Daniel Geiszler

@jjacky @vikings_ace Do you have a toy example and/or numbers for speed requirements? For that scale it doesn’t seem like an awfully hard problem.

@vikings_ace realtime search of api responses, i.e. you are querying your zendesk tickets in realtime with 1,000+ tickets, you want to be able to use semantic search to find the 10 most relevant tickets w/o the need to pre-process this is an actual problem i'm facing while building

whoever invents realtime vector/fulltextsearch without needing a pre-processing step, just like grep, is going to become a billionaire

Interesting point about this is that GitHub will effectively be distilling all the other coding models

@shashj I think this is the second one intercepted in Turkish airspace. At least one was intercepted over Syria on March 4th, with debris landing near Qamishli. middleeasteye.net/news/isolated-…

The new eLife model championed by @mbeisen has enabled an apples-to-apples comparison of machine and peer review. That's because the machine can review the submitted paper (preprint) which @eLife makes available, and the peer reviews are also published. biorxiv.org/content/10.648… 1/

@neelsomani It would be great to know what the replication rate here is if you did this again. Are the correct ones random hits, or do you always recover these “easy” ones?

You might be wondering if there are more Erdos problems that GPT 5.2 Pro or Deep Research can solve. I recruited a team of bright undergraduates to construct a dataset of ChatGPT responses to every open Erdos problem and validate the output. Here's what they found (all open source - repo below):

Exciting to hear that people can now test actual hypotheses with their data instead of publishing another superficial “atlas”

If someone had predicted before the last election that if Trump won, federal officers would be shooting Americans in the streets, he'd have been dismissed as an alarmist.

FDA sits on a treasure trove of data: past regulatory submissions, which often span tens of thousands of pages. These would be incredibly informative, but are locked away due to trade secret law. I propose leveraging AI & bankruptcy law to unleash this info & empower start-ups.

@RuxandraTeslo @FDA Do any of the dossiers from failed startups contain drugs that were approved? Or are we feeding a bunch of failed submissions into a model?

@FDA Thankfully, there is a workaround this, with relatively little money compared to the value it would bring. Set up a fund that buys the intellectual property of start-ups that go through bankruptcy. This has already been done and bankruptcy law in the US is set up to favour it!

@krishnanrohit @RuxandraTeslo @WSJ rdworldonline.com/how-much-does-…

@RuxandraTeslo @WSJ Hatred isnt the right reaction but worth noting pharma also spends more on S&M vs R&D

Good article in @WSJ. Why does everyone hate pharma companies? They are literally the most "useful" companies, yet continously vilified. It boggles the mind. wsj.com/opinion/donald…

@PracheeAC I’ve been saying that the real potential here is in methods development. Look at LASSO… ubiquitous in data science today, discovered and popularized by a stats professor in 1996. Except it had already been used for 10 years in geophysics. en.m.wikipedia.org/wiki/Lasso_(st…

I think a dramatically underappreciated feature of LLMs is the value-add of scientists being able to read and engage with science outside where we have the most specialized expertise. This isn’t to suggest they are currently replacing experts but it puts a much greater range of science on the table so we can be influenced by concepts in other fields…. increasing the ability to leverage discoveries that would have previously been siloed or waiting for the polymaths that happened to unlock the connections. Cross-disciplinary insights are more possible than ever and will only continue to improve

Assessing error control is fundamental in mass spectrometry-based proteomics. Authors introduce a theoretical foundation for entrapment experiments along with a method for more accurate evaluation of FDR control. nature.com/articles/s4159…

🚨 New preprint! 🚨 Presenting diaPASEF immunopeptidomics for bacterial epitope discovery. 💥 Showcasing DIA-NN immunopeptidomics using proteome-wide predicted HLA class I libraries! 🧬 Read more: biorxiv.org/cgi/content/sh… #Immunopeptidomics #MassSpec #DIA #HLA

@iskander @carlosbtcosta Firms are already incentivized to do this because shorter survival times mean shorter trials and longer effective patent life aeaweb.org/articles?id=10…

@carlosbtcosta Instead of surrogate endpoints, why not focus on most severe manifestations of a disease (eg cancers with <1y survival)?

Let’s dream a bit. How would you dramatically reorg biotech research / drug development to make dramatically faster progress towards curative therapies? (Don’t say “use AI”, hypothesis/candidate generation under current structure isn’t a bottleneck)

@stevehou Secret ballots en.m.wikipedia.org/wiki/Secret_ba…

I keep forgetting Australia is still there. What’s something in my life that’s impacted by Australia that I’m not thinking of?

@kareem_carr This is a bad example because LLMs are extrapolators, not interpolators. “We empirically and theoretically argue against those two points and demonstrate that on any high-dimensional (> 100) dataset, interpolation almost surely never happens.” arxiv.org/abs/2110.09485

Very simply, statistics is about taking two points you know exist and drawing a line between them, basically completing patterns. Sometimes that middle point is something that exists in the physical world, sometimes it’s something that could potentially exist, but doesn’t.

If you think about how statistics works it’s extremely obvious why a model built on purely statistical patterns would “hallucinate”. Explanation in next tweet.

@pitdesi Starbucks a successful Italian boomerang? There are 20x as many stores in Turkey as Italy, which is probably more relevant.

Chipotle is opening in Mexico. Was curious if any of these boomerang brands have been successful: Unsuccessful: Domino’s: Italy Taco Bell: Mexico (2x!) PF Changs: China Cinnabon: Sweden Successful Starbucks: Italy Pret A Manger: France Chili’s: Mexico

@TheStalwart Mark Carney went to sleep

The Treasury selloff last night cooled down

@JustinWolfers How many derivates deep can it go?

Yesterday's decline in market volatility was short lived. It's volatile volatility.

@i000 @owl_posting @mattszarz I think we’ll probably need to wait until the publication comes out to see if they’ve shared technical details, but the underlying tech could be more meaningful than this output.

@i000 @owl_posting @mattszarz I said it elsewhere, but I think the value proposition here is in high-throughput screening and feature selection, not really de-extinction. If it really looks and acts like a dire wolf with only 20 edits out of thousands of possibilities, that’s big.

its insane! i cant predict anyones take! i think the best differentiator i can find is whether or not you think the following statement is true: ‘biology as a field has barely meaningfully progressed in the last decade’ if you think its true, you’re probably pro wolf. if you say false, probably anti wolf UNLESS you believe there is a moral issue with doing undercharacterized genetic edits to an animal. then you’re anti wolf either way