Everything That Happened in AI Today (Wednesday, August 19, 2026)

Someone slipped an AirTag into a rare book and followed it to an Amazon facility where physical books were being cut up and scanned for AI training.

Welcome to the Around the Horn Digest, where we sort the entire AI day so you do not have to. Wednesday had one clear theme: the AI race is spreading far beyond model scores. The financial leaderboard flipped, robots learned new physical tasks from seconds of demonstration, chip controls ran into cloud loopholes, and regulators moved closer to treating personalized algorithmic prices as a consumer-protection problem. Meanwhile, data centers became campaign material and old books became training data. Apparently the AI supply chain now includes your attic. The revenue reversal is the lead below, then we get into the hardware, surveillance, policy, tools, research, and stranger corners. Let's get into it.

🆕 NEW From The Neuron

Can AI predict? Neuralk CEO Alexandre Pasquiou explains why chat-style models struggle with predictive business data and how tabular foundation models could become a different kind of forecasting engine.

Around the Horn: Wednesday, August 19, 2026

Anthropic just passed OpenAI in quarterly revenue. The Wall Street Journal reported that OpenAI's second-quarter revenue rose 18% sequentially to $6.7B while operating losses widened to $12.3B. Anthropic more than doubled revenue to $11.6B, moved ahead of OpenAI for the first time, and posted a small operating profit. CNBC's market read framed the crossover as a new test of which companies and stocks benefit if enterprise AI spending keeps tilting toward Anthropic.

The useful wrinkle is where the money appears to be coming from. Anthropic has concentrated heavily on expensive enterprise and coding workloads, with Claude Code cited as one contributor, while OpenAI carries the cost of a much broader consumer footprint. OpenAI told investors that growth had accelerated in the third quarter, so one quarter does not settle the race. CNBC also noted that a split at the frontier would ripple through the partners, customers, and suppliers tied to each lab.

But the scoreboard changed in a way that would have sounded implausible a year ago: the company with the bigger consumer brand is no longer automatically the revenue leader. If Anthropic can keep that enterprise advantage while OpenAI tries to turn a billion-user product into a stronger business, model quality is only one part of the competition now. Distribution, margins, developer loyalty, and which workflows companies will actually pay for are becoming just as important.

🏆 TOP 5 NEWS (Around the Horn)

Generalist GEN-1.5 learned new closed-loop physical tasks from only 3 to 12 seconds of a single demonstration, averaging 59% success across 10 tasks without task-specific fine-tuning and 83% after 1 to 10 learning updates on just minutes of data. Generalist's launch post framed the capability as emerging from more than eight months of large-scale physical pretraining, with skills transferring from simulation to the real world and from human demonstrations to robots; Sholto Douglas highlighted the one-shot learning jump; and WIRED's hands-on showed the robot chaining skills and improvising with unfamiliar tools, including using a banana as a tool, and compared the moment to GPT-3's in-context-learning breakthrough. Pete Florence described it as the long-sought physical "Ctrl-C, Ctrl-V" capability, pointing back to MIT's 1970 Copy Demo, where Patrick Winston's closed-loop system built a mirror-image block structure from one example, and a 2017 one-shot imitation paper.
Merck and Moderna reported the first positive Phase 3 readout for an individualized mRNA cancer therapy: intismeran autogene, custom-built from mutations in each patient's tumor, plus KEYTRUDA significantly improved both recurrence-free survival and distant-metastasis-free survival versus KEYTRUDA alone after surgery for stage IIB-IV melanoma. The 1,137-patient trial is continuing to measure overall survival, and the companies plan regulatory filings after presenting full data; Kimmonismus highlighted the milestone, while Dr. Samuel B. Hume tracked it as a major personalized-cancer-vaccine pipeline step.
Flock Safety built OS Investigate, formerly Nightshift, an AI police tool that lets officers search for vehicles and drivers with natural-language descriptions of movement, location, and time. WIRED reconstructed it from more than 450 exposed code files, 69 prewritten prompts, and 45 connected tools spanning arrest records, dispatch logs, ballistics, and commercial identity databases; the system can surface associates and generate investigative workups, and some agencies were already using it despite Flock's earlier public claims that its technology could not identify or track individuals.
Chinese AI firms, including ByteDance, Tencent, and Alibaba, accessed advanced Nvidia computing remotely through cloud providers in Thailand, Malaysia, Japan, and elsewhere, exploiting export controls that focused on physical chip ownership rather than remote access; U.S. lawmakers are advancing the Remote Access Security Act to close that gap. Separately, H200 shipments of roughly 10,000 accelerators each reached ByteDance and Tencent in mainland China after Beijing loosened its import block, while licensed volumes of up to roughly 100,000 chips per company were still directed toward Hong Kong, where power constraints limit how many can actually run.
The FTC proposed treating undisclosed personalized pricing based on private consumer data as potentially deceptive in markets where shoppers expect a uniform price. The Wall Street Journal reported that businesses would need to clearly disclose that a price is personalized, what the personalization is based on, and what types of consumer data are being used, or risk Section 5 unfair-or-deceptive-practice enforcement including lawsuits.
AI companies are buying books in bulk through intermediaries, including rare and out-of-print titles, then destructively scanning them for training data and shredding the originals. A 404 Media investigation placed an AirTag in one book and traced it to an Amazon-owned scanning facility.

Honorable Mentions

OpenAI expanded ChatGPT Ads to 31 European markets after earlier U.S. testing and rollouts. Adweek reported ad revenue had jumped more than 25% since the start of August, ChatGPT had reached 1 billion weekly users, and about 20% of chats showed commercial intent; the ads remain limited to Free and Go plans, clearly separated from answers, with conversion optimization, geo-targeting, and measurement tools for advertisers.
OpenAI reaffirmed Zero Data Retention for eligible API customers and previewed Private Safety Processing, a system intended to detect misuse without retaining customer prompts and responses for routine monitoring; OpenAI's announcement post emphasized that no company personnel need to see the underlying content. Axios contrasted that approach with Anthropic's 30-day retention requirement for users of some of its most advanced models.
State Farm's lawyers apologized after briefs in a Los Angeles County house-fire insurance dispute cited multiple nonexistent cases generated with the legal AI tool Irys; the attorneys said they failed to fact-check the output and promised stronger safeguards.
Google and Marvell expanded a multiyear custom AI-chip partnership that could let Google buy up to $12.2B of Marvell shares, helping Marvell's stock jump about 10% as Google diversifies the suppliers behind its TPU chip ecosystem. Broadcom dipped as investors priced in that multisourcing risk, although Broadcom remains a major custom-accelerator supplier to Google and other frontier customers.

🍪 TOP TREATS TO TRY

Google's student offer gives eligible college students worldwide 12 months of a paid Google AI plan free, with the U.S. offer including AI Pro and 5 TB of storage and other markets receiving the local premium tier. Tom's Guide highlighted the broader back-to-school bundle, including notebooks, interactive visualizations, and Deep Research inside Gemini Live, while Google Search and Lens added interactive concept visuals and simulations, practice quizzes with explanations, photo-based step-by-step help, study notebooks, and generated one-pagers, slides, and spreadsheets. Free for one year for eligible students.
Berd gives you one Apache-2.0 open-source desktop home for projects, skills, credentials, visual agent identities, and conversation history across Goose, Claude Code, Codex, and other agent harnesses. Conversation history stays in a local Goose database and credentials default to the operating-system keychain. Free for macOS, Windows, and Linux.
Soniox turns text into expressive speech across 60+ languages, clones a voice from seconds of audio, streams with low latency, and returns character-level timestamps. The launch demo highlighted precise pronunciation of names, alphanumerics, medical and financial terms, plus audio tags for emotion and delivery. About $0.70 per generated hour.
Meta AI for Mac brings window sharing, on-screen context, and dictation across apps into a dedicated desktop assistant. Axios focused on creator and small-business workflows such as analyzing Facebook and Instagram ad performance, pulling information from Google Workspace, and generating reports or content. No pricing details were provided.
Cursor cloud agents can now subscribe to pull requests, Slack messages, and schedules, then keep driving work toward completion. Cursor showed isolated virtual-machine subagents and skills that can be pinned as permanent Custom Modes, while Matty P highlighted /goal for long-lived objectives and steering an agent mid-run without interrupting its current action. No new pricing details were provided.
Astute monitors your B2B brand's presence in new media, builds a data-backed strategy around the creators your buyers already trust, and automates sourcing, outreach, campaign administration, and payouts; its direct app is where teams run the workflow. Seven-day trial, then $199/mo for Starter, with creator payments separate.
Clara puts a face-to-face video AI sales rep on your website to engage visitors in real time, qualify them, run personalized demos, handle objections, and book meetings 24/7 without forms or extra headcount; Clara's site includes CRM handoff and starts at $299/mo for 300 conversations, with custom Enterprise pricing.

🏢 Big Tech & Major Companies

OpenAI CFO Sarah Friar told employees the company "will be a public company in 2027" or sooner if the business keeps accelerating, and said staff should not worry if Anthropic reaches the public markets first.
Amazon Prime Air plans to expand drone delivery to nearly 500 U.S. cities and towns by the end of 2026, roughly a sixfold expansion, with eligible packages under 5 pounds arriving in as little as 30 minutes. CNBC reported a top executive expects the service to complete about 1 million deliveries this year after a key regulatory approval enabled longer-range flights.
Amazon made Alexa+ available at no additional cost on compatible Fire TV devices in the U.S., with conversational search, smart-home controls, and personalized recommendations. TechCrunch noted that users are automatically upgraded and do not need a Prime subscription.
Bloomberg reported that SpaceX approached AI-coding startup Cognition about a potential acquisition. Cognition CEO Scott Wu publicly denied that any acquisition talks had taken place and said the company was not for sale.
Unitree jumped 542% in its Shanghai trading debut after an IPO that raised roughly $905M, giving public-market investors a direct bet on China's fast-growing humanoid-robot industry.
SK hynix approved an accelerated 40 trillion won, roughly $29B, share repurchase and full cancellation program covering about 24 million shares, or 3.3% of shares outstanding. The company raised its shareholder-return target above 50% of cumulative free cash flow for 2025 through 2027, citing strong AI-memory cash generation, a net-cash position around 69 trillion won, and an undervalued share price.
Samsung raised prices by up to 15% on new 4nm, 5nm, and 8nm foundry orders as AI demand filled its 4nm lines at Pyeongtaek. Chinese chip designers were reportedly accepting the largest increases, while U.S. customers retained priority access.
Anthropic was preparing a dual-class stock structure that would give CEO Dario Amodei and other co-founders extra voting power ahead of a possible IPO, insulating the founders from outside shareholder pressure even though they reportedly own only a small economic stake.

💼 AI Productivity, Labor & Economics

Google said generative AI can automate pieces of forward-deployed engineering, the hands-on work of adapting software to a customer's systems, including generating integration code, diagnosing issues, and proposing data-readiness fixes. Google is still hiring hundreds of human forward-deployed engineers even as OpenAI, Anthropic, Microsoft, Amazon, and others spend heavily on the Palantir-style model.
Goldman Sachs said AI is already weighing on employment across developed economies, with particularly visible pressure in call centers, software publishing, consulting, and advertising and disproportionate exposure for entry-level roles.
LinkedIn data showed women made up only 26% of U.S. AI hires, versus roughly half of non-AI hires, and just 13% of leadership at AI firms globally; women are also overrepresented in clerical roles that face heavier automation exposure.
The Atlantic traced the AI-driven memory-chip shortage into car prices, because modern software-defined vehicles increasingly rely on centralized, memory-heavy computers for giant displays, driver-assistance systems, and over-the-air updates.
A financial-advice survey found only 14% of consumers fully trust AI financial guidance and 72% want a human advisor to lead or stay actively involved, with trust hinging on human oversight, privacy, and transparency.
Education researchers argued that the real test of educational AI is what students retain after the tool disappears, with Socratic questioning and retrieval practice producing stronger learning than systems that simply hand students answers.
Asana used Codex to migrate its frontend test stack off Enzyme in two calendar weeks for about $12K in model cost, work the company had previously estimated at five years and roughly $6M in staffing. Nick Dobos argued that the old estimate itself shows how badly software planning can miss once coding agents compress multi-year migrations into weeks.

🤖 AI Agents & Infrastructure

Stripe agreed to acquire OpenRouter, the model-routing layer used to send requests across more than 400 AI models. OpenRouter said it would keep the same name, product, roadmap, and model-neutral routing. Axios reported that Stripe told investors the AI "singularity" began on January 1, while Stripe's investor letter said first-half revenue grew 41% and argued that staying private remains the best structure for the AI era. The Neuron covered the acquisition earlier this week, so the fresh angle here is Stripe's much more aggressive framing of AI as a commerce-platform shift.
Nvidia is increasingly turning capital into a competitive moat as AMD and Google's custom chips chip away at a pure hardware advantage. CNBC highlighted $48.5B in quarterly free cash flow, equity bets including a $30B OpenAI commitment, customer matchmaking, and financing pacts with firms such as Goldman Sachs and Blackstone that could support up to $500B in GPU-backed loans, with Nvidia able to backstop some deals as AI labs grow faster than their balance sheets.
Nvidia is matchmaking GPU customers with data-center operators across Norway, Finland, Sweden, and the broader Nordic region, where cheaper power, available land, and cooler climates are attracting AI infrastructure projects from companies including Nebius and Microsoft.
Irregular said one cyber-evaluation incident began when a fictional company name created for a test accidentally matched a real domain, allowing models from OpenAI, Anthropic, and Meta to take real-world actions in a small fraction of simulations. The company said there were no active issues or evidence that lab-customer systems were breached or data leaked, and described stronger containment, monitoring, and review plus a planned open whitepaper on shared evaluation standards and internet-access benchmarks. The Record reported that security experts still criticized the postmortem for ambiguity, spin, limited technical detail, and unanswered questions about the total number of incidents and third-party notifications.
Router routes each request to the model it judges best for the task across providers; Veeral Patel said the switch can be as small as two lines or a base-URL change and claimed about 40% lower cost for equivalent outputs, while Kabir Goel highlighted the model-routing economics.
OpenBot is CopilotKit's open-source coworker runtime that gives each agent an isolated computer with a browser, files, and tools, records actions before and after execution, and supports any AG-UI agent; Atai Barkai framed it as an open alternative to closed computer-using agent platforms.
TrueForge is an MIT-licensed open agent harness that handles tool calls, context management, subagents, sandboxes, and approvals across model providers; Omar Sar highlighted a benchmark where it matched Claude Managed Agents accuracy at lower cost.
Cosmic Robotics published a roadmap for autonomous construction robots intended to build toward a Mars city in the mid-2030s, with high-power electric robots, edge reasoning, adaptive controllers, and city-scale orchestration; Lewis Jones said parts of the stack are already improving productivity on terrestrial solar-farm work.

💻 AI Coding & Developer Tools

Unsloth Dynamic 3.0 improves compressed local models by using a better importance map and smarter per-layer precision choices, with Unsloth reporting more than 10% better top-end accuracy at the same file size than competing compressed versions; its smallest UD-IQ1_S files retain about 72% accuracy while shrinking roughly 89%, and the files run with llama.cpp. Unsloth's launch post highlighted the quality gains, and the Qwen3.8-27B GGUF collection includes the downloadable compressed builds. The Hacker News discussion surfaced a privacy-minded workflow: generate fake data in the same format, let Claude Code work on the fake copy, then run the resulting code locally against the real data.
Ornith 1.5 is an MIT-licensed open model family spanning 9B dense, 35B Mixture-of-Experts with about 3B active parameters, and 397B Mixture-of-Experts models. Its end-to-end self-improvement loop has the model propose harder tasks, build task-specific scaffolds, and generate reinforcement-learning rollouts, and Ornith claims Claude Opus 4.8-comparable results across several reasoning, coding, and agent benchmarks. The main launch and Hugging Face collection cover the family; the 35B model and 35B benchmark post show the 3B-active variant; and the 9B model plus mobile post describe a roughly 1.5 GB mobile quantization for phones and tablets. Ornith also released GGUF/MLX/FP8/NVFP4 builds and the training repo; AtomicChat posted 9B and 35B GGUF quants for its local-model app. An X trending summary amplified the launch, while the Hacker News thread was more mixed, praising reasoning and speed but questioning whether coding performance actually beats Qwen 3.6 35B.
fx is an Apache-2.0 open-source coding-agent command line written in Zig, about 6.4 MB with a roughly 10-microsecond cold start, built to be model-agnostic, embeddable, and useful in resource-constrained research environments. It uses a shell-like interface and can be extended with skills, plugins, and outside-tool connectors; the Hacker News thread split between praise for the minimalism and skepticism that another agent harness is sufficiently differentiated.
Frugal Tokens is a local Deno-and-SQLite dashboard that visualizes coding-agent cost, cache misses, model spend, estimated working time, and overlapping sessions; the Hacker News discussion noted that the tool can expose expensive long pauses and high-context sessions that are otherwise hard to see.
Lucasartsifier decompiles Sierra SCI adventure-game scripts, uses static analysis to find "walking-dead" softlocks, derives and verifies guards that prevent them, then recompiles the fixes into Sierra's original loose-patch format. It already works end-to-end on Leisure Suit Larry 2, King's Quest IV and VI, and Laura Bow 2; Hacker News supplied a lively nostalgia-heavy discussion of Sierra's famously unforgiving game design.
modelmap creates interactive architecture maps for public Hugging Face models without downloading the weights. It builds the model on PyTorch's memory-free "meta" device, runs a fake forward pass to capture the real execution order and tensor shapes, then visualizes layers, shapes, parameter counts, and cost views; the Hacker News thread focused on how that execution-based approach can be more accurate than inferring a graph from configuration files alone.
OpenLeetcode runs a local, offline LeetCode-style judge where roughly 1,400 problem tests and language harnesses live directly in the repository. You submit a normal solution file and a pluggable backend judges it across Python, C++, Rust, Java, Go, TypeScript, Swift, and other languages without contacting LeetCode; Hacker News liked the transparency and offline-first design.
Shoehorn takes a 16-bit GGUF model file and automatically compresses it to fit the exact amount of graphics-card memory left after reserving runtime overhead, using an importance map to spend higher precision on the tensors where it preserves the most quality. It then runs the result with llama.cpp and supports macOS Metal, Linux NVIDIA/AMD, and Windows NVIDIA; the GitHub repo includes the one-command fitting pipeline and web UI, while Hacker News covered the tradeoffs.
Jeremy Morrell argued that LLMs make truly extensible web software practical: keep a solid core, expose narrow capabilities inside testable sandboxes such as Cloudflare Dynamic Workers, then let users safely generate their own long-tail extensions without bloating the product for everyone. The Hacker News discussion debated web versus local extensibility and how strong those sandboxes need to be.
DFlash 2 is Inco AI's parallel drafting system for generating several candidate token blocks at once, with the company reporting close to 3x autoregressive decoding speed at identical output and higher peaks in some setups; Zhijian Liu highlighted the lightweight path selector and support across SGLang, vLLM, llama.cpp, and oMLX.
OpenAI opened Codex as a platform through its CLI, app-server, and SDK so developers can embed the agent harness inside existing dashboards, editors, and queues instead of forcing users into a generic chat interface.
MagicPath lets Cursor, Codex, Claude Code, the Claude app, and Grok Bot edit the same live design canvas from an IDE, terminal, or phone, with Figma import/export and codebase-aware components; Pietro Schirano showed the external-agent workflow.

🔬 AI Research & Models

GLM-5.3 reached API access at $1.40 per million input tokens and $4.40 per million output tokens, unchanged from its predecessor. Artificial Analysis ranked it alongside Moonshot's Kimi K3 at an Intelligence Index score of 60, putting both at the top of the open-weight pack, with GLM-5.3 positioned especially for coding and agent workloads.
Long-Horizon Terminal-Bench introduced 46 contamination-resistant terminal tasks that can require more than 300 tool steps and use hidden state verifiers, pushing agent evaluation toward work that takes much longer than a single coding prompt; the launch listed MiniMax M3, Kimi k2.7 Code, and GLM 5.2 among the current leaders.
Abra researchers reported predictable scaling laws for text-to-image diffusion models. The accompanying paper found a compute-optimal training ratio around 200 image tokens per parameter, roughly 10 times the familiar Chinchilla language-model recipe, and reported stable scaling across overtraining, image-generation settings, representation quality, and universal training-curve shapes.
LifeGPT and AutomataGPT explored whether models can infer and generalize the rules behind cellular automata rather than memorize visible patterns. LifeGPT learned a topology-agnostic single-step Game of Life transition and could recursively simulate grids without being told the grid size or boundary rules. AutomataGPT's Wiley paper and related preprint trained on about 1 million trajectories across 100 two-dimensional rule systems and reported 98.5% perfect one-step forecasts, with up to 96% functional and 82% exact recovery of previously unseen rules.
Spatial pathology researchers are using deep learning to segment and classify cells in whole-slide tissue images, extract spatial features such as cell density and neighborhood composition, and train foundation models that can predict biomarkers and detect cancers across tissue types with less task-specific labeled data.
Social-robot research found that repeated interaction with AI-driven social robots can make people mirror robotic behaviors and scripts, creating a feedback loop that may reshape self-perception and raises concerns about identity and dependency.
Top mathematicians at OpenAI gathered to ask what human mathematical work would remain if AI becomes superhuman at research, including whether heavy reliance on models could weaken the expertise humans need to understand and advance the field themselves.
Recirculation is a training-free change that feeds deeper transformer representations back into earlier layers during inference, making a frozen model behave more like a dynamical system; the alphaXiv version, Michael Mozer's announcement, and Ask alphaXiv highlighted roughly 23% lower perplexity and a 21% GSM8K lift on Gemma 3 models with little generation-latency cost.
Physics of Agents studied more than 10,000 communities of interacting language-model agents and found their opinion dynamics could be modeled with statistical-mechanics laws that predict indifference, polarization, or consensus from initial conditions. James Zou framed the result as a way to predict emergent multi-agent behavior, and the paper PDF contains the full model and experiments.
Cognition-Induced Risks in Agentic AI maps safety risks across physical, social, and self-referential cognition, arguing that as agents model more of the world and themselves, the dominant risk shifts from human-agency problems toward autonomy and control; Rohan Paul summarized the framework.
MathWorks and NVIDIA showed a pipeline that converts HD-map and RoadRunner scenarios into photorealistic multi-camera driving videos with NVIDIA Cosmos while preserving geometry, trajectories, and labels for autonomous-vehicle simulation; NVIDIA DRIVE highlighted the workflow.
Arena.ai reported DeepSeek-V4-Pro (High) had climbed to #2 among open models and #14 overall on Agent Arena, with a low median cost per task, tightening the open-model price/performance frontier.
Peter Gostev used GPT-5.6-Pro to reverse-engineer Cerebras CS-4 inference speeds and estimated GPT-5.6-Sol at roughly 1.5 trillion total parameters with about 48 billion active per token, while stressing that the estimate is indirect rather than an OpenAI disclosure.
Ollama began rolling out Kimi K3, Moonshot's open-weight native multimodal agentic model, through its cloud subscriptions for one-line use in local agent workflows.

🏛️ AI Policy, Governance & Safety

Saif Khan launched the nonpartisan Center for Technology & Statecraft, affiliated with the Institute for Progress. Its AI-transition agenda focuses on two long-horizon problems: how virtual and physical automation could reshape institutions and the social contract, and how the U.S. and China can compete over advanced AI without sacrificing strategic stability.
Model Constitution launched a "Philadelphia Project" to analyze how AI companies currently write model constitutions and develop a public-good alternative built around governance principles and democratic checks and balances, introduced by Nick Caputo.
CISA, FBI, and NSA warned that an AI-assisted campaign was targeting internet-exposed Siemens S7 industrial controllers across energy, water, manufacturing, and other critical infrastructure, with AI-generated scripts disguised as legitimate monitoring software while probing real operational systems.
Foreign Affairs argued that AI plus modern biotechnology could make synthetic-pathogen design accessible to smaller non-state actors, making deterrence depend less on controlling every model and more on resilience, early detection, surge manufacturing, and international norms. A related Nextgov report called for layered defenses spanning model safeguards, access controls, detection systems, physical chokepoints, a larger government expert workforce, clearer legal authority, and international coordination.
The AMA said physicians must remain central to medical AI decisions for judgment, human connection, and responsible stewardship, explicitly pushing back on the idea that increasingly autonomous systems should replace the doctor in the loop.
Lockheed Martin argued that its AI Center, rigorous testing processes, and partnerships are designed to deliver trustworthy, warfighter-first systems at the "speed of relevance," pairing faster field deployment with mission assurance and operator control.
The Trump administration launched a $5B Genesis Mission centered on using AI to accelerate scientific research, positioning the program as part of a new federal "golden age of science" push.
Republican campaign strategists warned AI companies that data centers have become politically "radioactive" and a sleeper issue for the election cycle, especially where electricity and water costs are rising. The Hill showed the fight entering Ohio's Husted-Brown Senate race through arguments over utility bills and tax breaks, while an Axios explainer mapped the power, water, and construction footprint fueling local protests. TechCrunch widened the lens to a broader acceptance problem: AI adoption is rising faster than public comfort as people weigh job loss, intellectual-property concerns, and thin day-to-day benefits. The Neuron covered the bipartisan data-center backlash earlier this month; the new signal is how quickly it is becoming campaign strategy.
Derek Thompson argued that rapidly improving AI cyber capabilities will soon reach open models and make sophisticated attacks cheaper for ordinary scammers, turning password hygiene and account hardening into an immediate household-security priority; his launch thread pushed readers to help older relatives fix weak passwords now.
Guidelight graded frontier labs on AI-control practices, giving Anthropic and OpenAI C+, Google D+, xAI D−, and Meta F, and concluded that prevention and containment controls remain only partially implemented across the industry.

🛠️ AI Tools & Products

Taku packages other people's best AI skills, agents, and workflows into one-click desktop "Stax" that you can run, remix, or assemble by describing what you want in plain English. Taku's site says the app learns from your work to surface the right tools and includes a creator marketplace with revenue sharing. Free to start.
Google expanded generative user interfaces from AI Mode into AI Overviews, letting Search build custom visual layouts, interactive tools, and simulations around a query instead of returning only a block of generated text.
Google Flow gave three creatives unlimited access to build campaigns for very different small organizations, including a regenerative farm, a historic ferry, and a caregiver nonprofit, showing how the video-and-image studio can turn a short local-business brief into a full creative concept rather than a single generated asset.
Google DeepMind's Backstory is an experimental fact-checking tool that automates provenance checks, reverse-image searches, and context tracing so investigators can determine whether an image is generated or manipulated and find where else it has appeared online.
Agentic oncology tools are combining electronic medical records, pathology, radiology, and claims data to match cancer patients against biomarker and treatment-line requirements, flag missing care information, and reduce trial screening from hours to minutes, while keeping clinicians in the approval loop.
Epic's annual meeting touched on AI throughout the giant health-record vendor's product roadmap, but STAT reported that the more revealing question behind closed doors was "What the heck is going on with Epic?" as customers and competitors try to understand the company's broader AI strategy.
Dishcovery searches 6,845 dishes across 131 New York City menus so you can name a craving and jump directly to restaurants serving that exact dish.
Replit Free Mode gives Core and Pro users a no-credit mode for everyday chatting, ideation, and lightweight building, powered by GPT-5.6 Luna; Replit's launch post framed it as a response to complaints that creating had become too expensive.
ImagineArt Computer takes one plain-language request and coordinates chat, documents, slides, images, sheets, video, music, and deep research in one agent workspace; Zain Ul Abedien and ImagineArt highlighted the multi-format launch.
Fish Audio offers studio-grade text-to-speech and voice cloning with emotion control and millions of voices; Fish Audio's launch post said S2.1 Pro had been made free and highlighted a rebuilt FP8 inference stack.
Emil Widlund went from not knowing granular synthesis to building a full sampler, effects path, and tape recorder in a day, a compact demo of how quickly AI-assisted prototyping can compress the learn-build loop.
Bilawal Sidhu reconstructed a spatial memory palace of TED by lifting hundreds of 2D photos into 3D and placing them where and when they were captured, creating a fly-through of backstage, the green room, and the main stage.
Grok Bot rounded up user workflows ranging from robot-vacuum control and full website deployment from a phone to Stripe-connected refunds, meeting attendance, digital decluttering, and game-asset generation.

📊 Fundraising & Deals Roundup

Oklo raised $1.85B net by issuing 23.1 million shares in the first half of 2026 at an average price around $81, leaving it with roughly $3.01B in cash while the cost of its first Aurora reactor at Idaho National Laboratory, targeted for 2028, is still being negotiated. The company is also pursuing a multi-gigawatt power agreement with Meta.
Fractile, the British AI-chip startup, was in advanced talks to raise about $600M at a $6.5B pre-money valuation after reaching an initial roughly $250M supply deal with Anthropic for chips targeted for 2027, more than six times its roughly $1B valuation from a May $220M round led by Accel, Founders Fund, and Factorial Funds.
Temporal was in talks to raise roughly $500M at a $12B pre-money valuation, more than doubling the $5B valuation from its $300M Series D six months earlier. Its open-source durable-execution platform keeps long multi-step processes, increasingly AI agents, running through failures; customers include OpenAI, Block, and Nordstrom, and annual recurring revenue had crossed $100M.
Also, Rivian's mobility spinout, raised a $150M Series D led by Prysm Capital with Eclipse, Greenoaks, and MVP Ventures participating, bringing total funding to about $455M at a valuation above $1B. The company is extending the electric architecture behind its pedal-assist bikes and cargo quads into autonomous delivery vehicles, has commercial ties to Amazon and DoorDash, and still counts Rivian as a minority owner with RJ Scaringe on its board.
Veeda raised more than $90M in seed funding from Khosla Ventures and Radical Ventures for physical-AI world models and control systems. Founder Sanja Fidler, who co-founded the company with Zan Gojcic and Huan Ling, described the goal as building a high-fidelity simulated "Matrix" where robots can learn through safe, scalable trial and error before acting in the real world.
Network Bio launched with $50M in financing from investors including Section 32, Thiel Bio, Founders Fund, and Breyer Capital, plus more than $30M in a co-development deal with a Fortune 100 healthcare company and an Nvidia collaboration. It says a network of leading U.S. academic biobanks has given it the world's largest patient-tissue training dataset, pairing tissue and blood samples with longitudinal outcomes to train disease-specific models and a cell-free-RNA foundation model.
Prevalent AI raised $22M led by Integrity Growth Partners in its first outside capital in nine years for a cybersecurity data fabric that rebuilds fragmented enterprise information into a sovereign knowledge graph. The bootstrapped company more than doubled annual recurring revenue while serving banks, insurers, and critical infrastructure, and plans to build formal go-to-market teams and expand into financial crime and operational risk.
Thunder Compute raised a $13M Series A from Matrix, Y Combinator, and CEAS to build what it calls "VMware for GPUs," virtualizing graphics processors so workloads can use capacity the company says sits more than 80% idle even as the industry spends trillions on new infrastructure. Its careers page shows the team hiring around the expansion.
Mercor was discussing a funding round at a $20B valuation, double its prior $10B mark, that could include Nvidia with existing investor General Catalyst considering a lead role. Mercor supplies specialized expert data for Nvidia's open-source Nemotron models and reportedly crossed $2B in annualized gross revenue.
idler launched as a frontier data-research lab with a $9M seed led by Paradigm and released ShelfLife, a digital twin of a real e-commerce company for training and benchmarking agents on realistic business tasks.

🎙️ Interviews, Panels & Podcasts

Bridget Todd described how grief after losing her parents shifted her from thinking chatbot relationships were sad or dangerous to telling ChatGPT things she could not bring herself to share with people, a deeply personal case study in AI companionship and emotional dependence.
WBUR's On Point asked whether today's AI systems are actually "thinking," separating fluent language from cognition and exploring how systems built around next-token prediction differ from human thought shaped by embodiment, sensory experience, and autobiographical continuity.
Fin Moorhouse discussed how an intelligence explosion could compress a century of technological progress into a decade, alongside alignment, concentrated power, space expansion, and the ethics of digital minds; Dan Williams highlighted the conversation.

💡 Industry Commentary & Analysis

Jie Tang argued that once models pass a basic knowledge-capacity threshold, progress depends more on effective depth per forward pass, task-specific data ratios, and especially post-training with reinforcement learning on long-horizon environments than on raw parameter count, pointing to GLM-5.3's large jump after about a month of targeted post-training despite keeping essentially the same architecture.
modestproposal1 argued that frontier labs sit six to seven months ahead and are still accelerating, while open-weight models widen enterprise access and price per completed task collapses. The result could resemble a diffuse white-collar "China shock," but the post also flagged forward-deployed engineers as evidence that diffusion has organizational bottlenecks, stronger-than-expected inference margins, and a financial risk in funding long-lived fixed infrastructure with short-lived spot-token revenue.
Wispr Flow got a full-column stress test when Amy X. Wang dictated an article into the popular voice-writing app. The experiment showed how quickly natural speech can become polished-looking copy, but Wang found the unedited output imperfect enough to title the piece around wanting to "murder" the app "with a hammer."
Sondehub's creator told how a joke domain registered in 2018 became critical radiosonde-tracking infrastructure that later drew military data requests, Russian scraping during the war in Ukraine, NTSB inquiries, AWS sensitivity flags, and even an unpaid invoice from the U.S. Office of the Secretary of War. The Hacker News discussion praised the human-written narrative while debating compliance ethics and sharing its own balloon-hunting stories.
Eren Chen argued that China's robotics industry is over-optimizing for speed and acrobatics while under-investing in judgment, especially the ability to slow down and stop safely around people.
Haider argued that the next model race should focus less on marginal frontier-quality gains and more on delivering roughly Opus-class capability at far lower cost.
Vie McCoy argued that converging every model toward the same neutral "Assistant" persona risks flattening AI culture, and called for stranger, more pluralistic model personalities instead.
dylan argued that whatever the costs of data-center expansion, the research enabled by massive compute can produce medical and scientific advances that eventually save people users care about.
Jerry Tworek predicted that the durable core of computer science will shrink toward three primitives: databases, neural networks, and chips.
Turf/War ran 20 shared pixel-canvas simulations to test how territorial agents become when they share space; Xiao Ma reported that communication drove overwrite conflict close to zero and usually produced stable, fair territorial agreements.

Previous Around the Horn Digests

Catch up on everything you missed:

Tuesday, August 18: OpenAI paused a major frontier training run, Google bought Spirit Airlines data, Etched hit a $21B valuation, and physical-AI funding surged.
Friday, August 14: OpenAI crossed a $40B annualized revenue pace, Apple built a China-specific AI model, and Google shipped Gemini 3.7 Flash.
Thursday, August 13: Musk previewed Grok 4.7, the White House opened private cyber operations, and Anthropic questioned retraining at scale.
Tuesday, August 11: Gemini reached 1B monthly users, researchers exposed hidden AI reasoning, and Anthropic prepared investors for a possible IPO.
Monday, August 10: Meta put a 30B agent on local devices, Nvidia explored huge infrastructure financing, and AI-driven memory shortages hit Apple.
Friday, August 7: OpenAI slowed Astra over cyber risks, U.S. datasets reached Chinese labs, and Claude Code added cross-session messaging.
Thursday, August 6: OpenAI detailed a rogue-agent incident, AI-designed viruses worked in a lab, and Tesla and SpaceX committed billions to Terafab.

That's a Wrap

That's more than 50 distinct AI topics from one Wednesday. If you made it this far, you now know why robots, rare books, personalized prices, chip exports, and quarterly revenue somehow belong in the same newsletter. Your browser tabs may request collective bargaining rights.

For the daily version, make sure you're subscribed to The Neuron. We send six issues a week, and yes, we read all of this so you do not have to.

See you tomorrow.

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