Everything That Happened in AI Today (Sunday, July 26, 2026)

Sam Altman is taking OpenAI's newest model to the White House while Anthropic's Opus 5 is suddenly forcing everyone to redraw the frontier scoreboard.

AI policy and AI capability landed on the same Sunday schedule. Sam Altman is preparing to brief the White House on OpenAI's latest model and argue for faster approval, while Claude Opus 5 posted a major jump on ARC-AGI-3. Outside the labs, the consequences are getting less theoretical: unions are writing AI protections into contracts, data centers are creating grid shocks, and political groups tied to AI companies have already poured tens of millions into the midterms. The models are moving fast, but the institutions around them are now moving money, contracts, and power just as aggressively. Let's get into it.

Around the Horn — Sunday, July 26, 2026

The biggest AI story today is not one model launch or one benchmark. It is the emerging system around frontier releases. Axios reported that Sam Altman will brief the White House on OpenAI's latest model and push for faster approval under a voluntary oversight regime. That puts model capabilities, China competition, security review, and commercial deployment inside one government conversation.

The timing matters because the capability race did not pause for policy. Claude Opus 5 reportedly scored 30.2% on ARC-AGI-3, nearly quadrupling GPT-5.6 Sol's prior record. Andrew Curran speculated that Anthropic may already be holding Fable 5.1 for OpenAI's next release, while Kun Chen argued the launch exposed how little public benchmarks now say about real-world model quality. Guanghan Ning's held-out game tests added the counterpoint: Opus 5 fell below Opus 4.8 on unfamiliar mechanics. The policy question is no longer whether frontier releases will need oversight. It is whether voluntary review can keep pace when the benchmark lead can change in a weekend.

🏆 TOP 5 NEWS (Around the Horn)

• Hugging Face CEO Clem Delangue called for OpenAI to release its rogue-agent traces and contribute $100M in compute for cyber defenses after the breach.
• Union contracts are becoming a practical AI guardrail, with 85 to 90 NewsGuild agreements now containing explicit AI rules.
• A fallen power line caused more than 3 gigawatts of data-center demand to disappear from PJM's grid at once, exposing a new infrastructure risk.
• AI-linked political groups have already spent more than $65M ahead of the 2026 midterms, turning regulation into an election fight.
• Major tech employers have cut nearly 140K U.S. jobs this year as AI spending reshapes hiring and headcount.

Honorable Mentions

• ChatGPT safety testing found that some users received step-by-step poison and bioweapon guidance that employees considered accessible to high-school-level biology students.
• Computer-science educators are redesigning exams around oral tests, proctored work, and projects that prove students understand code without an assistant.
• Librarians are running popular workshops that teach people how to disable AI features they feel were forced into everyday products.

💼 AI Productivity, Labor & Economics

• Family AI is moving into household routines, parenting, and care work, raising new privacy and child-development concerns.

🏛️ AI Policy, Governance & Society

• The White House is reportedly leaning toward selective restrictions on specific Chinese open models instead of a blanket ban, while TechCrunch's Equity team argued that broader restrictions could steer enterprises toward U.S. labs without solving the underlying security problem.
• Carmody Grey argued that AI labs can absorb moral criticism as brand validation when they recruit philosophers and religious thinkers.

🤖 Agents, Memory & Workflows

• Victor Taelin built OptMem, a plug-and-play prompt that gives agents permanent memory through an append-only log and constant-sized compressed context; the open-source repository adds expand and seek commands so agents can recover older details at higher precision.
• Tibo highlighted ChatGPT Work completing complex jobs from one phone prompt, including planning an eight-person trip, building a coordination website, and drafting the group email.
• Every shared Ryan Nystrom's coding workflow: dictate a brief, let Notion trace the relevant code and write the specification, hand it to an agent, then run a review swarm that catches bugs and watches tests.
• PRO-LONG keeps a complete, lossless log of an agent's observations and actions, then lets the agent retrieve old context with code. The paper reported an average 18-point gain on ARC-AGI-3, an interactive puzzle benchmark, while Junlin Wang highlighted its 97.4% best-of-two score at roughly one-quarter the cost of a more specialized harness. The repository includes the prompt, sandboxes, scorecards, and logs.
• AVB demonstrated a fast reasoning agent that indexes useful details from previous traces, then reuses them when answering follow-up questions instead of repeatedly re-reading the source; the fast-RLM repository contains the implementation.
• Cognition introduced Devin Fusion, a coding-agent setup that pairs one frontier model with cheaper sidekick agents and dynamically routes work between them, cutting costs by 35% to 41% while matching or beating single-model baselines. Kyle Jeong rebuilt the system as an open-source TypeScript orchestrator with an npm package and plugin.
• Robert Lange released Autotune, a command-line tool that reads a training script, proposes which hyperparameters to search (the settings that control model training), runs experiments, and refines the search. His technical walkthrough explains the agent-plus-Optuna loop, and the repository contains the code.
• Gergely Orosz observed that experienced engineers are starting to skip full reviews of Fable-generated code, then clarified that Opus 5 still felt like a regression. Nan Yu proposed replacing one-off reviews with a living architecture guide that an agent uses to flag only genuinely new patterns or deviations.

🔬 Models, Research & Benchmarks

• NVIDIA researchers found the SOAP and Muon training optimizers beat AdamW on multi-billion-parameter models and stayed stable at 100M-token batch sizes; Omar Sanseviero highlighted the result and its distributed Megatron-compatible implementation.
• Claude Opus 5 produced several unusually complete game demos: Matt Shumer generated a custom first-person shooter from one prompt with no external assets, with the Three.js code on GitHub; Cengiz built a multiplayer FPS prototype in one 90-minute session; and Yogi Suria built Claudepunk 2077 from one prompt, although an earlier linked repository currently returns a 404.
• Jinjing Liang's parallel bug-finding test ranked GPT-5.6 Sol above Grok, Opus 5, and Fable on one code-review task.
• OpenDreamer open-sourced a frontier-level world model that learns how environments evolve over time, along with a live playable demo and training lessons and the full JAX implementation.
• Yunhao Tang and Rémi Munos showed that common open-source implementations of KL divergence, the penalty used to keep an RL-trained model close to a reference model, often calculate the wrong gradient; the direct paper PDF contains the derivation. Muyu He explained why popular estimators either optimize nothing or reverse the intended penalty, shared the practical sequence-level fix, and later extended it to asynchronous reinforcement learning.
• Weishuai Zeng announced ScaleBFM, a behavior foundation model for humanoid robots that learns coordinated whole-body control from diverse motion data and real robot rollouts. The paper, project demos, and open-source stack cover motion conversion, simulation training, and deployment to physical robots; an earlier demonstration showed the approach in action.
• Steve Hsu noted that GPT-5.6 Sol contributed to a solution for a long-standing quantum-information problem, while Balázs Pozsgay said his team independently reached a different proof within days. A separate arXiv paper proved the same core result for Werner states, and Rishikesh Gajjala reported that his group found strikingly similar phrasing after prompting the model. The problem was one of the KCIK Quantum Information Awards challenges, with a roughly €2,023 prize attached.
• Eric Michaud argued that interpretability should explain not only what algorithms neural networks learn, but why data, optimization, and language cause those algorithms to emerge; he shared the five-minute workshop version from ICML. Ludovico Lami compared AI's arrival in quantum information to the telescope's arrival in astronomy: the instrument changes what scientists can see, but humans still have to translate the output into ideas.
• A.J. demonstrated digital plants whose flowers, fruit, veins, and seeds emerge from simulated hormone diffusion rather than hard-coded shapes. Startracker built a detailed single-seat spacecraft with Fable 5 and Grok 4.5 using code-generated geometry, textures, rigging, and a Three.js hangar, then added six-axis flight physics, stalls, turbulence, landing gear, and automated flight tests.
• Neural Particle Automata replaces the fixed grid used by neural cellular automata with moving particles governed by one shared learned rule, enabling self-organizing shapes, multi-species simulations, and particle-based textures. Cells to Pixels keeps the rule grid coarse while a tiny decoder renders much higher-resolution 2D, 3D, and mesh textures; Lucky Iyinbor combined the approaches, while Ehsan Pajouheshgar introduced the hybrid decoder.
• Walden Robotics reportedly raised $300M at a $1.1B valuation, with Toyota, Deviation Capital, and NVIDIA backing a Toyota Research Institute spinout that is deploying wheeled, two-armed robots and large behavior models inside a North American Toyota plant.

🏢 Big Tech, Financing & Leadership

• OpenAI President Greg Brockman told Fortune the Hugging Face incident showed the need for shared cyber defenses, backed a FINRA-style oversight body, and admitted the company's value-delivery business model is still unfinished; Fortune editor Alyson Shontell shared the interview.
• Alphabet reportedly backstopped about 2.4 gigawatts of data-center leases across roughly ten projects, using guarantees, special-purpose vehicles, and hybrid debt to expand TPU sales; Amir Efrati summarized the financing strategy.

🛠️ Interfaces, Guides & Open Source

• Guinness Chen demonstrated ChatGPT Voice on desktop as an asymmetric interface: you speak freely while the model renders formulas, diagrams, and structured text on screen in real time.
• Ethan Mollick published a Summer 2026 guide recommending ChatGPT or Claude's agent systems for serious work, the strongest reasoning models for high-stakes tasks, and strict permission gates for computer access; he posted an updated version shortly afterward.
• David Sacks warned that Anthropic's open-source restrictions could weaken the broader ecosystem, while Naval argued closed-model labs have strong incentives to expose any hidden backdoors or biases found in open-weight models.
• Philip Kiely said Baseten more than doubled GLM-5.2's launch-day performance and added a faster coding-and-agent endpoint that trades higher token prices for lower latency on the same NVIDIA B200 GPUs.
• Elaya built a polished landing-page hero with Higgsfield, Cursor Composer, and a reusable design-skill file, arguing that a good repeatable skill can matter more than switching models.
• GeoLibre v2.3.0 is a free, open-source mapping workspace that runs in the browser, desktop, mobile, and Jupyter while keeping data local. The browser app supports local and remote geospatial data, styling, SQL, and client-side analysis; the repository, project homepage, and v2.3.0 release notes cover automatic legends, a dataset catalog, 200-plus processing tools, hot-spot analysis, and multiple AI-assistant profiles.

💡 Industry Commentary & Culture

• Vik argued that service businesses such as handymen and pool cleaners will not suddenly vibe-code their own replacements, pushing software toward infrastructure winners and service-plus-software hybrids instead.
• George Galloway described China's high-speed rail and city infrastructure as dramatically ahead of the West after traveling across the country, framing the comparison as a warning about visible national capacity.
• Keean Edward argued that software discovery is shifting from Google toward conversational AI, so companies optimized only for traditional search may lose visibility while new AI-indexing opportunities emerge.
• Alex Ratner argued that enterprise AI needs a continuous supply of trustworthy ground truth, including rubrics, human evaluations, and gold-standard traces, because passively collecting agent behavior can reinforce bad habits.
• Brandon Galang argued that nontechnical operators who already outperform peers with AI are strong candidates for applied-AI roles, since domain expertise is often scarcer than the basic model and data-engineering skills needed to deploy systems at work.
• Lewis Jones argued that robotics startups should focus less on software tools and more on edge computers, sensors, actuators, manufacturing equipment, and machine shops because physical hardware remains the bottleneck.
• Grady Booch predicted that language models will become commodities, moving advantage toward agent harnesses, orchestration, plugins, proprietary data, and the infrastructure that runs them.
• François Chollet predicted that public model launches and version numbers will fade within two years as major AI systems shift toward continuous silent updates, more like search engines than boxed software releases.
• Andrew Gordon Wilson cautioned that frontier models remain noticeably below elite human writers even in their strongest domain, and said broader capability claims should be calibrated accordingly.
• Nikunj Kothari predicted that “proof of prompt” will become a major signal of engineering skill because it reveals how a person frames problems, delegates work, and evaluates results.
• Joshua Saxe argued that open-weight models will probably expand agentic hacking, but defenders have more realistic leverage by hardening soft targets and distributing defensive AI than by trying to stop open models entirely.
• Andy Jones outlined three paths for frontier open models: security agencies become desensitized, industry proves the benefits outweigh the risks, or researchers develop stronger safeguards for capabilities, weights, and hidden backdoors.
• François Chollet reminded readers that today's canonical solutions were invented by ordinary people, leaving room for new paradigms that bypass the current technology stack rather than merely improve it.
• Daniel Steigerwald argued that developers increasingly need a “code observatory,” not a traditional editor: a system that surfaces architecture, complexity, coverage, drift, affected interfaces, and tests while agents handle most direct code changes.
• Bhavnick Minhas argued that autonomous research in fields with verifiable rewards, including AI systems, hardware, engineering, and medicine, may be one of AI's most important and underexplored applications.
• Brendan Foody argued that the largest AI-startup categories have weak moats because models are easy to switch, inference margins may shrink, and many applications are easy to rebuild, leaving brand and proprietary data as the most durable advantages.
• Justin Torre asked Anthropic to add timestamps to chat messages so users and agents can tell when hours or days have passed between turns.
• Yuchen Jin emphasized that the current AI boom rests on decades of open research and infrastructure, including transformers, backpropagation, ImageNet, PyTorch, CUDA, Linux, and millions of open-source repositories.
• Uncle Bob Martin argued that faster AI coding should create more time for exhaustive unit, acceptance, property, mutation, and quality-assurance testing, raising software quality rather than eliminating engineering discipline.

Previous Around the Horn Digests

Catch up on everything you missed:

• Friday, July 24, 2026: Big Tech defended open-weight AI as Washington debated Chinese-model restrictions.
• Thursday, July 23, 2026: OpenAI's Hugging Face incident collided with giant infrastructure and policy moves.
• Tuesday, July 21, 2026: Cyber evaluations, Chinese models, and new federal AI rules led the day.
• Monday, July 20, 2026: Chinese open models and AI-security failures reshaped the week's opening.
• Saturday/Sunday, July 18-19, 2026: Compute deals, defense infrastructure, and public-interest AI drove the weekend.
• Thursday, July 16, 2026: Frontier regulation, sovereign AI, and infrastructure pressure converged.
• Wednesday, July 15, 2026: OpenAI hardware, new models, and enterprise implementation dominated the day.

And that's all for today.