottleneck that has delayed data center buildouts. The order streamlines connection to the grid but does not address the underlying electricity supply shortage, meaning faster interconnection does not guarantee faster power delivery. For companies planning AI infrastructure, FERC’s order removes one bureaucratic bottleneck while leaving the harder constraint – actual power generation capacity – unresolved.

Amazon to sell AI chips to challenge Nvidia more directly #

TechCrunch

AWS is in talks to sell its custom AI chips (Trainium and Inferentia) to third-party data centers, expanding from cloud-only availability to direct hardware sales. CEO Andy Jassy has described this as a $50 billion opportunity. The move creates a three-way competition in AI silicon – NVIDIA GPUs, Google TPUs (cloud-only), and now Amazon chips available both in-cloud and on-premise – potentially giving teams outside hyperscaler clouds access to competitive AI hardware at lower price points.

Chip Industry Week In Review #

Semiconductor Engineering

This week’s chip industry summary covers Apple-Intel manufacturing discussions revived by the Trump administration, Amkor winning a major advanced packaging contract, Intel 18A-P process qualification, Amazon’s AI chip sales plans, and CHIPS Act funding updates. The convergence of geopolitical pressure (Apple-Intel), advanced packaging capacity (Amkor), and AI chip proliferation (Amazon) in a single week illustrates how AI compute demand is reshaping semiconductor industry structure across fabrication, packaging, and distribution layers simultaneously.

Threads to Watch #

Agent governance is crystallizing as a distinct engineering discipline. Google DeepMind’s AI Control Roadmap, the deontic governance framework for agentic systems, and MosaicLeaks’ privacy-aware training all address the same gap: agents are gaining real-world authority faster than the control infrastructure to constrain them. The common pattern is that retrofitting governance onto deployed agents does not work – controls need to be designed into agent architectures from the start, whether through runtime policy enforcement, training-time privacy objectives, or tiered capability restrictions.

The inference economy is attracting training-era capital. Baseten raising $1.5 billion at $13 billion, Amazon selling AI chips to third-party data centers, and FERC fast-tracking grid interconnections all point to the same shift: the bottleneck in AI deployment is moving from “can we train the model” to “can we serve it fast enough, cheaply enough, and close enough to where it’s needed.” The capital formation patterns that characterized the training compute buildout are now repeating in the inference layer.

AI talent and access are geopolitical instruments. Noam Shazeer’s move from Google to OpenAI, the SK Telecom export control escalation, Anthropic’s Seoul office opening amid Fable 5 restrictions, and ASML’s chip tool dispute all demonstrate that AI competition is no longer just about models and benchmarks – it is about who has access to which people, chips, and model weights, and which governments can restrict that access.

Sources Unavailable Today #

These sources could not be fetched today. Links point to their homepages so you can check them directly.

• Stanford HAI — scrape: content truncated