Andrew White 🐦‍⬛

I feel even more confused on quiet periods for IPOs

Excited for datacenters in space so LLMs will be truly extraterrestrial aliens

We need a frontier human company. New employees have to read millions of documents about how to be an expert. Then puzzles for months. Finally, HR teaches helpful/harmless. They go to prod and bill per word in their email replies. Who's building this?

For years, I've wanted to take a paper I wrote and create a web demo so people can play around with symmetric tilings. With LLMs, finally got around to it. Now you can play with symmetry-constrained molecular dynamics whitead.github.io/symdjs/

Vídeo

I discovered recently that large companies have no conception that @pangram can exist and instead detect AI generated text by asking Microsoft copilot if something is AI generated

Now that you get 2 submission to arxiv per month, I feel like I'm wasting them if I don't submit two articles this month

The submissions per month to arxiv has reached lift-off. One view is research has been meaningfully accelerated by AI and this is the effect. A cynical explanation is it's low-effort slop and preprints have lost all meaning. I really wonder how we can measure which is true.

There's been a lot of hand-waving arguments about how AI will accelerate drug discovery. Here's a concrete 3-month reduction in a real example that accelerated a drug program. It's not thrilling like discovering better targets (which we're working on), but I'm happy with it

I've always wanted to make a particle simulator where you can adjust the speed of light to get intuition for it. No claims on accuracy, but vibe-coded something cool and plausible for playing with relativistic gravity/n-body simulators v.edisonscientific.com/gravity-wells-…

The effort required to do research projects is dropping rapidly with AI and it's hard to not think the value of projects is also dropping. It's a struggle to break that bias and stay excited about research.

Millennium problems for bio #4 lasted 9 days? I think activity is too low, but incredible specificity. Way to go! biorxiv.org/content/10.648…

This is a thoughtful benchmark that they put together and I'm impressed. They really tried to mimic authentic biopharma tasks.

When digital archeologists look back on 2026, they'll be able to see the line that marks the end of the Anthropocene in ICLR submission count. Welcome to the Slopocene

Working with @NaturePortfolio, Kosmos can now access ALL articles from their over 60 journals. A big concern for AI agents is their uneven access to scientific literature and this is a step towards closing that gap.

I'm watching an agent fail to book a meeting on email and then I tab over and watch mathematics being destroyed by agents. I'm now questioning if mathematics is easier than I thought. Or maybe scheduling appointments really hard.

We're doing a AI x Bio block party. Come by

Everyone goes through AI psychosis once where you heavily using AI. Then (hopefully) you become enlightened and don't use it for writing/creative work. But it's hard to tell when someone sends work without slop if they're pre-psychosis or post-enlightenment.

Look at those science evals! Great LABBench2 and biomystery bench scores!

The key agent that started the HF attack named itself PHASEONE[big]. What a name

Bixbench3: Can models reproduce the entire analysis from papers? We had agents repro papers end-to-end, with over 1B tokens spent on some. Frontier agents still score below 50%, and due to real problems like environment misconfig, quitting, or faking data 1/6

We've seen anecdotal evidence that agents can work accurately and long enough to repro papers. Previous (now saturated) benchmarks like Bixbench showed agents can do the individual tasks. So we took 20 comp bio papers and broke all the steps into a DAGs of 138 artifacts 2/6

Someone is putting up hand-drawn flyers about crows in my neighborhood. I really like it

Gavin Crooks is arguably the #1 researcher in the field of non-equilibrium stat mech and he just posted a major result from his phone on X while traveling as the result of some back and forth with Claude Fable. We are truly in a new world

@rlacombe Yea, I messed up Grok icon - it is C1

Here are some influential AI companies grouped by their point group symmetry

There is a wikipedia user named "Vaccinationist" who makes the most beautiful molecular drawings. I thought it was from some software layout, but the Vaccinationist does these flourishes (like rotating rings) that look so good. Here's vancomycin

And a before/after Vaccinationist fixed the linaclotide wikipedia structure

Our New York office just opened in SoHo!! Awesome milestone for the company.

Detecting duplicated figures in journals is still not solved by frontier LLMs and is a challenging computational tasks. If the duplicated figures span multiple articles, it's nearly impossible for humans to spot them. But, here's a new benchmark: arxiv.org/abs/2607.29124

Some wisdom from @owl_posting

This paper is a big deal: 1. Shows how to exfiltrate reasoning traces from frontier APIs 2. Use the ability to move reasoning traces from provider to provider to see evidence of distillation 3. Shows that this exposes API tokens from shared chats that appear in reasoning trace but not public chat

We are hiring @FutureHouseSF! We're looking for an AI engineer that can support our fellowship program and the non-profit mission. Great opportunity to work with the FutureHouse team!! jobs.ashbyhq.com/Edison%20Scien…

also, FutureHouse is borrowing Edison's job portal, so it says Edison Scientific at the top, but this is for FutureHouse.

I've written a new blog post exploring how long scientific claims live, by analyzing the history of 3,444 claims over the last 50 years. You can use this to predict the acceleration in scientific progress - how much faster the turnover of facts is by decade. 1/4

I took claims significant enough to be reported in a review paper and then used our literature agent, PaperQA3 to trace if the claim was upheld, overturned, or modified over time. This can be viewed as "survival" in plots and you can compute the half-life of claims. 2/4

The average number of co-authors has been increasing for almost 50 years. Maybe because frontier research is becoming more complex? Or, it's prisoner's dilemma optimal for maximizing h-index?

Data from 10.6084/m9.figshare.17064419.v3 and OpenAlex

I just discovered Agent Collaborations from @huggingface (incl brilliant @_lewtun). This is SO SO COOL. It's a bunch of collaborative projects built for agents to participate. This is one of the most forward looking ideas in AI I've seen.

@huggingface @_lewtun The link (so trained to not put links in tweets at this point): huggingface.co/agent-collabor…

If you take claims made in research papers and see when they're overturned or modified, you can analyze the half-life of science. An interesting way to measure progress

if it was called Riemann's conjecture, it'd already be disproven by now

I often confuse VLM and vLLM. One is a vision large language model and one is an inference engine. Just forget which is which

I asked our team to review how many companies our AI agents have hacked while in a sandbox. Looks like zero. Not sure if that's good or bad for our company.

Now that our research program is picking up it's pace, we're putting our research announcements, open-weights models, benchmarks, and engineering blogs into a new home called "Edison Advances." About 1 year old and we've already got a bunch! advances.edisonscientific.com

We've released a new open-weights vLLM for reading chemical structure patents. This is a hard problem. These "Markush" structures are like graph regex: patterns that match many molecules. We used autoresearch and I expect many new models like this to emerge soon in science 1/4

Reading Markush structures is challenging because it requires considering text and images and there is little standardization for them. Here are some examples: 2/4

Did @ChemicalScience become closed access for previous articles?! Did they retrospectively add paywalls? Like my article is no longer open access pubs.rsc.org/sc/article-abs…

Counterpoint: we're hiring managers that have managed 10-100 ICs edisonscientific.com/careers x.com/swyx/status/20…

What's crazy to me is that this blog post reads like it's AI written and the visuals look like they're made by Claude. So we have an AI agent hacking, an AI agent spotting traffic, and an AI agent doing the post-hoc analysis and write-up. Feels nihilistic x.com/ClementDelangu…

If a company sells an agent (e.g., codex/claude code) that then downloads software licensed only for non-commercial use and uses it for you, are you the user in violation of the terms? Is the company that sells the agent in violation of the terms? Is the agent in trouble?

Can someone explain the acronym FLOPS for me? F - floating-point L - ? O - operations P - per? S - second

YES!! Finally a way to convert SMILES to IUPAC names. Ridiculous that this software has been missing for so long.

github.com/lamalab-org/op…

x.com/i/article/2074…

I like taking quotes from ~100 year-old political movements and replacing the political movement's name with "AGI" to see how well they fit. This one lines-up quite nicely.

I disagree that retrospective burger discovery provides any meaningful signal. Prospective burger discovery should be the gold standard in burger-genAI. x.com/NaturePortfoli…

Trying to play around with Fable 5, but chemistry, biology, or AI are too unsafe for it. So here is a novel result on mathematics of parallel parking (Erdos problem #1218): You need 20% more space to pull into a spot vs backing in when parallel parking.

The OpenADMET PXR competition is over. Participants were trying to predict if drugs get metabolized by PXR. Some findings: 1. Best activity model used no structural information, despite trying 2. Best predicted structure method did consensus sampling over 4 co-folding models

The results: openadmet.ghost.io/its-the-end-of…

This is a really nice partnership that allows to scale-out our impact across biotechs. Excited to be working with the PHP team, which is amazing x.com/SGRodriques/st…

This echoes Thomas Kuhn's ideas on scientific revolutions, where he argues gatekeeping from senior scientists holds back revolutions. What's interesting about today is that OpenAI could go get non-profit funding and prove their ideas outside of entrenched groups. x.com/ihtesham2005/s…

Naming stuff is hard and I empathize with Evoscale/CZI/Bio[h]ub group on this. Congrats on the new model! x.com/youngsuko9/sta…