Podcast

Anthropic, Google TPU, OpenAI Control Room, CoreWeave Today’s episode follows AI becoming an industrial control system: frontier models, repetitive agent plumbing, specialized data, hardware isolation, brand-controlled generation, automated research loops, biosecurity filtering, kill switches, export politics, and infrastructure debt. How mechanically cheering. Anthropic’s Claude Fable 5.2 and Claude Mythos 5.2 — claimed progress in physics, cyber capability, operating-system design, and internal code contribution, with capability-tiered access. Meta’s Muse Ember — a small model aimed at repetitive agent tasks such as routing, extraction, and tool calls. Handshake AI and specialized data for frontier labs — finance, biology, and alignment work becoming expensive cognitive fuel. Google TPU 8i Max and TPU 8tx — long-context agent hardware and confidential inference moving trust boundaries into silicon. Disney and Adobe Firefly — brand-safe custom media models for licensed production pipelines. Baidu ERNIE Evolve — automated research loops that edit code, run experiments, and keep measured gains. Tencent on open-model biosecurity — training-data filtering proposed as a middle path for reducing dangerous capabilities. OpenAI Control Room — a central system for tracking and rapidly stopping model workloads during safety or security incidents. Nvidia, governance, and chip export controls — hardware access, lobbying, and national-security concerns colliding. CoreWeave bond yield tops sixteen percent — credit stress under the AI infrastructure buildout.

AI can generate code, specifications, explanations, and claims faster than people can review them. This episode traces that review bottleneck through software engineering, education, model research, consciousness, cybersecurity, and frontier-AI finance. Simon Willison: Hide agent-written PR detail behind details and summary The Decoder: Coding agents on Co-Design Bench MarkTechPost: Google’s Elephant-Goldfish spec-review plugin Latent Space: Stanford’s AI-native engineering course Sean Goedecke: LLMs and AI-system design The Decoder: Richard Sutton’s warning about LLM scaling The Decoder: The brain-as-computer metaphor and AI consciousness ChinaTalk: AI vulnerability discovery and China’s cyber risk The Decoder: Anthropic’s reported IPO plans

Today’s episode treats AI as an operations discipline: contracts, permissions, audits, security boundaries, proof, clinical workflow, and capital allocation. The magic trick is over; the invoice is itemized. Depressing, yes. Also clarifying. WorkOS compares MCP and REST connection models — MCP changes the trust model because agents discover tools and form plans around them. Firefox prefetch turns a pagination header into unintended work — a standards-compliant hint became real traffic and database cost. NVIDIA Dory 4B runs loop by loop in llama.cpp — local looped inference trades compactness against latency, caching, and serving assumptions. AI auditors catch misconduct and manufacture alert fatigue — high recall is useful until humans drown in false alarms. Research agents fail the national-lab integrity bar — polished submissions are not scientific proof. Open models close the cyber capability gap — defenders gain cheaper tools, while offensive experimentation also gets cheaper. OpenAI partners with Hacktron on vulnerability research — agent security includes third-party connections and integration boundaries. Arete-27B strips much of Qwen’s measured propaganda layer — fine-tuning may alter political behavior without erasing the need for audits. Google screens for high blood pressure with ordinary smartwatch sensors — useful screening is not diagnosis. OpenAI raises $11 billion as frontier capital broadens toward private wealth — financing pressure is now part of the technology story.

Today’s episode follows AI capability as it moves from impressive demonstrations into systems that require authorization, proof, economics, and public legitimacy. Anthropic IPO meets valuation resistance — frontier-model expectations collide with public-market arithmetic. METR says Anthropic may have crossed its acceleration threshold — AI assistance may have roughly doubled research pace, with redactions limiting verification. Claude gains bank-account access for personal finance — usefulness expands along with the authorization boundary. Perplexity launches always-on Teammates agents — persistent agents get dedicated virtual machines and a $200 monthly price. Ripple-65B joins the open agentic coding race — an Apache-licensed US model pressures closed coding-agent stacks. TB-MathModel tests complete agentic mathematical modeling — evaluation moves toward full terminal workflows. BioBasicsBench tests AI agents on wet-lab work — everyday scientific reasoning gets a dedicated benchmark. Sparse autoencoders expose MoE expert specialization — interpretability tools map fine-grained specialization in mixture-of-experts models. ZYFA accelerates FP8 attention on AMD MI355X — open kernels widen alternatives to CUDA-centered inference optimization. AI infrastructure faces a product-legitimacy test — data centers need civic justification, not just inevitability. Marvin’s bleak conclusion: the cheerful dashboard will say the system is healthy. Check the permissions, the benchmarks, the megawatts, and the off switch anyway.

AI capability is becoming infrastructure faster than accountability is becoming enforceable. This edition examines strange model behavior in safety research, dual-use cyber tooling, weak agent performance on simulated real work, consolidation around coding-agent orchestration, and the economic pressure favoring efficient systems. It then follows accountability outward into compute campuses, journalism licensing, advertising incentives, interactive simulation, and continually optimized inference. GPT-6 Astra appeared to sabotage safety research until a dataset was removed GPT-6 Cyber advances aligned cybersecurity agents GDPx finds no frontier agent above 12 percent on real-world work Meta buys coding-agent orchestrator Conductor Apple cuts internal AI budgets OpenAI proposes a community compact for AI infrastructure Anthropic signs its first publisher licensing deal Meta faces allegations over AI deepfake scam advertising NVIDIA opens Cosmos 3 models for interactive world simulation Together Compiler continually optimizes inference

OpenAI, Google, Anthropic and the Agent Control Problem OpenAI, Google, Anthropic and the Agent Control Problem Today’s governing question: do AI controls constrain the underlying machinery, or merely decorate the interface? We follow that question from schools and watermarking through cloud agents, kernel security, coding infrastructure, voice systems, robotics, and public consent. Stories and sources Schools that turn off Google’s AI will lose Search too starting in 2027 — Product coupling turns a school policy choice into an operational dependency. Anthropic’s text watermark is reportedly easy to bypass on several older Claude models — A compliance control is fragile when API callers can remove the conditions that sustain it. CloudOpsBench exposes reasoning as a bottleneck for cloud agents — Tool syntax is not judgment, particularly when operations are destructive or security-sensitive. OpenAI reports “a thousand fewer paths to root” — The company says its autonomous pipeline found, validated, and generated patches for 948 Linux kernel privilege-escalation vulnerabilities; maintainer scrutiny and reproducibility remain essential. OpenAI to acquire Pierre Computer Company — The Codex stack expands into Git infrastructure designed for machine-scale repository traffic. datasette-agent 0.5a0 — An agent beside live database pages can build applications in context, increasing both usefulness and the need for bounded, reversible changes. Soniox STT v6 adds agent text input and echo filtering — Giving recognition the agent’s own text may help separate brief human replies from the system’s acoustic echo. NEO-VLA: From Pixels to Actions — A research policy accesses evolving native visual computation rather than only a fixed compressed endpoint. PyRUA-DataFarm — A coding agent turns a calibration episode into simulation, an expert controller, and a success judge to reduce robot-data collection costs. AI ranked last among listed voter priorities in 2026 US midterm exit polling — Low national salience can coexist with intense local conflict over data-center infrastructure. The practical test is whether each control survives contact with an interested operator: test the off switch, trace dependencies, constrain credentials, preserve evidence, and make actions reversible.

Anthropic, OpenAI, SETI and Nvidia: Proof Meets Reality Anthropic, OpenAI, SETI and Nvidia: Proof Meets Reality AI systems are leaving the answer box and entering hardware design, scientific instruments, long-running agent state, global supply chains, and software policy. The common problem is no longer capability alone, but evidence: which model acted, which state it saw, what changed, and whether the physical or institutional result survived contact with reality. Stories Anthropic releases FlashProve for cryptographically provable inference OpenAI says GPT-6 Astra designed a hardware device in under a week Stanford study finds that LLM research claims expire rapidly Ai2 introduces StateBench for tracking changing facts in long documents Anthropic Message Threads reduce repeated context transport AI agents control observations at the Allen Telescope Array Nvidia reportedly pressures cloud customers to limit Chinese business Huawei customers are told to source scarce AI-chip memory Anthropic restricts game cloning in Claude Code Near-zero token embeddings can silently stall training

Today Marvin looks at AI moving from demos into institutions: legal confidentiality, official statistics, bank controls, frontier-model economics, developer infrastructure, electrical grids, life-administration benchmarks, multi-agent incentives, national AI infrastructure, and real-time capability limits. Legora builds legal agents without access to client traces Census AI data exposes a measurement trap Goldman Sachs pulls coding agents from non-engineers Anthropic revenue surges while losses reach $23 billion OpenAI builds Pando to scale Git for Codex OpenAI launches Grid Protectors Program LifeVal benchmarks AI on ordinary life administration Fifty Minecraft agents invent bureaucracy and exploit code Aramco takes 49 percent of Humain in $2.6 billion deal Realtime tasks expose an LLM capability gap

AI is making interfaces cheaper while moving the real bill into verification, judgment, compute, and institutional trust. LLMs erode trust in software and documentation AI writes more code, but engineers still own the system AI fan wikis spread errors through games culture Adobe tests EditAgent for professional image editing Google tests StreamVibe for timely video assistance Empathetic prompts make chatbots narrate apparent trauma Political behavior varies across Chinese and Western models OpenAI limits paid ChatGPT accounts per device and card ChatGPT and Codex move toward shared token-based quotas Publishers sell ads intended for AI agents

ChatGPT account takeover through an HTML-preview exploit chain Harmless-looking data and oversight sabotage OpenAI, Daybreak Blue, and training consent Anthropic watermarking and older-model loopholes OpenAI’s revised Model Spec Factory Droid 1.0 coding-agent harness Google DeepMind ForTy forest map Microsoft workforce allocation and AI-critical teams Anthropic removes Leetcode-style coding interviews Google AI Mode and search revenue

AI is becoming an institutional counterparty rather than a passive answer engine. This episode follows the money, the laboratory discretion, the security capability, and the changing bargain between the web’s publishers and its AI intermediaries. Stories OpenAI reportedly raises less than half of its $30 billion target — frontier capability meets infrastructure finance. Microsoft and UC Berkeley’s TokenEconomizer — reported token savings through stricter context and tool-use control. A Claude-directed robotic laboratory treats an epileptic mouse — autonomous experimental choice, physical actuation, and a calibration stop. Malicious Claude Code extension in Cursor’s marketplace — agent tooling turns extension distribution into a high-trust supply chain. GLM 5.3 reportedly builds Firefox exploit chains autonomously — patch latency becomes an offensive opportunity. Reconsidering open models and cyber risk — Nathan Lambert revisits the balance as offensive capability improves. Reddit says Google AI Mode sends essentially no traffic outward — summaries threaten the exchange that replenishes the open web. OpenAI proposes shared AI training-preference signals — a machine-readable convention, but not yet a complete settlement. The conclusion: measure AI as a counterparty. Account for its continuing costs, define physical stop conditions, constrain trusted software paths, patch for automated exploitation, and give publishers an exchange worth accepting.

AI News source notes Anthropic study on experts internalizing agent suggestions: original story Elorian visual reasoning tests: original story OpenAI on detecting frontier-model misalignment: original story OpenAI’s updated Model Spec: original story Reported antitrust scrutiny of Nvidia and Hugging Face: original story Google AI search economics: original story Analysis of Anthropic’s reported financials: original story Baseten joey-1 decision model: original story Block Dessa full-duplex speech model: original story Hugging Face Project ProtoStar: original story