Super obnoxious to keep hitting you with LLM screenshots lmao. Medical doctors and lawyers must be inundated with exchanges like this
scientist at UC Berkeley inventing advanced genomic technologies lover of molecules, user of computers https://scholar.google.com/citations?user=63ZRebIAAAAJ&hl=en
405 followers 952 following 556 posts
view profile on Bluesky
Super obnoxious to keep hitting you with LLM screenshots lmao. Medical doctors and lawyers must be inundated with exchanges like this
it's also more than dexterity. awareness and observation can also be trained. understand the procedure, use those Schlieren lines
impressively desperate and gross
i wouldn't be surprised if free energy estimations are much less accurate for random sequences and de novo designed proteins than for natural proteins, as pLMs are learning coevolutionary statistics as discussed here. curious what you think! www.pnas.org/doi/10.1073/...
if instead the authors used a method like one of the below and drew the same conclusions, would that be more satisfying? for me, not really... www.biorxiv.org/content/10.1... academic.oup.com/bioinformati... www.nature.com/articles/s41...
ok this gets hairy. we don't have lots of structural data for mutations, so the backbone similarity used in this paper is maybe not the best way to extract information from the model. however there are lots of papers that look at correlations of model outputs against protein stability data
not very easy to experimentally verify. could be interesting to explore orthogonal protein DL models to explore the effect of model selection. that wouldn't solve the problem of training on natural proteins then comparing against random and designed sequences
now there's a title that catches the eye. paper's based on ESMfold predictions. it's an intriguing idea, but i fear the results may be an artifact of the training data. they may be seeing poor generalization of ESM rather than a biological signal
as a scientist, it's my job to figure out how to use the spam machine to cure cancer
too bad we can't tell if those big LLM usage numbers reflect responsible, productive activity. impossible to know! - cybersecurityasia.net/most-spam-ge... - apnews.com/article/fake... - www.theverge.com/2023/5/2/237... - www.frontiersin.org/journals/art...
i use LLMs in my work. i've found them to be only marginally useful for laboratory research. that information shouldn't be upsetting, and really it shouldn't be surprising. even for coding the evidence of utility is mixed, nuanced. the physical world is much more complex
raw usage doesn't distinguish between productive and destructive use. optimizing for usage often leads to negative outcomes. Better measures would be things like economic indicators, scientific productivity, business case studies, user surveys, and controlled comparisons.
I’m biased in favor of this technology. I want to use it to accelerate progress. I’m disillusioned because it doesn’t perform as promised
Usage numbers are a terrible proxy for benefit or productivity. I’m not dismissing value, I’m telling you my colleagues and I are getting practically zero gain in our work. Meanwhile, negative and neutral uses are so widespread it’s hardly worth listing them
i worry about a situation where these models are good enough for destruction, not good enough for productive work. as of today, LLMs are highly enabling for fraud, disinformation, surveillance. in all my attempts to use them for scientific research, it's a marginal benefit in anything
Caller ID was a real thing we had, but it… stopped unwanted phone calls… so now it’s gone
evaluating funding effectiveness is not easy, and it's totally possible that LLMs could have positive, negative, or neutral effects on the whole process
i will say, i'm personally not *that* worried about bias from LLMs compared to bias inherent to any granting process. and i like the proactive funding model as a complement to traditional calls. i'm not sure if an LLM could identify the best science/grants/recipients, but..
hey César, do you have a manuscript detailing your methods, or would you be willing to conduct simulations or real-world experiments to evaluate the methods? is this approach to finding grant recipients more effective than alternatives? more broadly, what is your goal in using LLMs for this task?
it's so bleak. their idea of success is also very dark. feels so naive to be like, wait, can we try to do good things instead?
any billionaires looking to invest?
Eighty billion dollars are you fucking kidding me. don't ask the deadbot CEO how much the deadbot company will be worth ten years from now.
Ezra Klein argued one reason to end the filibuster is so governing parties face the consequences of their own agendas
When it comes down it, we deserve this :(
Related: I am ok with people using LLMs to write grants and propose new research. If it’s good stuff, it’s good stuff. Think it’s gonna be a while before an llm can propose competitive projects tho. As the tech bros would say, they lack ‘taste’
Wow they must be using much better LLMs than I’ve seen. Stuff I’ve used isn’t close to being able to do this kind of task
i don't have a magical solution but in my view we've integrated journals into scientific careers as a proxy for quality, and we've arrived at a place where journals are a big drag on the enterprise
working in industry, journal access was a huge problem that small companies are poorly positioned to address. overall i think this introduces a lot of friction, removes agency from authors, and slows down R&D to prop up the for-profit publishing industry 2/3
not personally, but friends have used it successfully. my gripe isn't that access is impossible, but that the barrier is too high. when i'm researching a new field, i read dozens of papers and it's hard to know from an abstract if a paywalled paper is worth hunting down. 1/3
to me it seems like unfounded race science and a common trope, but i'm open to ideas and happy to be persuaded or proven wrong. i'm willing to give hypothetical bioterrorism plots their due level of concern, but i think it's quite misleading to suggest these are the proximal concerns with AI
the first "Biological AI Capabilit[y] of Concern" highlighted in this RAND report is: "design of proteins, genes, or genetic pathways that confer specific susceptibility of human ethnic groups to a pathogen, while maintaining pathogen fitness" does anyone think this is theoretically possible?
Alas, we’ve reached an impasse. My views on knowledge sharing are quite different
Authors or readers? As a reader I am still constantly stymied by paywalls. Authors should use available outlets for free dissemination, so I think we agree there
I think this misunderstands the purpose of open access. OA was a fix for the issue of access, not the broken journal system. Doing it through the journals was obviously going to end in journals charging more for the “service”. The problem is still the journals and the demand for quality proxies
LLMs are not theoretically or practically capable of peer review!
i like them both, worked with them both. they're addicted and a little too adjacent to silicon valley. same goes for a lot of bay area scientists, especially AI-adjacent ones (so, most!)
the software industry as a whole tries to have it both ways, all upside no downside. that's not how technology works. computer science, broadly, has the most infantile culture of safety because they think it doesn't apply to them. but they think chemicals are super dangerous
you shouldn't be able to do whatever you want with a computer. imo it's "insane" to think otherwise. there is no genie and no bottle! regulations around software are desperately needed, and the lack of effective regulation results in extremely negative consequences for society.
my argument is: 1. conduct, be it executing code or doing chemistry, is not considered speech under the US constitution 2. chemicals and software are both dangerous tools that are often misused 3. software, like chemicals, is subject to regulation including bans 4. my analogy illustrates the point
ya, still cringe
all these superlatives are fine (heroic, tour de force, magnum opus) provided they are rare praise coming from others. it's not important just because you said so
Is it not Davis? You consider yourself an independent contractor? Curious how you draw the distinction
That’s just laws. Regulation can shape development in productive directions, and provides for punitive action when people and organizations inevitably cause harm. My whole point is that coding is not special in these ways
Because you personally like coding and think it’s special and not dangerous? Because you are accustomed to extremely lax regulations on software? Because you haven’t given thought to dual use technologies, particularly scalable and distributable ones?
That depends entirely on the technology and its potential for productive and destructive uses. I don’t think LLMs should be banned. They must be regulated
agreed, but doesn't that undermine your argument that LLMs are a genie that can't be put back? it would seem that sensible regulation around LLMs could provide legal frameworks for ethical use and development, fitting squarely within existing precedent
because under our laws writing code is expression and executing code is conduct. see the case below which was about whether Corley's code was protected. No one was arguing that using his code for decryption was a protected act, because it's clearly not en.wikipedia.org/wiki/Univers...
What about running code on a 4080 at your house? Is that speech? I admit I wasn’t considering the case where someone would write code without the intention of executing it, and I should have been more clear. You were referring to my right to detail bomb making *instructions*
My mistake, I should have been clear I was referring to executing code as the original post discussed. Or are you claiming that executing code is also speech? Or just some hypothetical argument?
Lmao what are you talking about? Do you not understand the idea of tools and actions, and how those are different from speech? That doesn’t mean coding is illegal, it means it’s not a protected act under the us constitution. Which it’s not. Computer crimes exist
Neither is writing code
Promise?
Oo sorry that would mean cops enforcing traffic laws and they gave up on that like five years back
My crackpot theory is that animals evolved to farm bacteria in our guts
I work in bioengineering, an industry that tries to prioritize safety, even though it frequently slows progress. It’s crazy-making to see the disregard for safety exercised by tech broadly and the AI community specifically.
This is because we have intentionally and successfully limited experiments with the human germline. Also it’s super hard to do well
My text messages are over 50 percent spam and scams plus junk from companies I have otherwise legitimate businesses with
My work email is over 90 percent garbage. Personal is worse. Most retail transactions involve my personal phone number or email address. And we have to deal with the cookies crap too
It’s all based on the premise of exponential progress. Not steady development but regular step changes in capabilities. After a few cycles of stagnation the flaws are obvious and the value is very limited
The thing about society is people need to want to make it work. A critical mass needs to be invested in having a good society. That includes, perhaps most critically, managing the behavior of those who are not contributing to the project.
Trying not to think about all the mental hoops you gotta jump through to arrive at “code is speech” It’s not. By ANY reasonable and or legal definition, it is not.
Making bombs is speech. You can do it at home with stuff from the hardware store. Can’t ban making bombs. That’s how you sound
they think it's OK for an LLM to review a paper as long as it's disclosed. ther brains r cooked. may the field stagnate into mid-nacity
ngl I kind of think Sam Altman should be tried criminally? is that … feasible?
The only negative impact they consistently cite is bioterrorism, which they cynically use to give the impression of responsibility. Most of the harms they could cause are just literally part of their business model. They aren’t serious about safety.
The crazy thing is OpenAI is a world leader in natural language processing and they could definitely do a better job mitigating the negative impacts of their technology. But they actively choose not to.
“Reasoning” models lol
I could write a whole letter to the editor about how they are scared of their own imaginations, but it’s literally not worth the time! The idea is bollocks. Bunk.
Mirror life has no reasonable path to existence. It’s a non issue. We can’t even make synthetic life with standard chirality. Mirror life would be much, much more challenging. Some of the smartest people in the field think this worth their time. It’s nuts
i'm sorry the UCB occupation of 2016?
some of Trump's federal and military deployments: Kenosha, WI (USA) Portland, OR (USA) Seattle, WA (USA) Los Angeles, CA (USA) Washington, D.C. (USA) US-Mexico border (USA)
except language policing. THAT is a distraction and too many are falling for it
did a 'founders camp' with a VC firm and most of the advice was fairly anodyne except how to raise $$. VCs are all back-channeling with each other trying to figure out what the other ones are buying and racing to the same deals. the opposite of independent thinking
scaring off varmints?
for everyday interactions humanoids make sense iff they can also evaluate and plan at human level for production settings humanoids have never been the best solution
unholy flavours at blasphemous prices
one reason we’re not seeing more AI fraud is because it doesn’t work that well. These LLMs have only achieved porn level intelligence
In other words they are not leaders, but literal followers
So start that 24 hr timer people! Putin’s ready to back down any minute now!!!!!! /s
Just had the realization(?) that early career scientist means early in citation count since citations go up dramatically once you establish a research group. Still a pretty demeaning term but maybe a bit more hopeful
Interesting to think if trumpism would’ve faded or flourished in a trump loss. He was prepared to lose and take aim at democracy. The order of events would have been different, but the end result maybe not so much
My dreams are so fucked up sometimes. bizarre but realistic. Long, evolving plots. Resistant to detection. Just came up from the worst nightmare of my life and now I have some fake traumatic memories that weren’t there yesterday.
“Wealthiest cult in human history is writing fan fiction while their technology bleeds money”. Run that!
Don’t think you’re right on that one. It would be a big deal if an llm could design LIGO experiments, and the article intentionally uses ‘AI’ as a blanket term to give that impression. It’s deceptive imo
The truth wouldn’t get as many clicks though
Everyone who knows what they’re doing sees through this immediately. Everyone else is just a mark I guess
Hooray an entire article using the term ‘AI’ repeatedly to describe what appears to be a bespoke machine learning model. Deeply misleading
There can only be one
o my god when will 'accept all cookies' die
mmmmmmmmmmm nope. my TBI account is already overdrawn. gotta avoid brain cell bankruptcy
Ya arxiv.org/abs/2309.00236
legit hilarious
ya at least our government has the excuse of being run by fucking idiots. what's the university admin excuse?
isn't heritability the crux of epigenetics? it's *heritable* traits beyond the genome sequence? eventually people started muddling 'epi' with 'modification' and now words don't mean anything
As users, we also see way more overhead in the form of spam, over communication, push notifications, and constant prompts from the technology itself. We get less joy than we should because the products are using us and not the other way around. Which is my point
It’s not hard, m8! EVs, M series MacBooks, cell phone cameras better than standalone cameras, fitness trackers, biometric security, AirPods, insanely good and cheap TVs, internet so fast that physical digital media is dying.
a beaut! 🎸
what is the deal with releasing a preprint before the github repo? that's backwards!
ya it's a freaking nightmare and a danger to society. totally solvable with reasonable regulation
i think their investments are an abject disaster. a travesty of world historical proportions. i'm angry about it. i also think this stuff will be truly transformative in the long term