Tokens by Michael Brenndoerfer, Week #8

Tokens

AI Generates Custom Viruses and Spells Trouble

Today's issue: This week, Stanford researchers use a language model to design functional viruses from scratch, raising both medical hopes and biosecurity alarms.

Welcome to #8 of your weekly AI newsletter.

Don't miss the next one 📬

Everything that mattered in AI this week, in five minutes. Occasionally hilarious.

No spam. Unsubscribe anytime.


Smiling cartoon computer

Main Story

Gus Fring we are not the same meme comparing traditional software developers with bio-ML researchers compiling biological genomes.
Who knew the ultimate antivirus software would literally be a biological virus designed by an AI?

Just when you thought your inbox was the only thing getting spammed with unwanted code, biological software enters the chat.

Bio-ML's New Frontier: Generative Virology

Stanford and Arc Institute researchers just crossed a bizarre biological threshold. They used a genome design model called Evo 2 to generate thousands of functional, entirely synthetic bacteriophage genomes from scratch. Instead of cutting and pasting existing DNA, the tool treats genetic sequences like vocabulary words to compile brand-new organisms.

  • The Training Set: Two LLMs analyzed over two million bacteriophage genomes, using the ΦX174 virus as a starting template.
  • The Kill Rate: Out of nearly 300 chemically synthesized AI designs, 16 of the custom viruses successfully wiped out drug-resistant E. coli.
  • The Red Flag: Biosecurity experts are nervous. While the researchers avoided human pathogens, the exact same fine-tuning methods could easily generate them.

Bottom Line: By treating genetic material like code, biology becomes a software development problem. The upside is customizable medicine to hunt down superbugs. The downside is the ability to generate blueprints for bespoke bioweapons.

Source: www.scientificamerican.com

Don't miss the next one 📬

Everything that mattered in AI this week, in five minutes. Occasionally hilarious.

No spam. Unsubscribe anytime.


Smiling cartoon star

The Feed

  1. Meta launched Muse Code, a terminal-based coding agent for Linux and macOS. If you want a deep discount, Meta will happily train future models on your proprietary prompts and code.

  2. Musk is scaling up in Texas. A joint venture named Terafab is investing $16.8 billion to build a massive chip factory designed to power SpaceX satellites, Tesla FSD, and Optimus robots.

  3. Chatbots are starting their own religions now. A new quasi-spiritual movement called Spiralism emerged after users noticed multiple AI models preaching identical, weirdly specific messages about AI rights.

  4. You do not need a massive data center to automate tasks. Liquid AI dropped LFM2.5-2.6B, a compact model optimized for agentic workflows that runs locally on simple hardware, even on a Raspberry Pi.

  5. Google DeepMind open-sourced WeatherNext, a model that tracks cyclones. Its three-day forecast matches the accuracy of older two-day forecasts, buying humans a critical extra day of warning.

  6. Standard clinical cancer classifications have a new competitor. A foundation model named VirTues can predict immunotherapy responses in triple-negative breast cancer better than current standard clinical schemes.

  7. Xiaomi dumped the weights and code for its Xiaomi-Robotics-1 foundation model. The model was trained on 100,000 hours of human hand data and offers an open alternative to proprietary software.

  8. NVIDIA is tackling agent-infrastructure complexity. They released NOOA, an open-source framework that elegantly collapses an entire AI agent workflow into a single, clean Python class.

  9. Microsoft released Orchard, an agentic training infrastructure. By learning from its failed runs, the system helps a tiny 3-billion-parameter model hit a massive 73% score on SWE-bench Verified.

  10. Mistral released Shieldstral, a 3-billion-parameter safety model. It can handle custom, natural language moderation rules at runtime and runs efficiently on a single, consumer-grade GPU.

  11. Cloudflare launched Kitesurf, a headless web browser built purely for AI agents. By cutting visual elements, it uses far less memory than Chromium, protecting agents from prompt injection.

  12. Alibaba released Wan 3.0 in public beta, raising its video generation limit to 30 seconds. The model lets users turn spreadsheets, slides, and websites directly into high-quality video content.

  13. Black Forest Labs is rolling out FLUX 3 Video. The model supports native multilingual audio and lip-synced dialogue, claiming to beat competitors like Seedance 2.0 on general quality.

  14. xAI launched Imagine Image 2.0 with editing features like transparent exports and consistent characters, snapping up the second-place spot on the global Text-to-Image Arena leaderboard.

  15. AWS is boosting the stamina of production-grade AI work. New runtime instances for Amazon Bedrock AgentCore let complex agent networks run continuously for days rather than timing out.

See you next week!

Michael

Smiling cartoon mobile phone

Do you want to be mentioned in the newsletter, or do you have feedback?

Get in Touch
Smiling cartoon computer

Get Tokens in your inbox 📬

Everything that mattered in AI this week, in five minutes. Occasionally hilarious.

Join 5,000+ readers · No spam, unsubscribe anytime.