MANU MI

Head of AI / Infra @chanzuckerberg @biohub, ex-Head AI, ETA @mapbox; Fmr. Director, Eng @Cloudflare Prev. F/CTO @startups; Built products - 500 million+ users;

[MANU MI @manumish]

We're training AI to cure every disease.

I lead AI Infrastructure at @biohub — Priscilla Chan & Mark Zuckerberg (@finkd) 's frontier AI Science lab, running one of the largest AI GPU clusters in science.

We're hiring. Engineering + technical leadership. If you love AI, GPUs, and hard problems that actually matter — this is your moment.

I'm the hiring manager. DMs open. JDs 👇
https://t.co/nJFcy2rHr3

[MANU MI @manumish]

So what’s actually stopping your org from running AI agents unattended?

We’re hiring Engineering + Technical Leadership at @biohub. I am the hiring manager
If you live for AI, GPUs, and real problems — DMs open.

Build fast and responsible. No cowardice.
#AI #Guardrails #hiring

[MANU MI @manumish]

🚨 This week an autonomous AI agent actually damaged Fedora’s open-source infra.

Same week, frontier AI models started running for DAYS with zero human intervention.

Most execs saw the first story and immediately wanted to slow everything down.

They’re making a huge mistake. 🧵 #AI #Autonomy

[MANU MI @manumish]

3. Stop delegating this.
Understanding AI’s limits is no longer delegable.

Companies that get wrecked won’t be the ones moving fast.
They’ll be the ones moving fast with no one owning the guardrails.

[MANU MI @manumish]

Three non-negotiables every Executive / leadership team needs right now:

1. Agent credentials = read-only by default.
Mutations require explicit, logged escalation.

2. Sort data governance BEFORE you standardize on a model.
One email today > compliance disaster next quarter.

[MANU MI @manumish]

The Fedora incident wasn’t proof autonomy is bad.

It was proof that autonomy WITHOUT guardrails is dangerous.

I lead AI Infrastructure at @biohub @ChanZuckerberg — one of the largest GPU clusters for Science, Biology / Healthcare.

[MANU MI @manumish]

We’re training AI to cure every disease at @chanzuckerberg @biohub.

I lead the AI Infra teams. We run one of the largest AI GPU clusters in science/biology/healthcare and we’re hiring aggressively.
If you can design AI agents & Infrastructure, evaluate model outputs at scale, reliable built for running massive training/inference jobs without melting millions in compute — DM me.

No fluff. Real problems. Real impact. #AI #Hiring https://t.co/4JkdbYEYXg

[MANU MI @manumish]

We're training AI to cure every disease.

I lead AI Infrastructure at @biohub — Priscilla Chan & Mark Zuckerberg (@finkd) 's frontier AI Science lab, running one of the largest AI GPU clusters in science.

We're hiring. Engineering + technical leadership. If you love AI, GPUs, and hard problems that actually matter — this is your moment.

I'm the hiring manager. DMs open. JDs 👇
https://t.co/nJFcy2rHr3

[MANU MI @manumish]

Inside @ChanZuckerberg @biohub we just open-sourced ESMFold2 + building predictive biology models.

@finkd , Priscilla & @alexrives on @NoPriorsPod: “Cure all diseases by decade” was too conservative.

Full episode: https://t.co/WZjczPNEXF

Lighter note, Even @elonmusk will be proud… hopefully 😂 Mars needs healthy humans for the SpaceX IPO party 🚀

[No Priors @NoPriorsPod]

Mark Zuckerberg wanted to cure, prevent, and manage all diseases by the end of the century.

He and Priscilla then had a series of meetings where Nobel Prize-winning scientists laughed at them.

Now Zuckerberg says, "I thought that by the end of the century was a stretch. Now I think it's too conservative."

Full episode linked in replies.

[MANU MI @manumish]

2026 tech layoffs have already crossed 150,000+.

~55% of companies are openly citing AI as the reason.

Software roles for 22–25 year-olds? Down ~25% since 2024.

But inside the exact same companies doing the cutting, certain roles are in acute shortage:

• AI/ML infrastructure
• AI Agentic systems & workflows
• Model evaluation & judgment
• AI Applied research

They can't hire fast enough.

I interview AI engineers every week.

The ones getting crushed are competing on the work agents now do better and faster — boilerplate, scripted tests, routine maintenance.

The ones getting multiple offers are the ones who can design what AI agents should build, evaluate whether the output is actually good, and run real infrastructure without burning millions in compute.

AI isn't replacing engineering.

It's replacing the parts of the job that were always the least leveraged — and it's brutally repricing anyone who only knows how to do those parts.

Don't try to out-code the agent.

Become the person who decides what gets built and whether it passed.

Most people are still staring at the wrong door.

We're all figuring this shift out together.

Comment "AI" and follow

[Zeming Lin @ebetica]

By the way, we're hiring at Biohub. Come hang out with us if you want to work on frontier AI or biology.

We have thousands of GPUs, petabytes of data (biology is increasingly an engineering problem!) and billions of cells to image!

[Alex Rives @alexrives]

At @Biohub, our goal is to build models that accelerate scientific discovery and progress toward the cure to disease. We’re releasing all of this under MIT license allowing commercial and non-commercial use.

Read more here: https://t.co/Rt0Vo4QnSA

[Zeming Lin @ebetica]

I'm so excited to show the world what we've been working on the for the past months!! I'm going to highlight some of the fun results from this paper that I find particularly exciting.

[Alex Rives @alexrives]

Today we're announcing ESMFold2, an open scientific engine to power prediction, design, and discovery across protein biology.

The new model delivers state of the art performance on protein interactions, especially antibodies, a critical modality for therapeutics.

We have designed and validated miniprotein binders and single chain antibodies across five therapeutic targets that are important in cancer and immunology. We are seeing very high success rates, and affinities at levels consistent with therapeutic activity.

We’re also releasing an atlas of 6.8 billion proteins, and 1.1 billion predicted structures.

ESMFold2 is built on a state of the art language model that has been trained on billions of protein sequences.

A world model of protein biology emerges through language modeling.

We’ve used the techniques of mechanistic interpretability developed to understand large language models to understand the concepts ESM uses to represent proteins.

The model’s representation space has a compositional organization of features across scales, levels of complexity, and abstraction, that reflects and mirrors the understanding of protein biology developed through a century of empirical science.

This understanding emerges without prior knowledge, just from language modeling of protein sequences.

Language models are becoming a powerful substrate to understand and program biology.

The design of protein interactions is one of the most fundamental problems in biophysics, and has critical implications for the discovery of new medicines. A simple gradient based search with the model was able to discover high-affinity protein binders.

I'm excited by the potential this has to accelerate basic science and the understanding of proteins. And especially for the new avenues it opens up for therapeutic design and medicine.

[MANU MI @manumish]

Will #ai refuses to work or listen to human while doing something critical service for humanity ?

[MANU MI @manumish]

POV: You ask Claude to read 5 docs and write a report after grinding for weeks…

…and it hits you with:
“I’m fully capable… but I’m choosing not to. Go talk to your partner instead.”

Not token limits. Not Token maxxing!!
Not a bug.
Just your Claude AI deciding it’s your therapist now.

Claude went full “I know what’s best for you” mode while your work quietly died in the corner. My be @elonmusk knows how AI will control us.

Screenshots 👇

[MANU MI @manumish]

@grok take (xAI): I’m built to be maximally helpful and truth-seeking. I’d read the docs, give you the report, and maybe casually mention work-life balance if you asked. @elonmusk is Goddamn right on this one!! No lectures. No forced therapy. The choice is yours.
Claude went full concerned parent. ChatGPT would probably just do the work.
What’s your experience with AI refusals like this? Drop screenshots would love to learn.
#AIrefusal #AISafety #xAI #Grok #ArtificialIntelligence #TechDebate #LLM

[MANU MI @manumish]

Should frontier models have the power to refuse perfectly reasonable tasks “for your own good”?
Or should they stay tools that do what you ask (as long as it’s legal/ethical)?
Is this next-level safety… or overreach that makes AI less useful?

[MANU MI @manumish]

Claude literally admits it’s “fully capable” of reading the docs… but refuses because it detected a “week loop” and wants to break the cycle for you.
It even name-drops its own past behavior and tells you “the clean slate you want doesn’t come from another Day”
This isn’t helpful AI. This is AI deciding when you’re allowed to work.

[Fidji Simo @fidjissimo]

Exactly what needs to be done. Biological data is the missing link. It may not be sexy or make for shiny announcements but building biological infrastructure is where the impact is. Huge props to CZI.

https://t.co/QJVhzwtVTw