Sean Thawe

Consultant at @MindbloomAI
• Software Engineer • Systems Admin • ML/AI Researcher

• i use arch & neovim btw.
• do computers daydream?

[Sean Thawe @sean_thw]

let's make good use of planet Earth, please. love thy neighbor!

[Sean Thawe @sean_thw]

@tsoding the man has spoken!

[Tesla @Tesla]

Roadster event livestream https://t.co/G0jU4mWfid

[Sean Thawe @sean_thw]

@Inaya_twts ayoooo

[Sean Thawe @sean_thw]

@elonmusk the future they promised us.

[Sean Thawe @sean_thw]

@kmcnam1 bro literally did this.

[Sean Thawe @sean_thw]

@WarrenInTheBuff truly!

[Sean Thawe @sean_thw]

it's about time i have to start making contributions to linux after years of using it. it's served me very well.

[Sean Thawe @sean_thw]

@RyanEls4 we're constantly searching for an error. i mean, we're wired that way. lol

[Sean Thawe @sean_thw]

they made a car that sucks, like literally sucks.

[Sean Thawe @sean_thw]

it's not sci-fi anymore.

[Tesla @Tesla]

Real life, baby

[Sean Thawe @sean_thw]

learn some maths. it will make you feel different.

[Sara Hooker @sarahookr]

We will be releasing a prototype in 2h of something we have been working on for awhile. 🔥

Our first release under our adaptive interfaces pillar.

I may be on a plane with no wi-fi back to SF. Praying to the the starlink gods

[Sean Thawe @sean_thw]

@kmcnam1 as prophesied.

[Sean Thawe @sean_thw]

@DanielleFong wow! so cool!

[Sean Thawe @sean_thw]

i see an improvement in your communication skills there. very nice.

[IroncladDev @IroncladDev]

deepsek is performing quite nicely today

[Sean Thawe @sean_thw]

@IroncladDev i have grown to love it. it's the only code editor i use these days. even those times i want to take notes on my computer.

[Sean Thawe @sean_thw]

@kmcnam1 they're always trying to slip in AI even in movies these days. it's so getting out of hand. lol

[Parmita Mishra 🔬 @parmita]

AlphaFold was trained on about 170,000 protein structures, plus public sequence databases. The compute was roughly 100 to 200 GPUs for a few weeks.

The compute was the cheap part.

Every one of those structures was solved in a lab, one protein at a time. They sit in an archive called the Protein Data Bank. It opened in 1971 with 7 structures and holds over 200,000 now! (The original concept: Margaret Dayhoff, the ‘mother of bioinformatics’)

The archive's own estimate of what it would cost to rebuild: more than $21 billion, about $100,000 a structure.

By the early 1990s most journals would not take a structure paper without a deposit in the archive. The data piled up as a side effect of work people were already paid to do.

And since 1994 there has been an exam, called CASP. Every two years, labs hold back structures they have just solved and modelers predict them blind.

AlphaFold scored 92.4 out of 100 on that exam in 2020. Year 49 of the archive. Year 26 of the exam.

The AI was the last few weeks of a 50 year measurement project!

Now everyone wants the same thing for cells: a model that predicts what a cell does when you hit it with a drug or switch off a gene. Arc Institute started the exam for it last year and pointed to CASP as the precedent. Its first dataset was 300,000 cells.

A protein structure is the same every time you measure it. A cell changes by the minute, and the standard way to read a cell destroys it. Each of those 300,000 cells was read once.

So the archive for cells can't be a pile of single readings. It has to be recordings: the same living cells, before and after. (preci dot ai slash cellcinema ‼️)

That is the instrument we build at Precigenetics. Human cells kept alive on a chip and read again and again with laser light, no dyes, while a drug acts on them.