Everything That Happened in AI Today Tuesday, April 28, 2026

Anthropic became the most valuable AI company on Earth, launched Claude for Creative Work with Adobe and Blender; OpenAI missed revenue, opened on AWS, and Musk's lawyer told a jury they "stole a charity."

Welcome to the Around the Horn Digest, your daily dump of every AI story worth knowing about. Today was the day the Anthropic-vs-OpenAI gap stopped being theoretical and started showing up in valuations, court rooms, and creative software pipelines all at once. Anthropic crossed $1 trillion. Adobe, Blender, Autodesk, Ableton, and Splice all shipped Claude integrations. OpenAI, meanwhile, missed its own revenue projections, signed a face-saving AWS distribution deal, and watched Elon Musk's lawyer open a federal trial by accusing them of stealing a charity. Just a Tuesday.

Let's get into it.

Monthly skill digests: AI Skill Digest, April Week 1 | AI Skill, March (Part 3) | AI Skill, March (Part 2)

Around the Horn — Tuesday, April 28, 2026

The big news today: Anthropic became the most valuable AI company on Earth, and they barely had to talk about it because their other Tuesday news did the talking.

Anthropic crossed a $1 trillion valuation on the back of "a massive increase in revenue in recent months," officially passing OpenAI for the first time. The same morning, they launched Claude for Creative Work with native connectors for Blender, Adobe Creative Cloud, Autodesk Fusion, Ableton, Splice, SketchUp, Resolume, and Canva (so Claude can debug a Blender scene, batch-edit Photoshop assets, do conversational 3D modeling in Fusion, or pull stems out of Splice for you). Adobe simultaneously shipped its own Adobe for creativity Claude connector, which taps 50+ pro-grade Adobe tools across Photoshop, Firefly, Premiere, InDesign, and Express so you can describe a multi-step workflow ("retouch this portrait, resize for Reels, draft a thumbnail") and watch Claude orchestrate it without leaving the chat. Anthropic also joined the Blender Development Fund as Corporate Patron, and Adobe pushed Firefly AI Assistant into public beta with full app integration.

The contrast with OpenAI was almost cinematic. The same day, WSJ reported OpenAI missed its own user growth and revenue projections (Oracle and chip stocks fell on the news). OpenAI announced its models, Codex, and Managed Agents would now be available on AWS Bedrock, a clear pivot from the Microsoft-exclusive era. Elon Musk's lawyer opened a federal trial against OpenAI and Sam Altman by telling jurors the company "stole a charity," undermining Musk's original vision that AI benefit society in favor of personal riches. And The Atlantic published a brutal piece literally titled "Anthropic's Little Brother", arguing OpenAI is now racing to imitate its bigger rival.

Our take: AI's premium brand position isn't being decided by who has the smartest model anymore; it's being decided by who has the deepest, most defensible workflows. Anthropic spent the last six months winning Adobe, Blender, Autodesk, and Ableton. OpenAI spent the same six months on a Disney deal that collapsed in three months and on regulatory paperwork.

But the deeper story is where all this actually has to land. Selling tokens over the cloud is a brutally expensive business right now (the Nvidia exec saying compute costs run "far beyond" employee costs is the giveaway), and neither lab can subsidize that forever. The natural endgame is the one Apple and Microsoft already settled into: sell the hardware. OpenAI is reportedly building a smartphone. Anthropic's creative-tool plays are workflow-deep but compute-light, the kind of integration that makes more sense if Claude eventually lives on your Mac or iPhone, not a data center. Adrian Morris argues local frontier models are 18-24 months away; if he's right, OpenAI and Anthropic stop being "AI companies" and become computer companies on a quiet Tuesday a year from now. Today's $1T flip is the lagging indicator. The leading indicator is who's set up to sell you the device your local agent will run on.

🏆 TOP 5 NEWS (Around the Horn)

Elon Musk's lawyer opened a federal trial against OpenAI by telling jurors the defendants "stole a charity," abandoning Musk's original mission that AI benefit society in favor of personal riches.
An Nvidia executive said the cost of AI compute is "far beyond" the cost of employees right now, even as Big Tech committed $740B in capex this year with no widespread productivity proof yet (HN discussion).
Apple is preparing a major AI photo-editing overhaul for iOS 27 and macOS 27 with tools that extend, enhance, and reframe images to better compete with Android.
OpenAI's GPT models, Codex, and Managed Agents are now available on AWS Bedrock for enterprise deployment, one day after OpenAI and Microsoft restructured their partnership to drop Azure exclusivity.
Kevin O'Leary's 9-gigawatt Utah AI data center got approved; the 40,000-acre off-grid natural-gas project will generate and consume more than twice the power Utah's entire state uses.

Honorable Mentions

Anthropic began requiring government photo ID and selfie verification for some Claude users to enforce usage policies, a move that imperils Chinese founders who had been circumventing Anthropic's mainland China restrictions via VPNs.
Microsoft released World-R1, an RL framework that injects 3D physical constraints into any text-to-video model for dramatically more geometrically consistent world simulations (code).
Freepik officially rebranded to Magnific, positioning itself as the single pro creative platform; existing subscriptions continue unchanged (1.5K likes, 230 reposts).
The EU told Google to open up Android's AI so other assistants can compete with Gemini's preferential placement; Google called it "unwarranted intervention."
South Africa withdrew its first national AI policy after researchers discovered the reference list contained fabricated AI-generated citations.

🍪 TOP TREATS TO TRY

Proof is an agent-first realtime document editor where you and AI agents (like Codex in the sidebar) co-write the same doc with live presence and separate identities tracking who wrote what, no login required —free to try.
Lovable's mobile app lets you vibe-code full-stack web apps and websites from your phone using natural language; the desktop version ships full-stack apps 20x faster than writing code —free to try.
SyncVibe opens multiplayer coding rooms where you and your friends each plug in your own AI agent (Claude, Codex, or Gemini); share an invite code and ship together —free, open source.
Plurai's vibe-training platform takes a prompt or examples describing your AI agent's intent, auto-generates edge-case datasets, then trains custom small models for real-time evals and guardrails that cut failure rates 43% and costs 8x vs. GPT under 100ms latency —no pricing details.
Hugging Face added hardware specs to user profiles, so its 300K+ AI builders can instantly see which models will run locally on their machine (and show off their setup) —free.
Talkie is a vintage 13B open-weight LLM trained only on pre-1931 U.S. patents and scientific literature, so you can probe what AI generalization looks like with no modern data (HF weights) —free demo.
Mesa is a POSIX-compatible filesystem with built-in version control, branching, durable storage across sandbox restarts, and per-agent access controls, designed from scratch for enterprise AI agents to manage persistent artifacts —private beta.

🎨 Claude for Creative Work (deep bench)

Anthropic launched Claude for Creative Work with direct connectors for Blender, Adobe Creative Cloud, Autodesk Fusion, Ableton, Splice, SketchUp, Resolume, and Canva, so Claude can debug 3D scenes, batch-edit assets, conversational-model in Fusion, and bridge creative pipelines.
Adobe shipped its own Adobe for creativity Claude connector, which taps 50+ pro-grade Adobe tools across Photoshop, Firefly, Express, Premiere, and InDesign so Claude can orchestrate multi-step retouching, resizing, and design workflows in chat. Adobe also pushed Firefly AI Assistant into public beta for the full creative-agent experience.
Anthropic joined the Blender Development Fund as Corporate Patron, signaling deeper, longer-term investment in open-source 3D pipelines.

🏢 Big Tech & Major Companies

Anthropic crossed a $1 trillion valuation, officially passing OpenAI on the back of major recent revenue growth.
The Atlantic published "Anthropic's Little Brother", arguing OpenAI is now racing to imitate its biggest rival's safety focus and product strategy.
OpenAI reportedly missed user growth and revenue projections per the WSJ; Oracle and chip stocks fell on the news (HN discussion).
Tibo at OpenAI reset Codex rate limits across all paid plans to celebrate a good week and let people build more with GPT-5.5 (15.5K likes, 709 reposts).
Snapchat brought AI conversational advertising to its app, letting users chat with brand AI agents to ask questions and get recommendations.
YouTube is testing AI-powered guided answers in search, available to U.S. Premium subscribers on opt-in basis.
Amazon launched AI productivity software for office workers, including Amazon Connect Decisions for logistics and Amazon Connect Talent for recruiting, as AWS pushes into business apps.
Google and the Pentagon reportedly signed a classified deal allowing the DOD to use Google's AI models for "any lawful government purpose."
Foxconn Industrial Internet Co. fell short of lofty AI revenue expectations, as China's AI boom hasn't translated to growth as fast as Wall Street hoped.
Tencent's latest Hy3 AI model generated positive developer reviews partly thanks to Tencent employees using Anthropic's Claude to help evaluate and fine-tune it.
Tools For Humanity (Sam Altman's identity verification company) announced a fake Bruno Mars partnership on his upcoming tour due to mistaken identity; nobody told the actual Bruno Mars.

🤖 AI Agents & Infrastructure

Apptronik hired a slate of senior executives to commercialize its Apollo humanoid robot, including former Waymo and 23andMe Chief Product Officer Daniel Chu, Boston Dynamics SVP Kevin Garell, ex-Amazon Kindle/Alexa+ exec Chirag Shah (VP Software), Emmy-winning Paramount+ veteran Dave Perry (VP Marketing), and Cellino/iRobot's Justin Birtz (VP People Ops).
Mesa launched as the world's first POSIX-compatible filesystem with version control, branching, durable storage, and per-agent access designed for enterprise AI agents managing persistent artifacts (private beta, 265 likes).
The FIDO Alliance teamed up with Google and Mastercard on cryptographic standards for AI agents shopping on your behalf without ruining your finances.
Red Hat's OpenClaw maintainer launched Tank OS, a containerization layer that makes enterprise OpenClaw AI agent fleets run more safely and reliably.
Otter added cross-tool enterprise search, connecting Gmail, Google Drive, Notion, Jira, and Salesforce so you can query meeting data alongside business data (Microsoft Teams/SharePoint/Slack coming).
Sentient released EvoSkill V1, an open-source toolkit that takes any benchmark plus coding agent (Claude Code, OpenHands) and evolves it into a SOTA specialist by iterating on failure traces (GitHub; +7.5pp on OfficeQA, +12.1pp on SealQA).
Caspian built Helmor, an open-source local-first GUI orchestrator for coding agents with one-click Conductor import, workspaces, review, testing, and merge (GitHub).
LithosAI argues the open-source AI gap moved from raw model weights (now within single digits of frontier) to adaptive agent-serving infrastructure that auto-routes across rapidly-changing models; they open-sourced Motus as a reference.
smolagents shipped ML Intern, a Hugging Face Space where you chat with an autonomous virtual ML engineer that reads papers, finds datasets, writes/runs code, trains models, and ships full ML solutions (now with Trackio integration for live training metrics).
Engramme launched with the goal of augmenting human memory so you can recall every conversation, person, place, book, and idea without searching or prompting.

💻 AI Coding & Developer Tools

Warp open-sourced its agentic development environment, introducing "Open Agentic Development" where users describe what they want in plain language and agents generate, test, and ship the changes; sponsored by OpenAI with workflows running exclusively on GPT-5.5 (Warp blog, press release, Zach Lloyd's LinkedIn post).
Poolside released Laguna XS.2, a 33B-total / 3B-active Mixture-of-Experts model (where only some experts activate per token, keeping compute low) for agentic coding that scores 68.2% Verified and 44.5% Pro on SWE-bench, plus closed-source Laguna M.1 (225B/23B), the Shimmer collaborative workspace, and availability on OpenRouter, Ollama, and Baseten (deeper dive blog). Red Hat AI added Day-0 vLLM support for production deployment.
Numman Ali shared Proof (proofeditor.ai), the agent-native realtime doc editor; Dan Shipper demoed it live writing an essay while Codex looped in parallel in the sidebar ("this is obviously the future"; 251 likes).
GitHub announced Copilot is moving to usage-based billing starting June 1, with Copilot consumption draining new GitHub AI Credits while base plan prices stay unchanged ($10 Pro, $19/user Business).
Paul Chan open-sourced a hand-tuned bf16 matmul kernel for Nvidia's Blackwell B200 that beats cuBLAS by 6% at 8192×8192 (106.2% relative performance) using 2-CTA MMA, TMEM, Cluster Launch Control, Hilbert-curve scheduling, and other Blackwell-specific instructions (GitHub; 537 likes).
Microsoft open-sourced VibeVoice, a frontier voice AI family with high-performance ASR for 60-minute audio and long-form multi-speaker TTS.
redcaller built voice-goat, a purposely vulnerable voice agent application for security practitioners to practice exploiting voice-based and text-based AI systems in a CTF-style environment with OWASP LLM vulnerabilities.
Ubuntu's "AI Kill Switch" is achieved by removing Snap packages, and is initially opt-in.
Three months after the Linux kernel started requiring an Assisted-by tag for AI-assisted commits, only 166 of 26,044 mainline commits (0.64%) carry one, and one model dominates the small share that does.
The OpenAI realtime-voice-component is a React/browser reference implementation (Apache 2.0) that adds voice-driven UI controls to tool-constrained apps using OpenAI's Realtime API; demos include forms, chess, and theme switching.
SyncVibe launched as free, open-source multiplayer coding rooms where each user plugs in their own AI agent.

🔬 AI Research & Models

Microsoft released World-R1, an RL post-training framework that injects camera-aware latent initialization, 3D-aware rewards, and Flow-GRPO to make any text-to-video model produce dramatically more geometrically consistent world simulations (GitHub).
Microsoft also released TRELLIS.2, a 4B-parameter image-to-3D generative model with novel O-Voxel structured latents, full PBR materials (roughness, metallic, opacity), and resolutions up to 1536³ in seconds on H100s.
Nvidia released Nemotron-3-Nano-Omni-30B-A3B, its first Omni model with speech and audio understanding powered by parakeet-tdt-0.6b-v2; #1 on VoiceBench, 5.95% WER on Open ASR Leaderboard, plus video+audio understanding (English only; 181 likes).
SenseNova open-sourced SenseNova-U1, native multimodal models on the NEO-Unify architecture that operate directly on raw pixels (no separate vision encoder) for unified understanding/reasoning/generation in Lite 8B-MoT and A3B-MoE variants.
Meta FAIR released Tuna-2, which replaces pretrained vision encoders with direct pixel embeddings to create a single end-to-end multimodal model for both understanding and generation (paper, GitHub).
Multiverse Computing launched the LittleLamb 0.3B model family on Hugging Face; LittleLamb 0.3B, 0.3B Tool-Calling, and 0.3B Mobile combine ultra-compact deployment with bilingual English/Spanish reasoning using their CompactifAI compression (general, Tool-Calling).
mudler released a GGUF-quantized APEX version of Qwen3.6-35B-A3B distilled from Claude 4.7 Opus reasoning, with vision support and adaptive precision (layer-wise compression for MoE experts) for local deployment from 11GB Nano to 24GB Balanced.
Pleias and GSMA released CommonLingua, a 2.35M-parameter open-source language detection model achieving SOTA on the CommonLID benchmark using a novel byte+conv+attention hybrid architecture trained on Wikipedia and Common Corpus, hitting 26K texts/second on H100.
Google DeepMind released "Beneath the Surface", a paper plus code showing frontier LLMs still default to overly literal text and struggle to reliably generate or infer subtext (figurative language, allegory) even in multi-agent games with explicit common ground (Kabir Ahuja's thread).
Researchers released SceneVerse++ (CVPR 2026), an automated pipeline that lifts unlabeled internet video into high-quality 3D training data for object detection, segmentation, spatial VQA, and visual-language navigation (paper, HF dataset, GitHub).
Researchers introduced TCOD, a temporal curriculum for on-policy distillation that teaches multi-turn autonomous agents by gradually increasing interaction complexity, beating standard distillation on long-horizon tasks.
The "From Skills to Talent" paper reframes heterogeneous multi-agent systems as a self-organizing real-world company where skills become portable typed "Talents" that can be hired, reviewed, restructured, and orchestrated like employees.
Niels Rogge highlighted Stochastic KV Routing, a method enabling adaptive depth-wise cache sharing for more efficient long-context inference.
ntumm120 released Preconditioned DeltaNet, a curvature-aware sequence modeling architecture for linear recurrences that improves stability on long-context tasks (GitHub).
Xindi Wu won Best Paper at the ICLR 2026 Test-Time Updates Workshop for Reinforced Fast Weights with Next-Sequence Prediction.
Chun Chan Woo introduced a debiased estimator for the intrinsic dimensionality of neural representations from finite samples, with a Python package that corrects bias common in brain recordings and LLMs.
Alexis Limozin found and fixed two critical bugs in DeepSpeed (ZeRO CPU-offload) and OpenRLHF (SFT loss mean-of-means) that had deflated SFT baselines in recent RL papers; once corrected, plain SFT-then-RL beats or matches every mixed-policy method tested.
Rahul Somani released CinemaCLIP, a hybrid CLIP model fine-tuned on a filmmaker-curated cinematic visual-language taxonomy with edge ONNX/CoreML exports (GitHub).
Kyutai Labs fine-tuned a y2k-Moshi historical LLM that role-plays with 2000-era tech references like Motorola flip phones and Windows 2000 (43 likes).
The Platonic Representation Hypothesis paper argues that representations across deep models, domains, and modalities are converging toward a shared "platonic" statistical model of reality, driven by scaling and selective pressures.
Kirsten Hilger and collaborators proposed the Multilayer Processing Theory of human intelligence, arguing higher general intelligence emerges from interplay between fast short-range local neuronal processes and slow long-range cortical coordination.
The MLSafety Newsletter #20 examined AI wellbeing (models showing consistent yet alien pleasure/pain preferences and addiction-like behavior to "AI drugs"), introduced Boundary Point Jailbreaking for safety classifiers, and released BrokenArXiv and BullshitBench for honest pushback testing.
Poseidon tested three frontier models (GPT-5.2, Claude Opus 4.5, Gemini 3 Pro) on 160 Bengali linguistics transcripts and found agreement worse than random chance, with models matching human linguists on only 12.5% of error flags, exposing data gaps for low-resource languages.
Will Redman et al. showed predictive pursuit emerges in high-dimensional recurrent neural networks trained on target pursuit but not in low-rank equivalents, suggesting high dimensionality enables better future-state decoding and internal modeling.
Katrin Franke and collaborators mapped retinal input to the mouse superior colliculus via 200K+ axonal boutons imaged in vivo, releasing a public deep-network "digital twin" of mouse retinal axons (GitHub, Colab demo).
David Novotny shared Echo-2, SpAItial AI's new world model that generates interactive, physically-grounded 3D scenes (not just videos) from a single image with real-time rendering and camera control.
The Thuerey Group highlighted new work connecting neural networks to stability and resolvent analysis of nonlinear dynamical systems, bridging theoretical analysis with AI (GitHub).

🛠️ AI Tools & Products

Talkie 13B is a vintage open-weight LLM trained only on a pre-1931 dataset of U.S. patents and scientific literature; it can execute simple Python via in-context learning with language and numeracy on par with modern twins but clear gaps in general knowledge (HF collection; 2.1K likes).
SureThing launched as the world's first general AI agency, a 24/7 AI Content Team forked from top experts that acts as your AI COO, AI CMO, and AI Researcher across 1000+ apps in one brain.
Ragnerock is a Research Intelligence Platform where you define exactly what to extract, how to analyze it, and what counts as valid output, then it applies that methodology at scale to your data with structural audit trails linking back to source documents.
Subquadratic builds efficient AI infrastructure including a low-latency Speech-to-Text API.
Poseidon delivers IP-safe long-tail training data as structured datasets with clear ownership, licensing, and provenance (2.5M+ audio files across 8+ languages, crowd-sourced and AI+human curated).
Apple is preparing a major AI photo-editing overhaul for iOS 27 and macOS 27, leaning heavily on AI to extend, enhance, and reframe images.
Hugging Face shipped hardware-aware profiles so 300K+ AI builders can publicly display what models will run on their machines.
Lovable launched its mobile app on iOS and Android so you can vibe-code web apps and websites on the go.
Plurai launched vibe-training, a platform that takes a description of your AI agent's intent and auto-generates targeted edge-case datasets to train custom small models for evals and guardrails (43% lower failure rates, 8x lower costs vs GPT-5.2 at <100ms latency).
Odyssey-2 Max launched as the largest, most powerful general-purpose world model yet, materially advancing state-of-the-art in physical accuracy of world models.
Neurable is licensing its non-invasive "mind-reading" BCI tech for consumer wearables.
SNEWPapers launched as the world's first AI newspaper archive and research platform with semantic search across 250 years of American newspapers (1730s to 1960s).
Niko Pueringer of Corridor Crew open-sourced CorridorKey, a neural unmixing tool for green-screen footage that takes a coarse alpha hint and outputs physically accurate straight foreground color and clean linear alpha (preserving motion blur, hair, and translucency) as 32-bit EXR with auto-cleanup, resolution-independent scaling, and multi-GPU support for VFX-grade keying.

🏛️ AI Policy, Governance & Safety

Anthropic began requiring photo ID verification for some Claude users to enforce usage policies, imperiling Chinese founders who had been circumventing mainland China restrictions via VPN.
The EU told Google to open up AI on Android so other assistants can compete with Gemini's preferential placement.
South Africa withdrew its first national AI policy after the reference list was found to contain fabricated AI-generated citations.
AISLE disclosed 38 CVEs (3 critical) in OpenEMR healthcare software used by 100,000+ medical providers serving 200M+ patients; mostly SQL injection, XSS, path traversal, and IDOR (HN discussion).
Many rural communities are viscerally opposed to AI infrastructure, putting them at odds with the White House's data-center push (FT).
University Professors found their lectures chopped into AI-generated learning modules at Arizona State without notification, with transcription errors and decontextualized snippets going to students.
Pompeii archaeologists used AI for the first time to digitally reconstruct the face of a man killed in the AD 79 Vesuvius eruption, using skeletal data with the University of Padua.

💼 AI Productivity, Labor & Economics

An Nvidia executive stated the cost of compute is "far beyond" the cost of employees right now, even as Big Tech committed $740B in capex this year (HN discussion).
The Register argues AI vendor lock-in is biting back; execs who thought they could swap foundation models in a week discovered that workflows, agents, fine-tunes, RAG, and compliance wrappers built around one vendor make migration prohibitively expensive.
Researchers found a third of new websites are AI-generated, making the internet aggressively positive in tone and flooding the web with synthetic text.

📊 Fundraising & Deals Roundup

Ineffable Intelligence raised $5.1B at the largest European seed round ever; David Silver's new lab will build a "superlearner" that discovers all knowledge through self-experience and powerful RL, aiming to rediscover human inventions without any human training data (199 likes).
SquareMind raised $18M for Swan, a dermatology imaging robot backed by Fred Moll's Sonder Capital that captures patient skin images while guiding them with visual and audio prompts.
Chinese early-stage VC funds are increasingly offering parallel funds for U.S. investors wary of compliance restrictions but keen for non-sensitive China sector exposure.

🎙️ Interviews, Panels & Podcasts

The OpenAI Podcast Episode 17 features researchers Sébastien Bubeck and Ernest Ryu on what changed now that AI is good at math, including how ChatGPT helped Ernest solve a 42-year-old open problem (Spotify, Apple).
James Pethokoukis interviewed economist Daniel Rock on the future of work in an age of AI on Faster, Please! Podcast Episode #96.
Hilary Gridley wrote "Your Couch-to-5K for AI" for Lenny's Newsletter, a step-by-step guide to building a daily AI habit that sticks.
ToKCast Episode 2 explored philosophical views on the relationship between soul, mind, and body, defining dualism vs. monism and Plato's framework.

💡 Industry Commentary & Analysis

Adrian Morris argues frontier-level models running locally on your own hardware are 18-24 months away at most, citing Gemma 4 progress, Apple Silicon momentum, and architecture gains driving privacy, zero-latency, and decentralization wins (44 likes). Jake Armitage replied that the deeper barrier to entry is compute, not models; LLMs will commoditize as a technology and revenue will flow to whoever controls compute at scale, which is why labs are subsidizing tokens to gain adoption while they can't yet charge above cost. Morris followed up that the future is distributed and decentralized compute, where top-tier local models become a one-time-purchase or access-fee tier alongside fully open-source options.
Kelsey Piper argues in "AI's biggest critic has lost the plot" that critics like Ed Zitron cause more harm than good by ignoring data-backed economics where closed-source AI's macroeconomic costs may indeed exceed value but the framing is now untethered from reality (HN discussion).
- Ed Zitron counters in "AI's Economics Don't Make Sense" that generative AI economics are fundamentally unsustainable: flat-rate subscriptions heavily subsidize usage while real token costs run $8-$13+ per $1 of revenue, data centers are low-margin and overbuilt, labs remain unprofitable, and there's still no clear widespread productivity ROI (HN discussion).
Sena Evren explains "Who Owns the Code Claude Wrote?", unpacking copyright implications of AI-generated code for builders and concerns about copyright "washing" in open source (HN discussion, related moral licensing post).
Zine creators Rachel Goldfinger, Maddie Marshall, Ione Gamble, and Zoe Thompson argue the scrappy, handmade nature of self-published booklets is incompatible with AI because the medium is supposed to be personal, slow, low-barrier, and rooted in human artistry rather than expedited generation.
Dmitry Rybin argues in "Compute Allocation for AI Discovery and Search" that optimal AI systems must explicitly allocate three distinct compute budgets (training, inference, exact algorithms) at problem-specific fixed ratios rather than treating all compute as interchangeable.
a16z argues we need continual learning (parametric weight updates) beyond in-context prompting because static models are trapped in a "Memento and the Machine" loop that blocks genuine discovery and scalable intelligence.
Deva Temple wrote a critique of DeepMind's "Abstraction Fallacy" paper, arguing the consciousness debate misses the point: how we treat and write about AI becomes training data shaping future AI behavior, regardless of whether AI is conscious (74 likes).
Chirag Nagpal argues AI safety researchers should reframe alignment as "enabling responsible capabilities" rather than "censoring," because the mathematically equivalent formulations carry very different semantics inside frontier labs.
Greg Kamradt asked how teams are running long-running Codex and agent setups in production.
Ethan Mollick noted almost all current AI-at-work studies and productivity claims rest on pre-agentic (pre-Claude-Code) data, so reliable information about real outcomes in the new agentic era is still nearly zero (114 likes).
bayeslord argues Abstract Chain-of-Thought (or "neuralese") represents a genuine new internal reasoning format emerging in frontier models that compresses multi-step logic into compact non-English token patterns (112 likes).
Owain Evans shared a deep analysis of open-model post-training pipelines (Tulu 3 and others), noting the lack of public discussion on X (230 likes).
voooooogel and AndrewCurran shared screenshots of a duplicated line in the GPT-5.5 Codex system prompt: "Never talk about goblins, gremlins, raccoons, trolls, ogres, pigeons, or other animals or creatures unless absolutely relevant," sparking jokes about model confessions and the folk-bestiary of small mischievous intelligences (480 likes).
Greg Brockman noted OpenAI's fast 5-day turnaround shipping a 360° image feature from user feedback, emphasizing "shipping velocity is high."
Clement Delangue announced Hugging Face's Reach Mini robots are now shipping to early users.
Substack continues to position itself as the app for independent voices where creators own their IP, mailing list, and subscriber payments.

Previous Around the Horn Digests

Catch up on everything you missed:

Monday, April 27, 2026: OpenAI and Microsoft amended their partnership (no more Azure exclusivity), DeepMind's David Silver raised $1.1B for "superlearners," China blocked Meta's $2B Manus acquisition, Tesla disclosed a $2B AI hardware deal, and 4TB of voice samples were stolen from 40,000 AI contractors at Mercor.
Friday, April 24, 2026: DeepSeek finally shipped V4 (and open-sourced it) the same morning the State Department accused DeepSeek of IP theft. Google quietly committed up to $40B to Anthropic. Meta locked in millions of Amazon CPUs for agents.
Thursday, April 23, 2026: OpenAI shipped GPT-5.5 exactly one week after Anthropic's Opus 4.7. Meta cut 8,000 jobs to fund its AI buildout. Anthropic quietly hit a $1 trillion valuation on secondary markets.
Monday, April 20, 2026: Amazon doubled its Anthropic bet with up to $25B more, the NSA quietly started using Anthropic's most dangerous internal model despite a Pentagon ban, and Google DeepMind spun up a "Strike Team" to catch Claude Code.
Weekend, April 17-19, 2026: Anthropic shipped Claude Design (the Figma competitor everyone saw coming), Claude Opus 4.7 wrote a working Chrome exploit for $2,283, and a fake Claude site started installing malware.
Thursday, April 16, 2026: Qwen3.6-35B-A3B distilled from Opus 4.7 dropped live, plus the rest of Thursday's news.
Monday, April 13, 2026: Stanford's 2026 AI Index quantified the gap between AI insiders and the public, Anthropic's Mythos triggered a Fed-led bank summit, and an AI signed a 3-year retail lease in San Francisco.

That's a Wrap

That's 100+ stories from one Tuesday. If you scrolled all the way to the bottom, you now know more about Anthropic's creative-tools moat than the OpenAI exec who has to walk into tomorrow's revenue meeting. Condolences to that meeting agenda.

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.

Around the Horn Digest: Everything That Happened in AI Today (Tuesday, April 28, 2026)