😸 AI just decoded the 98% of DNA scientists gave up on

Welcome, humans.

Sundance premiered The AI Doc this week; sounds like its worth checking out once it hits streaming. The film lands halfway between doomerism and optimism ("apocaloptimism" they call it; we call it The Neuron's editorial mandate) and offers a concrete path forward for AI safety:

international coordination akin to Cold War nuclear agreements, mandatory disclosure of AI-generated media, legal liability for AI companies, an independent regulatory body, and rules that adapt as tech evolves. Shall we call them… generative regulations?

Here’s what happened in AI today:

• Google’s AlphaGenome can now read your DNA and improve science.

• Microsoft gained $7.6B from its OpenAI investment last quarter.

• Amazon laid off 16K corporate employees amid AI investments.

• Chrome added agentic browsing features and tighter Gemini integration.

LATER TODAY 🔴 (at 10AM PST | 12pm CT | 1PM EST | 6PM GMT): Join us LIVE as we run some new open-weight AI models in real-time with a special guest appearance from Ece Kamar, CVP at Microsoft Research, discussing their Fara-7B model.

Plus, Corey might also show off his local clawdbot install. Watch live here.

Google DeepMind Can Now Read 98% of Your DNA That Scientists Couldn't Understand

For decades, scientists have stared at 98% of your genome like it's written in an untranslatable language. The other 2%? That's the protein-coding region; the part we understand. The vast majority? Total mystery.

AlphaGenome, Google DeepMind's new AI model (released last June, paper dropped in Nature yesterday), might finally change that.

Here's the deal: Your genome is the "source code of life," as the team describes it. Small variations can determine whether you develop a disease, but we've been mostly blind to what those variations actually do in non-coding regions.

AlphaGenome reads up to 1M DNA letters at once and predicts what happens when you change even a single one. Think of it like spell-check for genetic mutations, except instead of flagging typos, it tells you if that mutation might cause cancer or a rare disease.

What makes it special:

• It sees the whole picture. Previous AI models had to choose: zoom in for detail, or zoom out for context. AlphaGenome does both: megabase-scale reading at single-letter precision.

• It models splicing for the first time. Splicing is how cells stitch together gene instructions. Screw it up, and you get diseases like cystic fibrosis. AlphaGenome can predict when that'll go wrong.

• It already works. Researchers fed it cancer-associated mutations, and it correctly identified which ones were driving the disease—matching what wet lab scientists found through painstaking experiments.

One team member put it bluntly: AlphaMissense (their previous model) covered the 2% of your genome that codes for proteins. "But what about the rest of the 98%?" That's what AlphaGenome tackles.

Where this goes next: Many rare genetic diseases remain undiagnosed because doctors can't figure out which mutation is causing the problem. AlphaGenome could narrow down the culprit, potentially unlocking diagnoses (and treatments) for conditions that have stumped medicine for years.

If you're a researcher, the model is available via API now.

FROM OUR PARTNERS

Agents that don’t suck

Are your agents working? Most agents never reach production.

Agent Bricks helps you build high-quality agents grounded in your data. We mean “high-quality” in the practical sense: accurate, reliable and built for your workflows.

Generic benchmarks don’t cut it. Agent Bricks measures performance on the tasks that matter to your business.

Evaluate agents automatically, and keep improving accuracy with human feedback. With research-backed techniques for building, evaluating and optimizing, you can turn your business data into production agents faster — with governance built in from day one.

See how Agent Bricks works

Prompt Tip of the Day

A new pattern has gotten popular versus trying to craft “the perfect prompt.” Make AI create a plan first, that you approve, that it then executes

This “plan/spec mode” (in tools like Claude Code) constrains where AI improvises, allowing you to define boundaries upfront instead of fixing problems after.

If you’re not ready to try a CLI tool like CC or Codex, you can actually recreate plan mode in ChatGPT/Claude/Gemini with prompting, since they don't have built-in plan modes yet. Here’s a starter version you can edit and work off of:

"Before starting [task], work in phases:

1. Ask clarifying questions about [request/context]

2. Propose your approach for [how to do it]

3. Review your initial plan to ensure it aligns with [my goals]

4. Based on your review, write a final plan showing: [what you'll do], [steps], [potential risks]

Keep it scannable but detailed. Wait for my approval before executing."

Want more? Check out our Prompt Tip of the Day Digest for January.

Treats to Try

1. Harvey researches case law for your arguments and reviews contracts by analyzing legal databases and your firm's documents (raised $160M at $8B valuation)—starts at $1,200/lawyer/month.

2. Rogo is like the Harvey for investment banking: it builds Excel models, investment memos, and pitch decks for your deals by pulling from your firm's systems and financial databases (raised $75M)—supposedly its about $275/month per user (so $330 a year), which is higher than typical SaaS but in line with specialized financial software.

3. Outtake finds and takes down phishing sites and fake accounts impersonating your brand across domains, social media, and app stores (raised $40M).

4. NVIDIA Earth-2 offers open-source models that generate 15-day weather forecasts in minutes (including local storm predictions) instead of the hours traditional supercomputers require—no pricing details (open-source)

5. Twin automates repetitive workflows by taking over your browser and completing tasks like a human, such as clicking buttons, entering data, navigating between apps (raised $10M); they say they are the “safe” Moltbot…

6. LemonSlice turns any photo into a FaceTime-style video call character that responds in real-time at 20fps (upload a picture of a person, alien, or cartoon and start talking) (docs, try Taylor, try alien)—paid only rn ($0.12-0.20/min).

7. skills.sh from Vercel teaches AI agents procedural knowledge through simple one-line commands—25K+ installs across 565+ skills, works with Claude Code, Cursor, and 14 other agents.

Around the Horn

1. Microsoft gained $7.6B in net income from its OpenAI investment last quarter, with commercial obligations jumping to $625B (45% from OpenAI alone).

2. Amazon laid off 16K corporate employees, its second major cut in three months; CEO Andy Jassy said AI-driven automation will mean “we will need fewer people doing some of the jobs.”

3. Chrome launched agentic “auto-browse” features that can shop, find coupons, and fill forms for you; rolling out to AI Pro and Ultra subscribers in the U.S.

4. Tesla invested $2B in xAI despite shareholders voting against it in November; the companies signed a framework agreement for AI collaborations including Optimus robots.

5. Zuckerberg teased agentic shopping tools and “personal superintelligence” on Meta's earnings call, with Meta capex jumping to $115-135B for 2026.

FROM OUR PARTNERS

Meet Slackbot, your personal AI work agent 

Want to work smarter, not harder? Slackbot streamlines how you handle your day-to-day work. Connected to your third-party apps? You bet. Knows your context, your tone, and can access your permissioned messages and files? Absolutely.

Thursday Trivia

One is real, and one is AI. Which is which? Vote below!

A.

B.

P.S: We’ll share everyone’s answers to this on Sunday, so write in what gave it away!

A Cat’s Commentary

Trivia answer: B is AI, and believe it or not, but A is real (at least according to the Reddit sleuths).