Trump administration pressed OpenAI to limit GPT-5.6 to select partners

The Trump administration pressured OpenAI to restrict the release of GPT 5.6 to select partners rather than the public, citing cybersecurity concerns – the first time a frontier model rollout has been shaped by direct government intervention. General Intuition raised $320 million at a $2.3 billion valuation to train AI agents on video game action data, while a cluster of ArXiv papers on agent reliability – including formal treatments of cross-module prompt interference, verification limits for coding agents, and parametric memory consolidation – signals that the research community is shifting from building agents to hardening them.

Regulatory & Policy #

The White House is asking OpenAI to slow roll the release of its new model over safety concerns #

TechCrunch

The Trump administration’s Office of the National Cyber Director and OSTP pressured OpenAI to limit GPT 5.6 distribution to a select group of partners rather than releasing it publicly, following the June 2 executive order directing AI companies to submit new models for government testing before public release. The restricted rollout mirrors Anthropic’s approach with Claude Mythos and establishes a precedent where frontier model distribution requires government coordination. For the AI industry, the practical implication is that “release day” for the most capable models is now a negotiation with regulators rather than a product decision – and that cybersecurity capability is the specific threshold triggering government intervention.

AI and Liability #

Bruce Schneier / Simon Willison’s Weblog

Bruce Schneier analyzes a German court ruling holding Google liable for factual errors introduced in its AI Overviews, rejecting Google’s defense that users should verify AI-generated information themselves. The ruling treats AI outputs as the deploying company’s own statements rather than as a neutral conduit, analogous to how publishers are responsible for editorial content. If this liability framework spreads – and Schneier argues it should – it creates direct financial incentives for accuracy that reshape the economics of AI deployment in law, medicine, and other high-stakes domains where errors carry consequences.

Funding & Business #

General Intuition’s $2.3B bet that video games can train AI agents for the real world #

TechCrunch

General Intuition raised $320 million at a $2.3 billion valuation from Khosla Ventures to train AI agents on millions of hours of video game footage from its parent company Medal, leveraging action data embedded in gameplay clips rather than inferring actions from video. CEO Pim de Witte demonstrated the same underlying model controlling both a Fortnite-playing agent and a quadrupedal robot that required only eight minutes of real-world fine-tuning data. The thesis – that game action data teaches spatial-temporal reasoning and causality more efficiently than traditional robotics training data – positions General Intuition as a foundational model provider for robotics and autonomous systems rather than an end-user product company.

Patronus AI lands $50M to build ‘digital worlds’ that stress-test AI agents #

TechCrunch

Patronus AI raised $50 million in Series B funding led by Greenfield Partners to build simulated digital environments that replicate websites and internal systems for stress-testing AI agents, bringing total funding to $70 million. The company, founded by former Meta researchers, reports that virtually every frontier AI lab now uses Patronus for agent evaluation. The investment validates agent testing as its own infrastructure category: as agents move from demos to production, the gap between “works in a prompt” and “works reliably in the real world” requires purpose-built simulation environments to close.

Anthropic’s Claude is winning over paid consumers, a market owned by ChatGPT #

TechCrunch

Credit card transaction data from Indagari shows Claude’s paid users and revenue grew approximately 75% since January 2026, while DataCamp reports that “Claude” is now the most searched term on its platform with consumer demand for Claude courses increasing 18x in 30 days and outpacing ChatGPT three to one among self-directed learners. ChatGPT maintains a commanding lead in absolute user numbers and market share. The data suggests that Claude is winning the segment of users who are willing to pay and invest in learning – the highest-value cohort for enterprise adoption – even as ChatGPT dominates the broader consumer market.

Adobe acquires image and video enhancement tool maker Topaz Labs #

TechCrunch

Adobe acquired Topaz Labs, an Emmy-winning AI-powered image and video enhancement company with over two decades of history, to integrate its upscaling and retouching models across Creative Cloud and the Firefly AI app. The deal, expected to close in H2 2026, continues Adobe’s strategy of acquiring specialized AI capabilities rather than building them in-house. For the creative AI market, the acquisition removes one of the few remaining independent enhancement tools and consolidates AI-powered creative capabilities further into Adobe’s ecosystem.

Netris raises $15M Series A from a16z to help AI neoclouds go live faster #

TechCrunch

Netris raised $15 million in Series A funding from Andreessen Horowitz for network automation software that runs on switches to help emerging AI cloud providers deploy GPU clusters faster. Operating at over 35 GPU clusters globally, Netris addresses the bottleneck where expensive GPUs sit idle while network infrastructure is being manually configured. The investment targets the specific operational gap between procuring AI compute and making it production-ready – a problem that grows more acute as every new neocloud discovers that networking, not GPUs, determines time-to-revenue.

Research & Papers #

Instruction Bleed: Cross-Module Interference in Prompt-Composed Agentic Systems #

arxiv:cs.AI

This paper formalizes compositional behavioral leakage (CBL): the failure mode where editing one prompt module in an agentic system silently shifts the behavior of others despite no shared variable or executable dependency. The root cause is architectural – transformer self-attention provides no formal boundary between concatenated modules sharing a context window. The authors demonstrate the effect experimentally and propose isolation primitives. For anyone building multi-module agent systems by composing system prompts, tool descriptions, and user instructions, the finding means that prompt engineering is not compositional: changing one module requires regression-testing all others.

The Verification Horizon: No Silver Bullet for Coding Agent Rewards #

arxiv:cs.AI

As foundation models develop stronger reasoning and engineering harnesses grow more sophisticated, generating complex candidate solutions is no longer the hard problem – reliably verifying them is. This paper argues that every verifier is only a proxy for human intent, and that each family of verifiers (static analysis, test suites, formal methods, LLM judges) has fundamental blind spots that cannot be eliminated by scaling. The practical implication for coding agent builders: verification diversity – combining multiple independent verification strategies – is the only path to reliability, and the verification budget should scale with the complexity of the generated code.

SKILL-DISCO: Distilling and Compiling Agent Traces into Reusable Procedural Skills #

arxiv:cs.AI

Agents repeatedly solve similar task instances from scratch, wasting reasoning compute on redundant work. SKILL-DISCO studies this problem in FSM-defined scenarios where successful traces can be viewed as paths in an unknown transition graph, formalizing skill extraction as recovering the shared procedural structure across successful executions. The framework distills and compiles these traces into reusable procedural skills that can be invoked without re-reasoning. For teams operating agents at scale, this addresses the economic problem directly: amortizing reasoning cost across repeated task patterns rather than paying the full inference bill every time.

Memory Depth, Not Memory Access: Selective Parametric Consolidation for Long-Running Language Agents #

arxiv:cs.AI

Long-running agents need more than retrieval – they need memory depth, defined as durable goal-conditioned tendencies written into a small parametric store that persist after the working context is unloaded. The authors introduce a loop-drift protocol that stress-tests whether an agent’s behavior drifts over extended operation even when the retrieval index remains intact and correct. The distinction between memory access (fetching facts) and memory depth (stable behavioral tendencies) is the kind of architectural primitive that separates toy agent demos from production systems that operate reliably over days and weeks.

When Does Combining Language Models Help? A Co-Failure Ceiling on Routing, Voting, and Mixture-of-Agents Across 67 Frontier Models #

arxiv:cs.AI

Multi-model LLM systems (routing, voting, cascades, mixture-of-agents) are used to beat single-model accuracy, but their gain is capped by a quantity the field rarely reports: the co-failure rate beta, defined as the fraction of queries where every model is wrong. The paper shows across 67 frontier models that average pairwise error correlation cannot bound this ceiling, meaning the standard diagnostic for model diversity is insufficient. For teams building multi-model pipelines, the actionable insight is to measure co-failure directly on your task distribution before investing in ensemble infrastructure – if beta is high, no combination strategy can help.

Localizing RL-Induced Tool Use to a Single Crosscoder Feature #

arxiv:cs.AI

RL fine-tuning enables tool use in language models, but the mechanistic basis has been poorly understood. This paper identifies a single crosscoder feature – one interpretable direction in the model’s representation space – that is both necessary and sufficient for RL-induced tool calling behavior. Ablating this feature disables tool use; amplifying it increases tool invocation. The finding suggests that complex agentic capabilities may be controlled by surprisingly localized mechanisms, which has implications for both safety (targeted capability removal) and efficiency (targeted capability enhancement without full fine-tuning).

Autoformalization of Agent Instructions into Policy-as-Code #

arxiv:cs.AI

Agent safety in high-stakes domains requires formal policy enforcement, but existing approaches either use probabilistic guardrails with no formal guarantees or hand-coded symbolic enforcement that does not scale. This paper presents an autoformalization pipeline that translates agent prompts, MCP tool descriptions, and natural language policy documents into formal executable policies – bridging the gap between what operators write in natural language and what can be mechanically enforced. For teams deploying agents with access to sensitive tools, this addresses the fundamental tension between the flexibility of natural language instructions and the rigor of formal safety guarantees.

Embarrassingly Simple Self-Distillation Improves Code Generation #

arxiv:cs.AI

A model can improve at code generation using only its own raw outputs – no verifier, teacher model, or reinforcement learning required. Simple self-distillation (SSD) samples solutions from the model, filters by execution correctness, and fine-tunes on the passing solutions. The technique is embarrassingly simple yet effective, suggesting that current models have latent code generation capability that standard training does not fully surface. For teams fine-tuning code generation models, SSD provides a zero-cost data augmentation strategy that requires only a test harness for filtering.

Security #

What happened after 2k people tried to hack my AI assistant #

Fernando Irarrázaval / Hacker News (174 points)

Fernando Irarrázaval invited over 2,000 people to attempt prompt injection attacks against his AI assistant through more than 6,000 emails employing techniques from authority impersonation to multi-language social engineering – and the secret never leaked. The experiment found that model choice is the dominant factor in injection resistance, with Claude Opus 4.6’s instruction-following robustness defeating all extraction attempts. For teams building agent systems that handle sensitive data, the finding provides actionable evidence that model selection, not prompt engineering alone, determines whether prompt injection defenses hold under sustained adversarial pressure.

Adaptive Evaluation of Out-of-Band Defenses Against Prompt Injection in LLM Agents #

arxiv:cs.AI

The dominant strategy for defending tool-using agents against indirect prompt injection – enforcing security outside the model with deterministic filters rather than training the model to refuse – is increasingly adopted but poorly evaluated. This paper presents an adaptive evaluation methodology that stress-tests these out-of-band defenses, finding that many deployed systems have bypass paths that standard evaluation misses. For agent builders relying on guardrail-style defenses, the implication is that defense evaluation must be as adversarial as attack evaluation – a fixed test suite does not capture the gap between a defense that passes benchmarks and one that survives deployment.

Infrastructure #

Amazon ups India bet with fresh $13B AI infrastructure investment #

TechCrunch

Amazon announced an additional $13 billion investment in India through 2030, bringing total committed spending to $48 billion, primarily to expand AWS data center capacity in Mumbai and Hyderabad for AI and cloud infrastructure. The investment competes directly with Microsoft, Google, and local firms racing to build AI compute capacity in India’s market. For the global AI infrastructure landscape, the scale of commitment signals that India is becoming the next major theater for cloud-AI infrastructure competition after the US, with multiple hyperscalers making multi-billion-dollar bets simultaneously.

Apple to skip high-end M6 Mac chips in favor of AI-focused M7 line #

Bloomberg / Hacker News

Apple will skip the high-end variants of its M6 chip (Pro, Max, Ultra) and instead develop an AI-focused M7 line, according to Bloomberg. The decision to leapfrog an entire chip generation for AI optimization reflects Apple’s assessment that the competitive landscape has shifted – local AI inference capability is now a primary differentiator for professional hardware, not just GPU throughput or CPU performance. For developers building on-device AI applications, the M7’s architecture choices will define the local inference ceiling for the Mac platform through the end of the decade.

Un-0: Generating Images with Coupled Oscillators #

Unconventional AI / TechCrunch / Hacker News (154 points)

Unconventional AI, led by former Databricks AI chief Naveen Rao, released Un-0, an image generator that replaces traditional neural network layers with simulated Kuramoto oscillators whose learned coupling strengths drive the generation process. The system achieves an FID of 6.74 on ImageNet 64x64 – matching early conventional generators – while using roughly 90% fewer parameters in its core dynamical component, with the company claiming the underlying architecture could reduce AI inference power consumption by approximately 1,000x. The energy efficiency claim warrants skepticism until demonstrated on actual hardware rather than software simulations, but if the oscillator-based approach scales, it addresses the fundamental constraint that AI scaling is increasingly energy-limited.

Chip Industry Week In Review #

Semiconductor Engineering

This week’s semiconductor roundup covers Onsemi’s acquisition of Synaptics, progress on 1nm MoS2 nanotubes, a $250 million CHIPS Act award, a new advanced packaging site, and humanoid robotics developments. The breadth of activity – from materials science to trade secret litigation to manufacturing capacity – reflects an industry where every layer of the stack is simultaneously under pressure from AI demand. For teams tracking AI infrastructure dependencies, the week’s density of deals and technical milestones underscores that the semiconductor supply chain remains the binding constraint on how fast AI compute can scale.

Developer Tools #

Run a vLLM Server on HF Jobs in One Command #

Hugging Face Blog

Hugging Face integrated vLLM with its Jobs infrastructure service, enabling developers to spin up a private, OpenAI-compatible LLM endpoint with a single command, billed per-second by hardware usage. The integration eliminates the infrastructure management overhead of running private model servers for testing, evaluation, and batch processing while maintaining access control through HF tokens. For teams that need temporary model endpoints for CI pipelines, evaluation harnesses, or development without committing to a persistent deployment, this reduces the operational cost of “just run a model” from hours of setup to a single CLI invocation.

Retrofit, don’t rebuild: Agentic overlays for transforming legacy enterprise services #

AWS Machine Learning Blog

AWS published a reference architecture for wrapping existing REST-based services with thin agentic overlay layers that enable Agent-to-Agent (A2A) interactions and expose REST APIs as tools compatible with Model Context Protocol (MCP), without rewriting business logic or duplicating code. The pattern addresses the gap between enterprises’ existing service architectures and the agent-native interfaces that multi-agent systems require. For teams building agents that need to interact with legacy enterprise systems, the overlay approach avoids the “rewrite everything as an MCP server” trap and provides a pragmatic migration path.

Notion killing Skiff-influenced email app since most users use AI agents instead #

Ars Technica

Notion is shutting down its Skiff-influenced email app and shifting to an agents-first approach for inbox management, stating that most users now prefer AI agents to handle their email rather than using a traditional email client. The pivot reflects a broader pattern where productivity tools are discovering that their users want agent-mediated workflows rather than improved traditional interfaces. For developer tool builders, Notion’s strategic decision provides a data point: when given the choice between a polished conventional interface and an agent-driven one, paying users are choosing agents.

TokenSpeed-Kernel: Portable APIs and High-Performance Kernels for Multi-Silicon LLM Inference #

PyTorch Blog / LangChain Blog

TokenSpeed-Kernel is an open-source subsystem that decouples high-level inference runtime code from hardware-specific kernel implementations through a registry-based selection mechanism where kernels declare their capabilities and the runtime transparently selects the best implementation for the target hardware. Benchmarked with GPT-OSS 120B on AMD MI355X, the system delivered 1.6x to 3.6x end-to-end throughput improvements over portable baselines. For teams serving models across heterogeneous hardware, the contribution solves the kernel selection fragmentation problem that currently forces inference engines to accumulate platform-specific conditional logic.

Open Source #

OpenKnowledge: open source AI-first alternative to Obsidian/Notion #

Hacker News

OpenKnowledge is a fully local, open-source markdown editor with direct integrations for Claude, Codex, and other AI agents, available as a macOS app or web UI with CLI. The project provides a “what you see is what you get” editing experience that bridges the gap between traditional note-taking tools and AI-native workflows. For teams looking for an Obsidian-like experience with first-class agent integration and no vendor lock-in, OpenKnowledge fills a niche that existing tools either ignore (Obsidian’s AI plugins are third-party) or gate behind subscriptions (Notion’s AI features).

Threads to Watch #

Government intervention in frontier model distribution is now operational. The White House’s direct pressure on OpenAI to restrict GPT 5.6 distribution, combined with the German court ruling holding Google liable for AI Overview errors, establishes two complementary precedents: governments can shape when and how frontier models reach the public, and deployers bear legal responsibility for what those models say. Together, these developments mean that model release strategy is no longer purely a product decision – it is a regulatory negotiation with liability consequences.

Agent reliability research is producing architectural primitives. Instruction Bleed formalizes cross-module interference in prompt-composed systems, Memory Depth distinguishes retrieval access from behavioral persistence, Semantic Early-Stopping replaces fixed iteration caps with meaning-aware convergence, and SKILL-DISCO distills repeated agent traces into reusable skills. These are not incremental improvements – they are named primitives that give agent builders a vocabulary for failure modes that were previously discovered only in production. The field is transitioning from “can we build agents” to “can we make them reliable enough to deploy.”

Action data is emerging as a training paradigm distinct from text and video. General Intuition’s $2.3 billion valuation is built on the thesis that video game action data – not video, not text, but timestamped input-output pairs – teaches spatial-temporal reasoning more efficiently than either modality alone. SKILL-DISCO’s approach to distilling agent execution traces into reusable procedures operates on the same insight from the opposite direction: instead of training on action data, it extracts action patterns from successful runs. Both suggest that the next wave of capable agents may be trained on what agents do, not what humans write about what agents should do.

Sources Unavailable Today #

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

• Weights & Biases: Fully Connected – scrape: content not extractable