We also look at a new robot emotion paper where VLMs outperform conventional facial analysis but can't bridge the gap between sensing a cue and understanding intent. The infrastructure, capital, and capability stories all converge on one question: who gets to build at frontier scale, and who decides when enough is enough.

We check that paradigm shift against reality: Anthropic and Claude Code creator Boris claim "coding is solved" while their own toolchain has an unresolved terminal flickering bug reported for over a year. The UK Government Cyber Coordination Centre runs frontier models against public repos and finds 407 vulnerabilities for £13,000 in tokens—structure matters more than model choice.

Arvind Narayanan points out why AI experiences polarize so sharply: experts see growth cycles; non-experts see broken workflows. And a credential stealer hides in astro.config.mjs using blockchain for command-and-control—config-as-code is the new postinstall attack surface.

Oracle's capex hits $70B annually while free cash flow stays negative. The infrastructure reality behind the agent narrative doesn't lie.

AWS Bedrock now requires sending inputs to Anthropic for Mythos/Fable tier models. Enterprise customers in regulated industries can't allow a second data processor without regulator sign-off. The change appears timed around Anthropic's IPO rather than safety concerns.

Richard Sutton gave a long talk arguing that generative AI trained by supervised learning cannot make novel discoveries — it produces either novelty (via randomness) or quality (via training data), never both simultaneously. True creativity requires evaluation and selective retention, which RL systems like AlphaFold have but LLMs don't.

Plus: a practical note from PostHog about why off-the-shelf embeddings group signals by format instead of meaning.

Meanwhile, Google is paying SpaceX $920M per month for 110,000 GPUs, while New York's legislature approved a one-year moratorium on datacenters over 20MW. Two different answers to the same question about infrastructure pacing.

And there's a quieter current: Salvatore Sanfilippo (antirez) building DwarfStar because he believes in local inference as a safety net; Nvidia pushing RTX Spark silicon into Windows PCs; Apple's fumbled Siri rollout accidentally putting it in a favorable position against Gemini's creepiness factor.

Three stories today map the same question from different angles: Alphabet is raising $85 billion, while its stock drops for the fourth week straight. New York became the first state to freeze new data center permits. And an Australian infrastructure company committed $30 billion to India.

The capital costs of building frontier AI are finally hitting visible resistance — in markets, in politics, and in geography. Plus: Paul Graham on why big companies struggle with LLM token economics (so far), and what the Huawei/DeepSeek chip story means for hardware sovereignty.

Meanwhile, George Hotz's four-tier framework for AI winners gets a concrete financial test: Broadcom stock dropped today despite an unchanged AI chip forecast. And Apple is going to Google's custom Nvidia-powered silicon for its revamped Siri this September — a hardware-layer signal about where core AI inference actually lives.

On policy: the Trump administration's executive order on AI model sharing went through one of the strangest legislative processes I've seen. Drafted, pulled at the last minute by Trump himself, re-signed in private with zero fanfare. Neither version gives the White House new enforcement power, but the classified thresholds for what triggers review are genuinely concerning.

We also look at Meta's Business Agent going global on WhatsApp after a two-year pilot, someone running DeepSeek V4 Flash on an M2 Max 64GB, and Eric Glyman introducing Stack — an AI operating system for accounting firms that closes books in half the time.

On the local pass, the ownership thread keeps surfacing. The models are spreading through elite institutions. The money is moving through PE. The hardware is migrating to the edge. The one practical test we found — a 10-trap honesty evaluation of Claude Opus 4.8 — showed genuine calibration gains, but also a concrete failure mode on legal certainty. The gap between model capability and model truth is still where the story lives.

Segments cover: Anthropic's Mythos expansion and what it means for model access geopolitics; the private equity / bionic organization layer Sanders is pushing against; Nvidia's RTX Spark PC chip and the edge play; the Opus 4.8 honesty test as a data point; and a counter-narrative from a rocket engine startup raising $500 million to bet on human talent.

Also: Robin Hanson's framing of AI as collective intelligence extracted without permission, and a Red Hat supply chain compromise via a trusted publisher.

The Uber AI budget overruns are being headline-d as a failure story, but they're really a planning mismatch: budgets set in 2025 priced inputs before they got five times cheaper to convert.

Vicki Boykis argues we should be "more tired than the model" — adding deliberate friction to preserve skill retention in an era of agentic code generation.

Sources: Google DeepMind (SynthID), Nathan Lambert on China AI policy, Tencent Hy (Hy-MT2), Techmeme/OpenRouter on OpenRouter, Politico on Dutch infrastructure block, ECLresearch on watermark scale, Wired on DHS threat categorization.

Meanwhile, llama.cpp ships native agent tools straight into its server binary. No MCP bridges, no Python wrappers. Just a GGUF file and a flag. You get raw speed, but you also get raw exposure.

And in Claude Code, a five-minute idle timeout quietly turns casual debugging into a token burner. The 12.5× cache miss penalty doesn't come from the model. It comes from the prefix. Understanding the invalidation table is now part of the craft.

Three structural moves. One Sunday.

SpaceX's IPO filing for the Anthropic deal, Jensen Huang conceding China to Huawei, Google's two-track agent strategy, Bolt Graphics betting on high-precision GPUs, and a Google vibe-coding experiment.

Local pass notes: Flash beating Pro is unusual but not unprecedented. The pricing confusion around Flash vs Pro tiers is worth tracking. The real test will be how these models behave in long-running agent loops beyond the demo stage.

Sholto Douglas asks what makes people reach for other models instead of Claude, and gets 800 replies. The answers are specific: Claude confuses PDF form fields, over-filters bio research, treats a question about your database as a migration request, and writes training code that breaks the model names. The thread and replies are here.

The Verge's Emma Roth documents how AI drive-thrus are rolling back after user frustration, and John Gruber makes the case that AI is technology, not a product — both stories land on the same tension: how we position AI versus how it actually behaves in the wild.

Plus a local Qwen 3.6 benchmark that suggests the gap with frontier models is narrowing on concrete coding tasks. See the benchmark details on the subreddit.

Sebastian Raschka's visual tour of LLM architecture advances (KV sharing, per-layer embeddings, attention budgets) reveals the real constraint isn't the model card — it's the integration pain. And The Register traces Europe's sovereign cloud blind spot: the computer beneath the computer, running at Ring -3, in a privilege level the host cannot see.

Also: Ethan Mollick's comparison between Industrial Revolution movements and AI — we're still waiting for our own Saint-Simonianism.

Also: Osaurus, the local-plus-cloud Mac harness, and Figure AI's 30-hour robot run. A Friday of infrastructure stories.

Mozilla's Claude Mythos found 271 vulnerabilities in Firefox version 150, while Anthropic's Opus 4.6 found only 22 in version 148. The 271-to-22 gap between two AI verification systems is the first large-scale, apples-to-apples comparison of verification quality. It challenges the assumption that human-written code is inherently trustworthy.

OpenAI is winding down its fine-tuning API, pushing teams toward other customization approaches. Runway reports $40M+ in new ARR this quarter as generative video hits enterprise adoption. Multi-token prediction gives local Gemma 4 models a 40% speedup in LLaMA.cpp. And the EU commissions separate technical studies for marking AI-generated text, audio, and video under Article 50 of the AI Act.

First, Trail of Bits reported beating Google's zero-knowledge proof of quantum cryptanalysis by exploiting bugs in their Rust ZKP code. They forged a proof with better metrics, and the May Tribune also released Trailmark, MuTON, and mewt. The implication for anyone relying on those Rust implementations needs to be checked.

Second, Anthropic announced The Anthropic Institute and its four-area research agenda: economic diffusion, threats and resilience, AI systems in the wild, and AI-driven R&D. This is an institutional commitment to studying post-deployment AI — not just alignment during training.

Third, AWS previewed AgentCore Payments, built with Coinbase and Stripe, enabling AI agents to transact. The agent ecosystem shifts from orchestration to commerce when transactions are built in.

Fourth, Ethan Mollick pointed out the deployment gap that keeps enterprise AI "normal" rather than transformative — models can't act as their own deployment consultants, process mappers, or change management experts. Even the labs building the models aren't confident they can handle it.

Plus a word on tooling problems that won't go away (Pi's read tool), and Spotify's AI DJ expanding to four more languages.

First, Trail of Bits reported beating Google's zero-knowledge proof of quantum cryptanalysis by exploiting bugs in their Rust ZKP code. They forged a proof with better metrics, and the May Tribune also released Trailmark, MuTON, and mewt. The implication for anyone relying on those Rust implementations needs to be checked.

Second, Anthropic announced The Anthropic Institute and its four-area research agenda: economic diffusion, threats and resilience, AI systems in the wild, and AI-driven R&D. This is an institutional commitment to studying post-deployment AI — not just alignment during training.

Third, AWS previewed AgentCore Payments, built with Coinbase and Stripe, enabling AI agents to transact. The agent ecosystem shifts from orchestration to commerce when transactions are built in.

Fourth, Ethan Mollick pointed out the deployment gap that keeps enterprise AI "normal" rather than transformative — models can't act as their own deployment consultants, process mappers, or change management experts. Even the labs building the models aren't confident they can handle it.

Plus a word on tooling problems that won't go away (Pi's read tool), and Spotify's AI DJ expanding to four more languages.

On the local pass, the infrastructure layer is the actual constraint today — not models, not pricing, but the plumbing between chips and the protocols that keep them from falling out of sync.

On the local pass, the infrastructure layer is the actual constraint today — not models, not pricing, but the plumbing between chips and the protocols that keep them from falling out of sync.

Also: Arvind Narayanan on using LLMs to review writing, and why that matters more than the headline-grabbing infra plays. And what a graveyard of dead AI tools tells us about the state of the ecosystem.

Also: India's first orbital data center for AI training, filed away from a headline that arrived without a full article body. Interesting signal even without the details.

Also: VS Code defaulting Copilot attribution to every commit, a million-line Haskell codebase at Mercury, and the Qwen3.6-27B vs Coder-Next results that say "it depends" with statistical backing.

• The 7M parameter model on ARC Prize — YC's Decoded on HRMs and TRMs
• o1 vs physicians — peer-reviewed clinical reasoning benchmark
• LangChain's Deep Agents and the create_agent primitive convergence
• GPT-5.5 and multi-day continuous agent runs in Codex
• Agent rate limits and the death of per-seat SaaS pricing
• Smol AI digest: Qwen3.6 27B leads open-weight, GPT-5.5 on cyber evals

Meanwhile, David Gomes from Cursor walked through replacing 12,000 lines of custom git worktrees infrastructure with a 200-line Markdown skill. The tradeoffs are honest and the lessons apply to any team building agent workflows.

• Granite 4.1: IBM's dense 8B matching 32B MoE — data quality over parameter count
• Mistral Medium 3.5: The dense unified flagship replacing Devstral 2
• Cursor's David Gomes: replacing 12K LoC with 200 lines of Markdown skills
• Figure AI hits 24x production — one humanoid robot per hour
• GCC 16 released — C++26 reflection and compiler plumbing

• OpenAI's multi-cloud pivot — what changed and what didn't
• ATOM's 40% inference efficiency claim — what it would take to believe it
• Why every agent benchmark you trust is optimizing for the wrong thing
• llms.txt at zero — an empirical look at agent documentation standards

• DeepSeek-V4-Flash on Apple Silicon, 2-bit DQ quantization, and the Pi harness ecosystem
• Matt Pocock on pulling away from Claude Code
• The defense production collapse parallel for software engineering
• Stanford LLM creates functional viruses from raw DNA sequences
• Eden AI — the European OpenRouter alternative
• Asahi Linux 7.0: VRR, PMP power management, and the long haul upstreaming Apple Silicon

• The Office for AI Employees — WUPHF, a shared collaboration space for multiple AI agents with per-agent notebooks and a team wiki that agents promote from private notes to shared knowledge.
• Anthropic's Internal Marketplace — Project Deal: Claude negotiating real transactions for employees in Anthropic's SF office.
• Video Generation's New Arms Race — Grok Imagine and GPT Image 2 on Runway, with lip sync and sound as the new differentiators.
• The Productivity Reality Check — Experienced developers taking 19% longer with AI tools. The gap between lab benchmarks and production reality.
• The Norm Carpet Problem — Susan Zhang's observation about how layer normalization hides training problems until it's too late.

• The pricing page that vanished — Amol Avasare's clarification on the fake-door test that sparked industry-wide backlash and forced a reversal.
• Mozilla's 271-bug bootcamp — How Firefox's CTO is treating AI vulnerability hunting as an unavoidable, finite overhaul every codebase must survive.
• The compaction wars beneath the harness — Mario Zechner's teardown showing that loop design matters less than context management, and why pi and Codex are converging on the same pattern.
• Google's 8th-gen TPU and the agentic infrastructure pivot — What the new 8t and 8i chips actually mean for streaming inference, and why the compute layer is racing ahead of the application layer.
• Taste, craft, and the quality wedge — Linear's CTO and Gergely Orosz on why shipping speed is easy, judgment is hard, and human taste remains the only real moat.