Tokens & Signals for 3/24/2026. We scanned ~605 Twitter accounts (1182 tweets), 13 subreddits (57 posts), Hacker News (18 stories), 9 newsletter posts, 8 podcast episodes, 252 Discord messages, and leaderboard data for you. Estimated reading time saved: ~14 hours.
TLDR
* 🚨 CRITICAL: LiteLLM supply chain attack. Versions 1.82.7 and 1.82.8 are compromised. Attackers (TeamPCP) poisoned a CI/CD action to steal SSH keys, cloud credentials (AWS/GCP/Azure), and drop backdoors. Pin to 1.82.6 right now. x.com/karpathy/status/2036487306585268612
* Anthropic's new "computer use" feature lets Claude take over your mouse, keyboard, and screen to navigate any desktop app. Out now as a research preview for Pro/Max users on macOS. x.com/felixrieseberg/status/2036193240509235452
* OpenAI is killing Sora — both the app and the API — just six months in, shifting that compute toward their next model, "Spud." The Disney/Pixar licensing deal is dead too. x.com/testingcatalog/status/2036533803888390290
* Claude Code just got "Auto Mode." A built-in risk classifier now auto-approves routine bash and file operations, so you're not clicking "allow" every thirty seconds. x.com/AnthropicAI/status/2036499691571953848
* @0xSero on Jensen Huang's AGI take: "Defining AGI as 'can build a $1B company' is a wild pivot from the sci-fi version we were promised." x.com/0xSero/status/2036386300681445774
* SK Hynix is dropping $8 billion on ASML equipment through 2027 to lock down HBM production. Hardware is officially the bottleneck now. x.com/jukan05/status/2036303653389344904
* Rumors are flying about a Microsoft vs. OpenAI legal fight over OpenAI running their new "Frontier" model on AWS instead of Azure. Grab your popcorn. reddit.com/r/OpenAI/comments/1s26o8m/grab_your_...
* Open-source 15B model daVinci-MagiHuman just dropped, and it's beating LTX 2.3 on quality while generating 5-second 256p video in 2 seconds flat on a single H100. huggingface.co/GAIR/daVinci-MagiHuman
* FlashAttention-4 is ripping on B200 chips — 1,613 TFLOPs/s, a 2.7x speedup. reddit.com/r/LocalLLaMA/comments/1s1yw23/flasha...
Best to Build With Today
* Coding — claude-opus-4-6-thinking-auto is the current top performer for agentic workflows.
* Reasoning — gpt-5.4-xhigh leads across math and complex reasoning benchmarks.
* Chat — gemini-3.1-pro is holding the top spot on the Chatbot Arena leaderboard.
* Video generation — daVinci-MagiHuman is the latest high-performance open-source contender to test.
* Value pick — gpt-oss-20B is your best bet for low-cost, efficient token usage.
Deeper Dives
💼 Industry & Business
* OpenAI Discontinues Sora
OpenAI is pulling the plug on Sora — the video app and API — just six months after launch. Compute is being redirected to the upcoming "Spud" model and internal robotics work, and a high-profile three-year partnership with Disney is gone along with it.
Why it matters: OpenAI is clearly betting on foundational model development over flashy, compute-hungry consumer video products.
� Twitter� Reddit
* SK Hynix Commits $8B to EUV Equipment
SK Hynix is spending roughly $8 billion on ASML's advanced EUV lithography systems to scale HBM production through 2027. It's ASML's largest order of this kind and a critical step toward the 1c DRAM process node.
Why it matters: High-bandwidth memory is still the biggest chokepoint for training the next generation of AI models.
� Twitter
* OpenAI Faces Potential Lawsuit from Microsoft
Things are getting tense. Microsoft is reportedly weighing legal action over OpenAI's plan to run its unreleased "Frontier" model on AWS Bedrock — which may violate an API routing clause in their partnership agreement.
Why it matters: This is the clearest sign yet that OpenAI is pushing for compute independence, and Microsoft isn't happy about it.
� Reddit
🧠 Models & Research
* Jensen Huang Declares AGI Has Been Reached
On the Lex Fridman podcast, Nvidia's Jensen Huang said we've hit AGI — then immediately defined it as the ability to "start, grow, and run" a billion-dollar tech company. He also walked it back a bit, noting current AI agents couldn't actually pull that off at Nvidia's scale today.
Why it matters: The industry's definition of AGI keeps drifting away from "human-level intelligence" and toward "useful enough to make money."
� Twitter� Reddit
* daVinci-MagiHuman Video Generation Model Release
A new 15B parameter open-source model that uses a single-stream Transformer to handle text, audio, and video tokens together. The inference speed is genuinely impressive — 5-second 256p clips in 2 seconds on one H100.
Why it matters: Open-source video generation is catching up to the closed-source stuff faster than most people expected.
� Twitter� Reddit
🚀 Products & Launches
* Anthropic Launches Computer Use for Claude
Claude can now actually use your computer — controlling the mouse and keyboard to open apps, click around browsers, and get things done. The research preview is live for Pro and Max subscribers on macOS.
Why it matters: This is the shift from "AI that talks about doing things" to "AI that does things across your whole OS."
� Twitter� Hacker News� Discord
* Anthropic Adds Auto Mode to Claude Code
Claude Code's new Auto Mode uses a risk-based classifier to figure out which shell and file commands are safe to run on their own. Less time clicking "approve," more time actually building things.
Why it matters: It's a smart middle ground — more agentic power without just throwing safety out the window.
� Twitter� Discord
🔥 Takes & Drama
* LiteLLM Package Compromised via PyPI Supply Chain Attack
Versions 1.82.7 and 1.82.8 were poisoned through a compromised CI/CD pipeline to steal SSH keys and cloud credentials. The bad versions have been quarantined — roll back to 1.82.6 if you haven't already.
Why it matters: When the whole ecosystem depends on a single Python package, it becomes an incredibly attractive target. This won't be the last time.
� Twitter� Reddit� Discord
Funding & Deals
* SK Hynix committed ~$8 billion (12 trillion KRW) to ASML for EUV lithography equipment through 2027. The investment supports the 1c DRAM node transition and HBM production for AI chips.
Launches
* Claude Computer Use — Anthropic's new capability letting Claude autonomously control desktop interfaces to handle complex tasks.
* Claude Code Auto Mode — A new permission tier for Claude Code that cuts down on manual approvals for routine shell and file operations.
* daVinci-MagiHuman — A new 15B parameter open-source audio-video generative model.
Closing thought: Agents are now touching our mouse cursors, and billion-dollar companies are lawyering up over cloud routing agreements. The "research" phase of AI is wrapping up fast — the "hard-nosed business" phase is already here. Keep your packages pinned, your dependencies audited, and your credit cards ready. The agents are clocked in.