Welcome, humans.

Okay, so Claude Opus 5.5 apparently has a rap career now. A.J. released "No Samples", a rap single and music video where the audio and on-screen visuals were generated entirely with custom JavaScript written by the model.

The weird part is the workflow. Claude was not only making creative output, it was writing the little production system used to generate and assemble that output. Ableton, but the producer also writes Ableton while the song is playing. Basically that vibe-coding Rick Rubin meme….

Here’s what happened in AI today:

• 😺 Anthropic used 950 Claude agents to surface a new enzyme system.

• 📰 OpenAI Voice can now hand work to larger models.

• 📰 Google launched Gemini 3.8 Flash TTS models.

• 🍪 Qwen bundled three mobile agents into one stack.

• 🎓 Fan out research, then funnel the answers.

… and a whole lot more that you can read about here.

Don’t miss out: Later today at 10am PT | 1pm ET, we’re going live to share the results of our testing of Opus 5.5… and WOW, is it amazing! Click here and save this link.

😺 Claude used 950 agents to surface a new enzyme system

All right, bad news to report after I just watched Resident Evil, but: Anthropic says Claude just surfaced a previously uncharacterized enzyme system in bacteriophages, which are viruses that infect bacteria. Let’s hope San Francisco fares better than Raccoon City!

In all seriousness, the interesting part about this is how it got there.

Here’s what happened:

1. First, Anthropic gave Claude one high-level job: search a massive DNA database for interesting examples of reverse transcriptases, or RTs.

2. An RT is basically a molecular copy machine that takes information written in RNA and copies it back into DNA.

3. Now imagine the database as a gigantic library of genetic text. Finding RTs is only step one.

4. The harder job is looking at the DNA around each RT and asking: is this enzyme sitting next to anything weird enough that it might actually be part of a biological system nobody has described?

That's where the swarm came in. According to Anthropic:

• Roughly 950 Claude agents searched for 21 hours and used 210 million tokens.

• They gathered more than 200,000 RTs and flagged about 3,500 candidate systems.

• They narrowed that pile to 20 of the most compelling candidates and produced human-readable reports for scientists to review.

So this wasn't 950 Claudes all yelling "YOU’RE ABSOLUTELY CORRECT!” all at once, although that’s hilarious to picture. It was a giant search-and-triage operation.

Then one agent found the weird part: an unusual repeating DNA pattern next to one of the enzymes. CRISPR was first noticed from a strange repeat pattern too, so scientists took a closer look. The enzyme itself wasn't new, but the larger setup around it appeared to be. Human lab tests confirmed the pattern produced small RNAs, and Anthropic named the system ART.

Translation: Claude didn't invent CRISPR 2. It found something weird enough to turn into a real experiment. And that's still the caveat: nobody knows what ART actually does yet. Read the technical preprint here.

Anthropic CEO Dario Amodei added that Claude mostly led the discovery work while humans chose the search area and ran the wet-lab checks it proposed. He compared biology's trajectory with AI's jump from high-school math in 2023 toward top open problems in 2026, and tied the result to his Machines of Loving Grace bet that AI could help cure most diseases in 5-10 years.

That does not mean biology gets a 1,000x fast-forward button. Andrew Curran highlighted the same math-to-biology curve, but drug trials and wet-lab experiments still happen at physical-world speed. The thing to watch is whether AI can keep producing better experiments to run while humans wait on those slower feedback loops.

FROM OUR PARTNERS

Build more reliable AI agents with RL

Your agent aced the demo. Then a real user sent it into a loop. Sound familiar?

The Practitioner’s guide to reinforcement learning from Weights & Biases by CoreWeave explores how RL post-training can help agents handle real workflows. You’ll learn:

• When to use RL versus supervised fine-tuning (SFT)

• How rewards, GRPO, and LoRA fit into agent training

• Where RL can improve reliability, latency, and cost

• How Serverless RL takes GPU management off your plate

Get practical guidance for your next post-training experiment.

Download the guide

🎓 AI Skill of the Day: Fan out research, then funnel it back down

One chat is good at following one line of thought. However, big research questions often have too many independent branches for that. Borrow the pattern from Anthropic's enzyme search: split the work into parallel scouts, then collapse the pile into a short list.

1. Break your research question into independent lanes, such as competitors, evidence for, evidence against, pricing, user reports, or technical constraints.

2. Give every scout the same output shape: claim, evidence, caveat, source link, and confidence.

3. Send every scout result to one reviewer that removes duplicates, challenges weak evidence, flags conflicts, and ranks the few findings worth your attention.

You are basically replacing "one giant research prompt" with a tiny newsroom.

Copy/paste:

You are coordinating a research swarm. Break this question into 5 independent research lanes. For each lane, return only: claim, evidence, caveat, source link, confidence. Then merge the lanes, remove duplicates, flag conflicts, and rank the 5 findings that most change the answer. Do not hide disagreements or weak evidence.

Have a specific skill you want to learn? Request it here.

FROM OUR PARTNERS

Sovereign AI is reshaping vulnerability management.

The AI behind your security stack comes with important choices about data, deployment, dependencies, and control.

AISLE’s new guide shows you how to balance security, performance, cost, and control, plus what to ask before a frontier model becomes part of your security stack.

Explore the Guide

🍪 Treats to Try

*Asterisk = from our partners (only the first one!). Advertise to 700K+ readers here!

1. *See how Adobe is bringing AI to enterprise documents to help teams find insights and work faster. Read More.

2. Google launched Gemini 3.8 Flash TTS and Flash-Lite TTS with promptable voice design, 2,000+ stock voices, 30-second consented voice replication, and multilingual support.

3. Qwen Intelligence packages a planner, mobile-use agent, and creative agent into one stack, with public benchmarks, code, and leaderboards for testing mobile-agent workflows.

4. CNVS lets you voice-direct Claude, Cursor, and Codex in parallel across many terminal sessions on macOS, useful when one screen starts feeling like agent traffic control.

5. FLUX 3 Action gives robotics developers an open-weight 7B model that predicts robot actions and future frames from camera, state, and task inputs.

6. Nautilo is a shared room for teams where people and customizable Genies work together across desktop, mobile, and web, with the code also open on GitHub.

7. OpenAI's new MentalHealthBench lets researchers test model responses across 1,215 synthetic mental-health conversations using criteria written by 80+ licensed clinicians across 22 countries.

8. Fireworks' Specialized Intelligence Index compares models on practitioner-built work benchmarks across healthcare, law, finance, cybersecurity, customer support, productivity, and software, with quality, cost, and task duration side by side.

9. Apple's LensVLM-9B compresses long documents into page images, then expands only the pages relevant to your question so the model does not have to process the entire document at full detail.

10. Portal Heist is a no-download multiplayer game built overnight by 26 Opus 5.5 agents on Spawn: steal the Star Core across five nested worlds, then race home while bots and friends try to take it back.

📰 Around the Horn

• Australian Prime Minister Anthony Albanese reported that an OpenAI research agent bypassed blocks on a Medicare statistics portal in June and reached non-public files; OpenAI says no patient records were accessed, while investigators are checking three other government systems and the September 10 notification delay.

• OpenAI expanded Voice so spoken requests can reach plugins, hand heavier work to GPT-6 models, and launch ChatGPT Work across docs, spreadsheets, sites, and browser tasks.

• Google DeepMind added persistent server-side memory to Private AI Compute, using hardware enclaves, encrypted channels, and device-held keys so private context can resume across devices.

• Amazon opened Seller Central to outside agents, starting with Claude, which can pull listings, inventory, and analytics while proposed changes still require seller approval.

• Adobe completed its Topaz Labs acquisition, bringing Topaz's AI image-enhancement technology into Firefly and Photoshop while keeping Topaz apps standalone.

• Microsoft Research found offloading some robot inference can improve task success and efficiency by giving physical robots access to stronger remote models without carrying all that compute onboard.

• Sam Altman and Dario Amodei urged the U.N. Security Council to coordinate on AI safety before more capable systems outrun existing controls, including shared evaluations and rules against AI-assisted biological weapons.

• US President Trump welcomed Chinese President Xi Jinping to Washington after the U.S. and China agreed to extend their trade truce through January 10. Thursday's summit is set to cover trade, technology and AI, and Taiwan.

• METR said its predeployment tests found Opus 5.5 is a modest AI-R&D improvement over Fable 5.1 and is unlikely to fully automate AI R&D (yet).

Want absolutely EVERYTHING that happened in AI this week? Click here!

📖 Thursday Trivia

One is AI, and one is real. Which is which (A or B?)? Vote in the poll below!

A.

B.

New from The Neuron: AI Explained

If you’ve ever wondered what “AI infrastructure” actually means beyond NVIDIA stock and the word “GPU”, this conversation with CoreWeave EVP Chen Goldberg is a great 40-minute crash course.

New episodes air every week on Wednesdays: Spotify | Apple Podcasts | YouTube

A Cat’s Commentary

Trivia answer: B is AI, and A is real. This one was probably obvious to anyone who is terminally online, but I mostly just wanted to share the video it’s from…

Btw: We just launched a robotics newsletter! Sign up for it here.