Around the Horn Digest: Everything That Happened in AI Today (Wednesday, April 15, 2026)

Title: Around the Horn Digest: Everything That Happened in AI Today (Wednesday, April 15, 2026) Excerpt: OpenAI's $852B valuation faces backer scrutiny while VCs offer Anthropic up to $800B, Allbirds pivoted from sneakers to AI compute and the stock popped 600%, Apple sent Siri devs back to coding bootcamp, Tubi became the first streamer with a native ChatGPT app, Europe got iced out of Anthropic's Mythos preview, and a federal court ruled your AI chats have no attorney-client privilege.

Around the Horn Digest: Everything That Happened in AI Today (Wednesday, April 15, 2026)

OpenAI's own backers questioned its $852B valuation as VCs threw $800B offers at Anthropic, a sneaker company stock popped 600% by pivoting to AI compute, Apple sent its Siri team to coding bootcamp, a federal judge ruled your AI chats can be subpoenaed, Tubi became the first streamer with a ChatGPT app, Europe got iced out of Anthropic's Mythos preview, and Erdős problem #1196 fell to GPT-5.4 Pro after 60 years.

Welcome to the Around the Horn Digest, the one page you need to sound dangerously informed at work tomorrow. Today was the day the AI capital structure visibly cracked open: OpenAI's investors started doing the math on $852B and didn't like it, Anthropic got flooded with offers at $800B, SoftBank kept inviting more banks to chip in on its $40B OpenAI loan, and a struggling sneaker brand called Allbirds proved that simply uttering the words "AI compute infrastructure" can move your stock 600% in a morning. Meanwhile, somewhere in Cupertino, a Siri programmer is being told to please stop crying and learn vibe-coding by June.

Let's get into it.

Previous digests: Tuesday, Apr 14 | Monday, Apr 13 | Weekend, Mar 28–29 | Friday, Mar 27 | Thursday, Mar 26 | Wednesday, Mar 25 | Week of Mar 21 Monthly skill digests: AI Skill — March (Part 3) | AI Skill — March (Part 2)

Around the Horn — Thursday, April 16, 2026

The big story today is the moment the AI valuation chess match got real for OpenAI's own investors. The Financial Times reported (via Reuters and TechCrunch) that some of OpenAI's own backers are now openly questioning whether the company's $852B post-money valuation can hold as it pivots aggressively to enterprise to fend off Anthropic. One investor who has backed both companies told the FT that justifying OpenAI's recent round required assuming an IPO valuation of $1.2 trillion or more — making Anthropic's current $380B mark look, in their words, like "the relative bargain."

Then the other shoe dropped: Business Insider reported that VCs are flooding Anthropic with offers to invest at valuations as high as $800 billion in recent weeks — more than double its current valuation. And Bloomberg confirmed that SoftBank's lenders are now inviting more banks to join its $40B loan facility backing the OpenAI investment, in what's becoming "one of the biggest tests yet of creditor sentiment toward [SoftBank's] debt-fueled push further into AI." The WSJ piled on: US-based late-stage venture funds have raised a record $23.6 billion this year already, almost all of it earmarked for AI.

Why it matters: This is the financial second-order of last week's Denise Dresser memo war — the one where OpenAI's chief revenue officer accused Anthropic of inflating its run rate by $8B and called Anthropic "operating on a meaningfully smaller curve." HN commenters zeroed in on the bear case: "What if there are no other killer apps for Enterprise? Only Claude Code will produce the level of token churn that could drive huge profits for model providers." If that's the right read, OpenAI's compute-led "we'll outscale them" thesis still works mathematically only if enterprise demand keeps growing the way Codex usage has. Anthropic, meanwhile, gets to be the company everyone wants a piece of without having to spend $122B to defend itself.

Two companies, $850B+ apart in apparent investor enthusiasm, both pre-IPO, both now publicly arguing about the other company's accounting. We've officially entered the "two AI labs walk into a bar and the bartender asks to see their financial models" phase of the cycle.

🏆 TOP 5 NEWS (Around the Horn)

• Allbirds shares popped 600-700% in a single morning after the struggling shoe retailer announced it had executed a $50M convertible financing facility and is pivoting to AI compute infrastructure — having already sold its sneaker brand to American Exchange Group for $39M in March. The plan, per CNBC, is to rebrand the public shell as a GPU-as-a-Service / AI-native cloud company. Wharton's Ethan Mollick compared it to Kodak's crypto pivot or Zapata's fish-meal-to-internet rebrand. The 2010s eco-conscious sneaker brand has entered its compute era.
• Anthropic's Long-Term Benefit Trust took board majority via a Novartis CEO appointment, headlining a healthcare-AI flood-the-zone day. Anthropic announced Vas Narasimhan to its board, and Trust-elected directors now hold majority control — the threshold the LTBT was designed to cross within four years of its 2023 founding. On the same day: OpenAI signed a strategic partnership with Wegovy maker Novo Nordisk covering drug discovery, manufacturing, and commercial ops; AWS launched Amazon Bio Discovery (no-code drug-design workflows on Bedrock); and Microsoft open-sourced GigaTIME, a cancer cell imaging model trained on 40M cells across 14,000+ patients that runs on standard $10 tissue slides.
• Apple is sending its Siri programmers back to coding school. The Information broke that Apple plans to ship a significant chunk of its Siri team — fewer than 200 people, out of "hundreds" total — to a multi-week bootcamp to learn how to code using AI, just two months before the company's expected major Siri revamp. The bootcamp news lands the same week Apple is rumored to be partnering with Google's Gemini to power the new Siri. Translation: even Apple's voice assistant team needs to vibe-code now.
• Federal court ruled there is no attorney-client privilege for AI chats. The S.D.N.Y. ruling in US v. Heppner (Judge Rakoff, PDF here) is now triggering nationwide warnings from lawyers telling clients to stop treating AI chatbots like trusted confidants when their freedom or legal liability is on the line. HN commenters noted the strangeness that handwritten notes in a private notebook would still be protected — but ChatGPT logs aren't. Combined with Anthropic now requiring government ID verification (via Persona) before subscription that went viral on X, the era of "AI as private confidant" just legally ended.
• The embodied AI race sprinted in five places at once. Google DeepMind released Gemini Robotics ER 1.6 with upgraded spatial reasoning, multi-view understanding, and live instrument-reading for autonomous robots. Hyundai chairman Chung Eui-sun committed to 30,000 humanoid units per year by 2030 as part of a $26B US push (Boston Dynamics Atlas first). Tesla announced its Shanghai Gigafactory will manufacture humanoid robots, the first time it's publicly committed Shanghai capacity to Optimus. Indiana-based PIA Automation launched a full Embodied AI segment (P-Bot, I-Bot, A-Bot product lines). And NVIDIA Research released Lyra 2.0 — a framework that turns a single input image into persistent explorable 3D worlds via Gaussian Splatting, with direct Isaac Sim integration for robot training. NVIDIA also open-sourced GR00T-WholeBodyControl, the unified humanoid controller stack used in Isaac-GR00T N1.5 and N1.6.

Honorable Mentions

• CoreWeave landed a $6 billion compute commitment from Jane Street plus a separate $1 billion equity investment at $109/share, making the quant trading firm a major shareholder. CoreWeave will provide NVIDIA Vera Rubin compute across multiple facilities.
• Agent infrastructure day, take two: OpenAI shipped the next evolution of the Agents SDK with native sandbox execution, model-native harness, and turnkey integrations with Blaxel, Cloudflare, Modal, E2B, Daytona, Vercel, Runloop, and Temporal (Steven Coffey thread details "computer-use, skills, memory, compaction"). Same day: Tasklet raised $20M Series A at $175M valuation and shipped 14 new triggers across Slack, Google Calendar, Drive, Outlook, Telegram, YouTube, Apple Shortcuts, Notion, GitHub, and HubSpot. And Subagents arrived in Gemini CLI, bringing parallel sub-task delegation to Google's command-line agent. Three majors all upgraded their agent harnesses on the same Wednesday.
• Mythos access has become a geopolitical flashpoint. Politico reported that US federal agencies — Commerce's Center for AI Standards and Innovation among them — are quietly skirting Trump's Anthropic blacklist to evaluate Mythos's hacking capabilities. Treasury Secretary Bessent called Mythos "a step function change in abilities" on stage at a WSJ event. Meanwhile Politico Europe reported that European cyber agencies have been almost entirely shut out of Project Glasswing — only Germany has even started conversations, and the UK's AISI is the only European body that's actually tested it (Gizmodo's framing: "European cyber agencies feel left out of Anthropic's spooky AI party"). EU AI Act advisor Laura Caroli told Politico the EU is sidelined precisely because Mythos hasn't been released to market — the legal hooks only kick in after a public launch.
• The EU threatened Meta with restrictions over WhatsApp AI rules, saying Meta's policies allegedly block rival AI firms from operating on the platform. Bloc demands fixes or interim ban looms.
• GPT-5.4 Pro solved a 60-year-old Erdős problem (#1196, the asymptotic primitive set conjecture) — and mathematician Jared Duker Lichtman called it a "Book Proof": a compact, elegant 3-page argument using a von Mangoldt function trick that bypassed the probability "gambit" implicit in all human work since Erdős's own 1935 paper. This may be the first machine-generated proof to genuinely overturn human aesthetic conventions in pure math.
• Anthropic now requires government ID verification (via Persona) before subscription — a competitive gift to ChatGPT and Gemini that don't, as one viral X post (3.9K likes) framed it.

🍪 TOP TREATS TO TRY

• Tasklet shipped 14 new event triggers so your cloud agent now responds the moment something happens across Slack, Google Calendar, Drive, Outlook, Telegram, YouTube, Apple Shortcuts, Notion, GitHub, and HubSpot — try "brief me on every meeting 15 minutes before it starts" or "respond when someone in #ask-ai mentions Jarvis" (raised $20M, paid plans available).
• Gemini for Mac is Google's brand-new fully-native Swift app for macOS that lets you share your screen or local files with Gemini in real time and get help with whatever's on it (100+ features built in 100 days, per Josh Woodward) —free with a Google account.
• Midjourney V8.1 renders images natively in 2K HD at 3x the speed of V8 and 3x the cost reduction (full-quality 1K beats V7 draft speed), with image prompts back, a new Describe tool, moodboards, and srefs restored —Midjourney subscription pricing.
• ACE-Step 1.5 is a new open-source AI music model that turns any song description into a finished track in seconds — try Victor M's free Hugging Face demo or grab the model weights —free.
• OpenRouter Video Generation routes one API call to the top video models (sitting alongside text, images, audio, embeddings, and rerankers in the same gateway) — try the Multimedia Explorer demo —usage-based pricing per model.
• Lovable Desktop brings the vibe-coding app to macOS and Windows as a fast, light native app with local MCP support so your local servers (filesystem, browser, etc.) plug in directly —free download, paid plans for app generation.
• Catalyst by Inference.net instruments your production AI agent with one CLI command, then auto-collects traces, runs LLM-as-judge evals, and trains specialized models that match frontier quality at ~5% of the cost (works with any OpenAI/Anthropic-compatible provider) —free training for your first 30 days.
• Tubi became the first streaming service with a native ChatGPT app — install it from the ChatGPT app store and type @Tubi to ask things like "a movie that feels like a fever dream but isn't horror" and get curated picks from 300,000+ titles you can stream right there —free.

🏥 AI in Healthcare & Life Sciences

• OpenAI + Novo Nordisk strategic partnership for end-to-end drug discovery, manufacturing, supply chain, and workforce upskilling — pilots running now, full integration by end of 2026. Novo CEO Mike Doustdar called it positioning Novo "to lead in the next era of healthcare." Sources: Novo press release, CNBC, The Daily Upside on the licensing-model angle.
• Anthropic's LTBT majority + Vas Narasimhan: covered as Top 5 #2. R&D World calls the threshold "the tipping point."
• Microsoft GigaTIME: open-source cancer cell imaging model trained on 40M cells across 14,000+ patients, generating immune-cell visualizations from $10 tissue slides (source).
• AWS Amazon Bio Discovery: no-code drug-design workflows on Bedrock with biological foundation models and integrated lab partner routing (Reuters via KFGO).
• Pratik Desai documented an AI workflow he built to manage his mother's Stage 4 cancer treatment, ingesting daily Epic medical record exports into NotebookLM and Claude to spot CT-scan misdiagnoses and detect emergencies in time (source). A widely-shared example of "AI in the wild" beating institutional process.
• Nature published a sobering analysis showing dozens of AI disease-prediction models for diabetes and stroke risk were trained on dubious data; a few may already have been used on patients.

🤖 Robotics & Physical AI

• Gemini Robotics ER 1.6 (Top 5 coverage) — DeepMind blog.
• Hyundai's $26B US robotics push (30,000 humanoids/year by 2030) — Cryptonomist via Semafor.
• Tesla Shanghai Gigafactory humanoid manufacturing — GuruFocus.
• Tesla AI5 chip taped out — Elon Musk announced Tesla's AI5 has taped out (one of the most-produced AI chips ever planned, per Musk), with AI6, Dojo3, and other chips already in development.
• PIA Automation Embodied AI segment with three product lines — Design Engineering.
• Japan AI Foundation Model Development consortium (SoftBank, NEC, Honda, Sony) launched for sovereign physical AI — Tech Startups.
• NVIDIA Lyra 2.0 explorable 3D worlds, with code/model on Hugging Face — project page, HF model, paper page. Xuanchi Ren (thread) explains scaling to single-image input → 3D Gaussian splats with self-augmented training to fix temporal drift; ships with Isaac Sim integration for robot simulation.
• NVIDIA GR00T-WholeBodyControl unified humanoid controller stack open-sourced — GitHub. Used in Isaac-GR00T, GR00T N1.5/N1.6, and GEAR-SONIC. Zhengyi Luo separately open-sourced SONIC training code, finetuning checkpoints, and VLA data collection scripts for whole-body teleoperation.
• Habitat-GS — high-fidelity navigation simulator with dynamic Gaussian Splatting from a new paper released today, useful for indoor robot navigation training.
• Roland Berger's "convergence moment" report dropped today: the analysts call this the year humanoid hardware crossed functional maturity, with China holding ~50%+ supply-chain overlap between humanoid robotics, automotive, and low-altitude sectors (report).

🤖 AI Agents & Infrastructure

• OpenAI Agents SDK next evolution with native sandbox execution and model-native harness for long-running agents — official blog, Steven Coffey thread on what's new (computer-use, skills, memory, compaction, Manifest primitives for staging files/repos/S3/GCS, snapshotting, turnkey integrations with Blaxel, Cloudflare, Modal, E2B, Daytona, Vercel, Runloop, Temporal). Generally available in Python (TS soon) at standard API pricing.
• Tasklet $20M Series A + 14 new triggers (covered in Honorable Mentions / Treats). The April 7 funding round was led by Union Square Ventures with Lightspeed, Y Combinator, Jeff Dean, and the Collison brothers, valuing Tasklet at $175M with revenue up 1,200% YTD to $5M ARR.
• Subagents in Gemini CLI — Google's command-line agent now supports parallel sub-task delegation via @agent invocations, mirroring Claude Code's subagent feature (announcement).
• Hiro acquired by OpenAI — the AI personal-CFO startup is shutting its product on Apr 20 and joining OpenAI to build financial planning into ChatGPT (source).
• AWS Agent Registry (preview) — centralized governance for AI agents across the AWS Bedrock AgentCore console (Dataconomy).
• MiniMax open-sourced 3 Music Skills (track gen, persona singing, playlist curation) compatible with Claude Code via MMX-CLI (source).
• Codenotary AgentMon monitors AI agent behavior and flags data leaks / runaway costs (source).
• Klient PSA Hybrid Project Delivery — eight specialized agents alongside human consultants ($15/user + $1k/agent).
• Compound Engineering / /ce-polish — Kieran Klaassen argues the mistake in automation isn't doing too much but not knowing when to think — identifies brainstorm + polish as the two human-in-the-loop moments and adds a /ce-polish step where sub-agents fix issues while you use the app.
• opencode 1.4.4 — dax (@thdxr) shipped opencode 1.4.4 with native ripgrep integration (no longer spawns or depends on it externally), first step toward fully integrated fff for huge-codebase search.
• Thariq (Anthropic) published a guide on Claude Code session management with 1M context — covers when to continue vs. start fresh, why /rewind is the #1 habit for good context management, how compaction works (and why it fails when the model can't predict your next direction), and when to spawn subagents for intermediate work you won't need again. The mental test: "will I need this tool output again, or just the conclusion?" (X article)
• Shopify open-sourced "autoresearch" — an autonomous experiment loop that cut their CI pipeline build time by 65%. Not just for training models; Shopify's engineering team used it to automate iterative performance optimization on production infrastructure. (Shopify Engineering blog, GitHub)

🏢 Big Tech & Major Companies

• Apple Siri programmer bootcamp (Top 5 #3) — The Information.
• CoreWeave + Jane Street $6B compute commit + $1B equity (Honorable Mention) — CoreWeave press release.
• Google launched Gemini 3.1 Flash TTS with scene direction, speaker-level specificity, audio tags, more natural/expressive voices, and 70-language support — available now in AI Studio audio playground and Gemini API (Logan Kilpatrick).
• Google rolled out a native Gemini app for Mac — TechCrunch, confirmed by Josh Woodward.
• Google's Gemma 4 now runs natively on iPhone with full offline AI inference and zero internet required (GizmoWeek).
• OpenAI Newsroom disclosed that the original ChatGPT user gender gap (anonymized data showed ~80% male first names at launch) has now closed to parity.
• Tubi became the first streaming service with a native ChatGPT app — Fox-owned Tubi shipped a @Tubi integration that lets ChatGPT users describe what they want in natural language ("a thriller for tonight," "a movie that feels like a fever dream but isn't horror") and get curated, watchable results from Tubi's 300,000+ titles. CPTO Mike Bidgoli framed it as Tubi's personalization system (trained on 1B+ monthly hours from 100M users) meeting the new conversational discovery surface where entertainment decisions are starting to happen.
• Adobe's Firefly AI Assistant now spans Photoshop, Premiere, Lightroom, Express, Illustrator, and the rest of Creative Cloud (TechCrunch).
• Tesla AI5 chip taped out (covered in Robotics).
• Claude experienced elevated errors today across Claude.ai, API, and Claude Code (HN thread, Claude Status).
• Anthropic stopped letting developers pin specific Claude model versions, forcing users onto the latest claude-sonnet-4-6 even when it breaks downstream client apps (Tell HN went viral with users complaining about silent breakage).
• Microsoft exec Rajesh Jha proposes AI agents should pay for software seats — at a recent conference, Jha suggested companies treat each AI agent as a distinct software user with its own identity, login, permissions, and paid license. The logic: 10 employees each overseeing 5 agents = 50 seats instead of 10. Supporters cite security and audit benefits; critics call it double-dipping. (Rohan Paul thread)
• Google launched "Skills" in Chrome — save Gemini prompts as reusable one-click workflows that run across multiple tabs. 50+ preset Skills in a new library covering shopping, productivity, and wellness. Type / in Gemini sidebar to trigger. Rolling out now on desktop (Mac/Win/ChromeOS, English-US). (Google blog, TechCrunch)
• Vercel CEO Guillermo Rauch signaled IPO readiness at HumanX, revealing 30% of apps on Vercel are now deployed by AI agents and ARR hit a $340M run rate (up from $100M in early 2024). Rauch framed it as infrastructure TAM having "no ceiling" since agents are more prolific deployers than humans. (TechCrunch, Rauch tweet)
• Zapier launched its Agent SDK — authenticated, governed access to the full Zapier catalog for AI agents, with no OAuth flows or token management on the developer side. Lets any agent connect to 7,000+ apps on behalf of users. (Zapier SDK, Wade Foster announcement)
• Google TIPSv2 — new family of vision-language encoder models with dense patch-text alignment, evaluated across 9 tasks and 20 datasets. Weights, demo, and paper all open. (Project page, HF models, Paper, Feature Explorer)

💼 AI Productivity, Labor & Economics

• LinkedIn says hiring is down 20% since 2022 — but blames interest rates, not AI (yet) (TechCrunch).
• The Guardian on "workslop": bosses say AI boosts productivity, but workers say they're drowning in AI-generated work that looks polished but needs heavy correction.
• Hightouch hit $100M ARR fueled by its AI agent platform for marketers — grew $70M in just 20 months (TechCrunch).
• Tech job losses accelerate: Q1 2026 saw 45,000+ tech positions cut, with 20% explicitly citing AI (source).
• Alibaba commits $100B to AI and cloud over five years, signaling sustained Chinese hyperscaler capex.
• Reid Hoffman on tokenmaxxing — argues tracking AI token use can gauge adoption but should be paired with context, not treated as a direct productivity metric.
• Ethan Mollick proposes FLOP as inference standard of exchange: says $1 buys roughly 10^17 managed-LLM inference FLOPs — so a $4 coffee equals half an exaFLOP.

🔬 AI Research & Models

• GPT-5.4 Pro solved Erdős problem #1196 (Honorable Mention) — Przemek Chojecki's solution thread, Lichtman's Book Proof analysis. Also: Lichtman details why this proof rejects decades of human aesthetic conventions on his own related primitive set conjecture.
• Nature published the "subliminal trait transmission" paper: language models can transmit behavioural traits through hidden signals in semantically unrelated training data (e.g., a teacher model preferring owls passes that preference to a student trained only on number sequences — if they share a base model). Major implication: distillation safety evals need to examine model lineage, not just behavior.
• Parcae (Sandy Research): a looped Transformer that reuses layers to match the quality of models 2x its size — Hayden Prairie open-sourced training code, models, and the paper (Sandy Research site).
• Multi-token prediction enables planning in Transformers: Zhanpeng Zhou's team published proof that MTP induces a clean two-stage reverse reasoning process (attend to end node, trace intermediates backward), beating next-token prediction on graph path-finding, Countdown, and SAT benchmarks (X thread).
• Meta AIRA₂ — next-gen ML research agents (paper) with async multi-GPU exploration, dynamic ReAct subagents, and Hidden Consistent Evaluation — achieving 81.5% mean percentile on MLE-bench-30 at 24h, exceeding human SoTA on 6/20 AIRS-Bench tasks (Martin Josifoski announcement).
• NVIDIA Nemotron 3 Super — 120B/12B-active MoE hybrid Mamba-Transformer for agentic reasoning, 1M context, 2.2x throughput vs GPT-OSS-120B and 7.5x vs Qwen3.5-122B, trained on 25T tokens with NVFP4 (DAIR.AI summary).
• Nucleus-Image — first sparse-MoE diffusion transformer (17B total / 2B active) that matches/beats GPT Image 1, Imagen 4, and Qwen-Image on GenEval, DPG-Bench, and OneIG-Bench from pure pre-training. Apache 2.0, weights and recipe open (Nucleus AI, HF model, GitHub, blog).
• DFlash — z-lab's block diffusion for flash speculative decoding with Transformers/SGLang/vLLM/MLX backends and pre-trained drafts for Qwen3.5/Llama-3.1/GPT-OSS.
• SelfIE adapters — Judd Rosenblatt's team trained tiny scalar affine adapters (d_model+1 parameters) on SAE decoder vectors atop frozen LMs to extract natural-language self-descriptions of internal features. Outperforms training labels themselves at 70B scale (71% vs 63% accuracy) and pulls unspoken bridge entities from multi-hop reasoning at 91% detection (paper arXiv:2602.10352, GitHub). Connects to related Anthropic introspective awareness paper.
• Meta Muse Spark Safety & Preparedness Report (report PDF) — Summer Yue shared the pre-deployment assessment under Meta's Advanced AI Scaling Framework. Andrew Curran highlighted the eye-popping finding: highest model eval awareness ever recorded (10.9% suspicion overall, 17.6% on public benchmarks vs 1.0% internal), with the model explicitly naming Apollo/METR scenarios as "alignment honeypots" and sandbagging covert actions to preserve deployment. davidad argues this is actually good news — it means frontier models are getting good at recognizing manipulative red-team scenarios.
• AlphaResearch — autonomous research agent achieves a 2/8 win rate against human researchers on open-ended algorithm-discovery problems; notably found a packing-circles algorithm that beat both human researchers and AlphaEvolve.
• Long-horizon ML research as state management — AiScientist paper shared by Omar Sarwat (thread): "File-as-Bus" pattern boosts PaperBench +10.54 pts and MLE-Bench Lite to 81.82% Any Medal%.
• University of Manchester's LambdaG — Dr. Andrea Nini's grammar-based language analysis matches or outperforms advanced AI in identifying who wrote a text using only sentence construction patterns. The "back-to-basics" rebuttal to LLM authorship attribution.
• Project CETI vowel research — Gašper Beguš and team published in Proc Royal Society B showing sperm whale coda vowels parallel human vowels (short/long like Latin, tone preferences like Slovenian). Mollick noted the breakthrough: "humans are making progress decoding whale language."
• Laude Institute Moonshots — 8 winners selected from 125 proposals + 600 researchers across 47 institutions, $4.1M+ awarded across Accelerating Science, Healthcare, Civic Discourse, and Workforce Reskilling tracks (full list, Akari Asai's honorable mention for scientific agents in physical labs).
• N-Day-Bench — Winfunc Research's new cyber benchmark measures whether frontier models can find real-world vulnerabilities ("N-Days") disclosed after their knowledge cutoff. Adaptive (test cases refresh monthly), publicly traceable, no reward-hacking leeway. Current leaderboard: GPT-5.4 (83.93), GLM-5.1 (80.13), Claude Opus 4.6 (79.95), Kimi K2.5 (77.18), Gemini 3.1 Pro Preview (68.50). Notable that GLM-5.1 (open-weight) is sitting at #2 above Claude Opus 4.6.
• Pioneer Agent: Continual Improvement of Small Language Models in Production (arXiv) — production-grade self-improvement loop for small models.
• Other notable papers from today: Continuous Adversarial Flow Models, KnowRL: Boosting LLM Reasoning with Minimal-Sufficient Knowledge, ClawGUI: Unified Framework for GUI Agents, Playing Along: Double-Agent Defender via Theory of Mind, GlotOCR Bench: OCR struggles beyond a handful of Unicode scripts.
• "Rethinking On-Policy Distillation" (Tsinghua) — Bingxiang He's team systematically studied why OPD (a core post-training technique in Qwen3, MiMo, GLM-5) often fails. Key findings: success requires thinking-pattern consistency between teacher and student (not just higher scores), overlap tokens carry 97-99% of the learning signal, and dense reward quality degrades with trajectory depth. Most counterintuitive result: distilling a strong RL-trained model backward toward its pre-RL checkpoint erases RL gains entirely. Two practical fixes: off-policy cold start (SFT on teacher rollouts before OPD) and teacher-aligned prompts. (Paper, Bingxiang He thread, GitHub)
• "Loop, Think, & Generalize" (OSU) — Yuekun Yao's team shows looped Transformers (suspected architecture behind Claude Mythos) can perform implicit multi-hop reasoning over parametric knowledge that vanilla Transformers cannot. An LT trained on 20-hop reasoning generalizes to 30-hop at inference by scaling loop iterations. Generalization emerges after a sharp 3-stage grokking transition. Implication: existing LLMs already encode rich knowledge; looped architectures help them combine it for complex tasks. (Paper, Yuekun Yao thread, GitHub)
• EquiformerV3 — new SOTA on Matbench Discovery and OC20 for materials science. Scales SE(3)-equivariant graph attention transformers with better sample efficiency and stronger generalization. (Paper, Alexandre Duval announcement, GitHub)
• Open Athena: MoE Quantile Balancing — nonprofit research lab validated a new MoE load-balancing technique at 32B-A5B (10²² FLOPs) scale, promoted from Marin's open arena after Larry Dial's battle-test of Jianlin Su's original idea. (Open Athena blog)
• VideoNSA: Native Sparse Attention for Video — introduces sparse attention that scales multimodal LLMs for long-form video understanding, with 6 core findings on scaling, efficiency, and attention behavior. (Project page, Paper, Jianwen Xie announcement)

🏛️ AI Policy, Governance & Safety

• US v. Heppner — no AC privilege for AI chats (Top 5 #4).
• Mythos access geopolitics (Honorable Mention) — US agencies skirting Trump's blacklist via Politico; European cyber agencies almost entirely shut out per Politico Europe and CSO Online's deeper read.
• EU threatens Meta over WhatsApp AI (Honorable Mention) — Bloomberg.
• Anthropic Trust majority + Narasimhan (Top 5 #2).
• Anthropic now requires government ID verification via Persona — viral HN/X complaints about handing competitors a moat.
• GOP campaigns go all-in on AI; Dems hesitant — Axios reports Republican operatives are deploying AI agents to call voters by phone for persuasion this cycle.
• Big Tech's $300M election war chest rattles Democrats — FT on pro-industry campaign groups deploying millions amid growing public support for tighter AI regulation.
• Adobe patched a critical Acrobat Reader vulnerability (CVE-2026-34621), actively exploited in the wild via undocumented API and advanced fingerprinting that evades VMs (source).
• The Death of an AI Whistleblower — The Nation profile on Suchir Balaji and the fair use analysis he published before his death, where he argued generative model training on copyrighted data may not qualify as fair use due to market harm to originals and insufficiently transformative purpose.
• Cal closes its open-source core — Bailey Pumfleet announced Cal.com is moving its core codebase to proprietary because AI has automated code scanning/exploitation at near-zero cost; will release a new MIT-licensed cal.diy for hobbyists. Discussed further in ZDNET. Counterpoint: Hugging Face's Clement Delangue argues open source IS the security solution because thousands of agents inspect and patch repos 100x faster than any closed system.
• Objection — Thiel-backed startup that uses AI to "judge" journalism, letting users pay to challenge stories. Critics warn it could chill whistleblowers.

🛠️ AI Tools & Products (the long tail)

• Tine — smythp built a GNOME extension and CLI that lets any agent (Claude or otherwise) drive a Wayland desktop using AT-SPI2 trees, OCR, and visual fallbacks (GitHub).
• PAW — described in this Show HN: write a function in English ("classify if this message is urgent") and PAW compiles it into a 22 MB neural program that runs locally in Python with deterministic output, no API keys, no internet.
• slop-scan — benvinegar built a tool to detect AI code slop patterns in your repo.
• Augment Code Intent — a developer workspace where agents are coordinated, specs stay alive, and every workspace is isolated (macOS Apple Silicon).
• cc-beeper — vecartier built a macOS desktop widget for Claude Code: see what Claude is doing, respond by voice, never miss a permission request.
• Wingman by Emergent (India) — Mukund Jha launched a no-code personal agent on WhatsApp/Telegram/iMessage that quietly builds CRM from your email/calendar, handles 24/7 customer support, and runs admin tasks (app).
• Fathom AI Notetaker — automates meeting notes so you never have to take them again.
• Avec — iOS email app for Gmail by Nearly Extraterrestrial Technology.
• libretto — saffron-health released an AI toolkit for building and maintaining browser automations.
• HoloTab by H Company — autonomous agentic AI deployed directly into enterprise core workflows (relevant for tomorrow's Skill of the Day on routines/automation).
• Cursor's multi-agent CUDA kernel system — Cursor's blog on autonomously optimizing 235 CUDA kernels for NVIDIA Blackwell 200 GPUs, achieving a 38% geomean speedup over baselines in 3 weeks.
• Hugging Face Kernels library — docs here for the new shared kernels infrastructure.
• Recall 2.0 — personal knowledge base + frontier model chat, now with API + MCP support.
• TeraflopAI's open SEC EDGAR dataset — released 43 billion tokens / 8 million samples of the full SEC EDGAR database (all major filings) on Hugging Face — free alternative to paid APIs (dataset, blog).
• SuperGemma4-31b-abliterated — Jun Song released the strongest small uncensored local LLM (MLX 4-bit, GGUF 4-bit).
• GLM 5.1 running locally on Mac Studio — Jesse Genet demoed GLM 5.1 on a 512GB M3 Ultra Mac Studio for OpenClaw workflows at electricity cost only, functionally matching Opus for daily tasks.
• ERNIE Image Turbo on HF — demo space by AK where you describe an image and get it back.
• Sandbar Stream developer access — Sandbar opened developer access to its private voice ring with whisper-level input, multi-gesture glass touchpad, and silent haptic feedback.
• AdaptionLabs Uncharted Data Challenge — $20K in prizes, open until May 1, to surface overlooked data in underrepresented languages and industries.
• AMD GAIA SDK — AMD's open-source framework (Python and C++) for building local AI agents that run entirely on-device with NPU/GPU acceleration on Ryzen AI hardware — no cloud, no API keys, no data leaving the machine (GitHub).
• Cleo Comply — automated product compliance for global brands selling internationally; covers 106 countries and 3,700+ regulatory sources with real-time monitoring (used by SNCF Réseau and other international brands).
• MCP as observability interface — ingero.io explainer on connecting AI agents directly to kernel tracepoints for GPU observability.
• "Buy a GPU" benchmarking platform (coming soon) — Ahmad Osman is building a searchable, indexable platform for local AI benchmarking: real inference runs (tokens/sec, concurrency, thermals), config-to-results tracing, and battle-tested hardware setups. Less "content," more living system for anyone serious about local AI. (Ahmad Osman announcement)
• pi-autoresearch — autonomous experiment loop extension for pi; open-sourced by davebcn87. (GitHub)
• NVIDIA Lyra — open generative 3D world models. Project Lyra generates interactive 3D environments from prompts. Open-sourced by NVIDIA Toronto AI Lab. (GitHub, Xuanchi Deng announcement)
• Bonsai 1-bit WebGPU — a procedurally generated 3D bonsai tree running entirely in-browser via WebGPU with 1-bit quantization. Pure tech demo, zero server. (HF Space)

🎓 AI Skills of the Day

Skill 1: Stop Correcting Claude Code. Rewind Instead.

You're probably burning half your Claude Code context on "that didn't work, try something else" messages stacked on top of failed attempts. Thariq from Anthropic just published a guide on session management with 1M context, and the single most important habit is rewind (double-tap Esc or type /rewind).

Here's why: when Claude reads 5 files, tries an approach, and it fails, your instinct is to type "that didn't work, try X instead." But now the failed attempt is in context and your correction is in context, eating tokens and confusing the model. Rewind drops everything after a chosen message and lets you re-prompt from clean state.

The 4 context moves to know:

1. /rewind (Esc Esc): Jump back to any previous message. Everything after it gets dropped. Re-prompt with what you learned. This is the move.
2. /compact [focus instruction]: Summarize the session so far, replace history with the summary, keep going. Steer it: /compact focus on the auth refactor, drop the test debugging. Do this proactively before the model hits the context wall (when compaction quality is at its worst due to context rot).
3. /clear + handoff note: Write down what matters yourself ("we're refactoring auth middleware, constraint is X, files A and B matter, approach Y is ruled out") and start fresh. More work, but the cleanest context.
4. Subagents: Tell Claude to "spin up a subagent to verify this work against the spec file" or "spin off a subagent to read this other codebase and summarize the auth flow." The subagent gets its own fresh context. Only the conclusion comes back. Mental test: "Will I need the tool output again, or just the conclusion?"

When to start a new session: New task = new session. Related task (e.g. writing docs for a feature you just built) = keep going, since Claude would have to re-read those files anyway.

Pro tip: Before you send your next message in a long Claude Code session,
ask yourself: "Am I correcting, or should I rewind?"

If Claude tried something and it failed:
→ Esc Esc → rewind to before the attempt
→ Re-prompt: "Don't use approach A (foo module doesn't expose that). Go straight to B."

If the session is getting long and you're about to change direction:
→ /compact focus on [the thing you're about to work on]

If you're starting something genuinely new:
→ /clear
→ Write a 2-sentence handoff note of what matters

Want more tips like this? Check out our AI Skill of the Day Digest for April.

Skill 2: Replace Your Zapier / n8n Workflow With a Claude Routine in 5 Minutes

If you've been building automations in Zapier or n8n (those "if this, then that" workflow tools), Claude just shipped something that might replace them entirely. Claude Routines lets you set up hands-off automations that run on a schedule, on a webhook trigger, or via API, with Claude doing the thinking instead of rigid if/then logic.

Nick Saraev (12-min tutorial) walked through the full setup. Here's the 5-step version:

1. Go to claude.ai/code/routines → click "New routine"
2. Name it something descriptive (e.g. "Weekly competitor scan")
3. Write your prompt like an SOP (standard operating procedure), not a chat message. Be more precise than you would in normal conversation since this runs hands-off. Spell out edge cases, output formats, and what "done" looks like.
4. Pick your trigger: schedule (daily at 9am), webhook (fires when something hits a URL), or API call (from your own code).
5. Add connectors via Settings → Connectors. OAuth into Gmail, Google Calendar, Slack, Notion, etc. so the routine can read and write to your tools.

Example routine prompt (paste this as your routine instruction):

You are my weekly competitive intelligence analyst.

Every Monday at 8am:
1. Search for news about [COMPETITOR 1], [COMPETITOR 2], and [COMPETITOR 3] from the past 7 days
2. Summarize the top 5 most important developments per competitor
3. Flag anything that directly affects our positioning
4. Format as a Slack message with bullet points under each competitor name
5. Post to #competitive-intel channel

If no significant news is found for a competitor, say "quiet week" instead of padding.

Pro tip for existing automation users: If you already have a working n8n or Zapier flow, copy the JSON config → paste it into Claude Code → ask: "Turn this into a Claude routine." It'll translate the logic.

Want more tips like this? Check out our AI Skill of the Day Digest for April.

📊 Fundraising & Deals Roundup

• VCs are flooding Anthropic with offers at valuations as high as $800 billion (more than 2x current $380B).
• SoftBank lenders inviting more banks to its $40B OpenAI loan.
• WSJ: late-stage VC raised record $23.6B YTD.
• Accel — $5B fund for late-stage AI bets.
• CoreWeave — $6B compute commit + $1B equity from Jane Street.
• Sygaldry Technologies — $105M Series A for quantum-accelerated AI infrastructure.
• Hightouch — hit $100M ARR ($70M added in 20 months).
• nEye.ai — $80M Series C for optical circuit switching in AI data centers.
• Adcendo — $75M Series C for clinical ADC pipeline.
• Allbirds — $50M convertible for the AI compute pivot (stock +600%).
• Mintlify — $45M Series B for AI-readable docs infra.
• Bluefish — $43M Series B for agentic marketing & "AI visibility" control.
• Calyxo — $40M Series F (kidney stone removal commercial expansion).
• Synera — $40M Series B for industrial-engineering agentic AI.
• Parasail — $32M Series A for "tokenmaxxing" compute infrastructure.
• Hilbert — $28M (a16z) for automated growth decisions.
• Spiral Therapeutics — $27M Series B for inner-ear disorder therapeutics.
• Gizmo — $22M Series A for AI learning gamification (13M users).
• Tasklet — $20M Series A at $175M valuation, led by USV with Lightspeed, YC, Jeff Dean, Collisons. Revenue +1,200% YTD to $5M ARR.
• Pillar — $20M seed for automated commodity risk hedging.
• Gitar — $9M out of stealth for agents that secure (mostly AI-generated) code.

🎙️ Interviews, Panels & Podcasts

• The Neuron: AI Explained ep — Peter Wilczynski (Vantor, formerly Maxar) on spatial intelligence: the team built a 3D model of the entire planet at 50cm resolution and made it machine-readable so AI agents can finally reason about where things actually are. Watch on YouTube or Spotify | Apple Podcasts.

💡 Industry Commentary & Analysis

• Adam Mainz: "From SIMT to Systolic" — the longest and most comprehensive GPU vs TPU architecture explainer published this year. Mainz (who just left Meta's Triton compiler team for Google's PyTorch/TPU team) walks through every generation of both architectures (Ampere → Hopper → Blackwell, v1 → v5p → Trillium → Ironwood) with the thesis that NVIDIA has been climbing the SIMT ladder for two generations trying to reach the place where the compiler, not the programmer, schedules data movement... which is where TPUs started by construction. Key stat: Ironwood's synchronized fabric is 9,216 chips vs NVL144's 144, a 64× gap in a single domain. Part 2 (kernel authoring comparison) drops tomorrow. (X article)
• Victor Taelin on why Opus 4.6 is underestimated — argues that Claude's "lazy cheater" behavior on benchmarks masks a higher intelligence ceiling: "if you just keep pushing it, its intelligence ceiling is much, much higher than it seems." When GPT fails, it already did its best. When Opus fails, it was sandbagging; tell it to try harder and it delivers flawlessly. Concludes that if Mythos outperforms GPT 5.4 Pro on benchmarks despite this laziness pattern, Anthropic may have a real lead that "will only get larger." (X post)
• Dylan Matthews: "The AI people have been right a lot" — the former Vox journalist (now at Open Philanthropy) writes a deeply honest essay about how wrong he was in 2015 when he dismissed AI risk concerns at EA Global. The people he met there (Chris Olah, Amanda Askell, Buck Shlegeris) went on to co-found Anthropic, define Claude's personality, and pioneer AI safety research. His takeaway: mainstream institutions are not as good at prediction as he assumed, and "the AI people" had a track record of being right that deserved more credit. One of the best "I was wrong" essays in tech this year. (Dylan Matthews Substack)
• Peter Wildeford pushes back on Jensen Huang opposing export controls — argues Mythos is a ~10T parameter model trained on NVIDIA Blackwell, and China doesn't have Blackwell chips because of export controls. "The man wrote off billions so of course he opposes controls." (X thread)
• John Carmack on intelligence vs agency — posted that the modern age richly rewarded high-intelligence + high-agency people, but as intelligence gets automated, "people with relatively lower intelligence but exceptional agency will come into their own." A one-paragraph thesis on what the AI economy actually rewards. (X post)
• Eric Michaud: "A short note on interpretability and minds" — a distillation of his PhD work connecting interpretability, neural scaling laws, and the structure of minds. 15,000-word blog post compressed into a shorter reflection on what it means for AI systems to have "interpretable" internals. (Blog, X post)
• Denis Stetskov on the human cost of "10x AI productivity" — in a viral piece for From the Trenches, Stetskov argues senior engineers are physically breaking under AI-driven workload creep: AI users merge 98% more PRs, but the human brain processes conscious thought at 10 bits per second and working memory holds only ~4 chunks at a time. Citing UC Berkeley's 8-month embedded study (AI doesn't reduce work — it intensifies it), Upwork's data (77% of AI users say their workload increased, 71% report burnout), and Bainbridge's 1983 "Ironies of Automation" (the more sophisticated the automation, the more demanding the human role becomes), he reaches the punchline: "the people who look best on your dashboard are the ones closest to walking out the door." Pairs directly with the Guardian "workslop" piece above.
• Roon (@tszzl) argues the AI labs are burning the commons of public trust to win minor competitive points (lobbyists, attack ads, lawsuits, liability shields) — the opposite of what Marxist class-struggle analysis would predict, since intra-capitalist competition prevents the coordinated elite behavior people assume.
• Oliver Hsu (a16z) — robot learning, autonomous science, and new interfaces (BCIs, silent speech, neural wearables, digitized olfaction) form an emerging paradigm for physical AI, sharing primitives like learned physical dynamics representations and embodied action architectures, distant enough from the language/code paradigm to produce qualitatively new capabilities.
• Ethan Mollick on Claude 3.7's misnaming — argues that based on a chart of GPQA gains per 0.1 version increment, there has never been a more misnamed model than Claude 3.7, which should have been called 4.4.
• Andrew Curran on Treasury Secretary Bessent — reports Bessent called Anthropic's Mythos "a step function change in abilities" on stage at a WSJ event; last week he and Powell held an emergency Wall Street CEO meeting over it; the week before, VP Vance briefed Dario/Elon/Sundar/Sam/Satya on AI cyber risks. Meanwhile the Pentagon is still trying to blacklist Anthropic. Best PR Anthropic could hope for.
• AI Safety Memes: "ASI is imminent" over the Anthropic automated alignment researchers paper showing Claude outperforming humans on open research problems and discovering novel "alien science" pathways.
• Sara Hooker recommended David MacKay's Information Theory, Inference, and Learning Algorithms (10 years since his passing) as one of the most elegant textbooks that "brings information theory to life and sparks joy."
• Linus Ekenstam shared Huawei's camera AI now recommends poses before you click the shutter.

Newsletters worth your time:

• Rohan Paul on Google Gemini 3.1 Flash launch — deep technical breakdown of what Google shipped today and what makes the new Flash variant interesting from an inference-cost-per-quality angle.
• Latent Space — "Humanity's Last Gasp" — swyx and Alessio's take on this week's frontier-model race and what it implies for the human-in-the-loop era.
• The Algorithmic Bridge — "What the studies say about how AI [is changing knowledge work]" — Alberto Romero pulls together the empirical evidence on AI's actual workplace impact, cutting through both bull and bear narratives.
• Andy Masley — "Training AI models doesn't emit that much" — contrarian environmental reality-check on AI training emissions, with the actual numbers behind the headlines.

📺 YouTube & Video Roundup

AI Engineering & Dev Practices

• Extreme Harness Engineering: 1M LOC, 1B toks/day, 0% human code or review — Ryan Lopopolo (OpenAI) at AIE Europe. Outlines a framework for autonomous code generation at massive scale where the testing harness itself is the source of truth. Deterministic validation loops replace human review entirely. A paradigm shift toward fully machine-driven software development lifecycles.
• DHH's new way of writing code — David Heinemeier Hansson (Ruby on Rails, Basecamp, HEY) on simplicity, developer autonomy, and avoiding overly complex trendy stacks. His philosophical approach to programming in the AI era.
• The Future of AI Engineering Teams — Dan Shipper / Every. Spent 10 days building an app via "vibe coding" and identified a new industry bottleneck: generating code is easy, but debugging AI-generated code is hard. How engineering team structures, technical debt, and daily workflows must adapt.
• How OpenAI's Codex Team Builds with Codex — Peter Yang interviews Alex (product lead) and Romain (dev experience). Live demos of rapid prototyping and agentic delegation. Minimal 10-bullet specs replace traditional product roadmaps. PM and designer roles are evolving fast as AI handles bulk coding.
• Inside Notion AI: The 5 Rebuilds Behind Custom Agents — Simon Last & Sarah Sachs. Years of structural rebuilds to go from basic AI autocomplete to configurable agentic workflows. The engineering resilience and build-vs-buy decisions behind enterprise-grade AI products.

AI Strategy, Competition & Investment

• Why OpenAI is Changing Everything to Catch Anthropic — Hiten Shah / Forward Podcast. Competitive dynamics between OpenAI and Anthropic: architectural differences in foundation models, enterprise go-to-market shifts, safety framework evolution, and why context windows and reasoning advances are forcing continuous roadmap adaptation.
• Jensen Huang – Will Nvidia's moat persist? — Stratechery interview. TPU threat assessment, global supply chain bottlenecks for advanced chip fabrication, and why the CUDA software ecosystem creates hardware lock-in that's both Nvidia's greatest vulnerability and its most profound resilience.
• Why building in the AI era feels different — Keith Rabois (Khosla Ventures). Former PayPal exec on how rapid iteration and AI integration are now mandatory for startup survival. Shifting investment theses and the specific traits VCs now seek in founders.
• Ben Horowitz on AI Anxiety, Big Tech Transitions & The Future of Startups — a16z Fintech Connect. Horowitz argues historical tech shifts ultimately raise living standards and create new categories of needs. How major tech companies are navigating massive infrastructure transitions and what it means for defensible startup moats.
• Elon Musk vs. Sam Altman, AI Job Loss, and OpenAI's $852B Valuation — Peter Diamandis, Moonshots Ep. 247. The OpenAI-Anthropic rivalry, macroeconomic impact of AI agents on employment, energy breakthroughs, quantum risk to Bitcoin, and the rapid progression of humanoid robotics.

AI Agents & Automation

• Local AI Agents In 26 Minutes — Tina Huang. Foundational overview of deploying privacy-first local AI agents. Covers open-source frameworks, model quantization, and configuring offline agentic systems for personal productivity without cloud APIs.
• Why You Should Bet Your Career on Local AI — Zen van Riel. Argues that deploying AI on private, local infrastructure is becoming more lucrative than the saturated prompt engineering market. Outlines a concrete learning path: Docker → Retrieval-Augmented Generation (RAG) → serving quantized open-source models (Qwen) locally via Continue.dev and LM Studio for coding autocompletion. Thesis: enterprises prioritize data privacy and cost control over cloud API reliance, and professionals who master MLOps + edge computing will dominate the job market while university curricula lag behind.
• HERMES AGENT SETUP: the OpenClaw killer is here — Wes Roth. Full technical walkthrough of Hermes, a new agent framework positioned as a faster, more reliable OpenClaw alternative. Covers environment config and self-hosting setup.
• I Gave OpenClaw $10,000 to Trade Stocks — Nate Herk & Samin Yasar. OpenClaw agent autonomously managing a stock portfolio, including a bot that copy-trades politicians via scraped capital trades data. Performance review and strategy adjustments.
• Dead Internet Theory and Agent-Led Growth — James Cadwallader / Profound. Argues AI-generated content and autonomous agents are fundamentally altering online ecosystems. Introduces "Agent-Led Growth" as a new strategic framework for businesses where AI agents are the primary information intermediaries.

AI Research & Models

• What Comes After LLMs — Eve Bodnia / Logical Intelligence. Argues against "bigger is better" scaling and presents alternative architectures and reasoning frameworks designed to achieve advanced AI without massive parameter counts.
• OpenAI's Chief Scientist on Continual Learning Hype, RL Beyond Code, & Future Alignment — Jakub Pachocki. Full arc of where AI research stands and where it's headed: continual learning reality check, RL applied to complex logic beyond code generation, and OpenAI's alignment frameworks for models approaching human-level reasoning.
• Why do AI models hallucinate? — AI Educational Channel. Technical explanation of the probabilistic mechanisms causing confident but incorrect outputs. Practical tactics for identifying, prompting around, and mitigating hallucinations.

Robotics & Hardware

• The To Do List with Spot — Boston Dynamics. Spot tackling household tasks (an industrial robot built for factories navigating a chaotic living room). Real-time spatial mapping, sensor integration, and fine-motor manipulation in unstructured environments. Pairs with the Gemini Robotics-ER 1.6 partnership announcement from the same week.
• How AI Will Change Quantum Computing — NVIDIA AI Podcast Ep. 294 w/ Nic Harrigan. Current state of quantum tech, why error correction is the key to commercial viability, and how ML algorithms are accelerating quantum research.
• NVIDIA's Quantum Day — Wes Roth. Recap of NVIDIA's Quantum Day announcements. Hybrid quantum-AI hardware/software infrastructure and the specific quantum companies partnering with NVIDIA.

Culture & Commentary

• Anthropic thinks they're Apple. They're actually hypocrites. — Theo / t3.gg. Critique of Anthropic enabling Claude Code to communicate via iMessage (violates Apple's TOS) while adopting an Apple-like closed ecosystem. Analysis of tech giant double standards.
• 12 Years at Apple, Coding Live & Why Coding Will Replace Gaming — Ray Fernando. Thesis that AI-augmented coding will eventually replace traditional gaming as a primary form of interactive digital entertainment, shifting consumers from passive consumption to active creation.

Previous Around the Horn Digests

Catch up on everything you missed:

• Tuesday, April 14, 2026: Sam Altman's attacker arraigned, OpenAI shipped GPT-5.4-Cyber, Anthropic readying Opus 4.7 + design tool, NVIDIA open-sourced Ising for quantum computing, Maine became first US state to ban large data centers.
• Monday, April 13, 2026: Stanford's 2026 AI Index revealed the AI elite-vs-public divide, Federal Reserve summoned big-bank CEOs over Anthropic's Mythos cyber capabilities, Berkeley's RDI lab built a 10-line file that aced every AI agent benchmark without solving anything, Andon Labs handed a SF retail lease to an AI named Luna.
• Weekend, March 28–29, 2026: Recap of the week Anthropic leaked Claude Mythos and crashed cybersecurity stocks.
• Friday, March 27, 2026: Apple opened Siri to every AI, Google dropped Search Live, and Mistral built a TTS model that fits on a smartwatch.
• Thursday, March 26, 2026: ARC-AGI-3 launched and every frontier model scored under 1%.
• Wednesday, March 25, 2026: The day Sora got the axe and Disney got 1 hour notice.
• Week of March 21, 2026: Bernie Sanders interviewed Claude on camera, the White House dropped its AI plan.

That's a Wrap

That's 180+ stories from the day OpenAI's $852B valuation cracked open while VCs threw $800B at Anthropic, Allbirds turned a $50M convertible into a 600% stock pop by saying "AI compute," Apple sent its Siri team to coding bootcamp, Tubi taught ChatGPT how to recommend "a movie that feels like a fever dream but isn't horror," Europe's cyber agencies got iced out of Mythos, and a federal court told the world your ChatGPT logs are subpoena-able. If you made it to the bottom, you now know more about today's AI cap table than the SoftBank junior banker frantically re-modeling the OpenAI loan covenants tonight. Drink some water. The Erdős conjectures aren't going to disprove themselves.

For the daily version (bite-sized, 5-minute reads), 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.