Billy Lau

@baym This is why the culture of being “scooped” is so toxic.

What I see a lot of is people so desperate to claim that they’re the one who made some big advance everyone knows is kind of inevitable that they don’t really do it but claim victory, and so when someone actually does it right and for real it’s treated as an also-ran validation

@arjunrajlab Maybe I’m too optimistic, but AI coding agents are a huge enabler for reproducible, readable, and maintainable code. And, the marginal effort required to implement sanity checks, controls, etc in the code means analyses can be more rigorous.

Transitioning to being a PI in the age of AI Computational biology is in a period of upheaval that is both exhilarating and terrifying. Rapidly, we are approaching a moment of “analytic abundance”, where basically idea you can think of (and several you didn’t) magically appear within minutes of you thinking of them. Of course, the central proximal challenge is the evaluation of the sheer volume results—how do we know they are right when we don’t have the time to check over every line of code? I think it’s very telling that when I talk to AI-pilled faculty, they are exhilerated, but many trainees seem more cautious and far more ambivalent. I think that’s because faculty often have been removed from the details for a long time and probably haven’t checked over a line of code in years. They are used to managing (rather than doing) analysis. Over time, they usually develop a sense for whether things seem right or wrong. In this day and age, this is the skill that you, too, must develop. How do faculty do it? I am guessing every faculty member has their own list of internal sanity checks, but here are a few of mine: * Checksums. I look for things that should add up correctly (percentages add to 100, etc.). If it looks even a little bit off, I ask questions. * Never let go. If something doesn’t make sense, I don’t let go until it does make sense. Never relent! * Explain stray datapoints. Always dig into outliers in the data. How did they come to be? Often, they reveal some hidden assumption or something unexpected about the data. * Do not tolerate warnings. If code gives you a warning, resolve it. Do not continue, do not pass go, until you either understand or eliminate the warning. * Track the number of datapoints. Even a single missing row can be a sign of some fencepost bug. And I’m sure many more that I’m forgetting right now. Basically, it’s a transition from a maker to an interrogator. I also feel it worth reiterating that this is a highly unsetting period of time. I have been fortunate (?) to have 16 years of time to make a transition that people are now being asked to make in months. Again, exhilarating and terrifying, all at once!

@JessicaSacher Amazon being an approved vendor is a game changer.

idk what's more mortifying — the fact that this small pack of foam is $50, or that my initial response was literally 'omg so cheap!' because i've been in science too long imagine how little grant money we'd need if real price competition was a thing in science

@wasserstein_rao Connecting CC and a webcam to an Opentrons OT2 might get so much more mileage out of it…..

Using claude code to directly control a liquid handling robot is such a crazy experience

@ghartoularos @BorisLenhard @anshulkundaje Yes, the whole point is that ORCID linkage means that it's not anonymized. Peer review should be open anyway. I don't see why we couldn't implement a Reddit-like system where good reviewers get upvotes and these reviews are prioritized.

@BorisLenhard @billytcl @anshulkundaje Linked to ORCID would not be anonymous (or at least harder to game, assuming reviewer has at least one paper tied to their ORCID). I'd think gamed praise would not be a big problem, but even then we could just randomize the reviewers and compensate for review.

Some of our papers have been in journal "peer review", response and "editorial input" ie purgatory for over 2 years. By the time they r "published" will be way past their life cycle. The preprints, code, data, evals heavily used by the community. The old system is dead.

@TheStalwart 99% of code is pure drudgery and just has to work regardless of style or “preference”. People read for style AND content.

I’ve seen this exact question asked a thousand times, and yet I’ve never seen anyone establish that the underlying premise is actually true. Has there been some poll?

@SashaGusevPosts For things like that, I would highly recommend using Cursor on their $20 plan and their Composer model. Very token efficient and gets the job done.

My Claude experience since Opus: 1. I describe the project 2. Claude: "Claude's response exceeded the 32000 output token maximum" 3. I increase the token maximum 4. Claude: "You have hit your token limit" 5. I switch to Codex.

@srikosuri @YSPTSPS @eLife I think the issue is more that there’s no real accepted standard of what is in/out of scope for what reviewers suggest. Plus very little of the review process is ever conveyed in the final document. IMO papers should be living documents with clear diffs and comments.

@YSPTSPS @eLife I think both papers could have been submitted to more glamorous journals. It's more that reviewers always want to ask for more things, and we just don't want to do those things for various reasons (or can't disclose some things). This way we get the reviews on what we presented.

I am loving these reviewed preprints in @eLife. Especially in a company, it's great to put stuff out that we've worked on into the world, but often times we really don't want to do a bunch more work; we moved on, and we just get to say... nah. elifesciences.org/reviewed-prepr…

@kirk3gaard @nanopore A100 AWS compute is cheap enough now that ONT could offer a cloud solution. Subscribe to get basecalling and alignment done for you. Recurring revenue!!! The only issue is that most people don’t have a big enough pipe to route 1Tb of raw POD5 onto AWS.

@billytcl @nanopore Could be considered up cycling if server farms allowed people to grab them for a good price and use them for basecalling ♻️. Would require a serious power supply to have P24 with live SUP using A100s 🔥

The majority wants SUP for their @nanopore data. Compute bundles should match that or we should see a model advancement that makes SUP level much more compute efficient (like we have seen with all other ai software). But honestly nobody should be shipping A100 GPU towers in 2026?

@vsbuffalo Same but with Cursor cloud agents. To me it feels way less clunky than a remote session.

I’m laying on the couch, watching a Claude Code session via Remote running on my Mac Studio re-prompt a dumber local LLM to do some task. It’s entertaining… I can practically feel Claude thinking “how is this model getting this wrong for the third time in a row?”

@EdReznik This is generalizable to probably every PhD program nowadays! The bar is now so absurdly high that I believe there is a significant number of admitted students who are way too overqualified to be graduate-school trainees.

I just want to say, for all those applying to PhD programs now in computational and systems biology: I would never have been admitted if I were to apply today. I don't even think I'd earn an interview. You are incredible and admission is not a judgement on your awesomeness.

@shae_mcl Mollie Stone’s is not even that good! I’m surprised they could charge those kinds of prices.

The Noe Valley Mollie Stones prices are destroying my family.

@OmicsOmicsBlog Do they still have the PDF version of the conference book? Just feed that into Claude.

Wish: Scientific Poster Snap Organizer App #AGBTGM26 omicsomics.blogspot.com/2026/02/wish-s…

@OmicsOmicsBlog Dude is just moving goalposts in his replies lmao

Narrator: “the market for Minimal Residual Disease testing was $1.3B last year & for Multi-Cancer Early Detection another $1.5B; rare disease testing may have been $2B”

@RuxandraTeslo I think an "AI-assisted" clinical trial where it handles a lot of the busywork is realistic. Things like formalizing data entry, analytics, etc. But there is so much institutional knowledge that is currently unknowable and really hard to replicate.

People overestimate the extent to which "automating labour" will solve complex operations in deeply regulated sectors like clinical trials. This is a field where tech-first (pre-AI) disruptors have repeatedly failed due to market structure & incentives (all tied to regulation).

@anselmlevskaya @Noahpinion Almost all of the chem/bio industry is locked up with KYC regulations. Not sure how some rando can buy a DNA synthesizer and associated chemicals without getting flagged. It's several orders of magnitude harder than cooking drugs in the basement.

@Noahpinion I was a synthetic biologist for ~two decades and have worked at a frontier LLM lab for 8 years. This scenario is wildly overblown. It’s just not easy to do novel synthetic virology in practice and LLMs don’t materially close the lab skill gap for amateurs. Bio’s not just code.

I'm mildly worried about mass unemployment from AI. I'm a bit uneasy about autonomous embodied superintelligence. But I am fucking TERRIFIED that some angry teenager will able to vibe code a doomsday virus. noahpinion.blog/p/updated-thou…

@shae_mcl Interesting that AI platforms are so worried about bio risk when that’s on a 20 year horizon at the most optimistic. Contrast that to world ending levels of social unrest due to mass unemployment from AI. They’re trying to be mindful but choose to ignore the proximal risks.

I'm making lentivirus with TRE-eGFP so I can make a reporter cell line for my experiment (extremely basic, innocuous bench work) and ChatGPT refuses to discuss the protocol with me. Turns out you cannot just build things.

@kirk3gaard @nanopore There are people using fast and hac?

What @nanopore basecalling is sufficient for your research?