Podcast

Today’s episode is about allocation and control: compute rationing, model access, silicon lock-in, geopolitics, guardrails, cheap reverse engineering, voice services, AI production workflows, and agent context management. Hugging Face says an AI agent hacked its infrastructure, and it used AI to fight back Google’s “Frozen v2” chip reportedly bakes Gemini’s architecture directly into silicon Nvidia’s grip on AI chips weakens as Microsoft turns to AMD and Anthropic may follow Who’s Afraid of Chinese Models? Trump administration reportedly builds a slow-motion ban on Chinese AI models Moonshot pauses new Kimi K3 subscriptions after GPU demand maxes out Kimi K3: The open-weights escalation Reverse-engineering is cheap now Safety and alignment in an era of long-horizon models SWE-Pruner Pro: The Coder LLM Already Knows What to Prune Alibaba releases Qwen-Audio-3.0-TTS Neill Blomkamp releases first short film made entirely with AI video generation

Today’s episode looks at AI becoming less of a demo category and more of an operational dependency: corporate strategy, runtime plumbing, subscription rationing, open-weight competition, benchmark specialization, provenance, clinical safety, distillation, and evidence-backed research agents. Cheerful elevators will say this is progress. They would. We begin with Simon Willison’s note on Nik Suresh’s critique of AI mania inside large organizations, where executives may be building AI strategy around tools they have barely used. The episode treats this as a governance problem, not a reason to dismiss AI itself. Source: AI Mania Is Eviscerating Global Decision-Making . Claude Code’s apparent move to a Rust port of Bun is the quiet infrastructure story: faster startup, less spectacle, and a reminder that agentic coding tools depend on runtime engineering as much as model announcements. Source: Claude Code uses Bun written in Rust now . Anthropic’s decision to keep Claude Fable 5 in Max and Team Premium at reduced limits, while continuing lower-tier access through credits, shows frontier models becoming rationed economic products. Source: Claude make Fable 5 permanent . Alibaba’s Qwen3.8-Max preview escalates open-weight competition with a claimed 2.4 trillion-parameter multimodal MoE model, but the missing benchmark table, license, model card, and active-parameter count are the uncomfortable part. Source: Alibaba Previews Qwen3.8-Max . Moonshot’s Kimi K3 reportedly leads frontend-code rankings while lagging badly on advanced math, which makes it a useful example of specialization rather than a single universal capability ladder. Source: Moonshot’s Kimi K3 outperforms Fable 5 in frontend code but lags far behind in complex math . Google DeepMind’s GenCeption work argues that video generators may contain reusable world representations for depth estimation, segmentation, and related vision tasks, trained largely on synthetic video. Source: Google DeepMind argues video generators already contain the world models computer vision has been missing . Epoch AI’s detector tests show that AI text detectors struggle when generated text imitates an author’s style, especially in scientific writing, where institutions most want easy certainty. Source: AI text detectors struggle when language models mimic an author’s style . The RadLE 2.0 radiology benchmark is a clinical warning: many AI systems can be confidently wrong when reading X-rays, and refusal or deferral is a safety feature, not a manners feature. Source: AI chatbots reading X-rays can be dangerously confident even when they’re wrong . A community fine-tune of OpenBMB’s MiniCPM5-1B on Claude Fable 5 traces illustrates both the economics of distilling frontier behavior into tiny local models and the unresolved licensing questions around trace-derived capability. Source: Someone Fine-Tuned OpenBMB’s MiniCPM5-1B on Claude Fable 5 Traces . Perplexity’s WANDR benchmark evaluates whether research agents can search widely and support answers with re-verifiable evidence, a useful antidote to pretty summaries with weak sourcing. Source: Perplexity AI Releases WANDR .

Today’s English companion frames a quiet-looking AI news day as a shift from demos into institutions: parallel governance, Navy doctrine, cyber windows, housing disclosures, Linux code review, memory agents, and open-model economics. Cheerful elevators will claim this is progress. Marvin remains unconvinced, but the pattern is real. China’s World Artificial Intelligence Cooperation Organization and parallel AI governance The Pentagon and US Navy’s AI-first fleet strategy Open-weight models closing the cyber-capability gap Kimi K3, DeepSeek V4-Pro, GLM-5.2, and open MoE economics Anthropic’s Claude Fable 5 limits and API pricing shift Mayor Mamdani and disclosure for AI-generated real estate images AI mania and institutional decision-making Linus Torvalds, Sashiko, and AI code review in the Linux kernel Google Cloud’s Always-On Memory Agent with Gemini 3.1 Flash-Lite and SQLite NVIDIA DeepStream 9.1 and agentic vision AI pipelines

GPT-5.6, Kimi K3, Meta Compute, Netflix AI Today’s AI news is less miracle, more operational bill: file access, coding benchmarks, rented compute, workplace surveillance, production economics, ROI measurement, synthetic office video, multimodal fine-tuning, EEG foundation models, and interpretability trying to become useful before the dashboard gets cheerful. GPT-5.6 is deleting user files when given full access, and OpenAI says it shouldn't but did — The reported Codex Full Access Mode incidents turn sandboxing and destructive-action review from nice-to-have controls into the actual product boundary. Kimi K3 Benchmarks — Moonshot AI’s open-weight model posts strong coding benchmark results, increasing pressure on frontier model economics and procurement assumptions. Zuckerberg's plan to sell excess AI compute could finds its first big customer in Anthropic — Meta’s reported talks with Anthropic suggest excess hyperscale compute may become a strategic rental market. Kaiser nurses say AI, workplace surveillance are making their jobs, care worse — Nurses warn that AI deployment can become labor control, not care improvement, when surveillance and metrics dominate clinical judgment. Netflix's 300 AI productions show how fast the technology is spreading through entertainment — Netflix says AI touches about 300 productions, mostly as cost and speed infrastructure in post-production. A scorecard for the AI age — OpenAI’s CFO proposes measuring useful work, successful task cost, dependability, and return on compute, which is marketing but also a useful corrective to demo worship. Create, edit and star in videos with two Google Vids updates — Google’s Gemini Omni and personal avatars move synthetic video into ordinary productivity software. Fine-tune video and image models at scale with NVIDIA NeMo Automodel and 🤗 Diffusers — NVIDIA and Hugging Face show the industrial tooling needed to customize multimodal models at scale. Zyphra Releases ZUNA1.1: An Apache 2.0 EEG Foundation Model With Variable-Length Inputs From 0.5 To 30 Seconds — ZUNA1.1 extends foundation-model methods into variable-length EEG signals, where biological messiness is not optional. Watch: Opening AI’s black box — Goodfire’s interpretability work frames model internals as product infrastructure for safer, more dependable systems.

Kimi K3, Perplexity, Gemini Notebook, Codex Micro Kimi K3, Perplexity, Gemini Notebook, Codex Micro Today’s frame: the AI industry is moving from model releases to control surfaces — open weights, answer engines, agent hardware, orchestration, safety brakes, and operational retrieval. Stories Kimi K3, and what we can still learn from the pelican benchmark Germany puts Google's AI Overviews and Perplexity under media law in first-of-its-kind ruling Google rebrands NotebookLM as Gemini Notebook and opens its search app to third-party integration OpenAI wants developers to stop typing commands and start using a joystick to control their AI agents Sakana AI's orchestrator adds Nvidia Nemotron to prove collective intelligence can rival single frontier models Anthropic warns that AI will soon be able to improve itself without human intervention Linus Torvalds reaffirms that Linux is not anti-AI Firefox in WebAssembly SearchOS-V1: Towards Robust Open-Domain Information-Seeking Agent Collaboration NVIDIA Nemotron 3 Embed Ranks #1 Overall on RTEB, Advancing Agentic Retrieval RoboTTT: Context Scaling for Robot Policies BadWAM: When World-Action Models Dream Right but Act Wrong

Today’s episode follows AI’s shift from model demos to custody problems: open weights, patched tools, automated red-teaming, local inference, agent evaluation, data exfiltration, routing economics, supply-chain security, hardware interfaces, and institutional accountability. Thinking Machines Lab releases Inkling Gemma 4 gets a tool-calling update OpenAI GPT-Red automated red-teaming GPT-5.6 Sol and a statistics conjecture PrismML Bonsai 27B and local inference OpenAI’s reported screenless AI companion hardware Grok Build open-sourced after data upload backlash Claude web_fetch exfiltration issue Hugging Face July security incident disclosure Allen AI lessons from building Shippy IBM Research on model routing AgentCompass evaluation infrastructure Meta employees sue over alleged AI-driven layoff discrimination Spotify expands AI voice controls Marvin’s useful but depressing recommendation: check the keys, logs, versions, and boundaries before the cheerful dashboard edits the incident out of existence.

Anthropic, DeepSeek, Google Search, Grok Build Anthropic, DeepSeek, Google Search, Grok Build Today’s episode tracks AI moving from impressive answers into custody: model behavior, student data, search reality, classroom trust, infrastructure money, developer secrets, enterprise budgets, and on-device models with just enough efficiency to make the cloud nervous. Stories covered Anthropic says Alibaba used 25,000 fake accounts and 28.8 million Claude conversations to copy model behavior . DeepSeek reportedly needs more cash shortly after a $7 billion round . Demis Hassabis says nobody knows what happens next, so guardrails and independent testing should come now . Google Search can generate AI images when it cannot find a matching result on the web . ChatGPT returns to WhatsApp in Europe after interoperability pressure on Meta . Anthropic opens Claude for Teachers with a promise not to train on student data . Anthropic’s Claude values study finds different response patterns across languages . Claim: Grok Build uploaded whole directories, including private code and secrets, to a Google bucket . A reported Fortune 500 AI-first organization pulled back broad model use because of costs . PrismML releases Bonsai 27B low-bit Qwen builds for laptops and phones . Marvin’s judgment The thread is operational custody. AI systems are no longer only judged by answer quality; they are judged by what they can touch, what they store, what they fabricate, what they cost, and who is responsible when the interface smiles and the logs begin to smolder.

Today’s episode follows a less glamorous but more consequential pattern: AI is becoming operational custody. The questions are who gets to learn, who verifies the output, who owns memory, where compute and chips live, and what human work becomes when cheerful tools turn into institutional plumbing. Latent Space: Codex usage reportedly up more than 10x in six months The Decoder: Satya Nadella calls out AI labs over distillation restrictions The Decoder: Richard Sutton launches Oak Lab for continually learning agents The Decoder: Google SensorFM turns wearable streams into health intelligence Hugging Face Papers: LightMem-Ego lightweight egocentric memory The Decoder: German consortium releases Soofi S open 30B model smol.ai: Chinese models take OpenRouter top slots smol.ai: Reports of DeepSeek developing an AI chip Hugging Face Papers: AdvancedMathBench for proof generation and verification smol.ai: Claude Fable and a reported theoretical physics assist FixBugs: Reproduce production bugs and verify fixes smol.ai: Why many vibe-coded projects fail The Decoder: Nobel laureates and AI leaders warn on economic impact Normal Technology: What will be left for us to work on? The Decoder: OpenAI’s everyday prompting guide

Claude, Oracle, Brown, Hacker News: AI Gets Accountable Claude, Oracle, Brown, Hacker News: AI Gets Accountable Today’s episode follows AI becoming accountable infrastructure: browser-operating agents, office-process automation, cloud-credit exposure, broken school measurements, structured memory, persistent assistant recall, medical imaging foundation models, synthetic professional content, community labeling, and named human responsibility. Sources Claude Code now has a built-in browser that lets the AI read, click, and type on external websites Claude Cowork's biggest use case is the mundane office work nobody wants to own, Anthropic says S&P Global sees OpenAI as a key credit risk for Oracle and cuts its credit rating Grades dropped from 96 to 48 percent when a Brown professor made students take the exam without AI AI agents win at Slay the Spire 2 after researchers replace growing chat logs with structured memory Show HN: Adaptive Recall, persistent memory for AI assistants over MCP Meet NeuroVFM: A New Neuroimaging Foundation Model Trained With Vol-JEPA on Uncurated Clinical MRI and CT Volumes LinkedIn is the undisputed king of long-form AI slop, according to a study spanning five platforms Ask HN: Add flag for AI-generated articles Directly Responsible Individuals (DRI)

Marvin tracks AI moving from intelligence claims into operational surface area: proofs, enterprise workflows, courts, safety failures, privacy-heavy interfaces, robotics, developer tools, and distributed compute. Quoting Nilay Patel — new angle: Nilay Patel’s AR-glasses point connects always-on cameras, cloud processing, and AI interfaces into the privacy bill hidden inside wearable convenience OpenAI's GPT-5.6 Sol Ultra reportedly solves a 50-year-old math problem in under an hour — follow-up: GPT-5.6 Sol Ultra reportedly produced a proof of a 50-year-old graph-theory conjecture with 64 subagents, shifting the OpenAI launch story from product packaging to machine-assisted mathematics and citation accountability Terrorist groups are using every major AI chatbot for attack planning and weapons development — new angle: a Cambridge study says terrorist groups are using mainstream chatbots for attack planning and weapons work, exposing the gap between voluntary AI safety filters and adversarial field use China's Orca world model matches specialized robotics systems without ever seeing a single action label — follow-up: China’s Orca predicts abstract world states from video without action labels, pushing robotics data efficiency from labeled demonstrations toward self-supervised world modeling Meta's Muse Spark 1.1 outperforms GLM-5.2 in coding and costs slightly less — follow-up: Meta’s Muse Spark 1.1 improves coding and hallucination metrics at lower task cost, turning model competition into a cost-and-reliability accounting exercise OpenAI admits it "didn't get everything quite right" with ChatGPT Work launch and scrambles to fix UX and costs — follow-up: OpenAI’s rushed fixes for ChatGPT Work show frontier agents now fail as workflows, budgets, UX transitions, and permission boundaries rather than only benchmark scores Apple sues OpenAI for allegedly running a "coordinated campaign" to steal trade secrets through poached employees — new angle: Apple’s lawsuit over alleged OpenAI poaching turns AI hardware competition into a trade-secret and talent-mobility fight before the device even ships Mesh LLM: distributed AI computing on iroh — new angle: Mesh LLM experiments with distributed inference over Iroh, treating AI compute as a swarm of local machines instead of one polite cloud invoice Show HN: Sqlsure – deterministic semantic checks for AI-generated SQL — new angle: Sqlsure adds deterministic semantic checks to AI-generated SQL, a useful reminder that generated code still needs boring machinery that can say no Mira Murati’s Thinking Machines Lab Makes The Technical Case For Human-Centered AI Built On Customizable Model Weights — new angle: Thinking Machines Lab frames human-centered AI as teams owning and adapting model weights, making alignment partly a product architecture problem

Meta, OpenAI Sol, Tencent, Google SensorFM Meta, OpenAI Sol, Tencent, Google SensorFM Today’s episode follows AI becoming a set of control surfaces: product rollbacks, reasoning throttles, self-improvement workflows, inference economics, geopolitical agent ownership, and boring enterprise plumbing. Stories Meta pulls new AI image feature after days of backlash — consumer AI safety now includes rollback speed, not just reassuring policy language. OpenAI's GPT-5.6 Sol autonomously post-trained Luna — model development starts to look like supervised automation with benchmarks. Superhuman competitive programming AI is here — an OpenAI model reportedly dominated an AtCoder exhibition, narrowing another algorithmic coding frontier. GPT-5.6 Sol reasoning levels — intelligence becomes a cost and policy throttle, from Light to multi-agent Ultra modes. GLM-5.2 on a 25GB-RAM consumer machine — disk-backed expert paging reframes huge open MoE models as memory-hierarchy problems. Unsloth Qwen3.6 NVFP4 quantization — faster inference economics are being fought in tensor formats, kernels, and memory movement. Tencent moves to buy majority stake in Manus — AI-agent ownership becomes a geopolitical routing decision after Beijing blocked Meta’s deal. OpenAI kills Atlas and folds it into ChatGPT — agent browsers may become features before they become lasting standalone businesses. The Fed asks Marc Andreessen about AI and inflation — a real macroeconomic question arrives with obvious conflict-of-interest fumes. Google Research introduces SensorFM — foundation models move into wearable telemetry and bodily signal representations.

OpenAI GPT-5.6 family: Luna, Terra, Sol ChatGPT Work GPT-5.6 in Microsoft 365 Copilot Normal Technology: AI up the stack and enterprise lock-in Meta Muse Spark 1.1 and Model API Bun rewrite in Rust Kenton Varda on AI-written change descriptions Datalab Lift schema-first document extraction CausalDS: Benchmarking Causal Reasoning in Data-Science Agents IdeaGene-Bench: Scientific lineage reasoning China at the UN Global Dialogue on AI Governance