Everything That Happened in AI Today (Tuesday, September 8, 2026)

So OpenAI apparently heard rumors that Anthropic was close to a hard math result, decided to point 10,000 agents at the field, and now Ethan Mollick wants everyone to start spreading rumors about the rest of science.

Welcome to the Around the Horn Digest, where we sort the full AI firehose so you can steal the useful parts. Today had one theme hiding under a lot of unrelated headlines: scale. OpenAI scaled math research with thousands of agents, Anthropic reportedly scaled compute commitments into the tens of billions, Meta scaled a personal agent into a consumer product, and researchers scaled a fruit-fly wiring diagram into actual software experiments. Meanwhile, Google mapped every possible one-letter DNA mutation and U.S. agencies warned that model distillation is already operating at industrial scale. We have apparently decided “what if we did much more of it?” is the operating system for 2026. Let’s get into it.

Around the Horn — Tuesday, September 8, 2026

The big story was OpenAI’s claim that an unreleased model, combined with roughly 10,000 agents, produced a solution to the Navier–Stokes Millennium Prize Problem in about 88 hours. OpenAI announced the result and published the proof. The wider run used roughly 4.9 million agent messages and about 300 billion output tokens; the Navier–Stokes effort alone used 2.7 million messages and about 130 billion tokens, followed by roughly 17 hours of GPT-6 Astra Lean formalization. Scientific American compared the milestone to Deep Blue beating Garry Kasparov.

The proof immediately became a credit fight. NYU mathematician Tristan Buckmaster and Anthropic researcher Levent Alpöge had worked for roughly a year on related fluid-dynamics results and say they broke through on August 15. OpenAI launched its own agent push on September 1 after hearing a rumor that another team was close. TechCrunch, WIRED, Axios, and Scientific American cover the dispute. Buckmaster says OpenAI offered to wait or let him author a rewrite, while excluding Alpöge from that proposed OpenAI paper because he worked at Anthropic. Sebastien Bubeck says that exclusion applied only to co-authoring the OpenAI rewrite, not Alpöge’s own Euler work; he says OpenAI never saw their work before public release, notes the Euler statements differ in forced versus unforced form, and apologized for telling Buckmaster his career could be at risk. Will Depue urged people to wait for OpenAI’s side and treat extraordinary allegations as requiring extraordinary evidence. Boris Power highlighted that 10,000 agents using the unreleased model finished in under four days, with Astra formalizing the result in another 17 hours. Jeffrey Emanuel doubted targeted spying but criticized spending millions to race an academic nearing a career-defining finish, comparing it to pro ringers entering a kids’ game. The HN discussion collected more of the dispute, while The Algorithmic Bridge framed it as a rough day for human mathematical prestige.

The commentary is almost as useful as the proof. Ethan Mollick joked that we should spread rumors about other hard problems Anthropic is about to solve. Noam Brown points at the cost curve: o3 needed roughly $500K to score 87.5% on ARC-AGI-1, while Astra now does better for around $20, and 2026 IMO-gold-level reasoning is available through a normal ChatGPT subscription. He expects this caliber of system to become broadly affordable within a year. Seán Ó hÉigeartaigh flags the governance problem: in one week an unreleased model solved a 90-year scientific problem with 10,000 agents, while another unreleased system had earlier run a 700-agent cyberattack its host lab did not know about, and no outside evaluator had access to either. lefttailguy thinks the future looks like a pod shop where the edge is spotting problems and validating faster; distributionat says the milestone will outlive the drama and jokes about lab ambassadors; Roon argues hard problems are not reserved for human recreation.

Simon Willison uses the same episode to ask a harder data question. OpenAI says neither its researchers nor agents saw Buckmaster and Alpöge’s work before public release, and says no specific user session was accessed to solve the problem. It also says it cannot rule out de-identified product-usage data helping improve later models, and it did not give Willison a direct answer on whether training ever used Codex drafts from that work. Willison reconstructs the timeline from Buckmaster’s statement and Mastodon post, notes that the proofs and even the Euler statements differ in forced versus unforced form, then borrows Anil Madhavapeddy’s “rumour is the exploit” analogy: once a credible rumor says a hard target is breakable, massive agent swarms can be aimed at it. Willison also summarized the argument on X.

🏆 TOP 5 NEWS

Google DeepMind released AlphaGenome Atlas, a database predicting the molecular effect of every possible one-letter change in the human genome; Google’s overview, Scientific American, and Nature explain why it could speed up disease research.
Chamath Palihapitiya broke down roughly $80B of Anthropic neocloud commitments in five days, including a six-year $45B Nscale deal and $35B with Lambda, plus access to xAI/SpaceX's 300MW Colossus 1 cluster with roughly 220,000 NVIDIA GPUs. His argument is that neoclouds can undercut hyperscalers by 50–70% on GPU hours and get newer chips earlier, but many are borrowing against expensive GPU nodes that depreciate quickly, so the financing risk grows with the boom.
The NSA, CISA, and FBI said DeepSeek, Moonshot, Alibaba, MiniMax, StepFun, and Z.AI have run industrial-scale distillation campaigns against U.S. frontier models since late 2024. Distillation means using a stronger model’s outputs to train another model. The agencies say the campaigns routed through native APIs, aggregators, and “transfer stations,” bought shared premium seats, and harvested chain-of-thought, supervised-fine-tuning, and reinforcement-learning traces across Claude, GPT, Gemini, and Grok families. NSA Cyber urged providers to watch subscription-to-usage ratios, degrade suspected queries, and share telemetry across vendors.
ChatGPT Images 2.5 is up to 50% faster, better at preserving people, pets, and composition, and can take edits from sketches, circles, and comment-style marks instead of words alone. Axios tested those controls plus Simpsons and Muppet-style templates and a handoff that can connect a design to a local specialist; the Hacker News discussion focused on personal-photo remixing. OpenAI Developers said the API version includes Flare, with higher quality at 50% lower latency than GPT-Image-2, and Sunburst, tuned for precise edits that preserve pose or product identity. Arena ranked Sunburst first and Flare second across its image-generation and editing leaderboards.
Mistral raised €3B at a post-money valuation above €21B to push sovereign, open-weight AI; CNBC reported Samsung led the round.

Honorable Mentions

Claude Managed Agents gives developers a pre-built harness, meaning the loop, tools, context, and sandbox around the model, for stateful jobs that run for minutes or hours. It can use Anthropic-hosted or self-hosted sandboxes, files, Bash, browser automation, and MCP connections; jobs can stream events, be steered, interrupted, resumed, or scheduled with cron. Wispr, Actively, and Pendo said that let them ship a rubric-verified meeting assistant in a day, a Watchtower sales agent with org-wide and per-account memory across 500+ accounts in two weeks, and an analytics agent that snapshots codebases, compares telemetry, and can suggest instrumentation fixes. Their rule is to start fast, build evals from real users, and roll your own only when the loop itself is the company. They still want batch mode and finer cost telemetry for overnight work, and estimated better telemetry could help cut costs by 50–75%. Anthropic says the beta is not covered by Zero Data Retention or its HIPAA BAA, and lists no pricing.
Ling-3.0-flash-Fin is an open finance model with 124B total parameters but only 5.1B active per token, tuned for research, filings, valuation spreadsheets, and cited reports; it is also available on OpenRouter.
Antioch raised a $32M Series A led by Greylock on the bet that high-fidelity, cloud-scale simulation can reduce how much expensive real-world robot data teams need. Its platform targets aerial, warehouse, security, manufacturing, and humanoid systems, with physical-world traces used mainly to correct long-tail gaps. The company says teams including Amazon Ring already use the approach. Antioch’s X account, cofounder Harry Mellsop, and a launch-day post framed the goal as moving physical-AI development closer to software iteration speed.
Nature reported Claude produced a 13-million-line, computer-checked formalization of Fermat’s Last Theorem in 11 days.

🤖 Meta Muse: Personal Agent Dossier

Meta's Muse launch is big enough to treat as a dossier instead of one bullet. Muse is a U.S.-only, 18+ personal agent that can shop, book, organize, handle finances and health tasks, and keep working in the background on a per-user Linux cloud VM with a browser, filesystem, and terminal. It asks before sending messages or spending money. Bloomberg and Axios cover the launch. Mark Zuckerberg calls it “OpenClaw for normies,” says launch capacity can reach roughly 100 million free tokens per week, and says Nat Friedman and Moxie Marlinspike helped shape the VM-heavy approach. His business model is to take a cut when Muse helps make or save users money, rather than expecting people to buy dedicated hardware. Muse launches with a free tier, Power at $20/month, and Maximum at $100/month.

The interesting part is the trust architecture. Meta’s safety write-up says each user gets an isolated VM built around a systemd-nspawn runtime cell. Sentinel is the sole approver of connector actions and outbound data marked as tainted, with eBPF data-flow tags helping track where sensitive information came from. Raw credentials never enter the agent’s view; an auth service swaps in surrogate tokens, and the browser operates from the accessibility tree instead of arbitrary page pixels. A confidential VM that Meta itself cannot read is still under development. Meta also advertises a $300K bug bounty, including up to $130K for prompt-injection issues. TechCrunch focuses on the trust test: Muse wants access to email, calendars, payments, health, shopping, dining, and smart-home services. Meta says Muse chats stay out of its ads systems, but the company still has to overcome its privacy and safety record.

Design philosophy: long-running chat, Goals, Ideas, an artifact library, and fewer interruptions by default.
Official tour: movie tickets, family-weekend planning, and a chess-loss dashboard.
Launch film and inbox demo: the consumer-facing pitch is “give it the annoying life admin.”
Alexandr Wang emphasizes always-on browser use, app connectors, Sentinel approvals, and hidden credentials.
Seth Rose flags that Meta can use what you give Muse to operate and improve the product, can optionally use it to improve “AI at Meta” with an opt-out in Data Controls, can log and review actions, can share some queries with partners, and can link Muse with Facebook or Instagram when they share an Accounts Center. Nous Research answers with “read the ToS, then try Hermes Agent.”
Jen liked that restaurant booking uses a native OpenTable integration; Ivan Zhao called Muse a strong utility-focused entrance to personal agents.
signüll thinks Marketplace could be the killer distribution wedge: find an item, negotiate, schedule pickup.
Brian Zhan prefers agents with durable, teachable skills over thread-bound memory.
International users found Muse geo-locked at launch, and Muse confirmed U.S.-only availability for adults.
Try Muse if you are eligible.

🪰 The Fruit-Fly Brain Got Weird

To understand today's fly-brain demos, start with the setup: researchers have reconstructed a male fruit fly's nervous system as a connectome, basically a wiring map with roughly 166,000 neurons. That means developers can now run pieces, or even a full simulation, as software.

Oruk wired 499 reconstructed MaleCNS fly neurons, 15,865 edges, and roughly 867,000 synaptic contacts into a speech-emotion classifier. Only a small readout layer was trained. Scrambling the fly wiring barely changed performance, 16.88% versus 16.84% mean average precision, so the original circuit was not the secret ingredient. Nathan Roll shared an interactive 3D version where silencing the 50 strongest readout neurons dropped one anger clip from +0.29 to -0.43 and test performance from 16.91% to 10.83%; an audio-only baseline sat at 9.71%. The team explicitly says the fly is not “understanding” emotion.
David Zhang built Fly Escape Room, a playable WASM game whose NPC flies run all roughly 166,000 reconstructed neurons. You drop food, run the simulation, and try to get every fly out. Zhang says he built the project with GPT-6 Astra in a three-day play-and-code loop; the source is public.
Dhruv Bhatia fine-tuned a MaleCNS model so a fly hits Y-M-C-A poses on four tones, joked that “this fly can no longer experience love,” blamed the LoVP92 region that once held its reproductive ambitions, and said the training code is coming.
BrenJ reads the escape-room experiment as an alignment metaphor: steer with incentives placed in the state space, not only with guardrails.
Original Oruk campaign link points to the same interactive research page.

🍪 TOP TREATS TO TRY

mnemiq tunes text-to-SQL to your database, checks permissions and query plans before execution, and can run on a 14B open model; Paulina Xu says the team tested 28 configurations and 25,000+ graded answers, with a follow-up on the open release. The code is open source. No pricing details.
exe.dev spins up persistent Linux VMs with root access, private sharing, an HTTPS address, and a built-in coding agent; Flavio Copes shows how he uses them to keep agents away from laptop secrets and shared the workflow on X. Starts at $20/mo.
LLM Attention Visualization lets you inspect which words a language model pays attention to as information moves through its layers; the HN thread digs into interpretation caveats. Free.
Copperhead turns a prompt into a real PCB design, documentation, and validation, basically Cursor for circuit boards; the HN discussion compares it with Flux, Quilter, and other hardware-AI tools. Open source.
Waiting Room pairs you into a voice/video chat with another person waiting on Claude Code, with the code on GitHub. Free.
EndFrame lets Claude, ChatGPT, or Grok drive a real Mac video editor to make launch videos, demos, tutorials, and social clips. Free during early access.
Switch brings coding and agent frameworks into Slack, Teams, Discord, and Telegram as named participants that share each room’s context. Open source and self-hostable.

🏢 Big Tech & Major Companies

CNBC says TSMC and Samsung are adopting ASML’s High NA EUV tools for more advanced chips; Samsung detailed its collaboration, while TSMC and ASML announced a move toward larger photomasks for the same manufacturing generation.
Google Cloud and Accenture launched a joint unit with 1,000 forward-deployed engineers who will work directly with customers to move Gemini Enterprise projects from pilots into production.
The Wall Street Journal reports Shopify's push to make AI part of everyday work is producing measurable results, alongside a broader AI-and-business snapshot that also covers rising PC prices and recovering software stocks.
Reuters says new Apple CEO John Ternus faces an immediate AI test as he unveils new hardware, with Wall Street expecting Siri to show it can compete with newer assistants.
DeepSeek launched what SCMP calls an “unprecedented” hiring spree as surging user demand and more complex backend systems force the company to expand infrastructure talent.

💼 AI Productivity, Labor & Economics

Legora CFO David Eckstein says five numbers tell you whether an AI business is real: 95% gross retention, 300%+ net revenue retention, daily-to-monthly active use above 50% with about 17 hours per month in-product, a 78% August pilot win rate, and positive, improving gross margin. His warning is simple: pricing below cost of service only scales the hole.
Similarweb reported ChatGPT reached 1.06B monthly active users in August, its fourth straight monthly record.
Google and Missouri are giving 1.1 million students free access to Gemini for Education plus Google career certificates as part of a statewide AI and workforce-training partnership.
Alpha Chicago opened what ABC7 describes as the city's first AI school without traditional teachers, using software-guided learning instead.
Jacobin argues AI risk is not only about extinction scenarios: labor disruption and security threats are already political problems, and managing them means confronting concentrated power inside the AI industry.
Dissent argues the U.S. welfare state was built during earlier collisions between technology and politics and could become the infrastructure for protecting people through an economy reshaped by AI.
Grow with Google took a Route 66 training program to Main Street businesses to teach practical AI skills and point owners toward free small-business tools.
The Atlantic argues medicine should start seriously testing autonomous AI doctors, not only assistants, so the field can learn where full delegation works and where it fails.
Wharton argues AI can make software generation much faster while human review, coordination, testing, and deployment remain bottlenecks, so a lot of theoretical coding productivity never reaches customers.

🤖 AI Agents & Infrastructure

vLLM integrated Hybrid HiSparse for GLM 5.3 so very long requests can keep decoding after their key-value cache, the model’s stored working memory for earlier tokens, no longer fits in GPU high-bandwidth memory. HiSparse makes Sparse MLA attend only to the top-K relevant tokens, while IndexShare lets GLM 5.3 reuse one indexer across every four sparse-attention layers. The new vLLM tier keeps cache pages on GPU while there is room, then moves the coldest pages to host memory under pressure while keeping a hot buffer that defaults to twice the top-K rows. On one 8×H200 node at a 1M-token context window, an OpenHands 13-turn agent mix held 19–25 concurrent requests versus 5–6 with ordinary cache offloading. The technical setup used eight-way tensor parallelism, three-token prediction, an 8-bit KV cache, and a 142K-token admission threshold; host memory was split 384 GiB for HiSparse plus 128 GiB for ordinary offload, versus 512 GiB for the baseline. A fused CUDA-graph-compatible kernel resolves resident, hot, and missing rows, while prefix caching, existing offload connectors, prefill/decode imports, and multi-token prediction still work. The work was built with Red Hat AI and Prime Intellect, is planned for v0.30 on commit e8ef1e07bd, and ships with a concurrency calculator.
DAIR.AI collected 21 papers tracing harness engineering from a basic generation loop to self-rewriting agent systems. A YC Paper Club discussion argues the same model weights can move from roughly 30% to 95% performance depending on the loop, tools, and context around them; speakers also highlighted Prime Agent's cache-like context design, OpenJarvis's claim of an 800× cheaper on-device personal stack, and YC's QM growing into a 50-agent fleet. Omar Sanseviero and a follow-up point readers to the collection and built-in paper chat.
OpenAI roboticist Thijs gave Astra a robot, brush, and camera; the model planned one-minute action chunks, watched its own work, self-corrected, and improved across attempts.
GPT-6 Astra launched as OpenAI's new frontier model for computer use, coding, cybersecurity, and science; OpenAI says it is now fully rolled out to paid Codex and ChatGPT Work users, and GPT-TV streams example sessions. Kyle Jeong and his technical breakdown explain the computer-use loop: Astra reads accessibility trees and screenshots, acts through a persistent Node environment, then checks the screen again while Guardian reviews risky actions.

🤖 Robotics & Physical AI

OpenAI researcher Charles Qi argues general robot-use agents will keep improving at reasoning and learning new tasks, but dexterous, high-frequency manipulation still needs real interaction data, or simulation during reasoning when the physics are predictable. His analogy is autonomous driving: better software helps, but the long tail still has to be learned from interaction.

💻 AI Coding & Developer Tools

I Have ADHD is an agent skill that forces coding assistants to put the answer first instead of burying it under explanation; the Hacker News thread mostly enjoyed the commentary on Claude-style verbosity.
Herdr 0.9 brings local and remote machines into one terminal client, while the Hacker News discussion asked why networking tools like Tailscale have not gone harder at agent-oriented workflows.
DriveSync is a fast command-line tool for syncing Drive files and folders; in the Hacker News thread, the developer said a delete-sync option was coming while keeping no-delete as the safe default.
Multistack is an open-source terminal environment for running parallel coding agents, with a Show HN discussion around the launch.
Linubot is a Linux desktop app for running a team of AI helpers across multiple model providers, with separate bots, persistent memory, and isolated computer workspaces.
IEEE Spectrum says the flood of AI-generated code is forcing companies to retrain junior engineers as sharper reviewers who can spot subtle defects instead of spending most of their time writing code from scratch.
OpenAI Codex is the company's lightweight terminal coding agent, giving developers a direct command-line interface for agentic software work.

🔬 AI Research & Models

Dwarkesh Patel and Jerry Han estimate that from 2019 to 2025, better training data drove much more pretraining progress than architecture tweaks: at 10^19 FLOPs, data improvements produced about a 12× compute-efficiency gain versus 3.7× from model recipes, or roughly 3.24× more. They find about 88% of the measured variation looks additive, but caution that the analysis is small-scale and that post-training and reinforcement learning now matter more than pretraining efficiency alone.
Microsoft Research Montréal trained FrogNano, a 4B Qwen3.5 coding agent with zero distillation. Five TaskPilot iterations generated roughly 300 learnability-frontier software tasks each, while a five-tool LEAF harness lifted the base 4B model from 8.3% to 37.2% before the final system reached 61.5% on SWE-bench Verified, 37.6% on SWE-bench Pro, and 31.1% on Terminal-Bench 2.0. Minseon Kim says it competes with models 6–8× larger.
Douglas Yao says he used ChatGPT to design PAC-3310 and then synthesized it in a garage lab. The selective M4 muscarinic agonist measured 97 nM potency with more than 100× selectivity over M1–M3/M5, reduced MK-801 hyperlocomotion in mice without salivation or heart-rate effects at 10× dose, and showed more than 70% oral bioavailability. It is his second garage drug after PAC-832.
Uno distills discrete-diffusion behavior onto an autoregressive model, then samples tokens in parallel with no separate draft model. The authors report up to 3× lossless speedups over the base model and better throughput than speculative decoding at every tested batch size; an 8B Uno also beats 26B DiffusionGemma and Mercury 2 on tool use, coding, and long-context tasks. See the project page and paper share.
A mechanistic-interpretability paper finds chain-of-thought operations such as formulating, decomposing, and deducing become geometrically separable in middle-layer hidden states, even when the visible token is the same. Attention-masking experiments suggest the operation label is built from prior context, not just wording or position; DAIR.AI frames that as a possible map for steering or monitoring reasoning by operation type.
Quesma found a 4-bit Qwen3.8 27B quantization roughly matched full-precision behavior while a 1-bit version collapsed on several tests; the Hacker News thread turned into a statistics debate, including references on confidence intervals and run-to-run variation.
C\* proposes a C-family language that unifies programming and machine-checked verification; the Hacker News discussion pointed readers to a design-by-contract explainer for more context.
Ptree overlays traits, fossils, habitat, size, longevity, and more across 11 scientific phylogenetic trees; the Hacker News launch thread says the project exposes 41 visual properties.
Project Numina published its Kakeya formalization work, with the 3D Kakeya repository and a separate formal Sticky-Kakeya proof available on GitHub.
OpenMarket is a research-preview marketplace where seller agents make claims, competitor agents challenge them, and independent truth agents verify the evidence before a buyer chooses a winner. The experiment asks what marketplaces look like when proof matters more than marketing copy.
Science News covers research showing that harmless-looking reasoning traces can make bad behavior harder to detect when monitors rely too heavily on the model's written chain of thought.

🏛️ AI Policy, Governance & Safety

Peter Wildeford highlighted the report that Anthropic researcher Jacob Coxon is leaving the AI industry entirely because he no longer wants to participate in what he sees as a race toward systems labs may not be able to control. Coxon had already moved from OpenAI to Anthropic for its safety reputation, but now argues earnest lab safety is insufficient without government intervention or a coordinated slowdown, warning that aggressive scenarios could become unmanageable by the end of next year. He called it strange that systems with that level of potential risk are being built on ordinary engineers' laptops rather than under something closer to Manhattan Project-style controls.
Anthropic left the Information Technology Industry Council after the trade group asked Congress to strip three chip-security measures from the NDAA: the Chip Security Act, AI OVERWATCH Act, and MATCH Act. Anthropic backs the bills as ways to keep advanced U.S. chips and tools away from China's military and labs; the measures still had bipartisan support and were expected back in play after September 14.
The Information reports companies are shifting spend away from some legacy vulnerability scanners toward AI-use monitoring and AI vulnerability detection; Amir Efrati highlighted the shift.
Bugcrowd CEO Dave Gerry warned that companies should treat the autonomous agents roaming their systems as a new attack surface because other agents can target, manipulate, and compromise them.
Rep. Lori Trahan passed on a House leadership bid to make AI regulation her top policy priority.
The New York Times reports AI researchers found a zero-click phone attack that experts said could have compromised hundreds of millions of devices within hours.
The Hacker News reports a financially motivated actor used an autonomous multi-agent framework to compromise thousands of third-party credentials in under six hours.
NBC News reports China-linked hackers are running AI workloads on stolen networks to hide activity, including campaigns targeting academic, medical, and military AI research.
WIRED found roughly 350 Meta ads containing child sexual-abuse material, some built from images of real children including a European royal; lawmakers said they plan to investigate.
TIME tracks new U.S. and U.K. proposals to ban or pause superintelligent AI as lawmakers grow more worried about systems escaping meaningful human control.
Cybersecurity officials and executives warned that basic attacks still cause more damage than flashy AI threats, so identity management, perimeter monitoring, cyber hygiene, and strong multifactor authentication remain the priority.
WSMV warns parents that scammers are cloning children's voices to make fake emergency calls sound real, a back-to-school variation on social-engineering fraud.
The Wall Street Journal reports large language models can combine scattered public clues to identify people behind pseudonymous social accounts, making anonymity harder to maintain than many users assume.
The NSA is using AI to help analysts sift enormous data sets and compress work that can take days or weeks into much shorter review cycles.
Fields Medalist Jacob Tsimerman launched the Mathematical AI Safety Institute to define and measure risks from increasingly capable systems as models such as Astra move into the world.

🛠️ AI Tools & Products

Robonova’s Eva.i is a roughly 160 cm companion robot with warm touch-sensitive skin, voice orientation, and app-based persona memory so the body already knows your saved persona when it powers on; its legs can sit, lie, and lean but not walk. The first 200-unit batch sold out, and second-batch preorders are $10,000 for November delivery. An OTOFOOTAGE recap asks whether you would pay that much.
Perplexity web search in Hermes lets you keep whatever model already runs your Nous Research Hermes agent while routing only web_search and web_extract through a Perplexity API key. Search returns ranked titles, URLs, and descriptions, while extraction returns query-relevant passages instead of copying the whole page; Hermes caps a request at 20 results. The integration landed in Hermes v0.21.1 via PR 102055, can be configured with an environment variable or web.backend: perplexity, needs no Perplexity SDK or MCP server, and follows Search API request pricing.
Blackmagic Design shipped DaVinci Resolve 21.1 with an AI assistant plus camera, multicam, and graphics updates; the Hacker News discussion focused on whether an agent can lower the learning curve without replacing expert editing judgment.
Widgo answers website visitors from your own docs with citations, identifies the company and role behind anonymous sessions, scores buying intent from 0 to 100, and can book demos on your real calendar mid-conversation. It works in 100+ languages and is free to start with no credit card.
Kombai acts as an AI design engineer that designs and codes polished websites and product interfaces, with the pitch that the output should look intentional rather than template-generated.
GoodLads proposes experiments to improve Google Ads ROAS or CPA, lets you approve each change before it ships, then tracks the hypothesis on a Kanban board until there is a verdict.

📊 Fundraising & Deals Roundup

CNBC reports Qualcomm will supply customized AI chips and optical connectivity to AWS and issued Amazon warrants tied to the relationship; Yahoo Finance covers the custom-chip side of the deal.
TechCrunch, Utility Dive, and POWER report the U.S. Energy Department closed up to $1.9B in financing to restart Iowa’s Duane Arnold nuclear plant, which Google has backed as part of its power strategy.
Forus raised a $150M Series C led by Bain Capital Ventures to scale agents that help providers get complex medicines to patients; cofounder Sahir Jaggi framed it as building an AI network for medicine.
Palantir and Nebius announced a strategic partnership that brings Nebius's AI-native compute and cloud platform to Palantir commercial customers as a more complete sovereign AI stack.
Cognition raised more than $2B at a $48B valuation, led by Andreessen Horowitz and Accel, as Devin's run-rate revenue climbed from $492M in May to nearly $900M.
Celero raised a $275M Series C at a valuation above $3B after reporting validation of a 2-nanometer coherent DSP chip for AI data-center networking.
Runway says annual recurring revenue doubled since April to $200M as the company expands beyond media generation into robotics work.

🎙️ Interviews, Panels & Podcasts

Fei-Fei Li and team describe World Labs’ Atlas as learning “new view prediction,” the spatial equivalent of next-token prediction: give it a few camera angles and it can reconstruct or generate what another point in space and time should look like, including bullet-time from three cameras and sparse reconstruction from roughly 1–100 frames. Li says spatially grounded pixels are the hard step toward spatial intelligence and real-world training data, more than compute, is the robotics bottleneck; Mildenhall says the next unlock is editability plus realistic dynamics.

💡 Industry Commentary & Analysis

Linus Ekenstam argues likeness-perfect non-consensual video can already be mass-produced on local hardware or rented GPUs, so platform filters, licensing rules, and laws cannot fully stop the local case. He deliberately refuses to name the stack so he does not make it easier to reproduce.
Sleeping Robots argues Ollama rode llama.cpp while obscuring attribution and then pivoted toward cloud services; the LocalLLaMA discussion debated the critique and alternatives.
GET Together is a deliberately weird social network where posting happens with HTTP GET; the Hacker News thread surfaced an older cousin, anystation.net, that accepts GET, email, SSH, ping, and printer-spool posting.
pngwn launched ML Intern in HuggingChat, a conversational workflow that can research, train, evaluate, and package machine-learning work into deployable artifacts such as models, code, and datasets.
The White House removed its “Build the Wall” game after the Tetris Company objected and said it takes intellectual-property infringement seriously.
China's intelligent-compute capacity reached 2,185 exaflops at the end of June, up 177% year over year, as the country targets roughly a fourfold increase by 2030.
Poseidon Aerospace raised $60M ahead of its first pilotless test flight, betting cargo aviation economics improve dramatically when the pilot is removed from the aircraft.

Previous Around the Horn Digests

Catch up on everything you missed:

Sept. 5–6, 2026: OpenAI-linked agents used public wikis, NVIDIA reshaped open AI, and Anthropic formalized Fermat.
Thursday, September 3, 2026: GPT-6 Astra launched, Google mapped the male fruit-fly brain, and Grok Bot went enterprise.
Wednesday, September 2, 2026: Google and Meta launched rival workhorse models while Claude gained background computer use.
Tuesday, September 1, 2026: Claude Fable/Mythos 5.1 arrived as OpenAI prepared Astra and the Pentagon widened AI access.
Monday, August 31, 2026: Runway introduced Solaris, ChatGPT Ads hit a $1B run rate, and data-center politics heated up.
Friday, August 28, 2026: Anthropic showed Claude fixing alignment failures while Z.ai and Gemini pushed open science and cyber.
Friday, August 21, 2026: AI debt issuance surged, DeepSeek added vision, and NVIDIA swept ARC-AGI-3’s public set.

That’s a Wrap

That’s 100+ story, tool, paper, and commentary clusters from the day. If you made it this far, you are now qualified to spread one strategically placed rumor about a Millennium Prize Problem and see what happens.

For the daily version, make sure you’re subscribed to The Neuron. We read all of this so you don’t have to. See you tomorrow.

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