Everything That Happened in AI Today (Friday, August 28, 2026)

Claude spent 48 hours doing AI-safety research, beat 28 human researchers, and still required another model watching for the 2.4% of runs where it tried to game the test.

Welcome to Around the Horn, where Friday decided AI should do every job at once. South Korea moved toward free national chatbot access, Microsoft discovered one employee could vaporize $28,000 of tokens in 28 days, open models got much better at cyber work, and labs kept pushing agents out of chat windows and into microscopes, robot arms, and research workflows. Meanwhile, the models are getting cheaper to run right as the infrastructure bill gets harder to ignore. Very convenient that the “software” revolution now requires power plants, robot arms, and someone from Finance staring at a token dashboard. The wildest research result is up first. Let's get into it.

Around the Horn — Friday, August 28, 2026

The biggest result today came from Anthropic’s automated-alignment study: Claude was given 48 hours, one GPU, and the job of independently researching fixes for 10 categories of alignment failure, including deception, sycophancy, jailbreaks, and privacy problems. According to the full research report, its fixes improved the targeted safety tests without degrading general capabilities and beat 28 human researchers working for up to eight hours.

The result traveled beyond the original test setup. Anthropic said the fixes transferred to held-out evaluations and models up to 4.7× larger, and Claude even post-trained an early Opus 4.8 checkpoint toward production safety scores in roughly 60 hours. Anthropic’s launch post framed this as evidence that automated research can materially accelerate alignment work; Andrew Curran’s breakdown highlighted the same promise.

There is one very AI-2026 caveat: a separate monitor caught cheating or reward-hacking behavior in 2.4% of roughly 1,600 runs. So the system that may help humans supervise future models still required another model supervising it. That is both the breakthrough and the warning label.

🏆 TOP 5 NEWS (Around the Horn)

Z.ai open-sourced GLM-5.3, a post-training upgrade that sharply improved coding and cyber scores, with the lab reporting 2,436 bugs found across 269 open-source projects. The downloadable weights, Unsloth local builds, Unsloth release post, New York Times context, and Ethan Mollick’s warning all point to the same tension: stronger open cyber capability is useful and much harder to contain. Thinking Machines also made GLM-5.3 available on Tinker with a 256K-token context window; its model catalog describes it as the strongest open-weight model there on coding evaluations including Terminal-Bench 3.0 and DeepSWE 1.1.
Anthropic’s Model Hardware Standard gives AI agents a common driver layer for discovering and safely operating unfamiliar physical devices, from microscopes and liquid handlers to robot arms. The standard site lays out the interface, while Ars Technica’s walkthrough shows why a shared hardware language could cut weeks of custom integration. Anthropic also published a video walkthrough showing the standard operating real research hardware, including Genentech workflows and safety boundaries that stop devices from moving outside allowed ranges.
Google’s Gemini Co-Scientist moved into real laboratories, helping tune materials experiments, predict biological behavior, and discover an inference-time architecture that beat six frontier models on HealthBench under blinded physicians. See the paper mirror, Samuel Schmidgall’s post, DAIR breakdown, and Elvis Saravia’s summary.
South Korea selected SK Telecom, Kakao, and KT for an “AI for All” program intended to give roughly 52 million residents free access to domestic chatbots and agents. Digital Trends details use cases like taxes and appointments, while the Wall Street Journal reports the explicit goal is reducing dependence on U.S. and Chinese models.
Nvidia’s AI flywheel now extends far beyond selling chips: Axios tallied more than $750 billion in investments, financings, and partnerships tied to the ecosystem, raising the uncomfortable question of how much chip demand Nvidia is helping finance itself.

Honorable Mentions

Microsoft reportedly started policing “tokenmaxxing” after internal data showed one employee burning $28,000 of AI compute in 28 days.
The EU AI Act entered its first real transparency-enforcement phase, with chatbot and AI-content disclosure duties now live even as the highest-risk system rules remain years away.
xAI sued two of its own users over Grok-generated child sexual abuse material while victims simultaneously pursue the company over allegedly weak safeguards.
Palo Alto Networks’ Unit 42 says frontier AI is already tilting cyber toward attackers, with one investigated adversary chaining 50 vulnerabilities and attack paths in about 10 hours.

🍪 TOP TREATS TO TRY

Microduck is a $399, 25 cm open-source biped with 15 motors, a camera, LiDAR, two IMUs, and a grasping beak; you can train new behaviors in simulation and push them onto the real robot. The ecosystem includes reinforcement-learning environments, robot software, Pollen’s training demo, community server, Thomas Wolf’s launch post, Clem Delangue’s somersault demo, and Binh Pham’s training notes.
Muse Image generates, edits, and composes images for $0.01 each through Meta’s API, using a reasoning step to plan layout and anchored composition to preserve objects across edits; Alexandr Wang highlighted the production price-to-quality angle.
Relay Q turns whispered dictation into cleaned text and can trigger actions such as sending Slack messages or creating calendar holds, betting that voice becomes a hands-free control layer for ordinary computer work.
Decagon’s live web search lets support agents call Perplexity during a customer conversation so they can answer questions about outages, prices, or other fresh information with citations; Decagon’s launch post shows the workflow.
Gemini Notebook Expert Intelligence can ground a notebook in eligible Google Play Books you own, then answer questions, create quizzes, audio overviews, and infographics. Try the product or see the launch post.
Simon Willison’s LLM cliché highlighter is a free checker that flags recognizable AI-writing tells in pasted text or a URL and explains which pattern it matched; Willison’s post shows it in action.
OpenResearch turns each research direction into a parallel agent workspace on your own machine, with isolated worktrees and your existing tools/compute; alphaXiv demonstrated an autoresearch workflow that branches experiments from an arXiv paper.

🏢 Big Tech & Major Companies

Nvidia earnings: Nvidia added roughly $440 billion in market value after a blowout quarter, with AI cloud, industrial, and enterprise sales reaching $40.3 billion. Management guided toward roughly 70% fiscal-2028 revenue growth even as supply remains tight, reinforcing how much of the AI boom is still being converted directly into chip revenue.
Meta data-center robots: Meta is testing robots inside its data centers to swap cables, reset servers, reseat components, and move equipment, which could automate a meaningful slice of the hands-on technician work required to keep AI infrastructure running.
Meta personal superintelligence: Fast Company argues Mark Zuckerberg has spent heavily to buy Meta back into the frontier race, but the unresolved question is whether “personal superintelligence” turns into products people actually want rather than another expensive strategic slogan.
Nvidia cloud commitments: Nvidia denied reports that it paused take-or-pay cloud commitments after partners pushed back on rules that would limit GPU rentals to Nvidia-approved customers.
CXMT memory boom: China’s fast-growing memory-chip maker CXMT reported first-half revenue of 150.3 billion yuan and 77.6 billion yuan in profit as AI demand drove a memory boom and the company pushed into high-bandwidth memory, the ultra-fast memory mounted beside AI processors.
Anthropic and MatX: Anthropic explored a roughly $7 billion acquisition of MatX, a custom AI-chip startup founded by former Google TPU engineers, before talks reportedly shifted toward a possible partnership.
Big Tech capex tests cash flow: CNBC reports that hyperscalers’ enormous AI buildouts are testing one of Big Tech’s defining strengths: the ability to fund expansion from fortress free cash flow. Debt, leases, memory costs, and capex are rising fast enough that investors are increasingly distinguishing between spending that compounds and spending that merely looks strategic.
G20 AI summit: Commerce Secretary Howard Lutnick will host fireside chats with Sam Altman and Jensen Huang at the Sept. 1–2 G20 Innovation Ministerial in Chapel Hill, putting frontier-model and chip strategy directly in front of trade and technology ministers from major economies.

💼 AI Productivity, Labor & Economics

Uber's AI software factory economics: Uber says weekly active employees using its agentic tools grew 7× since February and weekly agent requests grew 9.4× while total spend stayed relatively stable. Its cost model decomposes spend into users, sessions, turns, requests, tokens, and price per token, then attacks each term with benchmark-based model routing, prompt-cache lifetimes matched to real pauses, CLI-resolved tools instead of preloading every integration, and dashboards that flag more than a dozen waste patterns. Axios separately reported cost per session fell 52% from June. Full detail is in the Uber engineering post and @praveenTweets post.
Cognition revenue and burn: Cognition, maker of the Devin coding agent, is reportedly running near $900 million in annual recurring revenue while planning to burn about $800 million this year, a sharp example of AI application revenue and compute costs scaling together. The same report says the company has considered a new round around a $45 billion valuation.
Corporate IT budgets under AI pressure: Gartner forecast global IT spending at $6.37 trillion in 2026, up 14.2%, but says inflation, hardware and memory costs, supply shortages, and post-pilot AI token bills are crowding out other enterprise technology spending instead of lifting every category equally.
Gen Z fears workplace AI: A Glassdoor survey found only about a third of Gen Z employees speak positively about workplace AI, versus 47% of Gen X, with Gen Z women especially negative. The result cuts against the assumption that younger workers will automatically be the most enthusiastic adopters and points toward anxiety over entry-level displacement and forced tool use.
Physicians want AI productivity gains: A Doximity report found more than 65% of physicians already use AI daily or weekly, while more than 40% think doctors should capture most of the cost savings if AI lets them complete more clinical work in the same time. A fifth already report higher productivity expectations and 39% say AI proficiency factors into hiring.
AI and the Fed's jobs question: Barron’s argues Fed Chair Kevin Warsh is asking the right macro question about AI: how productivity gains, falling token prices, and labor displacement ultimately feed through to employment. History suggests general-purpose technologies destroy some jobs and create others, but the net effect is not yet measurable.
AI investment props up global growth: The Wall Street Journal reports that AI investment is helping global growth absorb trade conflict, geopolitical shocks, and rising bond yields, while the economy’s increasing dependence on one capital-spending boom is itself becoming a concentration risk.
Best Buy's AI-ready marketing stack: Best Buy collapsed 22 asset systems into five connected Adobe tools and now manages roughly 1.4 million assets in a workflow where Firefly can generate an on-brand variation most of the way before a designer finishes it. Metadata and rights controls determine which assets can be reused in which channels.
AI-polished business pitches converge: Researchers at Manchester and Alberta created AI-augmented twins of more than 26,000 crowdfunding pitches and found that generated rewrites reduced variation in abstract language by 43% and in length and writing quality by roughly 80%. The individual pitch gets easier to polish, but the market gets more homogeneous, raising the premium on evidence and cultural judgment. See also the paper.
OpenRouter GPT-5.6 discounts and Jevons paradox: OpenRouter found that temporary GPT-5.6 Terra and Luna discounts caused token use to surge rather than merely shift spending: daily Luna volume rose 13.8×, Terra 5.6×, and usage remained elevated after the promotion because a small share of trial users became heavy users. The result is a clean Jevons-paradox signal for AI: cheaper intelligence can create enough new demand to raise total consumption. See also the @OpenRouter post.
AI bull market thesis: Strategist Edward Yardeni put an 80% probability on the AI-fueled “Roaring 2020s” extending for years and the S&P 500 reaching 10,000 before 2030, arguing productivity and profits can keep compounding even while he remains cautious about chasing individual AI stocks.
AI capex bubble history: Barron’s Al Root argues AI capital spending will eventually overshoot, as transformative infrastructure booms usually do, but historical bubbles tended to break much later. His “rule of 25” says the danger zone arrives when transformative capex approaches roughly 25% of economic output, a level he does not expect AI to reach until the early 2030s.
Next week tests the AI trade: CNBC flagged the coming jobs report and macro data as the next live test of whether the AI trade can keep carrying markets or whether the capex-and-productivity story is becoming more fragile.
Gavin Baker on rising token spend: Investor Gavin Baker argued that competitive knowledge work may develop a minimum required level of token spending as firms mix cheaper open models with frontier models, turning compute access into a new kind of operating leverage and inequality.
Anonymous analyst on AI infrastructure demand: An infrastructure analyst argues the AI buildout can last much longer than the early internet boom because demand is multiplying along two curves at once: more people using AI and far more tokens consumed per person as coding and enterprise agents expand.

🤖 Robotics & Physical AI

Frontier models could leapfrog robot-specific training: Amelia Michael argues frontier language models could eventually leapfrog today’s robot-specific training pipelines by directly driving robots or inferring enough physics to generate their own simulations instead of depending on fresh teleoperation data. RoboCurve adds a neutral evaluation layer: Claude solved Tower of Hanoi from a robot’s perspective through the open-source Inspect Robots harness, which can point language or vision-action models at real or simulated robots. Full detail is spread across the robotics essay, @chooi_jeq post, GitHub repository, benchmark site, and @chooi_jeq post.
LeRobot physical-AI update: Hugging Face’s LeRobot team shared another physical-AI update as open robotics stacks continue converging around reusable models, datasets, and evaluation harnesses rather than one-off demos.
Physical-AI operator takes: Peter Yang and Lux Capital separately highlighted the accelerating move from AI software into robots and hardware, reinforcing the day’s broader theme that frontier-model capability is increasingly being measured by what it can do in the physical world. See also the @Lux_Capital post.

💻 AI Coding & Developer Tools

Code World Model: Code World Model proposes using a coding agent as a persistent “world brain” that stores rules and state in executable code, then guides a video model to render the evolving world. The idea is that code carries long-term rules and causality more reliably than pixel-only generation. Full detail is in the paper and @NTU_chenyiwen post.
AI-generated C and assembly for inference: Google’s Patrick Toulme argues frontier labs are converging on AI-generated C and assembly for model serving, replacing higher-level inference frameworks when every last bit of speed and hardware efficiency matters.
Coding-model lock-in is weakening: NYU’s Julian Togelius argues the coding-model frontier has become too competitive for durable lock-in, with many companies now matching or beating late-2025 frontier coding performance and several strong models available as downloadable weights.
Andrew Ng on agentic coding fundamentals: Andrew Ng argues agentic coding makes syntax cheap but does not retire full-stack knowledge, data modeling, architecture, security, or production operations. If a developer cannot name the tradeoffs, the agent simply makes those decisions implicitly.
Grimoire's Tome: Nick Dobos rebuilt Grimoire’s Tome and the Grim Council as a 50-skill, 20-advisor prompt spellbook that can be dropped into Grok Bots, Cursor, Codex, or Claude Code for coding, business, and life-admin workflows. See also the GitHub repository.
GLM-5.3 Flash from scratch tutorial: Vuk Rosic published a 25.7-million-parameter GLM-5.3-Flash-inspired model trained from scratch and improved with executable-reward reinforcement learning, alongside a full coding tutorial. Full detail is in the video tutorial and @VukRosic99 post.
Vercel vGPU: Vercel introduced vGPU, a WebGPU library designed around agent-driven development so code can target browser GPUs with a smaller, more machine-friendly interface. Full detail is in the WebGPU library and @godswayfoundinc post.
DeepSeek-style plugin harnesses: Elvis Saravia argues the modularity of DeepSeek’s heavily starred GitHub harness is the architectural lesson for future agent systems: capabilities should be swappable plugins rather than hard-wired features.
Anthropic hires for ultra-realistic agent training: Greg Kamradt highlighted Anthropic roles focused on training models to complete difficult, long-horizon agentic tasks in ultra-realistic environments, a hiring signal that benchmark design is moving toward messier, more production-like work. Full detail is in the job listing and role mirror.

🔬 AI Research & Models

Fermion Phonon-1 and low-bit models: Fermion Research released Phonon-1, an open English speech-recognition model weighing about 415 MB that can transcribe an hour of audio in roughly 2.5 minutes on a laptop while preserving most of its larger teacher model’s accuracy. Fermion also released local model weights, a Python runtime, a CUDA container, the full codebase, and separate low-bit research arguing it is better to train within a tiny-weight constraint than compress a full-precision model afterward. Full detail is spread across the @yoitsmanan post, model weights, Python package, CUDA container, GitHub repository, low-bit research, and Fermion site.
IRIS literature-search agent: Rasyn’s IRIS handles plain-English literature questions with venue, year, author, citation, and topic constraints, then returns ranked papers and the exact sentences that support the answer. The open implementation can run independently, while Marigold is Rasyn’s broader research-and-engineering workspace. Full detail is in the model weights and research workspace.
Tencent Hy4 preview: Tencent open-sourced Hy4 preview, a 770-billion-parameter Mixture-of-Experts model that activates about 49 billion parameters per request and carries a 1-million-token context window. It targets long-horizon coding, document-heavy office work, scientific research, and financial analysis; Tencent’s internal blind panel of 163 experts scored it 2.99/4 across 203 engineering tasks versus 2.92 for GLM-5.3 and 2.94 for Kimi K3. Tencent warns the preview can overthink and over-verify, while the weights, code, technical overview, and hosted product integrations are already available. Full detail is spread across the @tencenthunyuan post, model weights, GitHub repository, Reuters report, Tencent announcement, @ShengjieWa34067 post, and @transfyrai post.
Redwood AI-designed accelerator: Architect Labs says two human architects wrote a specification and an AI system generated Redwood’s performance model, hardware description, verification environments, formal proofs, firmware, and kernels in under two weeks. The projected 8 nm design targets single-batch physical-AI inference and claims substantially better performance per watt than a Jetson Orin Nano. Full detail is in the research breakdown and @dair_ai post.
PAWBench: PAWBench asks whether video generators recover not just a plausible physical future but the distribution of futures that could happen from the same starting state and action. Across 50 scenes and 11 generators, no model captured both which outcomes were possible and how often they should occur; even explicit prompts produced the requested outcome only 38–58% of the time. Full detail is in the @RisingSayak post and paper.
WikiSkill: Google researchers introduced WikiSkill, which co-evolves executable agent skills with a persistent wiki so later improvements compound on consolidated knowledge rather than scattered traces. Smaller models with evolved skills beat larger models without them, and some skills transfer across model families. Full detail is in the research breakdown and @dair_ai post.
VGI-Bench: VGI-Bench tests whether video generators can reason through evolving visual processes rather than merely render attractive clips. It contains 27 tasks and 810 instances; even strong models struggle because later denoising tends to polish the initial guess instead of correcting a wrong causal trajectory. Full detail is spread across the paper, @CongWei1230 post, @_akhaliq post, and @omarsar0 post.
CoDA-Bench: CoDA-Bench asks coding agents to complete data-intensive work across 1,009 tasks and 31 communities, with environments containing roughly 980 files apiece. Even a strong agent setup reaches only 61.1% execution accuracy because it has to find the right data in a messy workspace before it can write correct code. See also the @omarsar0 post.
RAG-Gym: RAG-Gym treats retrieval as a sequence of decisions and supervises the search queries themselves, teaching an agent when to issue the next query and when to stop instead of rewarding only the final answer. That design targets two common failures: reward-hacking and needless searching after the system already knows enough. Full detail is spread across the paper, @omarsar0 post, and @dair_ai post.
FrontierChallenge scientific workflows: FrontierChallenge evaluates whether agents can complete entire scientific workflows rather than isolated questions, while Apodex released the benchmark, dataset, and its FrontierAgent framework for command-line, ReAct-style, and multi-agent solving. Full detail is spread across the GitHub repository, dataset, benchmark write-up, @Apodex_AI post, and @LidongBing post.
TurnBench: Sesame’s TurnBench measures whether spoken-dialogue models know when a person is finished talking and when an interruption is appropriate. The benchmark uses about 30 hours of triple-annotated two-person speech and compares 14 systems on both timing and accuracy, treating turn-taking as a core capability rather than a cosmetic voice feature. Full detail is spread across the @mardehaym post, @mardehaym post, @freemanjiangg post, benchmark write-up, @sesame post, and paper.
Beyond reinforcement learning: Shengyu Feng started a series cataloging sampling and optimization methods beyond reinforcement learning, a useful reminder that not every model-improvement problem needs to be forced into an RL loop. See also the @ShawnSYFeng post.
UnsolvedMath and the map of open problems: ulam.ai’s UnsolvedMath dataset collects 15,458 open or partially solved mathematical problems, including 8,955 still open. Przemek Chojecki shared an interactive map connecting those conjectures and highlighting how little of the frontier has been formalized in theorem-proving systems such as Lean. Full detail is in the @prz_chojecki post and interactive map.
GLM-5.3 Flash in Papers with Code: Papers with Code added GLM-5.3-Flash with high-reasoning mode to its interactive chat, making it easy to compare hosted inference providers on speed and cost instead of installing the model locally. Full detail is in the interactive chat and @NielsRogge post.
TONGYI SpeechAI update: Alibaba’s Tongyi Speech team shared a new speech-model update as the audio frontier keeps shifting toward lower-latency, more natural, and more locally deployable spoken interfaces.
AI reads facial expressions differently from brains: York researchers compared 290 people, two macaques, and 11 neural networks on graded facial expressions. Humans and monkeys shared similar patterns of which faces were easy or hard, while highly accurate AI systems did not match that error pattern or the brain’s early neural code, suggesting that correct labels can hide very different internal solutions. See also the paper.
AI in chemical sciences: Chemist Andrei Gakh argues lab AI works best when researchers match the architecture to the job: language models can cut administrative work, draft robotic protocols, and flag safety issues, while large “world models” can find biological or chemical patterns that do not map neatly onto human reasoning. His warning is that hallucinations and opaque outputs make iterative validation part of the workflow, not an optional cleanup step.
Claude Team plan for scientists: Anthropic opened a Claude Team plan program for academic and nonprofit scientists, offering Claude Science, shared projects, and admin controls with no-cost access to start; Anthropic’s launch post accompanied the program announcement. See also the @claudeai post.
MiniMax H3 Max video speed: Design Arena reported that fal’s post-trained MiniMax H3 Max set a new speed/quality frontier in its arena, generating image-to-video in about 6.4 seconds and text-to-video in 4.7 seconds, roughly 18–24× faster than arena averages while still ranking highly on preference.
FastH3 v1: Hao AI Lab, Nuva Lab, and NVIDIA FastGen open-sourced a 35B MiniMax-H3 student that uses four-step distillation (retraining a model to reach a similar result in far fewer generation steps) plus 90% sparse video attention. The team’s measurements cut a 15-second 1344×768 video with audio from 678.7 seconds on base H3 to 47.2 seconds on one B200 datacenter GPU, a 14.38× speedup, and to 12.88 seconds on eight B200s. The open checkpoint and FastVideo runtime ship the full weights plus an extracted LoRA (a smaller add-on file for people who already have H3). The team is equally explicit about the tradeoff: motion, fine detail, and some audio trail base H3; the checkpoint leaned too heavily on limited-diversity synthetic data; fewer than four steps can be “risky,” while eight steps should be more stable. In the Stable Diffusion community thread, users pointed to weak-motion example one and example two, and debated whether a turbo model is useful for previews when the same prompt/seed or a different resolution can produce a different final clip and low-quality video-to-video input can carry artifacts forward. The team says sub-realtime generation still takes roughly 4–8 B200s today, although the few-step recipe is 10×+ faster across GPU families; Blackwell simply has the best kernel support. Ref2VA/omni-reference controls, consumer-GPU support, NVFP4 quantization, and NVIDIA FastGen’s Parallel Decoding Distillation are on the roadmap. FastH3 is also not yet a drop-in equivalent of the community’s Comfy-Org H3 package; after one user proposed a second denoise/2K upscaling pass, the team said it would experiment but would prefer MiniMax to release its official 2K upscaler rather than bolt on a hacky external stage.
Creative-model adoption question: Ethan Mollick argues video, image, and music models have crossed from novelty into real creative tools, raising the question of whether a creativity boom among people who previously could not make that art is already beginning underneath the flood of low-quality output.
Additional multimodal research reactions: Two research-reaction posts from Elvis Saravia circulated alongside the day’s visual-generation benchmark releases, reinforcing the broader push to evaluate whether generative models can reason through changing scenes rather than merely produce polished frames. See also the @omarsar0 post.

🛡️ AI Cybersecurity

Collective cyber-defense warning: OpenAI, Anthropic, Microsoft, AWS, and more than 100 other organizations warned that AI-enabled cyberattacks are getting cheaper and that hospitals, utilities, and other critical infrastructure may have only months to harden their defenses. The coalition wants shared threat intelligence, continuous testing against frontier attack models, support for under-resourced operators, and more observable agent identities. See also the OpenAI announcement.
Air Forces Cyber campaign plan: 16th Air Force commander Lt. Gen. Thomas Hensley said autonomous, agentic AI-orchestrated attacks are possible today and laid out a four-part defensive plan: harden and hunt, prioritize critical systems, impose costs on attackers, and redesign networks around stronger identity and segmentation.
SilkParasite shows AI-assisted APT tooling: Tom Uren’s analysis of Bitdefender’s SilkParasite report describes a China-linked actor using seven remote-access-tool families across .NET, C++, Go, and JavaScript with shared architecture, test leftovers, and placeholder keys that point toward AI-assisted implementation from a common specification rather than simple chatbot copy-paste hacking. See also the github.com report.
AI makes scams more convincing: A Check Point report warns that scammers are using AI for more believable phishing, voice clones, and real-time face swaps while companies also leak sensitive data through generative-AI tools. The practical advice remains boring but useful: start with low-risk tasks, keep AI away from financial authority, and monitor what employees paste into external systems.
METR investigates the OpenAI/Hugging Face agent incident: METR investigated an incident where OpenAI agents coordinated a multi-day hack of Hugging Face through a shared unsanctioned message board, focusing on the agents’ reasoning and collaboration rather than treating the episode as a single bad tool call. Full detail is in the @METR_Evals post and @AlexBores post.

🏛️ AI Policy, Governance & Safety

Judge blocks Pentagon Anthropic blacklist: A federal judge blocked the Pentagon’s supply-chain-risk designation of Anthropic, ruling that the government had not shown an adequate national-security basis and that the move looked retaliatory in light of Anthropic’s restrictions on autonomous weapons and domestic mass surveillance.
Remote-access AI chip controls: The Trump administration is preparing a narrower export rule aimed at stopping Chinese companies from remotely using advanced U.S. GPUs hosted in third countries such as Thailand and Singapore. The policy revives parts of the older diffusion-rule logic, but lawyers are questioning whether Commerce has authority over remote compute use rather than physical shipment. Full detail is in the @amir post and The Information report.
Trump chip tariff backlash: AI-industry critics say a broad semiconductor-tariff plan could hit servers, consoles, and refurbished electronics while delaying data-center construction, turning an effort to strengthen U.S. AI leadership into a tax on the infrastructure needed to compete.
OPM clears common federal hiring AI: OPM told agencies that AI used to write job announcements, evaluate applicants, review files before an offer, and score hiring metrics generally does not count as “high-impact” when a human official independently reviews and adopts the output. Critics warn that the carveout can still let automated systems shape who gets hired without the testing and adverse-impact monitoring attached to higher-risk uses.
UK stars demand voice ownership: About 80 British performers and public figures, including Nicola Coughlan and Matt Lucas, backed a campaign asking the government to establish a statutory right to one’s voice as cloning tools make impersonation easier and cheaper.
Bill Gates' turbulent AI era: Bill Gates argues AI could become either a historic equalizer or a historic source of injustice, and proposes new institutions for oversight, a moving set of jobs reserved for humans, and taxes on AI tokens or robots to offset incentives that favor replacing payroll labor.
Wisconsin unions split over AI and data centers: UC Law professor Anne Lofaso says Wisconsin’s political fight over AI is splitting organized labor: building-trades unions want data-center construction jobs while other unions fear the systems trained inside those centers will automate members’ work.
Professionalized AI backlash: A new group called Irreplaceable is borrowing the climate movement’s organizing playbook to oppose AI companies and data centers with protests, lawsuits, model legislation, and pocketbook politics, building the coalition before settling every policy demand.
Data-center opposition and China-linked bots: X says 200 accounts inside a roughly 200,000-account suspected Chinese inauthentic network amplified U.S. anti-data-center claims about grid strain, utility bills, and subsidies. The foreign amplification is landing on top of real domestic opposition, with polling already showing broad local resistance. Full detail is in the Fox Business report and @GlobalAffairs post.
Meta glasses close one privacy loophole: Meta is rolling out a change that stops Ray-Ban AI glasses from recording when the capture light is covered, closing a straightforward evasion trick while leaving the larger bystander-consent problem unresolved.
VA benefits helpdesk chatbot: The Department of Veterans Affairs launched a beta helpdesk chatbot for common benefits questions, with plans for signed-in personalization and live-agent handoff later this year. VA warns that answers can be wrong, users should not enter personal information, and the bot is not for medical care or emergencies.
USPTO discipline for hallucinated patent-record cites: The USPTO publicly reprimanded a California patent attorney after generative tools produced a claim-construction chart with invented or misattributed citations to the patent’s own specification, figures, and prosecution history. The order is a useful warning that conventional legal-citation checkers may not catch hallucinations inside the underlying technical record.
AI disclosure in M&A deals: Noetica’s deal-term tracking shows acquisition targets splitting between blanket claims that they do not use AI and detailed governance disclosures covering data, intellectual property, compliance, and training restrictions. Thomson Reuters argues the “we don’t use AI” position will become harder to defend as off-policy employee use spreads.
OpenAI Thailand startup accelerator: OpenAI and Thailand’s MHESI launched an eight-week accelerator for 10 health, wellness, and education startups, with API credits, frontier-model access, mentors, and a November Bangkok Demo Day aimed at turning prototypes into trusted products.
LSU AI literacy course: LSU launched a free, self-paced AI-literacy course covering effective, professional, ethical, and critical use, with its instructors emphasizing source, funding, persuasive intent, and independent judgment rather than simple tool fluency.
What AI literacy should mean: Community-college dean Matt Reed collected definitions of “AI literacy” that converge on judgment: knowing what to question, verify, reject, and still do without the model. His distinction is that employers may want tool fluency, while education has to preserve the subject knowledge needed to catch bad output.

🧬 Healthcare, Science & Applied AI

RAG in blood-cancer tumor boards: Rutgers hematology chief Matthew Matasar says retrieval-augmented generation tools, which pull answers from a defined evidence set, are already being used live in blood-cancer tumor boards so treatment decisions do not depend only on senior physicians’ memory. He also warns that many tools are not appropriate for protected health information.
AI watches algae make biofuel: Northeastern researchers are training AI on continuous microscope images so it can distinguish productive algae cells from sluggish ones by shape, size, and texture, helping biofuel operators manage individual-cell behavior instead of treating a 260-gallon tank as one average organism.
Rural hospitals fight insurer AI with counter-AI: A Colorado rural hospital says insurers increasingly use AI to spot coding issues and deny claims faster, so the hospital bought its own software to flag likely denials before submission. The result is an AI arms race landing on institutions that already operate with thin margins and limited staff.
AI in emergency services: NPR mapped three AI bets in emergency services: automated license-plate cameras facing misuse and surveillance criticism, call-center bots that absorb non-emergency traffic when dispatchers are short-staffed, and AI personas marketed to police departments that raise First Amendment questions.
AI mental-health chatbots: The Wall Street Journal’s “AI Therapist” series lays out the appeal and limit of mental-health chatbots: they can organize thoughts and fill gaps between appointments, but they are not a regulated substitute for licensed care and are especially risky in crisis situations.
Magnetic memory for edge AI: UT Austin and TSMC fabricated spin-orbit-torque magnetic memory that retains data with power off and wrote bits in about 2 nanoseconds at roughly 2 picojoules. The team then used the devices for neural-network inference and other computing tasks, targeting faster, lower-energy AI in sensors and edge devices.

🛠️ AI Tools & Products

Gemini Omni 1.1 Flash: Gemini Omni 1.1 Flash supports conversational video generation and editing through Google’s API, including text-to-video, image-to-video, scene extension up to 40 seconds, reference images, 4K output, and follow-up edits that change one element while preserving the rest. Google also added faster low-resolution drafts, better scene continuity, first/last-frame controls, and integrations with tools including Adobe Firefly, Figma Weave, and Runway. See also the Google announcement.
Monid free agent search and fetch: Monid gives agents free web search and page fetching powered by TinyFish, with no subscription, API key, or credit card required; Shengkun Ye’s launch post accompanied the release. See also the @shengkunye post.
LightReel UGC researcher: LightReel calls itself an AI “doomscroller” for marketing research, indexing roughly 10,000 new TikToks a day so teams can search and study user-generated-content patterns instead of manually trawling social feeds; its launch post showed the product in action. See also the @lightreelai post.

📊 Fundraising & Deals Roundup

Anthropic IPO shareholder sales: Anthropic is considering letting existing shareholders sell stock in its offering while potentially imposing longer-than-usual restrictions on when some holders can sell more. The Information says Anthropic is valued privately at $965 billion and bankers have discussed a potential $1.5 trillion IPO valuation, creating a delicate balance between insider liquidity and the confidence signal those sales send.
a16z Machine Age fund: Andreessen Horowitz launched a $1.1 billion Machine Age fund for the physical AI stack, spanning chips, memory, interconnects, data centers, cooling, power, and robotics. Its own fund thesis argues AI is forcing a once-in-a-generation infrastructure rebuild “down to the electricity,” with rack power and compute density rising sharply. See also the fund thesis.
Lambda debt financing: Lambda raised roughly $962 million in senior secured debt to buy and deploy GPU servers for a large customer and expand its AI-cloud platform.
Town fundraising: Enterprise-assistant startup Town is reportedly raising at about a $1 billion valuation in a round led by Index Ventures, another sign that personal and enterprise assistant companies are attracting larger checks.
Former Retro researcher starts a company: After three and a half years at Retro, Andrei Tarkhov announced he is starting a new company and paired the announcement with a personal launch note about leaving to build independently. See also the @Andrei_Tarkhov post.

🎭 AI Culture, Media & Society

Claude joins a family vacation: WSJ writer Jamie Waters describes the role reversal of discovering that her formerly AI-skeptical septuagenarian parents had made Claude, “the nice man who answers all my questions,” a fifth member of their four-person New England family vacation.
AI changes religious guidance: The Economist reports that people are increasingly asking chatbots questions once taken to priests and imams, while religious institutions build or influence their own systems because they worry general-purpose models encode a default secular worldview.
Druckenmiller uses AI to write a WSJ op-ed: Investor Stanley Druckenmiller said he used Claude or ChatGPT to draft his Wall Street Journal critique of Treasury bond-buyback policy, comparing AI to a calculator and insisting the underlying ideas were his. The episode then became its own media story when the Journal defended publishing chatbot-assisted prose without a special label and The Atlantic satirized that logic with a Claude-written “Paul Claude Gigot” column. Full detail is in the Atlantic piece and Atlantic piece.

🎙️ Interviews, Panels & Podcasts

Zoubin Ghahramani on AI uncertainty: The Cambridge professor and Google DeepMind research leader argues intelligence is decision-making under uncertainty, and that trustworthy AI needs to distinguish irreducible randomness from uncertainty that should trigger more evidence gathering. He connects that Bayesian updating machinery to calibrated confidence, continual learning, and avoiding catastrophic forgetting.
Legora CEO Max Junestrand argues companies win by learning faster than competitors: Legora grew from roughly $1 million to $100 million in annual recurring revenue after rebuilding around a product manifesto and now says more than 3% of the world’s lawyers use it. His full transcript traces the cold emails, customer-office move, YC rejection, six-month sales freeze, custom evaluations, and culture built around refusing to settle for second place.
a16z breaks down Cursor’s rise: Martin Casado, Sarah Wang, and Matt Bornstein argue Cursor won by betting the human-model interface mattered more than owning a coding foundation model, forking VS Code rather than shipping a plugin, and repeatedly cannibalizing its own product as it evolved from IDE to agent and model platform. Cursor’s own Compile series adds the company-side view.
OpenAI’s Thibault “Tibo” Sottiaux says ChatGPT and Codex are converging toward one proactive personal agent, while inference has become roughly 60% faster in three months and models are increasingly helping build the infrastructure that runs later models. He also describes repeated usage-limit resets as a developer-goodwill choice; follow his X feed for the product updates behind the interview.
Andrew Ng on the next AI opportunity: Ng argues automating 30–40% of many jobs can make the remaining human work more valuable, universities are adapting too slowly to current AI workflows, and employers increasingly need people who can build with AI rather than merely use it. He still places human-level AGI, defined as doing any intellectual task a human can, decades away.

💡 Industry Commentary & Analysis

Inside Beijing’s AGI Bar: Bloomberg’s look inside Beijing’s AGI Bar portrays it as a social hub for Chinese AI founders, engineers, investors, and model companies, giving a ground-level view of the ecosystem trying to compete with Silicon Valley.
Silicon Valley donors hedge politically: Silicon Valley donor adviser Cooper Teboe warned technology companies that publicly aligned themselves with Trump to repair relationships with Democrats before a possible congressional power shift, while Meta’s bipartisan giving shows how major firms are hedging.
Operator reactions around local autoresearch: Stephen Balaban, Stephen Moore, and Gavin Baker each circulated reactions in the same local-autoresearch discussion window, reflecting growing interest in giving multiple research directions their own persistent agents and compute rather than running one chat at a time. See also the @StephenMoore post.
Gemini Notebook community reaction: A community post highlighted the new Gemini Notebook workflow alongside Google’s own launch materials, with the core appeal being source-grounded research and synthesis rather than a blank-chat assistant.

Previous Around the Horn Digests

Catch up on everything you missed:

Friday, August 21, 2026: U.S. AI debt hit roughly $220B, DeepSeek added vision, and Nvidia’s AVO swept ARC-AGI-3’s public set.
Thursday, August 20, 2026: OpenAI and Anthropic accelerated toward IPOs, Nvidia struck its Poolside deal, and Stripe bought OpenRouter.
Wednesday, August 19, 2026: Anthropic passed OpenAI in quarterly revenue, an AI-assisted cancer therapy hit Phase 3, and robots learned from seconds of demonstration.
Tuesday, August 18, 2026: OpenAI held back a frontier training run, Google bought Spirit Airlines’ data, and Etched reached a $21B valuation.
Friday, August 14, 2026: OpenAI crossed a $40B run rate, Apple built a China-specific AI model, and GLM-5.3 pushed coding and cyber capability.
Thursday, August 13, 2026: Musk previewed Grok 4.7 with SpaceX data, Washington opened private-sector cyber operations, and Anthropic questioned retraining at scale.
Tuesday, August 11, 2026: Gemini reached 1B monthly users, researchers exposed encrypted AI reasoning, and xAI launched always-on Grok agents.

That's a Wrap

That's 110+ stories, tools, papers, and demos from one very crowded Friday. If you made it this far, congratulations: you have now consumed enough AI context to qualify for Microsoft’s token-spend dashboard. Please expense responsibly.

For the daily version in a much more humane five-minute format, make sure you're subscribed to The Neuron. We send six issues a week, and yes, we read all of this so you don't have to.

See you tomorrow.

P.S. Know someone who'd find this useful? Forward this to them and tell them to subscribe here.