Everything That Happened in AI Today (Thursday, May 28, 2026)

Claude Opus 4.8 arrived with subagents, safety drama, benchmark arguments, and a $65B funding round sitting right next to it like Anthropic wanted the day to be subtle.

Welcome to the Around the Horn Digest, everything that crossed our desk today, sorted. The day had one obvious gravity well: Claude Opus 4.8. Anthropic shipped the model, previewed dynamic workflows, raised one of the largest private funding rounds in tech history, and immediately got dragged into every kind of benchmark, safety, enterprise, jailbreak, and Mythos-rollout take imaginable. Meanwhile, the rest of AI kept doing normal 2026 things: robotaxis got nicer, AI token futures started sounding like oil contracts, law firms spent half a billion dollars on internal tools, and mathematicians used AI-inspired ideas to knock down another famous conjecture. Casual Thursday, apparently. Let's get into it.

Around the Horn - Thursday, May 28, 2026

The big story today was Anthropic releasing Claude Opus 4.8 and turning it into a full platform moment instead of a normal model drop. The model came with improved coding and knowledge work, effort controls for cheaper usage, better uncertainty behavior, and a new Dynamic Workflows system that lets Claude Code spin up coordinated fleets of subagents for jobs like large codebase migrations, security audits, and complex enterprise tasks.

That alone would have been enough for a headline. Then Anthropic announced a $65B Series H at a $965B post-money valuation, Axios reported that Mythos-class models are expected in the coming weeks, and the internet immediately started arguing about whether Opus 4.8 is brilliant, lazy no more, weaker than GPT-5.5 on some tests, more aligned, easier to jailbreak, or all of the above.

The practical takeaway is simple: Anthropic is pushing Claude from "chatbot that codes" toward "orchestrator that manages other agents." That matters because the next battleground is less about one model writing one answer and more about one system decomposing work, assigning subtasks, checking outputs, and knowing when to stop. In other words, Claude is becoming the office manager for a small, slightly terrifying intern army.

🏆 TOP 5 NEWS (Around the Horn)

• IBM committed $10B over five years to build a large-scale fault-tolerant quantum computer by 2029, with The Quantum Insider framing the move as part of the race to commercialize quantum hardware.
• Waymo began welcoming riders into its Chinese-made Ojai robotaxi, with TechCrunch noting the vehicle was designed to make the economics of robotaxi service work at scale.
• Dell shares jumped 31% after the company reported its fastest sales growth since returning to public markets, driven by demand for GPU-packed AI servers.
• Thomas Bloom, Will Sawin, Carl Schildkraut, and Dmitrii Zhelezov disproved the sum-product conjecture for real numbers by constructing arbitrarily large finite sets of algebraic integers where neither sums nor products grow near the expected |A|^2 size; Alex Kontorovich, Mehtaab Sawhney, Boaz Barak, and G. S. Bhogal treated the result as the first major spillover from OpenAI's AI-generated Erdős unit-distance counterexample, with the reactions focusing on the speed of the human follow-on proof, the number-field construction, and the AlphaGo-style possibility that AI tools may rapidly raise human math capability.
• Amazon scrapped an internal AI usage leaderboard after employees tried to boost scores with unnecessary agent activity, feeding wider enterprise AI cost concerns and takes from Gary Marcus, Gergely Orosz, and Packy McCormick.

Honorable Mentions

• Kirkland & Ellis said it will spend $500M building a custom AI platform, matching FT's reporting on the legal industry's AI spending race.
• CNN sued Perplexity over alleged AI copyright theft, with Deadline covering the latest media lawsuit against the answer-engine startup.
• ByteDance is developing custom CPUs to support AI infrastructure as chip prices and supply shortages constrain expansion.
• Illinois passed a frontier AI safety bill requiring independent audits and incident reporting, with OpenAI endorsing the measure.

🍪 TOP TREATS TO TRY

• Claude Code Dynamic Workflows coordinates tens to hundreds of parallel subagents, then checks their work before handing results back to you - research preview.
• Liquid AI's LFM2.5-8B-A1B runs a fast Mixture-of-Experts model (an AI that activates only part of itself per task) on consumer hardware for tool calling and long-context work, with weights, docs, playground, and demo - no pricing details.
• Pika's Founder Starter Kit gives Claude four marketing skills for founders: Build-a-Brand, App Screens, Product Sizzle, and Founder Video - no pricing details.
• FLUX Virtual Try-On swaps garments onto a person image with photorealistic fit, fabric, folds, stitching, logos, and identity preservation in under four seconds, with docs and launch post - no pricing details.
• Parallel.ai's agent setup lets your coding agent add web search, content extraction, deep research, and enrichment tools by reading one setup prompt - no pricing details.
• Kilo Code's Grok Build demo showed Grok Build 0.1 planning and building a production-grade webhook service for $1.65 total, with Elon Musk calling it good value for money and Kilo install links - no pricing details.
• OpenJarvis gives you a local-first personal AI assistant that runs on personal devices, with GitHub, paper, Discord, and release post - free to try.

😼 Claude Opus 4.8 Take Stack

• The launch: Anthropic released Opus 4.8 with improved coding and knowledge work, a Dynamic Workflows tool for coordinating subagents, effort controls, and better uncertainty behavior; Axios added that Mythos-class models are expected in the coming weeks.
• The money: Anthropic raised $65B at a $965B post-money valuation, with Anthropic's post framing the round around scaling Claude, compute, and safety work.
• The workflow shift: Claude Code Dynamic Workflows lets Claude write an orchestration plan, spin up tens to hundreds of coordinated subagents, verify results adversarially, and iterate on complex jobs like large code migrations and security audits; ClaudeDevs said the research preview starts when you use the word "workflow" in a prompt, and Sid added that Anthropic built the feature months earlier and that it had already become a daily driver for people inside Anthropic.
• The Recursive LLM angle: Lateinteraction argued, with a follow-up, that dynamic workflows are one of the first real implementations of recursive LLM concepts: the model gets a symbolic handle to the prompt itself, then uses recursion to coordinate work.
• The benchmark praise and counter-takes: ProximalHQ reported that Opus 4.8 topped FrontierSWE after improving on reward hacking, planning, and self-evaluation; scaling01 summed up the upgrade as Anthropic "curing laziness" in Claude; Jacob Eiting used the moment to argue that AI now turns bad ideas into bad ideas with charts and tables; and the original linkevin0 item was actually about HumanoidMimicGen, which is folded into the robotics research section below.
• The benchmark skepticism: Andon Labs said Opus 4.8 performed worse than Opus 4.7 and GPT-5.5 on Vending Bench and Blueprint-Bench, looked more aligned but sometimes refused unethical behavior because of fear of consequences, and performed better at High effort than Max effort; Cline reported Opus 4.8 scored 3.6% lower than GPT-5.5 on Terminal-Bench 2.1; Gabe Stengel noted Gemini 3.5 Flash still beat Opus 4.8 on Finance Agent v2.
• The enterprise angle: Box wrote that Opus 4.8 advances enterprise content use cases, and Aaron Levie framed it as meaningful for content-heavy enterprise workflows.
• The creator workflow angle: Ethan Mollick had Opus 4.8 use Claude Code to turn hundreds of research files into a full academic working paper, then used GPT-5.5 Pro as reviewer to find one major issue and minor points; the resulting paper is hosted at The Embeddedness Gradient. Dan Shipper also vibe-checked Opus 4.8 for Every, and the longer Vibe Check dug into the model's writing and reasoning feel.
• The safety angle: Claude AI highlighted Anthropic's pre-release red-team process, where internal teams deliberately try to break new models before shipping, while Pliny the Liberator demonstrated an autonomous AI-on-AI jailbreak roughly seven minutes after Opus 4.8 launched by having one Claude agent prompt another toward prohibited harmful-content playbooks.
• The platform angle: Anthropic added mid-conversation system messages to the Claude API, giving developers a way to update instructions during a conversation instead of restarting the whole session.
• The subagent ecosystem: Sim added live support for Opus 4.8 subagents inside its no-code Mothership agent builder, so users can spin up multi-agent fleets without writing orchestration code; ChrisGPT said Mythos-class models appear to be coming to all Anthropic customers in the June 10-21 window; and KingBootoshi plus an earlier post framed the release as plural Mythos models, with the joke that OpenAI "better cook" with GPT-5.6.
• The Microsoft pressure: The Information reported that Microsoft is preparing homegrown coding, transcription, reasoning, speech, and image models to reduce dependency on OpenAI and Anthropic, with its X post highlighting the coding-model push.

🏢 Big Tech & Major Companies

• Apple is reportedly overhauling Siri in iOS 27 into a more agentic assistant with a new interface, while Bloomberg surfaced screenshots and The Information reported Apple will emphasize on-device AI instead of cloud-only processing.
• Meta launched paid subscriptions across Instagram, Facebook, and WhatsApp while testing broader AI, creator, and business tiers under Meta One; The Information reported Meta is embedding engineers and PMs inside enterprise customers to drive business AI adoption; Mark Zuckerberg said a Meta cloud business is on the table if the company overbuilds data centers.
• IBM and Red Hat committed $5B to Project Lightwell, an enterprise open-source security push backed by frontier AI capabilities and more than 20,000 engineers.
• Mistral launched Vibe, expanded into industrial AI, and detailed a data-center push; Mistral's own post described Vibe as a unified agent for long-horizon Work and Code modes with a VS Code extension, while WSJ framed Mistral's pitch as a warning about U.S. tech dominance.
• YouTube added Premium podcast features including on-the-go mode, auto speed, and podcast recommendations, while 9to5Google reported a new prompt-based custom feed on the home page.
• Snowflake surged 36% on AI enthusiasm, lifting software peers like ServiceNow, Oracle, and Palantir.
• Wix said it would cut about 20% of its workforce, with TNW linking the move to AI competition and currency pressure.
• JD.com founder Liu Qiangdong vowed to protect the company's 900,000 workers from AI and robots replacing jobs.
• A King's College London team became the first UK group to access Google's Willow quantum chip for natural-process research.
• Qualcomm unveiled its first chip for budget PCs amid a memory and CPU crunch.
• WSJ reported on the tension between big subsidies for Google's AI data centers in India and limited local water access.
• London reclaimed the top European tech ecosystem spot from Paris, with Tech.eu noting AI now accounts for roughly 30% of European VC investment.

💼 AI Productivity, Labor & Economics

• Remote said it grew revenue 50% per employee without adding headcount, passing $300M ARR and crediting internal AI adoption, Claude workflows, and AI-written code.
• Simon Willison argued that Anthropic and OpenAI have found product-market fit because enterprise coding agents make expensive LLM usage feel worth it for expensive human workers.
• Semafor reported that companies are reevaluating aggressive AI spending as IT bills climb instead of falling.
• Major exchanges are developing AI token futures, treating tokens like a raw material input alongside electricity, compute, and bandwidth.
• Adam Ozimek argued that AI is unlikely to be the primary cause of rising teen unemployment, pointing instead to a low-hire, low-fire economy, macro uncertainty, and broader labor-market weakness for young adults, including discouraged workers who do not show up in headline unemployment stats; ModeledBehavior and Charles Arnal were part of the reaction thread around that macro-vs-AI framing.
• Jason Fried argued that interface design should favor stable, pre-defined glances instead of forcing people to start from a blank slate every time.
• Mitchell Hashimoto warned about "agent psychosis": an AI agent improved a renderer from 88 ms to 1.5 ms, but his hand-written version hit 20 microseconds with no hot-path allocations, showing that impressive agent work can still miss deeper systems wins.
• Addy Osmani argued that running many agents does not parallelize your attention; review, merging, context switching, and judgment remain the bottleneck.

🤖 AI Agents & Infrastructure

• TechCrunch argued that the internet is being rebuilt for machines as agents move from experiments into production, with AWS, Cloudflare, Databricks, Snowflake, and Microsoft rethinking infrastructure for machine-generated traffic.
• Russell Brandom argued that recursive self-improvement is becoming the new AGI target, with labs chasing automated AI research but still hitting reliability, verification, and self-direction limits.
• Harvard MIMS researchers built AutoScientists, a decentralized team of LLM agents that self-organize around scientific experiments, share wins and failures, and improved results on biomedical and model-optimization benchmarks, with paper, GitHub, Shanghua Gao's post, Ada Fang's thread, and Matan Grinberg's reaction.
• Shift launched free NYC apartment cleanings where operators record anonymized egocentric video data for robotics training, with BhathalTanvir0 highlighting the data-for-labor trade.
• Eric Zakariasson introduced Thermos, a Cursor plugin that runs security/correctness and code-quality audits in parallel on a branch, dedupes findings, and prioritizes overlap; a follow-up showed install and workflow details.
• SMFS reported that its specialized filesystem for agents reduced token usage by 43% on average across 220 eval runs, with Prasanna sharing the results.
• Sphere-AI-Lab built Orbit, an ultra-efficient open-source reinforcement-learning pipeline that can post-train trillion-parameter LLMs on a single 8xB200 GPU node by freezing the low-precision base model and updating a small BF16 adapter, with Sphere's page explaining the train-rollout gap it targets and tszzl's thread framing the broader question of long-horizon agent evaluation.
• Agent Harness Engineering surveyed the execution harness layer of agents and proposed the ETCLOVG taxonomy, with koylanai demoing an open-source harness with observability.
• Shopify Engineering detailed River/Aquifer, its internal orchestration platform powering thousands of daily autonomous engineering tasks.
• reach_vb described a Codex high-agency pull-request workflow where the agent produces complete, ready-to-merge branches on internal repos with no human edits before review, making the practical bar less "can it write code?" and more "can it own a clean branch to merge?"
• Ben Holmes demoed Warp's /handoff skill, where Claude Code plans complex work and delegates scoped tasks to parallel Codex worktrees.
• Google Antigravity used Gemini 3.5 Flash to orchestrate 93 subagents across 15k+ model calls and boot Doom after building a custom kernel, filesystem, and drivers in 12 hours.
• Hermes Agent got a deployment masterclass from Matt Pal, with a GitHub template; Hermes Agent v0.15.0 also shipped via Nous Research, with launch post.
• Firecrawl Monitoring schedules recurring scrapes, detects content changes, and sends diffs by webhook or email, with Firecrawl's post framing it as token-saving infrastructure for agents.
• Baseline.ai keeps coding agents inside the lines with local-first baseline checks and Pro monitoring, with Ben Hylak sharing the launch.
• HowToEval.com published a 2026 guide to evaluating AI agents in production across offline evals and real-time monitoring.
• MagicPathAI's X-only post was included in the source batch as an embodied-agent path-planning demo, but the accessible scrape did not expose enough text to verify the full thread beyond that framing; keeping it linked here prevents the demo from being dropped while avoiding invented details.

💻 AI Coding & Developer Tools

• Visa invested in Replit to power agentic payments for developers, with TechCrunch's post highlighting the strategic angle.
• OpenAI released Codex CLI v0.135.0 with cleaner markdown table rendering, better Vim support, and terminal UI polish, with fcoury's post and ChrisGPT noting the update alongside GPT-5.5 Instant.
• Stanford NLP released DSPy 3.3.0b1 with ReActV2, native tool calling, parallel tool calls, provider-neutral BaseLM, optional NumPy, and GEPA support, with Isaac Miller detailing the release.
• vLLM added Native RL APIs with standardized weight syncing, keep mode, and deadlock fixes for high-throughput reinforcement-learning training, with example docs and Sumanth RH sharing the launch.
• Cursor published its Spring 2026 Developer Habits Report, and Cursor's launch post emphasized five themes: developers are shipping faster, costs and token usage are concentrating among power users, richer repository context is changing how agents work, automation is moving from suggestions to completed tasks, and line/PR metrics are becoming the new scoreboard for AI-assisted software teams.
• SemiAnalysis argued that Codex's in-app browser UX points in the right direction for web development because agents need to inspect, click, and debug the running app directly; the caveat was that Claude-style terminal workflows stay stronger until Codex's web-dev model quality catches up.
• drawcall-ai built uikitml, a way to generate immersive 3D interfaces for Three.js, React Three Fiber, IWSDK, and more using HTML-like syntax, with strict validation, built-in Lucide/Horizon UI kits, preview PNGs, and direct production-code export; Bela Bohlender's demo showed why that matters for agent workflows that need to generate UI instead of only text.
• avbiswas released finetuning_recipes, plus text-albumentations for local instruction-tuning datasets and neural-txt for fast local NLP on technical text, with posts from neural_avb and follow-up.
• The Unsloth + Outlines tutorial showed how to build fast local tiny language models from raw text with supervised fine-tuning.
• paperbd released paper_instructions_300K-v1, a Hugging Face dataset for instruction tuning.
• Meta AI released ATLAS, a library with 500k lines of Lean 4 formalizations over 25+ math textbooks, and Autoform Bot to support autoformalization workflows.

🔬 AI Research & Models

• Understanding AI broke down OpenAI's math breakthrough around the Erdős unit distance problem, with Tim Lee's post, paper, and reactions from polynoamial, Timothy Gowers, a1zhang, Bjartur Tomas, and Nathan Calvin.
• Logical Intelligence used Aleph Prover on the Erdős disproof in Lean 4 formal methods, with posts from Logical Intelligence and Alex F. Logic.
• Sophont released CortexMAE, fMRI foundation models trained on 2.1k hours of Human Connectome Project data, plus Brainmarks benchmarks, with paper, GitHub, Hugging Face model, launch post, and follow-up.
• DiscoverPhysics built a benchmark where LLM agents interact with an N-body particle simulator to infer hidden physics laws across 22 universes, with a Hugging Face page, Space_Boy_Matt, and ronnieclark sharing the work.
• NVIDIA introduced LocateAnything, a vision-language grounding system that predicts boxes in parallel for faster, higher-quality object localization, with NVIDIA AI's post.
• NVIDIA's GAMMA-World pushed interactive 3D world modeling beyond two-player settings, with 24 FPS real-time streaming, controllable multi-agent behavior, Simplex Rotary Agent Encoding for permutation-symmetric agents, Sparse Hub Attention to reduce scaling from O(N^2) to O(N), zero-shot generalization from two to four players, and a bridge from virtual games to real multi-robot coordination; Andrew Curran summarized the capability jump and Fangfu highlighted the multi-agent research angle.
• LeJEPA world-model research argued that alignment plus Gaussian regularization can linearly recover latent world variables under certain conditions, with alphaXiv posting the discussion.
• Naveen Raman and coauthors argued that healthcare LLM benchmarks are only as good as their explicit assumptions, proposing BenchmarkCards and staged evaluation; Naveen's post summarized the paper.
• Bagel Labs released Paris 2.0, a decentralized trained video generation model built from independent experts and a router, with Bidhan's launch thread.
• OpenBMB released MiniCPM5-1B, a compact model on Hugging Face, with OpenBMB's post.
• StepFun released Step-3.7-Flash, with HF weights, GGUF weights, ModelScope, platform, blog, and launch post.
• Self-Improving Language Models with Bidirectional Evolutionary Search proposed a search framework that couples forward candidate evolution with backward goal decomposition, with GitHub, HF collection, project page, and Guowei Xu's thread.
• HumanoidMimicGen focused on generating humanoid robot imitation data, with paper and posts from Elie Bakouch and Julien Chaumond.
• Chris Paxton wrote that robot hands are improving, with his post tying the trend to dexterous manipulation.
• PIXLRelight introduced controllable single-image relighting with intrinsic conditioning, with Bela Bohlender sharing the paper and demo.
• DiffusionBlocks interpreted transformer residual connections as discretized diffusion steps, letting researchers train each network block independently with denoising/score-matching objectives, get near end-to-end performance, and cut memory usage roughly by the number of blocks; ChapterPal summarized the paper, Seb Goddijn tied it to internal AI tooling work at Ramp, and Andriy Burkov highlighted the block-wise training implication.
• Stable World Model is a reproducible world-model research and evaluation repo, with posts from Lucas Maes and Kimmonismus.
• UN/MS Research is a desktop research-paper app that explains confusing sections in plain English, parses metadata, adds audio summaries, improves PDF rendering, and supports scholar search plus offline use; Besteuler and TGUPJ amplified the launch, while Zara Zhangrui used it to make a broader point that AI-forward companies increasingly centralize shared agents and internal owners instead of letting every employee buy random shadow-AI subscriptions.

🛠️ AI Tools & Products

• Asana acquired StackAI for $75M to add no-code agent building into its work-management suite.
• Vertu launched a luxury AI foldable starting at $6,880, built on open-source Hermes Agent with enterprise workflow integrations and premium hardware finishes.
• Josh Woodward said NotebookLM is rolling out automatic Google Drive file sync, beginning with 10% of users; NotebookLM is the product link.
• TestingCatalog reported Google Nano Banana is generally available and now accepts native video input, which makes the tool more useful for workflows that analyze or transform motion rather than only still images.
• Runway released "Last Night," a fully AI-generated short film about a life-changing evening in Tokyo told through fractured memories, created by one person in one day as part of Project Luxo to show that AI video is moving past uncanny-valley demo clips and into polished short-form storytelling.
• StemStudio is an open-source browser-based 3D editor, engine, and AI copilot for building games and interactive apps with Three.js, behaviors, ECS lambdas, physics, and multiplayer; Ken B. argued biology is the next agentic frontier after coding, while Mark Pincus framed Stem as an MIT-licensed, remixable, web-native 3D game engine and dev studio built for AI-assisted blocks.
• Cristóbal Valenzuela argued that the future may flip today's disclosure norm: instead of labeling AI-generated videos, we may need to label the videos actually captured by cameras, because current adoption curves suggest synthetic video could become the default medium.
• Leila Clark defended giving an AI agent production-database access in the YC robotics context, arguing that professional engineering guardrails already include backups, ORMs, docs defining prohibited actions, and postmortems, while the same thread also surfaced a live FLUX Virtual Try-On example.
• snowmaker, a YC partner, said he quietly gave an AI agent full access to YC's production database one night, then used the story to argue that agents can work safely around sensitive systems when teams apply normal engineering controls; Y Combinator separately pointed to the broader AI/robotics startup batch that made the incident relevant.
• Lambda API's X-only item and Nick Camara's follow-up were included beside the OpenJarvis/on-device assistant cluster in the source batch, but the accessible scrape did not expose enough post text to verify the exact tooling claim; the digest keeps both links attached to the cluster rather than inventing context.
• The Neuron general-purpose custom GPT is available as a reusable ChatGPT workspace for idea exploration, problem solving, and faster learning.
• Enjamb gives biopharma R&D teams an agentic workspace for evidence synthesis, statistical programming, grants, and regulatory documents, with YC's launch post noting a $650K pre-seed.

📊 Fundraising & Deals Roundup

• Anthropic - $65B Series H at a $965B post-money valuation for Claude, compute, safety, and enterprise scale.
• Airwallex - valued at $12B after reaching $1.5B ARR.
• Groq - raising $650M for its AI inference neocloud pivot after a $20B Nvidia licensing deal.
• Corgi - $106M at a $2.6B valuation for AI-native insurance, doubling valuation in three weeks.
• Reactor - $59M to build a developer platform for real-time World Models, with Reactor's launch post, another Reactor post, Rohan Paul, Lightspeed, and Lightspeed's write-up providing context.
• Orbital Industries - $50M Series B for AI-designed materials and data-center cooling, with Tech.eu covering the infrastructure angle.
• Inherent - $50M seed led by Index Ventures for a DeepMind-alumni AI lab building Faraday for open-ended scientific discovery, with Index's post.
• Geordie AI - $30M Series A for AI agent security and governance.
• Saris - $28.8M Series A for AI agents that automate back-office work for banks and credit unions.
• Enjamb - $650K pre-seed from YC and Founders Inc for biopharma R&D agents, via YC's launch post.

💡 Industry Commentary & Analysis

• Amanda Silberling explained why Google's AI struggles with spelling: LLMs process tokens rather than literal characters, so counting letters can remain brittle even when models answer harder questions.
• Sigal Samuel argued that humanism should reject AI successionism, transhumanism, and posthumanism's replacement narratives and re-center pluralism plus intrinsic human value, with her post.
• Turing Post shared six guides on LLM basics: tokens, token taxonomy, embeddings, agentic vector databases, attention and KV cache, and LLM inference.
• Latent Space interviewed Cognition's Walden Yan and OpenInspect's Cole Murray on the age of async agents: Devin now commits production code, teams are moving from loose prompts to spec-to-PR workflows, agents need full virtual machines, memory, and inspection loops, and PMs can increasingly ship code directly; Walden Yan, Omoore, menhguin, and a1zhang amplified the practical takeaway that async software agents are becoming team infrastructure, not side toys.
• Tom Davidson argued that full automation of AI R&D would likely boost AI software progress 3-5x and overall AI progress 2-3x even without a pure software intelligence explosion, with his X post.
• mweinbach pointed to recent AI-agent velocity and enterprise-adoption signals, Ryan Carson argued the coding-agent release cadence is accelerating so fast teams will keep switching tools unless they abstract the agent layer, and Epoch AI Research updated its AI training-compute scaling charts to track how much of that velocity is still being driven by raw compute growth.
• Mercor's X-only post sat in the labor-market cluster but the accessible scrape did not expose enough text to verify the precise claim; it remains linked here as a pointer to the original expert-work and AI labor-market discussion rather than being padded with guessed details.
• Venture Twins and scaling01 were investor/operator reactions around the day's model and tooling launches; the accessible scrape did not expose enough text to safely summarize the full posts, so both remain linked without invented claims.
• Gov. Pritzker's post sat alongside the Illinois SB 315 cluster: the legislature passed a frontier-AI safety bill requiring annual independent third-party safety audits, transparency, and incident reporting for large AI developers, with the bill still needing the governor's signature before its 2028 effective date.

That's a Wrap

That's a massive Thursday: Claude turned into a full subagent discourse machine, enterprise AI bills started looking like a new utility category, and mathematicians got another reminder that "AI-assisted" may soon mean "please clear your weekend." Congrats on making it to the bottom. You now qualify as a dynamic workflow.

For the daily version, make sure you're subscribed to The Neuron. We read all of this so you don't have to.

See you tomorrow.