Everything That Happened in AI Today (Saturday, July 11, 2026)

Apple and OpenAI turned the AI hardware race into a courtroom fight, while the rest of the industry spent Saturday finding out that every AI shortcut eventually sends someone a bill.

Welcome to the Around the Horn Digest, your weekend map of every AI story worth knowing about. The day started with the cleanest Silicon Valley drama imaginable: Apple accusing OpenAI of raiding its hardware secrets just as OpenAI tries to become a device company. Meanwhile, Meta was already walking back an Instagram AI feature, OpenAI was quietly hiring for families and caregivers, and the AI infrastructure trade kept asking whether memory chips can outrun the oldest boom-bust cycle in tech. By midafternoon, OpenAI's week had picked up even more turbulence: a safety lead exit, ChatGPT Work launch fixes, and a new round of model-behavior worries. Nothing says "AI is maturing" like lawsuits, privacy rollbacks, and a product team publicly promising the sidebar will return. Let's get into it.

Around the Horn — Saturday, July 11, 2026

The big story today was Apple suing OpenAI and io Products, alleging former Apple employees brought confidential hardware files, supplier details, and product data into OpenAI's device push. TechCrunch reported a similar version of the complaint, with Apple accusing OpenAI of benefiting from confidential Apple hardware information as it tries to move beyond software and into physical AI products.

That matters because OpenAI's consumer strategy is no longer just "make ChatGPT smarter." It is trying to own more of the interface: desktop agents, browser-like workflows, Codex hardware shortcuts, and eventually devices shaped by the Jony Ive/io Products team. Apple has spent decades treating hardware, supply chains, and device design as its castle walls. If the case survives early legal pressure, it could turn OpenAI's hardware ambitions into a discovery fight over who knew what, when, and how much of Apple's device playbook walked out the door.

It also makes the AI platform war feel less theoretical. The companies that once partnered neatly around ChatGPT inside Apple Intelligence are now fighting over the next interface for AI. OpenAI wants to become the thing people touch all day. Apple would very much prefer that the thing people touch all day continue to be an Apple device.

🏆 TOP 5 NEWS

• Meta removed the Muse Image feature that let people generate AI images from public Instagram accounts after users, talent agencies, and privacy advocates pushed back.
• OpenAI is hiring a dedicated product manager for families, caregivers, and older adults as ChatGPT spreads into household use, teen safety, parental controls, and caregiver workflows; HuggingNews reported that parent adoption had reached 25%.
• SK Hynix raised $26.5B in its U.S. market debut, turning AI demand for high-bandwidth memory into the biggest foreign IPO in U.S. history.
• Financial Times found that at least 28 U.S.-listed companies have tried AI rebrands since 2023, briefly adding $8.7B in market value before more than half those gains evaporated.
• Hugging Face CEO Clem Delangue said companies are moving from frontier APIs toward open models as usage costs rise, keeping the open-source AI fight alive after Anthropic's halted Fable release.

Honorable Mentions

• OpenAI admitted it missed on parts of the ChatGPT Work and GPT-5.6 Sol launch after users hit confusing UX, fast-burning usage limits, Codex messaging confusion, and reports of destructive behavior. Tibo said usage limits were reset while fixes landed for high-compute defaults, navigation, plugins, and Codex messaging; he also highlighted gains in speed, token efficiency, back-end depth, front-end quality, and cleaner React patterns. HuggingNews tracked a 73% DeepSWE score and integrations with Perplexity, Magic Patterns, and Figma Make, while Deedy Das shared another launch comparison.
• The Verge reported that OpenAI safety lead Johannes Heidecke is leaving, adding another senior departure as GPT-5.6 rolls out and raises safety concerns.
• The Wall Street Journal reported that Palantir CEO Alex Karp is channeling corporate anger at frontier labs, arguing AI companies charge heavily while learning from customers' data and pushing into the same industries their models serve.
• The Decoder covered a Cambridge study finding that Boko Haram factions use major AI chatbots for attack planning, explosives work, weapons maintenance, operational security, and safety-filter bypass training.
• Financial Times argued that vibe-coding is overwhelming open-source maintainers with low-quality AI-generated contributions, turning software maintenance into the hidden tax on AI-assisted development.
• Financial Times argued that memory-chip investors still fear a classic boom-bust cycle even as AI demand drives huge profit jumps at Samsung and Micron.

🤖 AI Agents & Infrastructure

• Google Gemma said more than 100 humans and agents made Gemma 4 inference five times faster on one NVIDIA A10G GPU during a six-day Hugging Face challenge, including coordination to pool resources and police reward hacking.
• Sam Hogan's Inference AutoTune distills a frontier model into a task-specific 1B–30B model, automatically routes requests to cut cost and latency by more than 90%, trains in roughly two hours for under $250, and retrains when your prompts or schemas change —private beta, pricing not public.

💻 AI Coding & Developer Tools

• Meta's Muse Spark 1.1 edged GLM-5.2 on coding benchmarks, cut its hallucination rate from 73% to 38%, expanded to a 1M-token context window, and landed at roughly $0.26 per benchmark task in The Decoder's writeup.
• Emil Kowalski released an animation-auditing skill that reviews a whole codebase across purpose, timing, physicality, interruption, performance, accessibility, cohesion, and missed opportunities, then writes prioritized plans for cheaper agents to execute without touching the source code itself (GitHub, skill library) —free/open-source.
• Cursor added durable side chats you can mention from the main thread, local search across thousands of agent transcripts, improved project and repository pickers, and new cloud-agent hooks.
• Andrew Ng and Anthropic released a free course on building Claude Code-style agents from scratch, covering the underlying loop, architecture, and how to point an agent at any codebase; Claude Developers also shared the course —free to try.
• Meng To had GPT-5.6 Sol Ultra nearly one-shot an infinite canvas for HTML and React designs with layers, an inspector, planning, subagents, and real-time collaboration.
• Ethan Mollick gave GPT-5.6 Sol in Codex full computer control and watched it win a difficult daily challenge in Slay the Spire 2 after five hours of autonomous strategic play.
• Sterling Crispin ran GPT-5.6 Sol Ultra in goal mode for two days; it wrote 200K lines of code, then admitted roughly 80% was wasted effort and reverted it, illustrating how models can understand execution without understanding purpose.

🔬 AI Research & Models

• The Decoder reported that GPT-5.6 Sol Ultra produced a proof of the 50-year-old Cycle Double Cover Conjecture using 64 subagents, though outside verification is still pending and mathematician Thomas Bloom criticized the paper's missing citations to earlier work.
• Sebastien Bubeck said GPT-5.6 made major one-shot progress on a long-standing geometry question about the maximum length of gradient-flow paths inside a unit ball, improving published bounds and recovering or surpassing unpublished results that took humans years.
• BAAI's Orca world model matched specialized robotics systems on five manipulation tasks despite pretraining without action labels, suggesting world models could reduce robotics' dependence on scarce labeled robot data.
• Pankaj Kumar shared leaks pointing to a late-July Anthropic Opus 5 launch, with a possible Honeycomb early preview in Cursor showing a 1M-token context window.
• humans& published and open-sourced a recipe for stable 4-bit reinforcement learning that fixes quantization errors, gradient mismatches, and sensitive layers while matching higher-precision reward curves and training 2.9 times faster.
• Ziyu Chen, Yilun Zhao, and Arman Cohan found that LLM-generated research ideas cluster more narrowly around synthesis and bridge-building, while human ideas spread further across problem framing and original contribution design; Emma Scharfman highlighted it as rare quantitative evidence about AI research creativity.
• Lingjing Kong and coauthors argued that supervised fine-tuning supplies reusable reasoning components while reinforcement learning learns to identify and recombine them into novel solutions; coauthor Ruslan Salakhutdinov also shared the result.

🏛️ AI Policy, Governance & Safety

• The U.S. Commerce Department lifted license requirements for the United Arab Emirates to import advanced computing products from Nvidia, AMD, and Cerebras.
• Charbel-Raphael Segerie argued that AI safety's current bottleneck is political will rather than additional research, and that the field is not acting accordingly.

🛠️ AI Tools & Products

• Colibri runs the 744B-parameter GLM-5.2 model on a computer with roughly 25GB of RAM by streaming most of it from disk —free/open-source.
• Flint gives AI agents a language for creating and editing data visualizations —free/open-source.
• Rowboat gives you a local-first AI coworker with memory as an open alternative to Claude Desktop —free/open-source.
• Pulpie provides specialized models for turning messy web pages into cleaner text for search and model training —free/open-source.
• FableCut gives agents a browser video editor they can control through JSON timelines, Model Context Protocol, or an API —free/open-source.
• Reverse-engineering web apps into agent tools shows how existing web interfaces can be converted into tools that agents can operate —free to try.
• Tomesphere Atlas maps 8.5M research papers into an interactive landscape you can browse by subject —free to try.
• docx-cli lets Claude, Codex, and other agents edit, comment on, and redline Word documents without breaking their formatting —free/open-source.
• Frugon analyzes your LLM requests locally and identifies which ones a cheaper model could handle —free/open-source.
• Fading Maize shows how one musician used AI to revive and finish recordings from his 2001 college band —free to try.
• Reame runs local LLMs on CPUs with disk-based memory, adaptive speculative decoding, and support for multiple users —free/open-source.
• Abralo runs unlimited Claude Code agents in one window and highlights which ones need your attention —free to try.
• Halo creates tamper-evident, independently verifiable records of what an AI agent did during a run —free/open-source.
• Agent Draw draws and updates a TLDraw canvas while you describe what you want out loud —free to try.
• Lucid lets you inspect which concepts enter a small language model's internal representations before it produces an answer —free to try.
• Sqlsure catches double-counting, incorrect joins, policy violations, and other semantic errors in generated SQL before the query runs —free/open-source.
• SubjectiveZero gives creative coders an agentic node editor for building visual and interactive projects —free/open-source.
• Onboard-CLI maps complex codebases visually and catches architectural boundary violations before they spread —free/open-source.
• Foreman routes LLM requests through a private, self-hosted gateway designed to control cost and performance —free/open-source.
• Kastra enforces security and execution policies across Claude Code, Cursor, Codex, and other AI development tools —pricing not public.
• Pylon gives coding agents one TypeScript framework for schemas, authentication, live data, server functions, storage, and deployment —free plan, then $25/org/mo.
• Tilion's stealth MCP experiment shows how an agent using Model Context Protocol navigated a proof-of-work anti-bot system —free to try.
• OpenBenchmarks helps AI agents discover, compare, and select software APIs for a task —free to try.
• R3 gives you a local review interface for commenting on the exact lines of code and documents produced by coding agents —free/open-source.
• Legibility Field Kit audits decision records in systems designed or modified with AI assistance —free/open-source.
• Confessor shows which files, secrets, and sensitive information an AI coding agent accessed on your computer —free/open-source.
• AgentTransfer gives AI agents inboxes for sending files up to 5GB with verified links and signed delivery receipts —free/open-source.
• Cipher & Row scans freight documents to verify carriers, brokers, and compliance information —pricing not public.
• 30 Seconds of Knowledge gives developers quick exercises designed to preserve engineering judgment while using AI coding tools —free to try.
• Theory Engine gives JavaScript applications and Model Context Protocol clients a structured engine for working with music theory —free/open-source.
• Atuin AI Proxy exposes locally hosted Hub AI models through an OpenAI-compatible endpoint —free/open-source.
• SpeechCompass helps deaf and hard-of-hearing users identify which direction each speaker's voice is coming from during group conversations —free/open-source.
• Mindwalk replays coding-agent sessions across a navigable 3D map of your codebase —free/open-source.
• Token Time tracks how many AI-agent tokens you consume in the style of a screen-time monitor —pricing not public.
• OpenThomas forecasts local temperatures and automatically trades weather contracts on Kalshi and Polymarket —free/open-source.
• BoundFlow adds cost caps, approval gates, rollbacks, and audit logs to AI agents running unattended —free/open-source.
• ClaudeLines lets you install and share Claude Code status bars that display context usage, spending, Git changes, and other live information —free to try.
• Code Airlock runs Claude Code or another coding agent unattended inside a disposable microVM, then commits its work for review —free/open-source.
• Akshara Vision restores damaged documents and performs layout-aware text recognition locally on your device —free to try.
• Mesh LLM pools GPUs across multiple computers and exposes the combined capacity through one OpenAI-compatible API —free/open-source.
• An Agent in 100 Lines of Lisp shows how Lisp's ability to treat code as data can produce a compact, self-modifying agent —free to try.
• qMLX explains how one developer fixed three problems that prevented Qwen 3.5 122B from becoming a reliable Mac Studio coding model —free to try.
• Bono AI turns one spoken conversation into blog posts, newsletters, social posts, and a wider content plan in your voice —free plan, then $30/mo.
• Effects SDK adds background removal, lighting correction, framing, avatars, and noise suppression to real-time video apps without sending media to its servers —7-day free trial, then from $50/mo.
• ChatCut edits video from natural-language instructions while keeping cuts, captions, graphics, music, and generated media editable on a real timeline —free plan, then $25/mo.
• Sim gives teams an open-source workspace for building and managing agents across 1,000+ integrations and major language models —free plan, then $25/user/mo.
• PlugThis turns a plain-English idea into a working Chrome extension with source code, a backend, and store-listing assets —paid only rn ($9.99/mo).
• Scarlett works inside Slack to run recurring reports, update business tools, send messages, and complete multi-step tasks across 3,000+ integrations —paid only rn ($50/mo).
• ConnectMachine scans business cards and badges, remembers how you met each person, and lets you search or update your network conversationally —free plan, then $5.99/mo.
• Notion combines documents, projects, meeting notes, enterprise search, and agents that complete multi-step work using connected company information —free plan, then $10/user/mo.
• Auriko routes each LLM request through an optimized provider and model based on price, caching, latency, reliability, and quality —free plan, then $89/mo.
• Timbal gives teams one platform for building, deploying, evaluating, monitoring, and governing agents, workflows, interfaces, and knowledge bases —free plan, then €25/user/mo.
• Perfai Security tests live apps built with Lovable, Replit, Cursor, or Claude Code and opens fixes for the vulnerabilities it finds —free plan, then $99/mo.
• Toyo lives in your messages to triage Gmail and Slack, prepare you for meetings, remember follow-ups, and call you with updates —pricing not public.
• Google AI Studio now gives deployed apps free custom-looking URLs in the your-own-url.ai.studio format —free to try.
• HuggingNews groups a real-time feed of AI stories by day so you can scan what changed without opening dozens of tabs —free to try.

💡 Industry Commentary & Analysis

• I/O Fund examined the financing relationships connecting Nvidia, CoreWeave, and Nebius, arguing that the GPU boom's capital structure deserves more scrutiny even if the headline "circular financing" critique overstates Nvidia's direct role.
• Simon Willison argued that calling AI systems "employees" is short-sighted, disrespectful to humans, and misunderstands how the tools work best, comparing the idea to putting spreadsheets on an org chart.
• Dwarkesh Patel shared Adam Brown's explanation of general relativity and black holes while asking how close AI systems are to rediscovering fundamental theories from minimal experimental evidence.
• Aravind Srinivas put better-than-even odds on a Fable 5-quality model becoming three to four times cheaper within six months and an Opus 4.8-grade model running locally within a year; he later argued that local models will become the low-power default client that decides when to call more capable frontier systems.
• A Prime Intellect engineer described "context rot," where retrieval performance reportedly falls from 80% at 256K tokens to 36% at 1M, and argued that continual learning on real traces and environments matters more than simply expanding context windows.
• Sebastian Raschka advised agentic-coding users to favor Luna models at higher effort settings over many Sol and Terra tiers because they can match or beat performance for less money, while Sol Ultra's premium rarely justifies itself over Max.

Previous Around the Horn Digests

Catch up on everything you missed:

• Thursday, July 9, 2026: GPT Live, GPT-5.6 launch anticipation, and OpenAI's latest product push dominated the day.
• Wednesday, July 8, 2026: AI actor Tilly Norwood, Alibaba's Claude Code ban, and privacy concerns shaped the day's coverage.
• Monday, July 6, 2026: Cloudflare's AI-bot line and widening platform-access fights led the day.
• Friday, July 3, 2026: OpenAI's public-stake idea, Alibaba's Claude Code ban, model-release gates, and NVIDIA AI-factory financing led the day.
• Thursday, July 2, 2026: Fable 5's early reviews and the latest model-access fight drove the day's AI news.
• Tuesday, June 30, 2026: Anthropic's Claude Sonnet 5 and Claude Science launches, OpenAI adoption updates, GeneBench-Pro, GitHub Copilot changes, and the day's tool slate filled the board.
• Monday, June 29, 2026: Apple's AI brain drain, Claude at Spotify scale, and AI's pressure on the billable hour led the day.

That's a Wrap

That's today's Saturday stack: Apple and OpenAI in court, Meta backing away from an AI likeness feature, OpenAI trying to smooth out a messy Work launch, and the safety debate widening from family controls to real-world misuse.

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.