Good morning. It's Wednesday, October 21, and we're covering who should control superintelligent AI, Claude’s scorpion toxin research, and a breach exposing OpenCode users’ AI keys.
Plus: yesterday’s poll results on your preferred AI assistant, a new Real or AI challenge, and an article on how AWS is challenging NVIDIA’s CUDA.
Attackers breached an older system used by OpenCode Console, an online service connected to the open-source OpenCode coding agent. The incident affected 1.4% of Console users and exposed some API keys, leading to unauthorized AI charges that reached hundreds of dollars and, in a few cases, more than $1,000.
The U.S. government ordered civilian agencies to stop buying DeepSeek products and remove the Chinese company’s AI from federal systems, citing data and national-security risks. The order also covers third-party services built with DeepSeek technology, although agencies may seek exceptions.
Muse, Meta’s AI companion app, passed 10 million U.S. downloads in its first 40 days. Only 11 other apps have reached that mark as quickly, according to the data cited by TechCrunch, and Muse had 2.5 million daily active users on October 16. The next test is whether people keep returning after the launch surge fades.
POWERED BY FACTORY
Make Your Software Improve Itself
Factory is a self-improving system for your software development lifecycle. Droids write, review, validate, deploy, document, and monitor software across the full SDLC.
Enter Factory’s giveaway for a chance to be selected to win a free month of its $200 Max plan and put AI Droids to work in parallel — planning, coding, reviewing, and validating software. → Enter to win
FORWARD FUTURE ORIGINAL
Torch This: Can AWS Finally Bridge NVIDIA’s CUDA Moat?
Somewhere in Santa Clara, NVIDIA CEO Jensen Huang is probably standing atop the ramparts of NVIDIA's magnificent castle, surveying the enormous and technically deep moat his company has spent nearly two decades digging around its GPU empire. It's deep, wide, stocked with crocodiles named cuBLAS and cuDNN, and guarded by thousands of software engineers who speak fluent CUDA — NVIDIA’s proprietary set of software products that translate PyTorch code into AI-ready applications. → Read the full article here.
AI GOVERNANCE
Altman Lays Out Three Endgames for Powerful AI
OpenAI CEO Sam Altman told employees that the AI industry has discussed three ways to control systems more capable than humans, according to The Information. The options include an international regulator modeled on the nuclear watchdog IAEA, a shared research organization similar to CERN, or one dominant company with enough economic power to deny computing resources to rivals—an outcome Altman called terrible.
The discussion is becoming practical rather than theoretical as the U.S. government begins building a system to oversee superintelligence developers. The choice will determine who can inspect frontier models and whether competing labs are expected to share control before any one of them becomes too powerful to contain. → Read the full article here. (Paywall)
BIOTECH
Claude Designs Proteins to Block Scorpion Toxins
Anthropic asked an internal Claude research model to design 318,484 minibinders—small proteins intended to attach to scorpion neurotoxins. Researchers tested 211 designs in the lab, and 17 met their criteria against individual toxins; the strongest binders attached at single-digit nanomolar concentrations, a promising early laboratory result.
This is not an antivenom yet: Anthropic has not shown that the proteins neutralize venom in cells, animals or people. The next step is biological testing, but the project shows how an AI model can narrow an enormous design space into a small set of candidates that scientists can test, potentially shortening the first stage of drug discovery. → Read the full article here.
HEADLINES
In Other News
Mistral Launches Voxtral 3 Voice Agents: Mistral released Voxtral 3, a voice-agent toolkit with speech models for more than 50 languages.
Kling AI Adds Longer Videos and Stereo Audio: Kling 4.0 supports 30-second videos, ten keyframes, stereo audio and multilingual lip sync.
Electric Plant Raises $11.2M for Plant AI: Its Bombadil model interprets electrical signals from about 125 plant species.
Yelp AI Calls Restaurants for Wait Times: Yelp's voice agent calls restaurants and reports current wait times in the app.
Judge Orders Meta to Name AI Training Decision-Makers: Meta must identify executives who approved book downloads and produce related logs.
REAL OR AI
Can You Still Tell What’s Real?
POLL RESULTS
Go-To Personal AI Assistant?
Here's how you voted: Nearly half haven’t picked a personal AI assistant yet, while Grok Bot and Claude Cowork tied at 19% each.
🟨🟨⬜️⬜️⬜️⬜️ Grok Bot – 19%
🟨⬜️⬜️⬜️⬜️⬜️ Muse – 9%
🟨⬜️⬜️⬜️⬜️⬜️ Dots – 8%
🟨🟨⬜️⬜️⬜️⬜️ Claude Cowork – 19%
🟩🟩🟩🟩🟩🟩 Other / I haven’t started yet – 45%
