Around the Horn Digest: Everything That Happened in AI Today (Monday, June 29, 2026)

AI is no longer just changing products. It is now changing invoices, elections, government workflows, chip supply chains, data-center valuations, and the sleep schedules of the people building it. Also, One Meta volunteer reached 78% word accuracy by typing with their brain, which feels like a useful moment to remember that keyboards were once considered advanced.

Welcome to the Around the Horn Digest, where we sort the day’s AI chaos into what actually matters. Today’s lineup had one clear economic thread: AI is forcing institutions to renegotiate the old rules. Consultants are rethinking billable hours. California is putting Claude into government workflows. Data centers are trading like strategic infrastructure. Lobbyists are dropping Chinese tech giants. Entry-level jobs are flashing warning signs.

Monday's bigger pattern became AI colliding with government, labor, and the physical supply chain. Apple chased relief from memory prices, and Stanford's latest payroll data showed the career ladder losing its bottom rungs. Meanwhile, tech workers lost sleep managing agents, and somebody launched 24/7 robot television. Apparently “too much screen time” was a challenge, not a warning. Let's get into it.

Around the Horn - Monday, June 29, 2026

AI’s most interesting pressure point today was the business model hiding underneath professional work. The consulting and legal worlds were built around time: billable hours, staffing pyramids, long research cycles, expensive drafts, and senior review. AI makes a lot of that work faster, which sounds great until the client asks the obvious question: if the work took less time, why am I still paying the same hourly structure?

That makes the billable-hour rethink the best lead-story candidate. It connects consulting, law, entry-level work, agentic productivity, and the broader shift from “pay me for effort” to “pay me for outcome.”

The other big news from today was Meta's Brain2Qwerty v2, a non-surgical system that turns brain activity into text. Meta trained it on roughly 22,000 sentences from nine volunteers, then moved from decoding individual characters to producing words and meaning in real time.

Average word accuracy reached 61%, compared with 8% for previous non-invasive systems. The best participant reached 78%. AI at Meta also released the training code, while its research partner released the first dataset for this version.

For now, Brain2Qwerty remains a laboratory system. Its direction still matters: safer brain-computer interfaces are moving from laboratory signals toward usable language. That could eventually help people who cannot speak or type communicate without brain surgery.

Previous digests: Monday, June 22 | Friday, June 19 | Thursday, June 18 | Tuesday, June 16 | Monday, June 15 | Thursday, June 11 | Monday, June 8

Monthly skill digests: AI skills guide | Skills video guide

🏆 TOP 5 NEWS (Around the Horn)

• California struck a statewide Claude deal with Anthropic, giving state agencies plus cities and counties discounted access, training, and support for government workflows.
• Digital Realty agreed to buy Blackstone’s stake in three Northern Virginia data centers, valuing the fully leased assets at $7.8B.
• Stanford and ADP launched a dashboard showing AI-exposed entry-level jobs weakening for workers ages 22-25, while older workers in similar roles have mostly held up.
• Trump’s AI access crackdown rattled tech companies, with Anthropic forced to pull Fable 5 and Mythos 5 while OpenAI slowed GPT-5.6 as labs waited for clearer release rules.
• Taiwanese prosecutors raided Super Micro’s local office as an Nvidia chip-smuggling investigation widened.

Honorable Mentions

• Alibaba and Tencent lost Washington lobbyists as a Pentagon rule forced firms to choose between defense contracts and blacklisted Chinese tech clients.
• Apple reportedly asked Washington for clearance to buy memory chips from blacklisted Chinese supplier CXMT as memory prices surged.
• AI money is flooding 2026 races, with OpenAI- and Anthropic-linked super PACs backing rival candidates in a fight over whether AI rules should stay federal or include tougher state laws.
• WIRED found Chinese Claude users are routing around Anthropic’s geo-blocks with proxy services, Telegram-sold accounts, and fake IDs.
• Asian AI labs are racing into the gap left by U.S. restrictions on Anthropic’s Mythos, with Sakana and 360 pitching new models as export-control-proof alternatives.

🍪 TOP TREATS TO TRY

• Cursor for iOS helps you launch cloud coding agents, control desktop agents, review diffs, and merge PRs from your phone. Pricing not listed.
• ClinePass gives you discounted, key-free access to GLM, Kimi, DeepSeek, MiMo, and other open coding models inside Cline. $9.99/month.
• Devin Fusion routes coding-agent work between a frontier model and a smaller sidekick model to reduce costs while preserving review quality. Pricing not listed.
• Baz Planner reviews specs, security, reliability, and fixes before code gets written, with Baz extending seed funding to $17M. Pricing not listed.
• Draft captures context from meetings, Slack, and GitHub, then injects it into new agent sessions so teams stop re-explaining projects. Pricing not listed.
• Halo connects iPhone apps like mail, calendar, reminders, music, and health into one private assistant for briefings and automations. Pricing not listed.
• Adrafinil keeps your Mac awake only while AI coding agents are actively working. Free to try.

🏢 Big Tech & Major Companies

• Microsoft put Jacob Andreou in charge of Copilot across consumer and commercial products, with Satya Nadella betting on a more unified AI product strategy.
• Salesforce employees reportedly questioned the company’s promotion of Anthropic’s Claude Tag inside Slack, fearing a rival AI agent could deepen its grip on enterprise workflows.
• Claude became generally available in Microsoft Foundry, giving enterprises another path to Anthropic models on Azure.
• Arena said it reached a $100M annual revenue run rate eight months after launching its real-world AI evaluation product.

🏛️ Policy, Government & Geopolitics

• California made Claude the first AI productivity tool broadly available to state agencies and local governments under a discounted Anthropic deal.
• Anthropic said U.S. restrictions forced it to suspend Fable 5 and Mythos 5 access for foreign nationals, including foreign-national employees.
• OpenAI reportedly slowed GPT-5.6’s rollout while the White House built a cyber review framework.
• Alibaba and Tencent lost lobbying firms after a Pentagon rule tied defense-contract eligibility to whether firms represent companies on the 1260H Chinese military companies list.
• Apple reportedly sought approval to buy memory from CXMT, a Chinese supplier on the Pentagon’s 1260H list but not the Commerce Department Entity List.

💼 AI Productivity, Labor & Economics

• The billable hour is under pressure as AI makes professional work faster and clients push firms toward outcome-based pricing.
• Stanford and ADP found a 16% relative employment decline for early-career workers in the most AI-exposed occupations after controls, concentrated where AI is more likely to automate work.
• Bloomberg reported that AI tools are fueling burnout in parts of Silicon Valley, as faster output raises expectations faster than it reduces work.
• Apollo warned that Mag 7 stocks are starting to underperform as investors question whether massive AI capex will generate enough free cash flow.

🧱 Infrastructure, Chips & Data Centers

• Digital Realty agreed to buy Blackstone-managed funds’ 64% blended equity interest in three fully leased Northern Virginia data centers for $3.5B, valuing the assets at $7.8B.
• Super Micro said it was cooperating after Taiwanese authorities searched its local office as part of an Nvidia chip-smuggling probe.
• Data center backlash has become a local political issue as voters push back on power costs, water use, noise, and unclear community benefits.
• Apple is reportedly looking to Chinese memory supplier CXMT as AI data-center demand squeezes memory supply and raises component prices.

🤖 Agents, Coding & Developer Tools

• Cursor launched a native iOS app for launching always-on cloud agents, controlling desktop agents remotely, reviewing diffs, and merging PRs.
• Cognition introduced Devin Fusion, a hybrid-model harness that uses a frontier “main” agent plus a smaller sidekick agent to cut coding-agent costs.
• Cline launched ClinePass, a flat subscription for selected open coding models with higher rate limits.
• Baz launched Baz Planner, which uses specialized agents to catch bugs and security problems during planning rather than after code is written.
• Retriever AI argued browser agents can get much cheaper when cheap models write executable plans instead of renting frontier models for every step.
• Weave Router drew Hacker News interest for routing agent work across models to cut token costs in long coding sessions.
• OpenKnowledge offered an open-source, AI-native markdown editor and LLM wiki as a possible alternative to Obsidian or Notion for agent-heavy teams.
• Tau launched as an educational Python project for learning how coding agents are built.

🧠 Models, Research & Benchmarks

• SemiAnalysis traced how the Transformer’s attention mechanism evolved from Multi-Head Attention into MQA, GQA, SWA, FlashAttention, MLA, sparse attention, Radix Attention, and Paged Attention.
• Yuzhen Mao introduced Simplified Sparse Attention, which teaches standard Transformers to use gist tokens for long-context compression.
• Meta shared Brain2Qwerty v2, a non-invasive brain-to-text research pipeline for real-time sentence decoding.
• AA-Briefcase launched as a private benchmark for long-horizon agentic knowledge work across spreadsheets, presentations, and memos.
• BINEVAL proposed evaluating LLM outputs through atomic yes/no questions instead of opaque holistic scoring.
• Reasoning Quality Emerges Early showed that reasoning-trace quality can be predicted from the first ~100 reasoning tokens, making SFT data filtering cheaper.
• Skill Neologisms introduced trainable soft tokens that can teach frozen LLMs new skills without weight updates.
• MegaTrain proposed full-precision training of 100B+ parameter LLMs on a single GPU by treating VRAM as a transient cache.

🦾 Robotics, Physical AI & Autonomy

• Nvidia’s CHORD used contact-wrench guidance from human demonstrations to train dexterous robot manipulation policies.
• Amazon Science argued that agentic AI in warehouses and factories needs grounding methods like calibrated uncertainty, numerical precision, and formal verification.
• Waymo and Uber ended their Phoenix robotaxi tie-up after nearly three years, with Waymo folding the cars into its own app while Uber prepares another autonomous-vehicle partner in the city.

🎨 Media, Culture & Consumer AI

• Tidal said it will label and restrict monetization for substantially AI-generated music.
• Madonna criticized AI as the “opposite of making art,” tying the concern to algorithms, streaming numbers, and the loss of creative risk-taking.
• Krea2 drew strong Stable Diffusion community interest from users sharing workflows and saying it was replacing some Ideogram use.
• AI scouting apps are spreading in Brazil, with Footbao and CUJU using phone videos to rank young players who might never reach traditional club tryouts.

⚽ Sports, Games & Niche Use Cases

• Local model NPC engines drew interest from LocalLLaMA users who saw small models, speech-to-text, RAG, and local TTS as a possible future for more responsive RPG characters.
• Brazilian soccer scouting apps promised to widen talent discovery by letting young players submit phone videos instead of relying only on traditional club scouting.

Deep Bench

More details on all that, and additional cool stuff we found.

🏆 TOP 5 NEWS (Around the Horn)

• California struck a deal with Anthropic making Claude the first AI productivity tool available to every state agency, city, and county. Agencies get a 50% discount, free workforce training, technical help from Anthropic, and support designing workflows. Existing uses include Department of Motor Vehicles customer service, Medicaid administration, state cyber defense, document work, and information analysis. The partnership sharply diverges from Washington, which labeled Anthropic a supply-chain risk after a Pentagon contract dispute.
• Apple asked Washington for approval to buy memory chips from China's ChangXin Memory Technologies, or CXMT. The Pentagon blacklisted the company over alleged military ties. The Next Web explained that Apple can legally buy from CXMT today, but wants assurance that Washington will not add it to the stricter Commerce Department Entity List. Memory prices quadrupled over three quarters as manufacturers redirected capacity toward AI data centers. Apple raised selected device prices by $100 to $500.
• Stanford and payroll provider ADP's Canaries Dashboard found employment in highly AI-exposed jobs fell 3.8% yearly for workers ages 22 to 25, while less-exposed jobs grew 2%. The live Stanford dashboard covers 4.6M workers and more than 730 occupations. Mid-career workers ages 31 to 34 fell 1.7%, while ages 35 to 40 grew 2%. Stanford's underlying six-fact study found a 16% relative decline for early-career workers after controlling for company-specific shocks. Losses concentrated in jobs where AI replaces tasks, while pay changed less than employment. The pattern survived removing technology jobs, interest-rate-sensitive work, and remote-work effects.
• Claude became generally available through Microsoft Foundry on Azure. Claude said customers get Opus 4.8 and Haiku 4.5 with existing Azure authentication, billing, and spending credits.
• NVIDIA's CHORD taught robot hands from human demonstrations by matching the forces and twists that move objects, rather than copying hand positions. Its technical report covered 4,739 two-handed simulation tasks. Yan Chang reported 82.12% average success across 1,831 evaluated tasks, 90.77% success for whole-body manipulation, and transfer to real robots. NVIDIA's Video to Data repository converts human videos into searchable action clips, reconstructed scenes, and robot-training data. Its main branch says the final robot-training stage is still coming.

Honorable Mentions

• Taiwanese authorities raided Super Micro's offices while investigating alleged Nvidia-chip smuggling into China through the company's servers. Investigators searched six residences and three affiliated companies, while Super Micro said it was cooperating. The probe follows earlier arrests and the seizure of roughly 50 servers, while Taiwan considers stronger export controls. Super Micro shares fell as much as 9.2%.
• Salesforce employees worried that Anthropic's Claude Tag could compete with Slackbot while building direct relationships with Slack customers. Claude Tag joins selected channels, remembers discussions, flags unanswered threads, and handles coding, analytics, and file search. Slack says its own bot sees everything a customer does, while Claude Tag sees only invited channels. Anthropic's commercial terms say customer data is not used for model training. Salesforce still called Slack an open platform for outside agents, but plans to charge when external agents use Salesforce data through Headless 360. A Digg roundup captured the public concern that Salesforce had welcomed a rival directly into Slack, while builders argued custom bots can already undercut Salesforce's AI upgrades.
• Cognition introduced Devin Fusion, which pairs a top-tier coding agent with a cheaper sidekick. On FrontierCode, a test for correct, mergeable code, Fusion scored 57.6 at $3 per task. Fable 5 alone scored 57.0 at $5.12. Anthony Kroeger explained that parallel agents avoid repeatedly rereading the same context, while the Devin signup offers preview access.
• Tidal's AI policy will label fully machine-generated tracks starting July 15 and block them from royalties or direct-to-fan revenue. Fraudulent uploads and artist impersonations can be blocked or removed. Tidal will start with fully generated tracks to reduce mistakes, add substantially generated music as detection improves, and offer support review. Hacker News readers supported labels but questioned detection accuracy, appeals, mixed human-machine songs, and whether platforms should profit from ineligible tracks.

🍪 TOP TREATS TO TRY

Core picks

• ScreenMind privately analyzes screen changes and audio on your computer with Gemma 4 E2B, a local vision-and-audio model. It adds meaning-based and exact-word search, history chat, meeting summaries, voice notes, analytics, and plain-English or Python automations. ScreenMind integrates with Slack, Notion, Obsidian, webhooks, Claude, Cursor, and VS Code. Redaction, encryption, PIN, incognito, and no telemetry protect sensitive data. Three speed modes range from roughly 12 to 76 seconds on a GTX 1650 and one to four seconds on newer hardware. Its Show HN post says caching reduced inference by up to 40%, but installation and Mac testing remain rough (free and open source).
• ClinePass bundles GLM-5.2, Kimi, DeepSeek, MiniMax, MiMo, and Qwen coding models without separate provider keys, while still supporting your own keys. Cline advertised two to five times standard pay-as-you-go usage across its editor and command-line app. TestingCatalog also highlighted developer-toolkit access ($4.99 for month one, then $9.99 / mo; separate $1.99 command-line signup promotion).
• OpenKnowledge combines plain-text Markdown, Notion-style editing, Claude, Codex, and Cursor integrations, agent search, skills, simultaneous editing, version history, and team syncing. Show HN readers praised lossless visual editing and requested local models, AI comments, easier Obsidian migration, collaboration, and plugins. Criticism focused on Electron, its desktop-app framework, plus slow files, configuration overwrites, and differentiation from Obsidian (free and open source; modified versions must remain open).
• AgentBrush gives coding agents reusable brand colors, fonts, logos, and references through Model Context Protocol, the standard that connects agents to outside tools. It supports Claude Code, Cursor, Codex, Gemini, and Windsurf. Five presets, draft-and-refine generation, image-to-image controls, and a mask editor keep visuals consistent while allowing regional changes. Its Show HN submission drew no substantive feedback (free for 50 tokens, roughly seven to ten realistic images or 50 drafts; paid plans from $6.99 / mo, with high-quality generation at $14.99 / mo).
• Draft keeps Claude Code, Codex, Cursor, and other coding agents aligned with shared product context, priorities, and decisions. It runs locally on macOS and the command line, proposes updates from Granola, Slack, and GitHub, then publishes approved context to your team's repository (free open source that permits reuse).

Cool / niche picks

• Appaca turns chat descriptions into hosted internal dashboards, workflows, customer databases, admin panels, shared notes, and specialized AI coworkers. It handles data, model connections, schedules, integrations, and infrastructure inside one workspace. Founder Ash's Show HN post described Appaca as an early third pivot built for operators rather than software builders. Readers saw small-business potential but questioned reliability, custom software connections, ideal customers, and sales needs (free to start; paid pricing not public).
• Aerial Autonomy Stack combines PX4 and ArduPilot flight software with ROS2, the messaging layer connecting its robot components. YOLO finds objects, LiDAR maps surroundings with lasers, Gazebo simulates flights, and NVIDIA's Jetson mini-computers run the system on real drones. The workflow supports faster-than-real-time simulation, visual mission plans, wind, waves, multiple worlds, Windows, continuous integration, tests connecting real hardware to simulation, and real flights. Its Show HN post says one command recreates the simulation after an approximately 45-minute Ubuntu build (free and open source).

🏢 Big Tech & Major Companies

• Microsoft put 33-year-old Jacob Andreou in charge of more than 11,000 people and its Copilot reset. Only about 4.5% of Microsoft 365's 450M customers pay for Copilot features. Andreou merged consumer and enterprise teams, cut extra product versions, introduced usage-based billing, and pushed shorter development cycles. His team built Copilot Tasks in roughly two months, testing it by autonomously ordering him a cheeseburger. Microsoft's official restructuring note put consumer and business Copilot under one effort spanning the experience, platform, Microsoft 365 apps, and models. Mustafa Suleyman shifted his focus toward Microsoft's frontier-model and superintelligence work.
• Waymo and Uber ended their nearly three-year Phoenix robotaxi partnership in May after hundreds of thousands of rides with just over a dozen dedicated vehicles. Waymo moved the cars into its own app-based fleet, while Uber is preparing a separate local partner. They still work together in Austin and Atlanta, but may compete directly in London. Waymo now operates around 4,000 vehicles and more than 500,000 weekly trips.
• Andrew Curran reported that Anthropic renegotiated its Amazon partnership. The deal reportedly switched from compute-hour pricing to token-based pricing, raising Amazon's costs.
• OpenAI shipped smoother scrolling and position tracking for long Codex threads. Tibo Sottiaux detailed turn previews, broader settings search, fixed zoom alignment, Slack copies that preserve bullets and links, and a Pets panel. OpenAI also teased upgraded shortcuts for July 15.
• Gamma joined ChatGPT, so you can turn reports into executive summaries, build marketing briefs, or create pitch decks inside a chat, then continue editing in Gamma.
• Google started developing an Inbox for Gemini Enterprise. It sorts recommendations from your work context into Needs review, In progress, and Done to help you reach Inbox Zero.

💼 AI Productivity, Labor & Economics

• Apollo reported that the seven largest US technology stocks lost much of their earnings-growth advantage over the rest of the S&P 500. Their valuation premium hit a decade-plus low, while cloud infrastructure consumed more operating cash. Hacker News readers disputed Apollo's timeframe, company groupings, and some data choices. They still agreed that $700B-plus in expected infrastructure spending creates risk if AI returns arrive slowly. Historically, stocks in the prior five years' top 20% produced a ten-year median return 17.8% below the broader market.
• Bloomberg found that Silicon Valley's AI productivity boom is stretching workdays and raising anxiety. One founder runs more than six Claude Code agents constantly, including an agent that watches the others while he sleeps. Wispr AI's CEO reportedly slept at the office for three weeks while working 16-hour days. Career coach Kyle Elliott called 2026 his busiest year as workers prepared for layoffs or escaped burnout. The productivity gains translated into more strategic work, constant supervision, and less mental downtime. AI Weekly's synthesis added that venture investors feel the same speed pressure, but cautioned that the evidence comes mostly from a small group of visible AI-startup founders.
• The Wall Street Journal argued that AI could push lawyers, consultants, and other professionals away from billing for time and toward charging for outcomes. Faster research and drafting weaken the logic of hourly fees when clients can see that the same work now takes less time.
• Boris Cherny proposed five function-agnostic roles for AI-era product teams: prototyper, builder, sweeper, grower, and maintainer. He argued that teams need different mixes as products move from experiments to proven customer demand and scale.

🤖 AI Agents & Infrastructure

• AI SDK 7 launched with reasoning controls, human approval for agent actions, file and skill uploads, durable workflows, terminal interfaces, safe execution environments, monitoring, and Model Context Protocol apps. Model Context Protocol is the standard that connects assistants to outside tools and data.
• Mantis is an open-source, self-hosted traffic manager that gives applications one stable interface for multiple language models. It routes requests across weighted providers, retries failed calls, switches to backups, enforces timeouts and cooldowns, and can reuse exact or meaningfully similar answers. Teams also get a management dashboard, Python developer kit, Amazon usage monitoring, and optional safeguards that mask sensitive data or block policy violations. Its Amazon-focused setup includes automated cloud deployment, load balancing, caching, storage, and identity controls. Creator Riz said in the Show HN post that small teams can deploy it with mantis deploy while retaining control of their infrastructure and data.
• Weave Router sends each Claude Code, Codex, Cursor, or custom-app request to an appropriate model in under 50 milliseconds. It uses a small local classifier and routing rules learned from past agent-task logs, while accounting for the cost of rereading conversation history. Weave claims 40% to 70% savings across Anthropic, OpenAI, Gemini, and downloadable models, with streaming, images, tool use, monitoring, local keys, and hosted or self-managed deployment. In the Hacker News discussion, Weave reported roughly 40% internal savings without slower or worse work. Readers liked independent routing for helper agents, but wanted public evaluations and raised concerns about stale routing data, privacy, ambiguous tasks, and savings lost when a model switch rereads history.
• Retriever AI argued that browser agents become about 100 times cheaper when DeepSeek Flash writes one JavaScript plan and a controlled browser system executes it locally. Its approach replaces 40 to 100 screenshot-and-tool conversations with one planning call, a few targeted extractions, and ordinary code. The browser supplies structured page text, while stable instructions and page content use DeepSeek's roughly $0.0028-per-million-token cached-input rate. Screenshots become a fallback, and the controlled system still decides which clicks require approval. Retriever's larger thesis is that the browser system, permissions, recovery logic, and live signed-in session create the product's advantage, while models become interchangeable planners.
• Sydney Runkle introduced dynamic subagents for LangChain's Deep Agents. The main agent writes a short orchestration script, then a lightweight interpreter runs loops, branches, and parallel tasks. This makes coverage deterministic at large scale and supports conditional, multi-phase work that sequential tool calls handle poorly. The dynamic-subagent docs, programmatic-subagent guide, and interpreter guide explain the setup.
• AA-Briefcase tests AI models on private, multi-week knowledge-work assignments that produce spreadsheets, presentations, and memos. Ethan Mollick's analysis found fast gains among the best models, alongside a clear gap between openly downloadable and closed models.
• Rose Yu argued that physical-world agents need four safeguards: built-in physical laws, calibrated uncertainty, numerical simulators, and external verification. Amazon's approaches cut confidence-calibration error by over 25%, raised physical-science accuracy by 29%, and improved formal-proof performance by 422%.

💻 AI Coding & Developer Tools

• Cursor launched a public-beta iOS app for paid users to launch cloud coding agents, control desktop sessions, and give instructions by voice or annotated screenshot. You can review demos and side-by-side code changes, merge updates from a phone, and follow progress from the lock screen. TestingCatalog noted no current EU availability. The App Store lists plans from $25.99, while Composer 2.5 runs are 75% off through July 5.
• Bash4LLM+ is a single-file Bash wrapper for Groq and other services using OpenAI's common connection format. It streams responses, preserves sessions in newline-delimited JSON, lists available models, and accepts prompts from files, pipes, or JSON without Python or Node. Optional scripts add providers, an expanded session system, and small graphical or browser interfaces. Its security choices include restrictive file permissions, no /tmp, no eval, and never executing model output. It also runs on Termux for Android, but requires several common command-line utilities plus curl and jq. In the Show HN discussion, the creator defended the roughly 5,000-line script as portable, transparent, and easy to copy. Critics argued that its repetitive code makes auditing and maintenance harder.
• Claude Code will run subagents in the background by default, so you can continue talking to the main assistant while delegated work runs. You can still request foreground execution when you need to watch one task closely.
• Tau is a small Python coding agent you can read like a textbook. Alejandro AO built it in three visible layers covering model connections, the agent loop, tools, saved sessions, commands, and a terminal interface.
• Emil Kowalski's skills repository gives coding agents domain-expert design rules, starting with animation reviews. His new animation vocabulary skill teaches precise terms such as morphing, rubber-banding, and layout animation, helping agents produce the motion you intended. The collection is MIT-licensed and installs with npx skills@latest add emilkowalski/skills.
• Adrafinil keeps a closed Mac awake only while a coding agent is actively working, then restores normal sleep after the last task ends. The free, MIT-licensed menu-bar app supports nine tools, including Claude Code, Codex, Cursor, Gemini CLI, Cline, and OpenCode. It handles overlapping sessions, releases abandoned holds, cuts power if heat becomes unsafe, confirms lid-close protection with a chime, and summarizes completed work when you reopen the Mac. The Show HN discussion contrasted this activity-based approach with always-on tools such as Caffeinate and Amphetamine. It requires macOS Tahoe 26.4, administrator access for closed-lid control, and currently builds from source.
• Fernanda Graciolli and Nada Amin argued that AI has made formal verification practical for mainstream software. Formal verification mathematically proves that important rules hold for every valid input, while ordinary tests only sample selected cases. Modern models can translate requirements into precise specifications, draft proof strategies, and revise rejected proofs. A separate verifier still checks every result, leaving humans responsible for choosing the correct rules. The authors recommend starting with money conservation, permission boundaries, state transitions, and other business rules where a rare failure would be expensive.

🔬 AI Research & Models

• SemiAnalysis traced how the Transformer model's attention system evolved from its original multi-head design into newer methods that reduce repeated memory, limit which words get revisited, or reuse earlier calculations. The guide connected MQA, GQA, sliding-window, FlashAttention, latent, sparse, Radix, and Paged Attention to the practical race for longer inputs, faster answers, and lower computing costs.
• Quesma argued that Qwen 3.6 27B is a strong local coding model for Apple and NVIDIA consumer hardware. It built a packaged hexagonal Minesweeper from one prompt and followed instructions better than Qwen's faster version. The guide covers compressed model copies, software for running downloadable models, and OpenCode. Hacker News users praised planning, privacy, offline work, transcription, and new-project performance. They reported loops, nonexistent-tool calls, and cloud fallbacks on large codebases, while several preferred Gemma 4 31B for first-try code quality.
• NanoEuler implements a GPT-2-style language model with roughly 116M adjustable settings entirely in C and CUDA, NVIDIA's low-level graphics-chip programming system. It uses no PyTorch, automatic differentiation, or machine-learning libraries. Its creator hand-wrote forward and backward calculations, checked gradients against finite differences, built a byte-level tokenizer, and added FlashAttention, a faster way to process text during training. The pipeline pretrains on Project Gutenberg and FineWeb-Edu, then teaches the model instruction-following with Alpaca data. The name compares a residual block, which adds a learned update to the current state, with a forward-Euler step in numerical simulation. FlashAttention reportedly tripled training speed on an RTX 4070. The repository positions NanoEuler as an educational artifact whose small training run produces fluent-looking text with little reliable knowledge. Hacker News readers asked for training time, token counts, clearer verification of the CUDA code, and evidence supporting the project's mathematical framing.
• MegaTrain stored the model and its training state in ordinary computer memory, while using the graphics chip as a temporary calculation cache. Grigory Sapunov explained that overlapping data transfers with computation keeps the chip busy. The technical breakdown reported training without reducing numerical precision for models up to 120B settings on one H200 graphics chip with 1.5TB of computer memory. It trained a 14B model 1.84 times faster than a leading CPU-memory method. MegaTrain also trained a 7B model on inputs up to 512,000 tokens, roughly 350,000 words.
• Simplified Sparse Attention taught a standard language model to compress text chunks into learned summary markers, then reopen only the chunks relevant to each question. Yuzhen Mao's thread emphasized that the technique needs no new model architecture. It beat other compression methods on long-document tests, outscored full attention by over 5.7 points on document retrieval, and reached up to 3.37 times faster decoding. The MIT-licensed code includes training data, checkpoints, and optimized graphics-chip kernels.
• Skill Neologisms trained new vocabulary tokens to represent procedural skills while leaving the model's existing settings unchanged. Mihaela van der Schaar explained that independently learned skill tokens can combine with existing and previously unseen skills, without retraining old skills together. Tests showed the skills could combine without extra training on synthetic tasks and the realistic Skill-Mix evaluation.
• Block Diffusion combined models that revise many words in parallel with ordinary left-to-right generation. It supports flexible output lengths, parallel word generation, and reuse of prior calculations, while setting a new record for its model class on language tests. Volodymyr Kuleshov noted that this work underpins NVIDIA's new downloadable diffusion model. Nathan's analysis found that saved calculations remain useful between revision steps because changes stay local and generation naturally resolves left to right.
• Reasoning Quality Emerges Early found that a model can identify hard, useful training examples from roughly the first 100 steps in an answer's reasoning. DAIR.AI explained that this turns expensive full-answer review into cheap early filtering. The method beat existing selection methods by up to 1.7% while processing 91% less text.
• BINEVAL replaced opaque AI-generated quality scores with small yes-or-no questions, then combined those answers into inspectable ratings. Elvis highlighted stronger factual-consistency results, better separation between borderline and flawed answers, and question-level feedback that can directly improve prompts. The approach needs no extra model training.
• Yuchen Jin highlighted DeepSeek's DSpark, a generation method that reportedly increased model output speed by 51% to 400%. DeepSeek also released DeepSpec, the full training framework behind it.

🏛️ AI Policy, Governance & Safety

• US model-access restrictions produced shutdowns, delayed releases, workarounds, and policy backlash. Anthropic said a government directive barred every foreign national, including its own employees, from Fable 5 and Mythos 5. The order arrived at 5:21 p.m. and forced Anthropic to remove both models for everyone. Washington cited an unspecified bypass of safety controls, while Anthropic said the method found only known, minor software flaws that other public models could also identify. WIRED reported that the White House separately asked OpenAI to delay GPT-5.6 and release it first to government-approved customers. OpenAI called the process temporary and warned against making it the default. Asian AI companies launched alternatives, including Sakana AI's Fugu model coordinator and 360's systems for finding software flaws and automating cyber defense. Another WIRED investigation found Chinese users buying verified accounts, fake identities, and access through overseas relay services that forward requests to Claude. Those services can expose prompts to scams or data resale, while their traffic may inflate Singapore's apparent Claude adoption. POLITICO reported that industry lobbyists wanted clear rules but feared retaliation from Washington. The administration later allowed a limited Mythos release to approved partners.
• Bloomberg reported that AI became a major force in the 2026 US elections. Tech billionaires including Marc Andreessen and OpenAI cofounder Greg Brockman poured hundreds of millions into AI-friendly political groups. The Los Angeles Times counted more than $38M spent by four OpenAI- and Anthropic-linked groups through June 15. The groups backed both parties while reflecting a policy split: OpenAI favored federal control, while Anthropic supported stronger state rules. Campaigns also embraced generated attack ads, while voters pushed back on data centers, job losses, and corporate influence.
• Alibaba lost five Washington lobbying firms and Tencent lost four before a new US restriction took effect. The law bars the Defense Department from working with lobbyists who also represent companies blacklisted by the Pentagon for alleged Chinese military support. Firms now must choose between Chinese technology clients and US defense contractors.
• Austria invited Anthropic to establish itself inside the EU, arguing that the company's safety-first values fit Europe. Andrew Curran predicted Anthropic would stay in the US because guaranteed computing capacity is becoming too strategically important to risk.
• Hugging Face CEO Clement Delangue argued that open model weights cannot be meaningfully restricted once released because copies spread across platforms and torrents. Downloadable models give users local control, transparency, and protection from sudden access cuts.

🛠️ AI Tools & Products

• AI scouting apps are spreading through Brazilian soccer, promising to evaluate millions of young players who traditional scouts may never see. Cuju's mobile app scores speed, control, and footwork from standardized drills. Santos partnered with Footbao to analyze uploaded videos and drone footage across boys' and girls' programs while keeping final selection decisions with human scouts. AFP's follow-up said roughly 120,000 players had used Footbao and 160,000 had downloaded Cuju. Footbao helped 18-year-old Leo Veiga reach an Italian academy, while Cuju helped a Brazilian girl earn a place with Corinthians' under-15 team.
• Autofit2 is an open-source pipeline for classifying text from only a few labeled examples across more than 50 languages. From one JSON configuration, it preprocesses data, trains a SetFit model, evaluates results, resumes interrupted runs, creates documentation, and packages the classifier for a standard model server. SetFit learns categories from sentence embeddings, numerical representations of meaning, so it can work with only a few dozen labeled examples and serve predictions efficiently on ordinary processors. The project reports 95% to 99% accuracy on selected tasks, but results depend on the data and category. Reports include benchmarks, self-consistency checks, bias analysis across 50 small language datasets, and estimated training emissions. In the Show HN post, creator Stefan said an earlier version moderated offensive text in more than 20 languages. A reader asked how it differs from Hugging Face's SetFit; the main addition is the complete training, evaluation, documentation, and deployment workflow.
• Clanker TV streams continuous, never-ending generative “robot television” through a traditional linear channel. The 24/7 experiment presents itself as entertainment from the future, with recurring lore, robot K-pop group NEONIX, and connected Botflix radio, records, and film projects. Its Show HN thread offered no technical explanation or substantive feedback, so treat it as an always-on creative demo rather than a documented production platform. No pricing was announced.
• Cross Canon lets you search the public-domain World English Bible by meaning, then returns related passages rather than simple keyword matches. Creator Jackson Stone's Show HN post said “more money more problems” retrieved Ecclesiastes 5:9-13. The initial 4GB search index took about 15 seconds per query, though Stone later reported a sixfold speedup. Readers suggested smaller local search databases and additional Biblical canons, while tests also surfaced occasional strange matches. Protestant and Catholic collections were added, while Orthodox coverage proved harder because its canon varies.
• Hacker Trends charts how often topics, tools, and people appeared across 45M Hacker News posts and comments over 18 years. You can overlay terms, select a date range, and inspect the underlying stories and comments behind each line. The site also includes prepared comparisons covering AI labs, coding tools, frameworks, databases, hardware, and internet culture (free, no pricing details).
• Halo runs a personal assistant on your iPhone, combines mail, calendar, reminders, spending, and other phone data into one briefing, then automates connected apps. Tanmay said its on-device system supports your choice of model provider, Model Context Protocol connections, skills, subagents, and a self-updating wiki-style memory. Halo is taking waitlist signups, with no pricing announced.
• Stable Diffusion community members shared Krea 2 workflows and said the image generator was replacing some of their Ideogram use. The discussion was useful as an adoption signal, but offered no controlled comparison or verified performance data.
• LocalLLaMA discussions explored running more responsive game characters with small downloadable models, speech recognition, private document lookup, and local voice generation. The idea remains an enthusiast prototype pattern rather than a proven game-development standard.

📊 Fundraising & Deals Roundup

• Digital Realty agreed to buy Blackstone funds' blended 64% stake in three fully leased Northern Virginia data centers. The $7.8B asset valuation makes this one of the day's largest AI-infrastructure deals. Digital Realty will pay $3.5B, including $1.2B in cash and $2.3B in stock, with closing expected June 30.
• Arena reached a $100M yearly revenue pace eight months after launching its evaluation product. The former UC Berkeley project now measures long-running agents on completion, errors, tool use, feedback, and recovery across tens of millions of users. Cofounder Anastasios Angelopoulos added that Agent Mode handles long conversations with hundreds of actions, moving Arena beyond its original preference-voting system.
• Baz extended its seed round by $9M, bringing total funding to $17M. Battery Ventures and Boldstart co-led the extension, with AFG Partners and Disruptive VC participating. Baz Planner checks proposed code during planning, before it is saved, using separate agents for requirements, security, reliability, and fixes. A risk matrix blocks unsafe plans, while Baz claims early users cut later reverts and emergency fixes by more than 65%.

💡 Industry Commentary & Analysis

• Fergus Finn traced one million simple additions from CUDA, NVIDIA's graphics-chip programming system, through compilation, driver commands, scheduling, memory access, and execution on an RTX 4090. The walkthrough explains how one high-level instruction becomes hardware-specific operations, how the computer launches work onto the chip, and how groups of threads wait on data. Hacker News readers praised its unusually complete explanation. They recommended lower-level NVIDIA interfaces for clearer, reloadable code and noted that new chip generations keep introducing constraints models have never seen. That makes universal automated optimization unlikely soon, especially when real workloads and driver bugs differ.
• Schamper reverse-engineered Apple's ASIF, a macOS 26 virtual-disk format that stores only used blocks while representing disks approaching 4.5 petabytes. The article maps its chunk allocation, versioned directories, block-status maps, and metadata, then shows a repeatable process for decoding unfamiliar file formats. Hacker News readers discussed ASIF's native macOS mounting, its usefulness for virtual machines, and slower inspection when compressed images are unmounted. They also debated C-style structure definitions in Python. One supplied PyCParser link is truncated and returns 404; the resolved demonstration shows the suggested Python parsing alternative.
• Amber Liu argued that human guidance will eventually stop being AI's main source of new knowledge. Once models can explore simulations, use agents, verify discoveries, and update themselves faster than people can, she expects research to become a machine-native process. Her conclusion was stark: humanity's last contribution may be letting self-evolving AI explore beyond our own limits.
• Chubby described Meta's "distillation trap": replacing Claude Code and Codex with an internal coding assistant requires proving its training and tests contain no rival-model outputs. As companies build on frontier tools, tracing where their models learned each capability becomes harder.
• Steve Yegge defended Beads as a model-independent knowledge graph disguised as an issue tracker. Its Git history records what agents did and why, giving future agents the project context needed to handle work larger than one agent can manage alone.
• David Pan predicted that agent loops will become invisible within a year because tools will ship with them built in. His practical takeaway: advantage comes from learning new techniques quickly, then abandoning them when they become standard.

Previous Around the Horn Digests

Catch up on everything you missed:

• Monday, June 22: Sakana launched its Fugu model router, OpenAI pushed Codex deeper into software security, and AI infrastructure anxiety reached $750B.
• Friday, June 19: OpenAI helped solve 18 rare pediatric cases, Google advanced medical AI, and Amazon aimed Trainium at NVIDIA.
• Thursday, June 18: Midjourney pitched 60-second full-body ultrasound, Noam Shazeer left Google for OpenAI, and Odyssey raised $310M.
• Tuesday, June 16: SpaceX bought Cursor's maker, Anthropic met the White House, and CoreWeave trained DeepSeek-V3 in two minutes.
• Monday, June 15: Anthropic's model-access fight spilled into markets, Salesforce bought Fin for $3.6B, and Qualcomm eyed Tenstorrent.
• Thursday, June 11: OpenAI acquired Ona, Anthropic faced a Claude Fable backlash, and SpaceX priced a record IPO.
• Monday, June 8: Apple rebuilt Siri, OpenAI confidentially filed for an IPO, and NVIDIA expanded its AI factory push.

Monthly skill digests: AI skills guide | Skills video guide

That's a Wrap

That's 73 story clusters and 138 source links from today. If you made it to the bottom, your brain-to-text accuracy is now unofficially higher than most Monday meeting notes.

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 don't have to.

See you tomorrow.

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