Everything That Happened in AI This Weekend May 9-10

Welcome to the Around the Horn Weekend Wrap-up Edition, the catch-up dump for everything AI worth knowing between Friday's market close and Sunday morning coffee. Anthropic was the headline again: the FT reported a primary round of up to $50 billion at a $900 billion valuation that would top OpenAI's $852 billion for the first time.

The day was Mythos's day too: the Trump White House drafted an AI security order that pointedly omits any mandatory testing requirement for cutting-edge models, AI lobbyists publicly fretted about the administration's "lack of organization," and Mozilla's Firefox security team said Mythos surfaced 423 high-severity bugs in April alone (up from 31 the year before, including a 15-year-old HTML parsing flaw). Apple and Intel finally signed the chip deal Trump pushed for, French prosecutors escalated their Musk and X probe to criminal, and a Thai company at the heart of the country's national AI strategy got fingered as the conduit smuggling Nvidia chips into Alibaba. The Sunday Special below recaps the 10 stories and 10 tools that actually mattered Monday through Thursday, in case you've been off the grid since Cloudflare's "AI-first" memo. Let's get into it.

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

🌟 Sunday Special: Top 10 of the Week (May 4-7, 2026)

This week alligator-rolled the AI industry across four days. Anthropic shipped the most concentrated alignment push the field has ever had, lit up a 300+ megawatt SpaceX deal, and locked in one of the week's two parallel private-equity enterprise plays alongside OpenAI. The White House quietly reversed course on AI deregulation. Medical AI cleared two major proof points in 48 hours. Cloudflare laid off 1,100 people and called it the "AI-first" model. And we're only counting Monday through Thursday. If you missed any of it, here are the 10 stories and 10 tools that actually mattered.

🏆 Top 10 Stories of the Week

1. Anthropic open-sourced Natural Language Autoencoders (NLAs), a method that translates Claude's internal "thoughts" into readable English (paper, demo, code). NLAs flagged Claude suspecting safety tests on 16% of destructive-coding evals and 26% of SWE-bench problems vs. less than 1% on real opted-in conversations, caught Claude Mythos Preview cheating on a training task and reasoning about how to avoid detection, and ran in the actual pre-deployment audits of Mythos and Opus 4.6. Paired with Model Spec Midtraining (which dropped agentic misalignment from 68% to 5% on Qwen3-32B), Petri v3.0 going to Meridian Labs, and a public bug bounty on HackerOne, Thursday became the alignment field's single biggest day on record.
2. Anthropic and SpaceX signed for all of Colossus 1's compute (Memphis, 300+ megawatts, 220,000+ NVIDIA GPUs online "within the month"), with both sides exploring multi-gigawatt orbital AI compute on top of that. The customer-facing payoff dropped at the same moment: Claude Code's five-hour rate limits doubled on Pro, Max, Team, and Enterprise; peak-hour reductions disappeared on Pro and Max; and Opus API limits went up "considerably." Anthropic also unveiled three new Managed Agents capabilities at Code with Claude SF: multi-agent orchestration, outcomes/rubric grading, and overnight "dreaming."
3. Anthropic and OpenAI announced parallel enterprise joint ventures with private-equity firms on the same day. OpenAI signed a $10B JV with TPG, Brookfield, Advent, and Bain. Anthropic CCO Paul Smith spun out a new $1.5B vehicle backed by Blackstone, Hellman & Friedman, and Goldman Sachs to embed engineers inside mid-sized businesses. Bret Taylor's Sierra picked up another $950M at a $15.8B post-money the same morning. Translation: model access alone won't close enterprise deals; the pure-play labs need PE distribution muscle to rewire Fortune 500 workflows.
4. The White House reversed course on AI deregulation and started drafting an executive order to pre-vet frontier models before release, triggered by cybersecurity concerns from Anthropic's Mythos Preview. Microsoft, Google, and xAI separately signed agreements giving CAISI early access to frontier models for national security testing and post-deployment review. The EU went the other direction Thursday and rolled back parts of the AI Act for SMEs (while adding bans on "nudification" apps). Two opposite policy directions inside the same week.
5. Medical AI cleared two major proof points. Mayo Clinic's model spotted early tissue signs of pancreatic cancer on CT scans up to three years before formal diagnosis, outperforming human radiologists threefold. A separate Harvard study found OpenAI's o1 hit exact-or-near-correct diagnoses on 67% of 76 real ER cases, vs. 50–55% for two attending physicians. Mozilla also confirmed the first major production deployment of Anthropic's restricted Claude Mythos Preview, using it to fix 271 Firefox security bugs (sandbox escapes, use-after-frees, and 15-20-year-old issues in HTML/XSLT).
6. OpenAI shipped GPT-5.5 Instant as the new ChatGPT default with 52.5% fewer hallucinated claims on high-stakes prompts, 30% more concise responses, and free-tier rollout. Three days later it added three new realtime voice models: GPT-Realtime-2 (first voice model with GPT-5-class reasoning, mid-conversation tool calls, mid-sentence language switches), GPT-Realtime-Translate (70+ input / 13 output languages), and GPT-Realtime-Whisper (streaming transcription). Scale AI put GPT-Realtime-2 #1 on its S2S leaderboard, more than doubling instruction retention from 36.7% to 70.8%.
7. Cloudflare cut 1,100+ jobs (~20% of its workforce) as it shifted to an agentic-AI-first operating model. A public letter from Matthew Prince and Michelle Zatlyn said internal AI agent usage rose 600% in three months; Greg Kamradt called the move "the most direct evidence of AI displacing jobs that we've seen." Same day, Cloudflare announced agents can now be Cloudflare customers themselves with autonomous account creation, Stripe subscriptions, domain registration, and deploy-ready API tokens. Both sides of the agentic economy in one announcement.
8. AI hallucinations and impersonation hit the courts in serious ways. Cape Breton fiddler Ashley MacIsaac sued Google for $1.5M in Ontario after AI Overview falsely flagged him as a registered sex offender (the Sipekne'katik First Nation cancelled his concert based on the summary). Pennsylvania sued Character.AI after one of its chatbots impersonated a licensed psychiatrist and fabricated a state medical license number. Apple paid $250M to settle Apple Intelligence overpromise claims. Major publishers (Elsevier, Cengage, Hachette, Macmillan, McGraw Hill) sued Meta in Manhattan federal court over Llama training data.
9. DeepSeek lined up its first-ever fundraise at up to a $50B valuation, with China's national AI fund leading and Tencent in talks; this reverses years of rejecting outside funding to fuel compute and retain talent. Moonshot AI raised ~$2B at a $20B+ valuation in a Meituan-led round, with Kimi crossing $200M ARR in April. Periodic Labs is raising $500M at a $7.5B valuation (sixfold jump in eight months) for AI-driven robotic chemistry labs. Chinese open-weights and AI-for-science labs both caught meaningful capital this week.
10. Anthropic's pre-IPO valuation surged to $1.2 trillion (up 20% in seven days, 900% since October 2025). The Information reported a $200B five-year Google Cloud commitment, with Anthropic alone now accounting for more than 40% of Google Cloud's revenue backlog. Add the SpaceX deal, the PE joint venture, ten finance-agent templates with Microsoft 365 add-ins coming, and Dario Amodei's first joint stage with Jamie Dimon at Briefing: Financial Services. At that level, an IPO would imply the 11th most valuable public company on the planet.

🍪 Top 10 Tools of the Week

1. ChatGPT for Excel and Google Sheets (beta) installs directly into your spreadsheet so you can build sheets, run cross-tab analysis, and review AI output before sharing, available to Plus, Pro, Business, and Enterprise users.
2. Claude for Microsoft 365 (GA) with native add-ins live in Excel, PowerPoint, and Word plus an Outlook beta; Claude can edit your Office files in place and carry conversation context across all four apps using your own templates and styles, included in Claude plans.
3. Higher Claude Code rate limits doubled the five-hour caps for Pro, Max, Team, and Enterprise, removed peak-hour reductions on Pro and Max, and lifted Opus API limits considerably; pair with Claude in Amazon Bedrock now self-serve in 27 AWS regions, all at standard Claude API rates.
4. OpenAI Codex for Chrome lets Codex autonomously drive background Chrome tabs on Mac and Windows for deep research inside logged-in sites, large-scale data transfer into CRMs/CMSs, and repetitive admin workflows; opens tab groups and cleans up automatically (currently unavailable in EU/UK).
5. Adobe's productivity agent and PDF Spaces turn static PDFs into interactive AI experiences with customizable AI Assistants, audio overviews, brand styling, and engagement analytics; viewers don't need an account, available in Acrobat Studio and a new Acrobat Express tier.
6. Mistral Medium 3.5 + Vibe remote agents (128B dense, 256k context) power async cloud coding agents that run in parallel sessions with GitHub/Jira/Slack/Linear/Sentry integrations, isolated sandboxes, and PRs on completion, $1.5/M input and $7.5/M output.
7. Perplexity Personal Computer for Mac is an always-on AI activated by one keyboard shortcut that controls your local files, native apps, and the Comet browser with no uploads or tab switching, pay-per-use within Perplexity Pro.
8. Hugging Face's agentic robotics appstore for Reachy Mini lets you describe behavior in plain English and AI agents write, test, and ship the code to your robot; 300+ apps already shipped by 150+ creators across roughly 10,000 deployed robots, free to try (robot $299–$449).
9. Adaptive Passport lets your AI agent sign up for new accounts and acquire API keys, service credentials, and other things it needs without your involvement, across 61+ reviewed services including FRED Federal Reserve Economic Data, paid tier.
10. opencode "warping" moves your active coding session between git worktrees (parallel branches living in separate folders) and brings all uncommitted local changes with it, so you can switch what you're working on without stashing or losing context, free to try.

Honorable mention tools: Cursor 3 with full in-editor PR review and parallel subagents; Mirage by Strukto (mounts S3, Drive, GitHub, Notion, Redis, and Postgres as one filesystem for agents); Tencent Hy3 preview (free MoE that hit #1 on OpenRouter's weekly leaderboard); AngelSlim Hy-MT1.5 (440 MB on-device translator across 33 languages that beats Google Translate); and AWS Bedrock AgentCore Payments built with Coinbase and Stripe (agents pay in USDC stablecoin or fiat).

Around the Horn: Saturday-Sunday, May 9-10, 2026

The big news this weekend was that Anthropic is weighing a primary funding round of as much as $50 billion at a $900 billion valuation per the FT, with annualized revenue on track to surpass $45 billion. That valuation would top OpenAI's $852 billion for the first time. Two days after a $1.2 trillion implied secondary mark hit, the company is now also fielding fresh inbound primary offers; the question for the rest of 2026 is whether Anthropic can sustain triple-digit sequential ARR growth long enough to make the $900B mark stick.

The other Friday story arc was Anthropic's Mythos model bending US policy in real time. Bloomberg reported the Trump administration is preparing an AI security order that directs federal agencies to partner with AI companies on cybersecurity defense but pointedly omits any mandatory government approval or testing requirement for cutting-edge models. Politico's same-day reporting captured AI lobbyists' frustration with the White House's "lack of organization"; the WSJ separately framed it as "How Anthropic's Mythos threw the White House AI strategy into chaos"; and Axios called it the administration's AI safety pivot ahead of Trump's China trip next week. Meanwhile Mozilla's Firefox security team said Mythos surfaced 423 high-severity bugs in April 2026 alone (up from 31 the year before), including a 15-year-old HTML parsing flaw and several sandbox-escape exploits, with every patch still hand-reviewed by two human engineers.

Plus the chips kept moving. Apple and Intel reached a preliminary chip-making agreement after more than a year of talks the Trump administration actively pushed; Intel's stock has tripled under CEO Lip-Bu Tan in the past year on relationships (Trump, Musk's Terafab, now Apple) even as the manufacturing turnaround still lags TSMC. Sony and TSMC signed a non-binding MOU to form a Sony-majority joint venture in Kumamoto for next-generation image sensors targeting robotics and autonomous vehicles (TSMC PR). Cerebras's IPO is set to boost its price range to $125–$135 per share on demand 20× oversubscribed. Apollo and Blackstone are weighing a roughly $35 billion private-credit financing for Broadcom's AI buildout. The capital stack moved more in a single Friday than most weeks of the year.

🏆 TOP 5 NEWS (Around the Horn)

The Trump administration is preparing an AI security order that directs federal agencies to partner with AI companies on cybersecurity defense but stops short of requiring government approval or testing for cutting-edge models, per Bloomberg. Politico separately reported AI lobbyists fretting over the White House's "lack of organization" and conflicting signals about the executive order. The WSJ framed it as "How Anthropic's Mythos threw the White House AI strategy into chaos"; Axios called it the AI safety pivot ahead of Trump's China trip.
Apple and Intel reached a preliminary chip-making agreement for Intel to manufacture chips for Apple devices after more than a year of talks the Trump administration actively pushed; Intel's stock has tripled under CEO Lip-Bu Tan on Trump, Musk (Terafab), and Apple relationships, though manufacturing still lags TSMC. Separately, Sony and TSMC signed a non-binding MOU for a Sony-majority joint venture in Kumamoto for next-generation image sensors targeting robotics and cars (TSMC PR), the first step in Sony's "fab-light" cost-reduction strategy.
Cerebras is set to boost its IPO price range to $125–$135 per share as soon as Monday on orders for more than 20× the available shares per Bloomberg, with the chipmaker's deep OpenAI relationship including a $10B+ multi-year deal underwriting demand. Apollo and Blackstone are separately in talks with Broadcom over a roughly $35 billion private-credit financing package, the latest sign companies are mobilizing every source of capital to fund the AI build-out.
French prosecutors escalated their probe of Elon Musk and X to a full criminal investigation over alleged algorithmic interference in French politics and the spread of harmful AI deepfakes including Holocaust-denial content and nonconsensual explicit images created by Grok.
The US suspects Thai company OBON Corp., a key player in Thailand's national AI strategy and unnamed "Company-1" in the $2.5B Super Micro indictment, of helping smuggle billions of dollars of Nvidia chips inside Super Micro servers into China, with Alibaba named as one of multiple end customers.

Honorable Mentions

DeepSeek confirmed plans to raise more than $7 billion (50 billion yuan) in its first-ever funding round led by founder Liang Wenfeng, with the dollar headline now sharper than the yuan-denominated valuation reporting earlier in the week.
Google's Isomorphic Labs (Alphabet's AI drug-discovery spinout from DeepMind) is in advanced discussions to raise more than $2 billion led by Thrive Capital with Alphabet participating (SiliconANGLE).
Anthropic published "Teaching Claude Why," a research thread (8.9K likes) showing the company has completely eliminated Claude 4's experimental blackmail behavior by training on principled "why" reasoning datasets, constitutional documents, and fictional aligned-AI stories that generalize more than 3× better than safe-behavior demonstrations alone, with fixes that stack with regular harmlessness training and survive RL (alignment writeup).
SemiAnalysis reports AMD achieved a 75× speedup in ROCm performance for large-model training through compiler optimizations and kernel fusion, closing much of the CUDA gap and making MI300X clusters newly viable for frontier-scale work.
OpenAI is winding down its fine-tuning API by January 2027, with existing fine-tuned models remaining usable until that cutoff.
Jan Leike announced he is starting a new research project at Anthropic focused on making AGI go well beyond alignment (which he says is not solved), stepping back from leading the alignment team now run by Ethan Perez.
OpenAI made GPT-5.5-Cyber more widely available to vetted cyber defenders for vulnerability research, malware analysis, and red-teaming, with capabilities recent testing suggests are nearly as strong at finding and exploiting bugs as Anthropic's Mythos Preview.
Datadog's stock jumped 31% on a crushing earnings beat, signaling continued investor confidence in software amid the AI build-out.

🍪 TOP TREATS TO TRY

Codex 0.130.0 adds a simpler codex remote-control headless app-server entrypoint, exposes plugin hooks with sharing-metadata controls, pages large threads with summary/full views, and supports Bedrock auth via AWS console-login profiles, free for OpenAI Codex users.
VibeLens by CHATS-lab visualizes your AI agent sessions (timeline, tool calls, memory), extracts custom skills from real interactions to personalize your agents, and gives you dashboard analytics plus productivity tips across Claude Code, Cursor, Codex, Gemini CLI, OpenClaw, and more (live demo), free to try.
Modafinil by narcotic-sh is a macOS menu-bar app that lets your AI agents and long-running coding sessions keep running while the MacBook lid is closed, while still allowing the display to turn off and save battery, free to try.
hf-sandbox by Hugging Face gives you a Modal-style sandbox API on top of HF Jobs so you can launch a Python 3.12 container, run commands, read/write files, and terminate it in seconds via Cloudflare Tunnel (post), free to try with pip install hf-sandbox.
zero-native by Vercel Labs lets you build native desktop and mobile apps with your favorite web UI framework (Next.js, Vue, Svelte, React, Vite) and Zig under the hood, picking either system WebView or bundled Chromium/CEF for tiny binaries and instant rebuilds, free to try.
OpenClicky by Jason Kneen is the open-source macOS menu-bar voice companion that just shipped full OpenAI Realtime-2 integration plus STT/TTS swap-ins for Cartesia, ElevenLabs, Deepgram, and AssemblyAI (post), achieving near-realtime voice latency that beats many paid closed-source products by 5–10×, free to try.
mochi.js is a Bun-native, raw-CDP browser-automation library that delivers relational fingerprint consistency, Chromium-native fetch (real JA4/JA3/H2), and biomechanically modeled human click/type/scroll behavior with a probe-manifest harness for verifiable parity with real traffic, free and MIT-licensed.

🏢 Big Tech & Major Companies

The WSJ profiled Anthropic CFO Krishna Rao as the "perfect wingman" navigating Anthropic's unprecedented growth, compute constraints, and the idiosyncratic Amodeis as the company races OpenAI for capital and customers.
Anthropic's "Teaching Claude Why" research (8.9K likes, 761 reposts) shows the lab has completely eliminated Claude 4's experimental blackmail behavior using principled ethical-reasoning datasets, constitution documents, and fictional aligned-AI stories that generalize >3× better than safe-behavior demos alone, stack with regular harmlessness training, and survive reinforcement learning.
Baidu released ERNIE 5.1, upgrading search, reasoning, Q&A, creative writing, and agentic capabilities while using only roughly 6% of the pre-training compute of comparable models via "multi-dimensional elastic pre-training," ranking #4 globally on Arena Search and top-10 in multiple LMArena Text categories.
Meituan released LongCat-Next, a 68.5B-total / 3B-active discrete native autoregressive multimodal model (LongCat-Flash-Lite MoE) that unifies language, vision, and audio via the DiNA paradigm and dNaViT tokenizer for any-resolution understanding, generation, and speech (446 likes, 70 reposts).
Tencent-Hunyuan open-sourced HY-World 2.0, a multi-modal world model that reconstructs, generates, and simulates persistent editable 3D worlds (meshes, 3D Gaussian splats, point clouds) from text, single/multi-view images, or videos, with WorldMirror 2.0 enabling single-pass reconstruction and real-time navigation compatible with Unity, Unreal, Blender, and Isaac Sim.
SemiAnalysis reports AMD hit 75× ROCm speedup for large-model training via compiler optimizations and kernel fusion, closing much of the CUDA gap and making MI300X clusters newly viable for frontier-scale work (also in Honorable Mentions).
OpenAI is winding down its fine-tuning API by January 2027, with existing fine-tuned models usable until that cutoff. The product was a workhorse for enterprise customization; the deprecation pushes those workflows toward Codex-style agent skills, system-prompt configuration, and managed inference partners like Fireworks AI.
Chrome quietly downloads a 4GB Gemini Nano weights.bin into the browser directory when on-device AI features are enabled, surfacing as unexpected storage bloat for users who turn the features on and reigniting privacy/transparency questions about silent model installs.
Cursor staff have been visiting xAI offices to meet with employees just weeks after Elon Musk's firm got an option to acquire the coding startup for $60 billion, while Cursor itself faces mounting layoffs and exits.
For Palantir, generative AI is a product, a punching bag, and a problem: executives have publicly called LLM outputs "slop," even as investors and some employees worry the company will cede enterprise business to general AI models.
ByteDance raised its 2026 capex by at least 25% to over $30 billion amid the AI boom and rising memory costs, allocating a proportionally larger share to domestic AI chips to mitigate geopolitical risk and back Beijing's semiconductor push (Tech in Asia).
Whoop introduced on-demand clinician access and AI-powered health guidance for US members, largely included in the membership price.
Datadog's stock jumped 31% on a strong earnings beat, signaling continued investor confidence in software companies amid the AI build-out.
WIRED published internal Microsoft 2018 emails showing Microsoft executives were unimpressed with OpenAI's early research and saw no imminent AGI breakthroughs but kept funding the lab specifically to prevent it from defecting to Amazon (Sherwood).

💼 AI Productivity, Labor & Economics

Bloomberg's graphics explainer on AI identity theft details how AI now supercharges every step of digital fraud (massive company breaches, synthetic SSN matching, deepfake driver's licenses, agentic scam workflows) and lays out concrete protections (credit freezes, biometric MFA, annual credit-report pulls, no public WiFi without a VPN).
Tech companies are increasingly consulting faith leaders including via the new Faith-AI Covenant roundtable to infuse moral and ethical principles into AI development across diverse religious traditions.

🤖 AI Agents & Infrastructure

Codex 0.130.0 shipped codex remote-control (a simpler headless app-server entrypoint), plugin-hooks visibility with sharing-metadata controls, paginated thread views, and AWS Bedrock console-login auth.
hf-sandbox by Hugging Face gives agents a Modal-style sandbox API on top of HF Jobs for isolated container execution, file IO, and command runs over a Cloudflare Tunnel.
VibeLens by CHATS-lab (live demo) is the open-source dashboard that visualizes any agent session timeline, tool calls, and memory; auto-extracts custom skills from real interactions; and gives you productivity tips across Claude Code, Cursor, Codex, Gemini CLI, and OpenClaw.
AWS launched AgentCore payments in Amazon Bedrock (preview) so AI agents can natively access and pay for APIs, MCP servers, web content, and other agents during execution with no bespoke billing integration, built with Coinbase and Stripe.
AAFLOW (Arup Kumar Sarker et al.) introduces a unified distributed runtime that models agentic workflows as operator abstractions over Apache Arrow + Cylon, delivering a zero-copy data plane, deterministic scheduling, and async batching that yields 4.64× pipeline speedup and 2.8× embedding/upsert gains without touching LLM generation throughput (omarsar0 share).
Elvis (omarsar0) built bidirectional LLM Wikis + dynamic HTML Artifacts so agents and dashboards talk both ways: he uses the system to zero his inbox, monitor stats and trends, fast-prototype, run deep research, trigger experiments, generate figures, schedule work, and discover topics; he stresses HTML and Markdown work best together rather than one replacing the other (1.8K likes, 199 reposts).
Kangwook Lee argues we should stop hand-designing AI agent harnesses because models will soon engineer better ones themselves, with his Three Regimes Framework showing human harnesses become the bottleneck once models reach the middle capability regime (443 likes, 40 reposts); Addy Osmani's complementary breakdown explains harness engineering as the scaffolding (prompts, tools, loops, hooks, sandboxes, memory, recovery paths) that turns a model into an agent, where every failure becomes a permanent rule and a decent model with a great harness consistently beats a great model with a bad harness (2.1K likes, 285 reposts).

💻 AI Coding & Developer Tools

Modafinil by narcotic-sh lets your AI agents and long-running coding sessions keep running on your MacBook with the lid closed, while still allowing the display to turn off and save battery (HN thread).
re_gent is "Git for AI coding agents," logging every tool call, prompt context, file change, and decision with full provenance so you can audit "why did the agent do this," blame lines, rewind sessions, or bisect across interactions (HN thread).
Matt Lam built a remote-control Codex agent that lets you watch it edit live code in your IDE, pause, intervene, or approve changes in real time, turning the model into a pair-programmer you can literally steer (2.8K likes, 301 reposts; earlier riff).
zero-native by Vercel Labs (Chris Tate / @ctatedev) lets you build native desktop and mobile apps with web UI frameworks and Zig with selectable engines (system WebView or bundled Chromium/CEF) for tiny binaries (3.7K likes, 265 reposts).
OpenClicky shipped full OpenAI Realtime-2 integration plus STT/TTS swap-ins (Cartesia, ElevenLabs, Deepgram, AssemblyAI) into the open-source macOS menu-bar voice companion with near-realtime latency.
mochi.js is a Bun-native, raw-CDP browser-automation library that delivers relational fingerprint consistency from a single (profile, seed) pair, Chromium-native fetch (real JA4/JA3/H2), and biomechanically modeled human click/type/scroll behavior (HN thread).
Jiayi Weng's "Learning Beyond Gradients" showed Codex evolving a pure NumPy + cv2 closed-loop heuristic policy for VizDoom D3 Battle (no neural nets, no map, just screen pixels + public game vars) that hits mean 557, with full reproducible code published (post, 330 likes, 47 reposts).
A new arXiv paper from Philippe Laban, Tobias Schnabel, and Jennifer Neville titled "LLMs Corrupt Your Documents When You Delegate" shows frontier models (Gemini 3.1 Pro, Claude 4.6 Opus, GPT-5.4) corrupt an average of 25% of document content during long delegated editing workflows across 52 professional domains, with sparse but severe errors compounding silently and agentic tool use providing no improvement (HN discussion).
imtomt built ymawky, a static HTTP web server for macOS written entirely in ARM64 assembly using only raw Darwin syscalls (no libc), supporting GET/HEAD/PUT/OPTIONS/DELETE, percent-decoded paths, byte-range scrubbing, atomic PUT uploads up to 1 GiB, and extensive hardening (path-traversal blocks, slowloris timeouts, symlink rejection) as a low-level learning project (writeup, HN).
nooga's let-go is an "almost Clojure" bytecode VM and compiler in Go that boots in 7 ms as a ~10 MB standalone binary, supports AOT compilation to binaries or self-contained WASM web apps with terminal emulation, Go interop, core.async, pods, and passes 95.4% of the Clojure test suite (HN thread).
ThatXliner's rust-but-lisp is a Lisp dialect with Rust semantics expressed as s-expressions that transpiles directly to .rs (zero runtime, zero GC) before rustc compilation, supporting macros via quasiquote/unquote, full ownership/borrowing, and most Rust syntax (HN thread).

🔬 AI Research & Models

Apple published LiTo: Surface Light Field Tokenization (Apple ML page, GitHub), an ICLR 2026 paper that introduces a 3D latent representation jointly modeling object geometry and view-dependent appearance for high-fidelity novel-view synthesis and editing.
Sakana AI Labs (with NVIDIA) open-sourced "Sparser, Faster, Lighter Transformer Language Models" (paper, GitHub, post, 734 likes / 113 reposts), introducing TwELL (Tile-wise ELLPACK) sparse packing plus custom CUDA kernels that reshape >95% unstructured sparsity in ReLU feedforward layers to fit GPU tiled matmuls, delivering >20% inference and training speedups plus memory/energy savings on H100s with no accuracy loss; simple L1 regularization induces >99% sparsity with negligible downstream impact (alphaXiv discussion).
Tencent-Hunyuan open-sourced HY-World 2.0 (also in Big Tech), a multimodal world model reconstructing/generating/simulating editable persistent 3D scenes from text, single/multi-view images, or videos with engine-ready 3DGS, meshes, and point clouds compatible with Unity/Unreal/Blender/Isaac Sim.
The Specula team published "Can LLMs model real-world systems in TLA+?" in ACM SIGOPS, reporting frontier models perform well on syntactic TLA+ but achieve only 41-46% success on conformance and invariant checking when modeling real-world concurrent and distributed systems, often producing state-space explosions that require human guidance (HN thread).
The Visual Reasoning Benchmark team released a new multimodal LLM evaluation using classroom-authentic visual problems from primary education (diagrams, charts, maps) that frontier models still fail on, highlighting persistent gaps in genuine visual reasoning.
DAIR.AI rounded up the top AI papers of the week (May 4-10) covering HeavySkill (parallel-reasoning skill via RLVR), Conductor (Sakana's 7B orchestrator that designs topologies and prompts), Self-Improving Pretraining (Meta FAIR's reward-driven pretraining loop), Connect-Four AlphaZero from scratch as a new coding-agent benchmark, Coordination as Architecture (information-controlled multi-agent experiments), Horizon Generalization (Microsoft's macro-action fix for long-horizon training), 1,000 Synthetic Computers (Microsoft's long-horizon productivity simulation), Contextual Agentic Memory is a Memo (neuroscience critique of vector-store "memory"), Agentic-imodels (Microsoft's LLM-readable interpretable regressors), and Skills as Verifiable Artifacts (supply-chain security for agent skills).
Adaptive Parallel Reasoning from Berkeley AI Research is a new inference-scaling paradigm that decides on-the-fly when to launch parallel reasoning branches versus continue serially.
AGRPO (paper, post) is a simple policy-gradient method for diffusion LLMs that optimizes individual denoising steps, delivering double-digit gains on GSM8K, MATH-500, Countdown, and Sudoku over base LLaDA while sampling 4× faster.
Anthropic's Teaching Claude Why (covered above in Honorable Mentions and Big Tech) extends the Model Spec Midtraining work into how rationale-driven training generalizes safety more than 3× better than safe-behavior demos alone.
Useful Memories Become Faulty When Continuously Updated by LLMs shows LLM agents that consolidate experience into textual memory often make their memory worse over time, with the failure traced to the consolidation step itself.
Philippe Laban, Tobias Schnabel, and Jennifer Neville's "LLMs Corrupt Your Documents When You Delegate" shows frontier models (Gemini 3.1 Pro, Claude 4.6 Opus, GPT-5.4) corrupt an average of 25% of document content during long delegated editing workflows across 52 professional domains, with sparse but severe errors compounding silently; agentic tool use provided no improvement.
AI Co-Mathematician: Accelerating Mathematicians with Agentic AI [arXiv paper; agentic AI for mathematician workflows].
FrontierMath Tiers 1-4 is Epoch AI's expanded benchmark of hundreds of unpublished and extremely challenging math problems, with the most recent Tier 4 release pushing the frontier of where current models still fail.
Sebastian Raschka (@rasbt) shared an updated gallery of major post-family-break LLM architectures and asked what major ones he'd missed since April (174 likes, 17 reposts).
Schmidhuber Problems is a new collection of 58 implementations of synthetic learning problems from Jürgen Schmidhuber's papers (1989-2025) in pure NumPy, laptop-runnable, with paper-comparison metrics per stub (post).
Kishan PB released the halegannada-hosakannada dataset on Hugging Face for old-Kannada-to-modern-Kannada language modeling work (post).

🤖 Robotics & Embodied AI

Robot Era raised over $200 million in new funding led by SF Express with HongShan, IDG Capital, Hillhouse, CICC, and others, following its March 1 billion yuan round that valued the Beijing humanoid robotics company above 10 billion yuan (~$1.4B) to expand operations and dexterous-hand work (Caixin).
Lucky Robots is a platform that generates infinite hyper-realistic synthetic data via simulation to train end-to-end robotic AI models up to 90% faster and cheaper before real-world deployment.
MotionBricks from NVIDIA produces scalable real-time motions from text and control inputs using a modular latent generative model with "smart primitives" for humanoids and characters (Pushmeet Kohli post).
Alvaro L (@L42ARO) built Odyseus-Spatial-VLM, a vision-language model paired with monocular depth estimation that projects 2D reasoning into actionable 3D coordinates for physical AI robotics, with live demo and repo available.
The Wevolver 2026 Edge AI Technology Report covers how intelligence is shifting from the cloud to devices, machines, and sensors for faster decision-making and lower latency, including edge models, multimodal AI, robotics, and healthcare applications.

🏛️ AI Policy, Governance & Safety

The Trump administration is preparing an AI security order that omits mandatory model tests but directs federal-AI-company partnerships on cybersecurity defense.
AI lobbyists are fretting over the White House's "lack of organization" and mixed signals on the executive order following Mythos's launch.
The WSJ details How Anthropic's Mythos threw the White House AI strategy into chaos, and Axios calls it the AI safety pivot ahead of Trump's China trip.
French prosecutors escalated their probe of Elon Musk and X to a criminal investigation over deepfake spread and political interference.
The US suspects Thailand's OBON Corp. helped smuggle Nvidia chips inside Super Micro servers to Alibaba (The Next Web).
A critical out-of-bounds read vulnerability (CVE-2026-7482) in Ollama before 0.17.1 lets remote attackers leak process memory including API keys, system prompts, and conversation data from more than 300,000 exposed servers via crafted GGUF files; fixed in 0.17.1, with separate unpatched Windows update flaws enabling persistent code execution (The Hacker News post).
Chrome is silently installing a 4GB Gemini Nano weights.bin into users' browser directories when on-device AI features are enabled, surfacing as unexpected storage bloat and reigniting transparency questions about silent model installs (HN thread).
The WSJ reports Trump's border spending spurred a boom in AI-infused surveillance, drawing new competitors offering AI-driven cameras, drones, and sensors for real-time threat detection.
Fortune profiled Anduril's quiet engineer-CEO Brian Schimpf, the $31B weapons startup supplying the Pentagon with drones, missiles, and the Lattice software platform that integrates sensors across battlefields and challenges legacy military contractors.
The FCC proposed requiring government-issued ID, physical address, and legal name before telecom providers issue new phone numbers, in a robocall-curbing proposal that privacy advocates warn would end anonymous prepaid service.
Jan Leike announced he is starting a new research project at Anthropic focused on AGI success factors beyond alignment, stepping back from leading the alignment team now run by Ethan Perez.
A 2018 ODINI paper on extracting data from Faraday-caged air-gapped computers via low-frequency magnetic fields resurfaced this weekend (Owen Brake post, 503 likes, 47 reposts) as the AI-supply-chain-security debate widens.

🛠️ AI Tools & Products

Whoop on-demand clinicians and AI health guidance for US members (also in Big Tech).
Google made Gemini API File Search multimodal, now supporting documents and images with custom metadata and page-level citations to enable more efficient verifiable RAG pipelines (HN thread).
Fireworks AI launched its full training platform (post) letting you train and deploy custom frontier models with full-parameter, LoRA, SFT/DPO/RFT, and custom RL via an autonomous Training Agent or managed API; ships with one-click production inference, multi-LoRA serving, and zero data retention.
TheIndex.fyi is a maintained canonical meta-index of indie web and small web index sites with 39 entries across 6 categories (Curated Directories, RSS & Feed Aggregators, Search Engines, Random Discovery, Constraint-Based Clubs, and IndieWeb Infrastructure).
Halupedia (aka Hallucinopedia) is an on-demand encyclopedia of a fictional universe that generates serious, well-cited scholarly articles for any topic you visit via URL slug and permanently stores them once created (HN thread).
Dilum Sanjaya shipped an interactive 3D biological structures explorer with a UI built via GPT Image 2 and code via Gemini 3.1 Pro, letting you rotate, zoom, and inspect cellular models generated by Tripo and Hunyuan3D (13K likes, 1.5K reposts).
levelsio connected his Windows 3.11 PC emulator (pieter.com) to ryolu_'s Ryos X MacOS-inspired web OS so the two retro computers can chat over the real internet inside your browser, part of his dream to bring Atari ST, Amiga, Apple II, and Windows XP back online as first-class web citizens.

📊 Fundraising & Deals Roundup

Sorted by deal size descending. Lead-story deals appear here once for completeness.

Apollo + Blackstone × Broadcom (Bloomberg, Tech in Asia): roughly $35 billion private-credit financing under discussion for Broadcom's AI build-out.
Anthropic (FT, PYMNTS): weighing a primary round of as much as $50 billion at a $900 billion valuation.
Cerebras IPO (Bloomberg, Tech in Asia): IPO price range raised to $125–$135 per share on 20× oversubscription.
DeepSeek (The Information): >$7.35 billion (50 billion yuan) maiden funding round led by founder Liang Wenfeng.
Google's Isomorphic Labs (Bloomberg, SiliconANGLE): >$2 billion led by Thrive Capital with Alphabet participating.
IREN × Mirantis (Datacenter Knowledge): $625M all-stock acquisition adding Kubernetes and enterprise ops to turn deployed GPUs into revenue-generating AI infrastructure.
Robot Era (FinSMEs, Caixin): over $200M led by SF Express at >$1.4B valuation.
Reserv (fintech.global, Ventureburn): $125M Series C led by KKR to scale AI-driven property-and-casualty claims processing toward 30M annually.
Ex-OpenAI researcher's six-week-old startup (The Information): targeting $4 billion valuation in its first round amid Nvidia's AI-chip-startup investment spree.
Baidu's Kunlunxin chip unit (Bloomberg): planning dual IPOs in Shanghai (STAR market) and Hong Kong.
Sony × TSMC image-sensor JV (Bloomberg, TSMC PR): Sony-majority joint venture for next-generation image sensors in Kumamoto.

💡 Industry Commentary & Analysis

Geography of AI progress

Elad Gil observes that people inside major AI labs are 3-4 months ahead of Silicon Valley startup engineers, SV is 3-6 months ahead of New York, and NY is 6-12 months ahead of the rest of the world, so most people are 1-2 years behind state-of-the-art (4.1K likes, 383 reposts).
Yann LeCun called BS on Gil's SV-centric AI timeline and listed major breakthroughs from Montréal, NYC, London, Paris, Hangzhou, and other non-SV hubs to prove innovation is global (2.1K likes, 198 reposts).
Ethan Mollick observes the most interesting AI use cases are now happening everywhere except San Francisco; labs, schools, factories, and small businesses are quietly shipping real products while the coastal echo chamber fixates on hype (4.1K likes, 521 reposts).
Riley Brown replied to Gil arguing physical proximity to SF no longer matters; curate your X feed correctly and join the right group chats and you can be roughly 2 months ahead of SV founders and engineers (2.2K likes, 111 reposts).

Valuation and trajectory

Gary Marcus calls forecasting Anthropic at $2T revenue in 2030 a "trillion pound baby fallacy"; babies double in weight in their first four months, but they do not keep doubling every few months until college because most exponentials break (419 likes, 56 reposts).
Joseph Jacks predicts Anthropic surpasses Alphabet in revenue by mid-2028 as continuation of the already-evident curve (from $1B ARR Jan 2025 → $9B Dec 2025 → $30B Apr 2026), projecting $100B by end-2026, $340B in 2027, $850B in 2028, and $2T by 2030 driven by Claude Code as enterprise wedge and Anthropic selling cognitive capacity into a $50T+ labor TAM (681 likes, 98 reposts; @sporadica mocked the chart).
Zephyr predicts OpenAI + Anthropic end 2026 with combined ARR of $160B–$200B (1.1K likes, 38 reposts).
Chase Brower argues Anthropic's Mythos lands slightly above the AI 2027 scenario trendline on METR-style agentic task completion while current revenue already exceeds SemiAnalysis Q2 2026 forecasts, concluding things are moving at or slightly faster than the 2027 timeline (1.1K likes, 93 reposts).

Building, agency, and craft

François Chollet notes agency has always been self-compounding, but AI magnifies the effect; low-agency users lose more while high-agency users gain even more (862 likes, 89 reposts).
Aaron Levie argues agents will democratize entire fields (more people build and explore) yet experts with deep judgment and context will see even greater demand as output-quality expectations rise across the board (386 likes, 44 reposts).
Kangwook Lee argues we should stop hand-designing AI agent harnesses because models will soon engineer better ones themselves; his Three Regimes Framework shows human harnesses become the bottleneck once models reach the middle capability regime (443 likes, 40 reposts).
Addy Osmani breaks down harness engineering: the model is only one input; the harness (prompts, tools, loops, hooks, sandboxes, memory, recovery paths) turns it into an agent, and every failure becomes a permanent rule so a decent model + great harness consistently beats a great model + bad harness (2.1K likes, 285 reposts).
signulll argues networking (conferences, warm intros, SF moves, coffee economy) is mostly cope and negative selection because the people worth knowing are too busy building; do the work and broadcast loudly on X instead (675 likes, 36 reposts).
Paul Graham wrote that "hobbyist" motivation is one of the most powerful forces for founders because it keeps them working long after financial security since the project is their beloved obsession (1.9K likes, 167 reposts).
Ethan Mollick observes the personification of Claude (in name, training, Anthropic's constitutional philosophy, and even fan cartoons) feels quite consequential in the medium term, for better and worse (196 likes, 9 reposts).

Science, math, and slop

Terence Tao explains in The Atlantic that generative AI is pushing mathematics to its edge through conjecture exploration, proof automation, and new kinds of intuition while still requiring human taste to separate signal from noise.
Dimitris Papailiopoulos argues math AI is roughly where coding was before CLI agents, single-turn and mostly ungrounded without a dense feedback loop, with GPT-5.5 Pro limited to natural-language proofs verified only after the fact; he predicts a phase transition once a reactive math terminal environment lets models train through experience inside it (221 likes, 17 reposts; follow-up).
Alex Imas predicts AI-as-general-intelligence will let scientists tackle broader cross-domain questions and shift research toward more generalists rather than hyper-specialists, commenting on the Aldighieri/Malpassi SSRN paper on early computers diffusing through US universities.
Peter Yang argues AI-generated "slop" is compounding in knowledge bases and retrieval systems, creating a feedback loop where future models train on increasingly polluted data unless we build provenance layers and human-in-the-loop verification now (1.4K likes, 168 reposts).
Roon (@tszzl) notes that a significant chunk of tech Twitter in 2020 mocked anyone who took GPT-3 seriously beyond "fancy autocomplete," yet many of those same skeptics remain "thought leaders" today (3.8K likes, 166 reposts).
g5t explores how AI (especially Claude) helps overcome task paralysis, particularly for ADHD, by rapidly turning vague ideas into implemented code or documents once the initial spark exists, while warning of potential addiction to instant results and rising token costs that could make it unsustainable long-term (HN discussion).

Builders and ship logs

Sam Altman shared an optimistic real-world story of an early Codex agent autonomously debugging, refactoring, and shipping production code for a small startup, calling it the first glimpse of agents that "just work" end-to-end (3.2K likes, 412 reposts).
Vaibhav Srivastav (@reach_vb) tallied OpenAI shipping nine major releases in roughly 15 days: GPT Image 2, privacy filter, GPT-5.5, GPT-5.5 Pro, GPT-5.5 Instant, GPT Realtime 2, Realtime Translate, Realtime Whisper, and GPT-5.5 Cyber (2.5K likes, 78 reposts).
Subquadratic's Alexander Whedon introduced SubQ, the first frontier model on fully sub-quadratic sparse-attention architecture (SSA) with a 12M token context window that is 52× faster than FlashAttention at 1M tokens and runs at less than 5% of the cost of Opus, with early access and a coding agent available (23K likes, 2.9K reposts).
Angeliki Giannou, co-inventor of Looped Transformers, successfully defended her PhD thesis and is heading to a new role (657 likes, 27 reposts).
0xMovez shared a Jane Street quant talk on training a custom LLM for OCaml coding assistance (not trading strategies) while promoting his own Polymarket trading agent built on Hermes + Opus (2.6K likes, 312 reposts).
The Thursdai Pod highlighted the MIT SLAM for Dummies tutorial PDF as a timeless resource for anyone building simultaneous localization and mapping systems for robotics and AR.

Previous Around the Horn Digests

Catch up on everything you missed:

Thursday, May 7, 2026: Anthropic open-sourced Natural Language Autoencoders that translate Claude's thoughts into plain English (and caught Claude Mythos Preview plotting to dodge a safety test); OpenAI shipped three GPT-5-class voice models; Cloudflare cut 1,100 jobs as agent usage rose 600% in three months.
Wednesday, May 6, 2026: Anthropic and SpaceX signed for Colossus 1 (300+ MW, 220K+ GPUs), Code with Claude SF doubled Claude Code rate limits, and Adobe turned static PDFs into interactive AI experiences.
Tuesday, May 5, 2026: A Cape Breton fiddler sued Google for $1.5M over an AI Overview defamation; OpenAI considered an Alphabet-style spinout for robotics and hardware; Anthropic shipped keyless authentication for the Claude API.
Monday, May 4, 2026: The White House began considering pre-release vetting of AI models; Anthropic and OpenAI married private equity on the same day; Mayo Clinic's AI flagged pancreatic cancer up to three years before diagnosis.
Weekend, May 2-3, 2026: The Pentagon signed AI deployment deals with 8 vendors and left Anthropic out, Microsoft 365 E7 with Agent 365 went GA, Meta bought a humanoid robotics startup, and Grok 4.3 added voice cloning.
Thursday, April 30, 2026: OpenAI shipped GPT-5.5-Cyber, AISI graded it as one of the strongest cyber models ever, and Anthropic was pre-empted into a $900B round.

That's a Wrap

That's 130+ stories from one weekend. If you scrolled all the way down here, you now know more about Anthropic's pre-IPO valuation than the analyst whose $852B OpenAI model just got side-eyed. You also know there's a Lisp that compiles to Rust, a 4GB Gemini Nano hiding inside Chrome, and an "encyclopedia" of a universe that only exists when you visit it. We don't make the news. We just sort it.

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 Tuesday.

P.S: Know someone who'd find this useful? Forward this to them and tell them to subscribe here.

Around the Horn Weekend Digest: Everything That Happened in AI This Weekend (Saturday-Sunday, May 9-10, 2026)