might go for this look

Yes. LLM output is inherently political! So is social media output, search results, other generative computer interfaces. I don’t know how to govern this stuff but we’ve got to figure out better internet regulations

not to dunk on this conclusion, but it's an insidious mindset that has become pervasive. how you spend your time, how you make your money, is one of the most important ethical and political choices you can make. convincing ourselves otherwise is arguably The Whole Problem in American society today

employers can flip the script and turn damaging layoffs into an AI productivity story. satisfying for now, maybe, but it's the kind of behavior that'll exacerbate a potential future collapse

a race draws maximum investment early on. it’s an effective lever against the USG to deregulate US companies and restrictively regulate Chinese ones. A race implies a winner, and the prize is Control. It is not looking like an accurate description of AI progress.

big moves! so that's Parse -> Qiagen Scale -> 10X Genomics Fluent -> Illumina next-gen single-cell cell omics, capable of processing 10^6 to 10^8 cells i wonder if what the near term looks like - will commercial kits stagnate while the buyers rebrand?

o wow that James Zou paper was preprinted a year ago! i wonder if training on synthetic wrong answers could give an abundant, although sparse, training signal? also, it looks to me like performance is not consistent with model scale or capability? arxiv.org/abs/2410.21195

gonna have a hard time with this one at current LLM capabilities because www.nature.com/articles/s42...

if you have used these systems for challenging work, I don’t know how you can come to this conclusion. and the New Yorker is not your venue of choice. You want to do a blog post and try to use it to scruff up another $100B

In which we all learn the phrase “ballistic microscopy”:

AGI is such a funny concept to adopt. "no i mean, like, actually AI. like, actually"

yes, this is the width of the bullseye hyperscalers need to hit to avoid bursting the bubble. They need LLMs to become much more effective (to create value) while keeping compute costs high (to capture value)

so who's the uh Albert Speer in this situation?

I often feel like OpenAI doesn’t understand the bet they’ve placed.

This isn’t how you get to AGI. This is the path to a $10B LLM financial consultant, not a superintelligence worth trillions. This training regime doesn’t even make sense in a world with AGI on the horizon.

the quality of an idea is very poorly correlated with how long it took you to come up with it fairly well correlated with how much time you spend working on it

we should try to make mirror life

you have got to be kidding me. (also love the effect on the sign that makes it look like a party invite)

our communications system is terrible for every user. completely co-opted to blast spam, scrape info, sell ads, and run scams. so many of America’s problems encapsulated in this one slice of life

Dwarkesh-Katpathy interview was quite good. Andrej is such a clear thinker, very well positioned to offer an honest and informed AI outlook. I share his view on the state of the field, but I’m more pessimistic about the intersection of big tech, AI, and American capitalism

allowing adult content means LLMs don't have the juice

mess around in the lab or on the computer?

you know i'm down bad when i'm scrolling the ResearchGate forums. just completely out of my element. flailing.

Using LLMs for science is like scrolling posts on some niche labrat internet forum. and not a good one like SeqAnswers. i'm talking some dicey ResearchGate advice. some r/DIYbio. like you almost wish you hadn't read it.

"I’m talking about using AI to perform, direct, and improve upon nearly everything biologists do." "my basic prediction is that AI-enabled biology and medicine will allow us to compress the progress that human biologists would have achieved over the next 50-100 years into 5-10 years." Dario Amodei

my goodness what a fear-mongering, over-the-top hype headline. really leaning into telling lies to sell newspapers

web developers using a stochastic generative algorithm: omg it's alive u guys. it has rights

LLMs have been bricked for biomedical research. OpenAI deactivated my account ('weapons'). Claude won't engage question: in the last decade, how many hospitals have been held for ransom with chemical, biological, or radiological weapons? and how many have been held for ransom by cyberattacks?

it's weird that the web browser has taken over desktop but phones have been colonized by apps. are apps better for thumb-based GUIs? or gathering more user telemetry? or maybe the desktop is just a portal for search and cloud access?

between global researchers choosing the EU over the US, Europeans doing the same, and Americans making the jump.. the EU is going to be flooded with researchers. it's a historical opportunity and Europe would do well to take full advantage

it doesn't have to be like this.

Cryo-electron tomography reconstructs polymer in liquid film for fab-compatible lithography pubmed.ncbi.nlm.nih.gov/41027883/ #cryoem

the Gemini button on Chrome browser. you tell it to go away it becomes a slightly smaller button worst demand vs capital investment of any technology in history

mesmerizing wagon-wheel effect using hi-speed camera and droplet generator. droplets enter the channel from the top, flowing counterclockwise around the bend, a mixing feature. droplet generation rate is ~1kHz, droplet size ~80 microns

email is awful. whether you prefer email depends on if the work you are doing is better or worse than awful.

Super excited to get this out. This collab started a few years ago and is the first paper from it. Here, with experimental and computational approaches we: 1. establish that cell villages can be just as accurate (one might argue more accurate!) than arrayed-based designs bsky.app/profile/bior...

an LLM bubble doesn’t mean LLMs aren’t useful. It means investment has outstripped ROI. This has happened. if capabilities don’t improve rapidly, the bubble might burst.

once you see how often software developers assume the language of science and engineering, you can't unsee it

cool new paper on designing proteins against immunogenicity by incorporating human k-mers and hidden k-mers (not presented on MHC-1) into proteinMPNN designs. i've been thinking about this for a while - cool to see an implementation! pmc.ncbi.nlm.nih.gov/articles/PMC...

it was confirmation that we will be stuck in this cycle for a long time. MAGA will be competitive with uhh objective reality for the foreseeable future and our national nightmare will continue. post-truth is a structural outcome. the margin by which it wins or loses is secondary

enjoyed the Dwarkesh-Sutton interview. i largely agree with Sutton re: continual/experiential learning and goal orientation Dwarkesh showed more of his incomplete grasp of human biology, tendency to apply it to AI. Sutton course-corrected. think D is learning and i'm happy for him.

sure.. but losing to Trump, for the second time, after a farcical pandemic response, an insurrection, 34 felony convictions, and everything else.. warrants deep reflection on our national identity and political-economic system

Google's mission statement has been on my mind: "organize the world's information and make it universally accessible and useful" so much depends on whether LLMs can be wrangled to be both generally useful and highly reliable, or if that tension is irreconcilable.

GPUs last 2-4 years, data centers 15-20. To justify trillions in ephemeral hardware, AI must deliver immediate, world-changing returns. Big tech’s “success” scenario isn’t growth. it’s breaking the economy and rebuilding it for themselves.

You’ve heard of ubiquitination, meet deazaguanylation: Doug Wassarman in our lab discovered phage defense pathways have co-opted Q nucleobase biosynthetic enzymes to catalyze a new form of protein conjugation chemistry @science.org www.science.org/doi/10.1126/...

New method 🚀: scEpi2-seq lets us read DNA methylation+histone marks in the same single cell. Found that methylation “memory” after replication depends on chromatin+histones, and TF promoters repressed by H3K27me3 carry distinct methylation. So stoked to share this as first+co-corresponding author 🎉

some cool papers on the challenge of creating train-test splits for protein deep learning models Constructing benchmark test sets... journals.plos.org/ploscompbiol... Cracking the black box of deep sequence-based protein–protein interaction prediction: pmc.ncbi.nlm.nih.gov/articles/PMC...

The other end of this is super interesting to me: with custom architecture, performant protein structure/design models can be trained on just 10k structures, <10% of PDB! protein sequence-structure is an incredibly rich datatype

a performant model without protein specific architecture. One comment: they trained on 9M structures, >95% generated by deep learning models that DO use protein specific architecture. I’d call this knowledge distillation and challenge the core claim of the paper. Will read more deeply

From descriptive to quantitative biocatalysis “The 1913 study ‘Die Kinetik der Invertinwirkung’, by Michaelis and Menten, marked a pivotal advancement in enzymology by illustrating the application of mechanistic models and quantitative kinetics to biocatalysis.” doi.org/10.1038/s419...