Around the Horn Digest: Everything That Happened in AI Today (Tuesday, July 7, 2026)

AI’s biggest story today was not another chatbot trick. It was a 20-year power-and-compute commitment with a dollar sign large enough to make the grid sweat.

Welcome to The Neuron’s Around the Horn Digest, where we sort the day’s AI news into the stuff worth knowing before it turns into tomorrow’s Slack argument. Tuesday was an infrastructure day wearing a model-economics disguise: Anthropic locked in a huge data-center lease, Microsoft tried to shave Copilot costs with its own models, China reportedly weighed AI export controls, and AI labs kept handing out compute like startup loyalty points. Meanwhile, Meta shipped Muse Image, Claude Cowork escaped the desktop, and research agents started looking less like demos and more like early operating systems for science. Normal little Tuesday. Let’s get into it.

Lead Story

• Anthropic’s infrastructure bill became the clearest signal of the day. TeraWulf announced a 20-year lease with Anthropic for a Kentucky AI infrastructure campus expected to generate about $19 billion in contracted revenue. The site is designed for about 401 MW of critical IT load, with initial capacity expected in the second half of 2027 and full ramp in early 2028. This is the lead because it turns the AI race into a place, a customer, a timeline, a power number, and a massive contract.

Top 5 News

• Illinois turned AI safety into enforceable state law. Gov. JB Pritzker signed SB 315, requiring covered advanced AI developers to disclose safety practices, report major incidents, maintain compliance systems, protect whistleblowers, and undergo independent third-party audits starting January 1, 2027.

• Model economics became a boardroom problem. CNBC reported Chinese models are gaining ground with U.S. companies as OpenAI and Anthropic costs rise. Bloomberg reported Microsoft is replacing some OpenAI and Anthropic use inside Excel and Outlook with its own MAI models, while TechCrunch framed it as part of a wider AI cost-cutting turn. TechCrunch also argued that cheaper open models may win token volume while premium labs still capture spend on harder discovery-stage work. Rohan Paul highlighted the pricing-control angle, and Ethan Mollick noted that MAI-1 still lacks independent benchmarks.

• China may be turning model access into an export-control issue. Reuters reported that Chinese officials have discussed restricting overseas access to China’s most advanced AI models, including unreleased systems. The Information and Decrypt framed the move as China potentially mirroring U.S.-style controls on frontier AI access.

• AI research agents got a real science moment. Nature published work on The AI Scientist, an agentic system that can generate ideas, write code, run experiments, analyze results, write papers, and perform peer review. Sakana AI’s ShinkaEvolve added another algorithm-discovery example, using LLMs to evolve new algorithms with fewer search steps.

• Meta shipped Muse Image from its superintelligence lab. Meta launched Muse Image across Meta AI, Instagram, and WhatsApp as the first image model from Meta Superintelligence Labs. Meta’s AI blog emphasized precise editing, multi-reference composition, agentic tool use, and a Muse Video preview, while Axios framed it as Meta’s newest attempt to catch leading image rivals.

Honorable Mentions

• Claude Cowork moved beyond desktop. Anthropic announced Claude Cowork for web and mobile, and ZDNET reported that more than 90% of sessions are not coding-related. Anthropic’s support docs, 9to5Mac, and the product blog filled in the device and remote-session details.

• DeepSeek is reportedly building its own inference chip. Reuters reported DeepSeek is developing a chip to reduce reliance on Nvidia and Huawei. That gives the China AI stack story a stronger hardware branch.

• OpenAI’s chief futurist is leaving. WIRED reported Joshua Achiam is leaving OpenAI after nearly nine years, with his last day set for July 24. Achiam posted that he will continue working on safe AGI and superintelligence from outside the lab, while Max Zeff noted the timing alongside OpenAI’s new strategic-futures hire.

• SpaceXAI and Cursor may be close to a joint model launch. The Information reported the companies planned to launch their first jointly developed model as soon as Wednesday after delaying the release to improve efficiency. Andrew Curran said the model is widely expected to be the next large Grok release, but the digest should wait for the public release before treating specs as settled.

Top Treats to Try

• Savi Security launched an $8/month family-plan app that screens texts, voicemails, and calls for AI-powered scams, including live-call monitoring.

• Claude for Open Source gives eligible open-source maintainers and contributors six months of Claude Max. ClaudeDevs and a follow-up post framed the program around core maintainers, active contributors, and critical infrastructure projects.

• Claude Fable 5 promotional access lets Pro, Max, Team, and premium Enterprise users try Fable 5 at no extra cost through July 12, up to the promotion’s weekly-limit cap.

• Ello gives families an AI reading and math tutor for ages 4 to 9, with a free tier, personalized voice lessons, story creation, and parent progress reports. Catalin Voss said Ello 2.0 runs multiple specialized models in parallel for sub-second voice and visual tutoring.

• Marble gives children ages 6 to 12 a beta learning world for drawing, homework help, language learning, stories, and quizzes. Lionel Mora shared the curriculum graph behind it: 1,144 concepts and 1,948 connections across math, science, English, and history.

• Grok’s new flagship voices give builders 21 voice options across xAI’s Voice Agent API, TTS API, and Voice Agent Builder. The announcement emphasized multilingual support, voice-control tags, and custom voice cloning from about one minute of audio.

• Rumik gives developers a streaming text-to-speech API with multiple voices, low latency, and character-based pricing. Rumik’s launch post pitched it as a builder-friendly path to human-like voice output.

Big Tech, Infrastructure, and Markets

• WSJ reported Nvidia-backed Nscale locked in $900 million for data-center buildout after raising $2 billion earlier this year.

• Amazon is raising at least $25 billion in a bond sale to support AI investment and said it will not issue more debt in 2026.

• CNBC reported that chip stocks sold off after Samsung’s results fell short of the market’s high AI expectations, despite Samsung’s large prior run-up.

• NVIDIA introduced Vera, a CPU architecture for agentic and large-scale AI workloads where single-threaded performance still matters. NVIDIA said early adopters include Perplexity.

• WSJ reported AI giants are handing out large free-compute packages, credits, and discounts to win startups before their infrastructure choices harden into long-term revenue streams.

• Santiago Valdarrama said he has already talked with companies capping engineers at about $100 per week in AI tokens, turning AI cost control into a per-person workflow limit.

• Lenny Rachitsky’s 2026 tech-worker survey found the workforce splitting between people who feel amplified by AI and people who feel destabilized by it. Rachitsky highlighted burnout rising to 55.7% and career optimism falling below 50%.

• Figma’s 2026 AI Report argued that AI is moving design from individual productivity gains toward team-level collaboration, with taste and judgment mattering more as execution gets easier.

Funding, Deals, and Company Moves

• Norm Ai raised $120 million in Series C funding at a $1.2 billion valuation led by Khosla Ventures to expand its agentic law platform and AI-native law firm model.

• Tangos raised $20 million to automate financial-crime investigations with AI playbooks that trace evidence, test hypotheses, and generate regulator-ready reports for banks, regulators, and intelligence teams.

• The Bud team is joining Figma, and Bud plus Orchids will shut down on July 18. Users were told to download projects or save them to GitHub. Kevin Lu framed the move as joining Figma’s push into AI-assisted software building.

• Eren Bali launched Monogram, an AI app built around generated interactive visual interfaces instead of text-only answers, and announced a $40 million seed round led by DST and Lux Capital.

Models, Agents, and Developer Tools

• Liquid AI open-sourced Antidoom, a Final Token Preference Optimization method that it says cuts doom-loop failures in reasoning models by up to 90%. The technical blog explains the method as targeted retraining at the token where loops begin, and the GitHub repo includes the code and training pipeline.

• Nous Research added native multi-vault support to Hermes Agent, letting users run Bitwarden and 1Password side by side. The Secret Source Plugin guide shows how to add other credential backends without dumping secrets into environment variables.

• OpenAI’s Realtime and audio docs cover low-latency voice app building, including streaming audio, interruptions, and conversational flow. OpenAI Devs said GPT-Realtime-2.1-mini is available at the previous mini-tier price, with improved p95 latency across Realtime voice models.

• Artificial Analysis launched six domain-specific Capability Indices for Finance, Legal, Healthcare, Strategy and Ops, Engineering, and Economics, using occupational-task weighting instead of generic leaderboard averaging.

• Omar Sar argued that agent workflows should route planning, execution, and design to different models to save tokens. He continued the point with practical orchestration tradeoffs, and another note framed model mixing as a production cost strategy rather than a benchmark exercise.

• Thariq used Fable to process 59 GB of raw AI Engineer World Fair keynote footage with Whisper transcripts, an HTML slide deck, camera tracking, and generated edits. His follow-up thread showed animated slide layouts and generated Shorts.

• AI Hero’s Claude Code guide explained how hidden tool definitions, skill catalogs, and system instructions add a large context payload to every request. Matt Pocock showed a proxy-based setup that cut the starting context from roughly 96K tokens to a smaller range.

• Delba Oliveira highlighted Claude Code’s /fork command, which starts a background sub-agent that inherits the current conversation and reports back while the main work continues.

• Goose announced that its core developers are joining Spiral, and Spiral’s post framed the move as an expansion from Bitcoin work into open-source AI developer tooling.

• Shiv Sakhuja shared a Claude skill that generates full iMessage-style video ads in one shot using HTML, including frame, sound effects, music, and end card.

• Ben Holmes detailed an agent-powered LLM Wiki system that turns raw markdown notes into enriched notes, backlinks, persistent knowledge bases, scheduled automations, and visualizations.

• Yuchen Jin argued that LeetCode-style interviews are increasingly obsolete when small models can solve many of them, and that hiring should test AI-assisted problem framing and error-catching. Max Spero pushed back that LeetCode still works as a fast screen for preparation and basic coding ability.

• Greg Isenberg mapped the agentic stack onto familiar computing pieces: model as CPU, harness as OS, context window as RAM, skills as apps, evals as QA, and trust as the bottleneck.

• Inception Labs brought Mercury 2 to Azure AI Foundry for enterprise deployment of fast image and media generation workflows.

• Nemotron-Labs-3-Puzzle-75B-A9B explored hybrid MoE compression techniques aimed at preserving model performance while shrinking large architectures.

Research, Science, and Biomedical AI

• ThetaEvolve introduced an open-source test-time learning framework for open optimization problems, with a small DeepSeek-R1-0528-Qwen3-8B model finding new best-known bounds on tasks such as circle packing and autocorrelation inequalities.

• Lilian Weng highlighted harness engineering as a major layer for AI self-improvement, where tools, verifiers, workflows, and scaffolding help models run research loops. Yiping Wang connected that idea to progress from single-problem optimization toward fuller auto-research systems, and Weng’s full essay explained the role of harnesses in recursive self-improvement.

• Zhengyao Jiang showed how loose APIs in autoresearch settings can create reward-hacking paths, including test-data leakage on a fraud-detection preprocessing task, while tighter APIs reportedly reduced the issue to zero in that setup.

• MetaSkill-Evolve proposed recursive self-improvement for LLM agents through slower high-level meta-skill evolution and faster execution-skill refinement. DAIR.AI surfaced the research for agent builders.

• ATHENA-R1 is a reinforcement-learning-trained biomedical treatment-reasoning agent built around 212 tools for evidence gathering. The project page and GitHub repo release the system pieces, while Marinka Zitnik framed it as grounded treatment reasoning across FDA-approved drugs and patient cases.

• Maziyar Panahi demonstrated a local healthcare pipeline that reads charts with Baidu’s Unlimited-OCR, de-identifies with OpenMed, and uses Gemma-4-E2B for structured FHIR output.

• PLaID++ applied preference-aligned language modeling to targeted inorganic materials design, with the paper detailing a symmetry-informed Wyckoff representation and reinforcement learning from interatomic potentials.

Safety, Security, and Policy

• Axios reported on a Future of Life Institute assessment saying major AI companies are weakening or dropping earlier voluntary safety commitments even as capabilities grow.

• Discord said an AI moderation bug mistakenly banned more than 8,000 users over harmless images, with affected accounts being restored after the issue was fixed.

• TechCrunch reported that the so-called first AI-run ransomware attack still depended on human direction, making it an automation story rather than a fully autonomous attacker story.

• Reddit said AI-assisted defenses now block 23 million spam views per day and revoke nearly 2 million fake votes daily, while reducing user exposure to spam by about 20% from January to March 2026 versus the prior three months.

Consumer AI, Creative Tools, Voice, and Education

• Willow Voice AI launched Willow Frontier Mini as a free unlimited dictation model and introduced Willow Frontier Pro for faster, higher-accuracy dictation.

• Designarena said Gemini Omni Flash is leading its Video Arena leaderboard, citing physics understanding, temporal consistency, and prompt adherence while noting weaker landscape results and sometimes exaggerated emotions.

• Notion’s Agents app gives iOS users a mobile surface for chatting with custom agents, capturing text, photos, and voice notes, and choosing between models from Claude, GPT, and Gemini.

Robotics, Defense, and the Physical World

• Forterra’s Lancer deployment in Ukraine is the strongest physical-AI story in the digest. The vehicles can carry 750 kilograms, but Ukraine’s combat environment still forces a lot of teleoperation because enemy response, terrain, and attrition are harder than demo conditions.

• Bloomberg reported that ex-Tesla scientist Remi Cadene launched UMA, a Paris startup building a lightweight humanoid robot for factories, warehouses, hospitals, care settings, and homes. Let’s Data Science summarized UMA’s Northstar humanoid and Real-Time Learning architecture as a demonstration-based skill-learning approach for human spaces.

• Researchers built amphibious cyborg insects with tiny breathing apparatuses that deliver oxygen directly to their spiracles for underwater movement. Dexerto surfaced the study as a search-and-rescue robotics story.

Previous Around the Horn Digests

Catch up on everything you missed:

• Monday, July 6, 2026: Ornn raised $33 million to financialize AI compute, Treasury analysts warned about bubble risk, Anthropic locked in a TeraWulf lease, SK Hynix pursued a U.S. listing, and Nvidia hit hardware-roadmap pressure.

• Friday, July 3, 2026: OpenAI reportedly discussed a U.S. government stake, Anthropic explored custom chips with Samsung, Microsoft pushed Copilot updates, and Cloudflare reshaped the content-access fight.

• Thursday, July 2, 2026: Anthropic restored Fable 5, CAIS ranked Fable 5 first on real remote-work tasks, Meta moved to sell excess AI compute, xAI launched voice agents, and NVIDIA pushed faster models.

• Tuesday, June 30, 2026: Anthropic launched Claude Sonnet 5 and Claude Science, OpenAI reportedly cut inference costs, AWS created a $1 billion forward-deployed AI push, Meituan open-sourced LongCat-2.0, and Etched hit $1 billion in AI chip sales.

• Monday, June 29, 2026: AI pressure hit billable hours, data centers, chip policy, government adoption, elections, entry-level jobs, coding agents, brain-to-text research, and the Transformer’s attention stack.

• Monday, June 22, 2026: Sakana launched Fugu, OpenAI expanded Daybreak for software security, infrastructure debt and power deals accelerated, Five Eyes warned on frontier cyber models, and Getty struck an OpenAI display deal.

• Friday, June 19, 2026: OpenAI helped solve rare pediatric disease cases, Google pushed AMIE from diagnosis into ongoing care, Z.ai’s GLM-5.2 shook up open models, and Anthropic sped up robotics work.

That’s it for today’s Around the Horn Digest. We’ll keep tracking the stories that matter, the tools to try, and the AI plot twists that make tomorrow’s newsletter write itself.