YHK - yet another hacker news clone

The technology has the capability, they're just gating who has access.

I know it’s not the models themselves, but their preflight safety classifiers are a dumpster fire and serious impediment to performing regular work.

That situation can’t go on forever if they’re going to be a competitive service.

Tell that to my (now solved) health issue, which was correctly diagnosed by Gemini, while a real doctor failed to find the cause.

Diagnosis is the one thing where today's AI is actually very good.

    > Neural networks in drug discovery: have they lived up to their promise?

     "Over the last decade neural networks have become an efficient method for data analysis in the field of drug discovery. The early problems encountered with neural networks such as overfitting and overtraining have been addressed resulting in a technique that surpasses traditional statistical methods. Neural networks have thus largely lived up to their promise, which was to overcome QSAR statistical problems. The next revolution in QSAR will no doubt involve research into producing better descriptors used in these studies to improve our ability to relate chemical structure to biological activity"

    > The use of neural networks in drug discovery started at the end of the 1980's when they were applied to various chemical problems including the study of quantitative structure–activity relationships (QSAR) 

2002.

https://link.springer.com/chapter/10.1007/3-540-45984-7_6

Want more robots and silicon? That's been a thing too!

    > Computer-aided drug discovery and development (CADDD): In silico-chemico-biological approach

     In biomedical arena, computer-aided or in silico design is being utilized to expedite and facilitate hit identification, hit-to-lead selection, optimize the absorption, distribution, metabolism, excretion and toxicity profile and avoid safety issues. Commonly used computational approaches include ligand-based drug design (pharmacophore, a 3D spatial arrangement of chemical features essential for biological activity), structure-based drug design (drug-target docking), and quantitative structure–activity and quantitative structure–property relationships. Regulatory agencies as well as pharmaceutical industry are actively involved in development of computational tools that will improve effectiveness and efficiency of drug discovery and development process, decrease use of animals, and increase predictability. It is expected that the power of CADDD will grow as the technology continues to evolve.

Biologists aren't slouches. The thought about using computers to process biological data has been "obvious" for as long as there has been computers. AI and biology have been a thing since the Dartmouth conference. A lot of things that could have been tried have already been tried, but I guess we'll keep trying until something works?

There ain't much room for miracles here.

I wish that wasn't the case. I really do. But it is what it is.

https://podcasts.apple.com/gb/podcast/black-box-the-chatbots...

I don't know if he is in that video or not, or ever, personally, but just in general, it's the Ockham's razor approach any time you see something like this. It's a much more direct and clear explanation than the various more convoluted theories people keep throwing around as to why everyone sounds like a methed-out megalomaniac all the time.

"it is a literal and useful description of anthropic that it is an organization that loves and worships claude, is run in significant part by claude, and studies and builds claude..."

With Holmes, some of it was intentionally fake and manipulative (especially the voice), half of it was believing her own BS. With SBF it was mostly the latter, I think.

Amodei is a smart guy, but there are aspects of his behavior that almost seem like he's doing an over-the-top exaggeration of neurodivergent mannerisms as a kind of act.

I don't see how one can seriously argue that Amodei was similar for holding an optimistic view of AI's progress in the next 5-10 years.

Amodei might be an even bigger optimist than I am, but he's not lacking credentials on neither the biology nor AI side.

Accusing him of fraud or mania like some here do seems unwarranted.

Agree with this if it was something a random person said over coffee or lunch.

But when its the CEO of a company that is about to IPO saying it, its very very different and more concerning.

A wildly absurd claim, especially for some of the most studied diseases in human history. This has only ever been true for a tiny handful of very specific diseases, such as scurvy, in all of human history. Expecting that cancer can be cured or prevented using already existing drugs that we've just not been noticing is complete quackery.

> Even current AI can solve a LOT of tasks if it has quick feedback loop and can verify results.

IF it has a quick feedback loop. You know what definitely DOESN'T have a quick feedback loop? Research into medicine, and biology in general. It takes raw time to find out what the effects of a substance on an organism are, often weeks or months.

> So is it that unrealistic to assume AI can find a correlation on massive amounts of data and figure our what leads to Alzheimers? (where the solution could be not cure but prevention). I don't think so.

Yes. Because we've already run all of the analyses, and we have plenty of observed correlations. What we don't have is the kind of data that can lead to deducing causation - because it's a disease that stays hidden for a long time, it progresses glacially, and the people who have it do or don't do huge amounts of things in the period before getting it and during early onset that are impossible to track or verify - so the data is both gigantic, and extremely incomplete.

Compute with the power to solve literally any medical issue is also compute that can create bioweapons that instantly kill any human.

Might as well fart into the microphone, in my humblest opinion.

The architecture of ML as it currently stands does assist us in modeling things like protein folding. But that's what they are; models. And before some person chimes in by saying "look how the AI pessimists move the goalposts", see my first point. Biological systems have high entropy and discoveries in biology often do not follow from preexisting discoveries whatsoever and are completely novel. Barring some quantum leap in ML design and architecture you cannot predict stochastically or deterministically how exactly a new enzyme will function in vivo. There is absolutely no evidence that Anthropic has solved this, even if they gobbled up every ML tool used in drug design and trained their models on every biomolecule known to man, it would not work because of fundamental gaps in what can be observed. They are a for-profit company trying to juice their IPO.

Now could LLMs and ML potentially speed up the pace of this research? Absolutely. But it's like saying a Formula 1 car is faster than a NASCAR. Yes, one is very clearly faster, but it's not as if the NASCARs are slow.

Yes. The mistake you're making here is that you're not understanding where the actual bottlenecks are in drug development, and why some targets like Alzheimer's have been so elusive.

The problem with Alzheimer's in particular is that we don't have a working model of how Alzheimer's works. The dominant model--the amyloid hypothesis--can pretty conclusively be chucked in the bin at this point because of the number of amyloid clearing drugs that just simply have no effect on the progression of Alzheimer's, but there's no model that actually looks any better. Without a proven model, the only to verify effectiveness of an Alzheimer's drug is try it out in humans [1] and wait several years to get a success-or-no-success readout. That's the rate-limiting step for Alzheimer's drugs.

Okay, so maybe you're going to suggest that we can use AI to come up with new Alzheimer's models. Okay, but the rate-limiting step is still proving those models, which because of the human-only nature of Alzheimer's, requires developing the drugs and testing them and you're going to wait a decade to get back your answer as to whether or not the theory was correct. And what data are you going to use? If you're trawling scientific literature, you're going to come up with only the same ideas people have already had for alternative Alzheimer's models (there are a lot of those). If you're trawling medical data, well, medical data just doesn't exist in the level of detail you need for that enterprise, and handwaving away the legal issues with collecting it, it would still take years just to collect that data (and probably still not be good quality!).

[1] There's another complication of Alzheimer's, which is that only humans get Alzheimer's, so we can't even do anything like testing in animals with shorter lifespans to speed this part up.

[0] Anthropic's CEO says that in 3 to 6 months, AI will be writing 90% of the code software developers were in charge of. https://www.businessinsider.com/anthropic-ceo-ai-90-percent-...

90% by LOC, or by semantic meaning/value?

So re-interpreting it now as "Software engineers will be using AI for 90% of their code" seems like a pretty major retcon.

I don't think it's impossible that he's right. I'm ultimately buying Anthropic stock regardless of this discussion. But I would not take it particularly seriously if the CEO of Moderna took a break from their amazing work applying mRNA vaccines to diseases, and told the world that soon mRNA will solve climate change.

But across the wide world of software? Everyone I see is just using AI as a compiler. They are still writing the code.

(Yes, compilers also write code, but I don't think that was the intent of the message. We've been doing that for decades. 90% of the software has been written by AI since we discovered the optimizing compiler if that was the intent.)

You might not be, but I'm certain that the CEO of Eli Lilly, the 4th largest biomedical firm in the world, knows more about what it takes to cure diseases than the leader of the Claudebot cult does.

Openness and sharing is precisely why the Internet was so successful. Anthropic, by contrast, is a proprietary SaaS vendor with monopolistic ambitions that’s openly seeking regulatory capture and actively supports the swarms of AI bots that are killing what remains of the open Internet by driving everyone’s hosting costs up to an unsustainable level.

LLMs, while definitely a useful tool have not massively replaced engineers.

If we can’t even have huge paradigm shifts in the area it’s specifically tailored toward, there’s a low probability it will eliminate entire fields that require physical testing and extreme domain knowledge.

We don’t have self driving cars, solid state batteries, or humanoid robots. None of this has been sped up by AI in the way Amodei is suggesting.

But that's more like super-knowledge-connector. Intelligence would require, at the minimum, be the ability to discover/create things that don't logically follow from anything else. So for example - advancing some esoteric field of mathematics by connecting things in a previously unexplored way = super-knowledge-connector. 'Inventing'/discovering mathematics from nothing = intelligence.

On the other, we're still very much in the time of headlines which should say "in mice" but don't because the uncertainty is that level of bad.

Uncertainty will have to come down a long way before we can even do without animal testing, let alone do without human testing.

If a level of intelligence far beyond humans is achieved, it probably could eliminate most of the uncertainty about new treatments.

OK. If you already have perfect confidence in a treatment, what would a trial do for you?

Broken clock doesn't apply here

that’s funny, my customers are completely different than they were a year ago

people that would have never paid for coding because it was too expensive and risky are now getting their MVPs validated, earning and investing into more

Sure LLM is good to develop a feature but in 90% of cases in my career the problems I was working on weren't just coding and in that part, LLMs can't do everything.

It's also important to know that a typical dev spends 400-1000$ on API tokens that aren't still profitable for them.

Sure AI's useful and it will reduce the demand for software engs. to a point, but it won't eliminate them. At least I hope. Same with other fields... let's hope?

My previous company has a huge department where LLM coding is banned. The rationale is they work on really important industry secrets, as well as deal with customer's data. The risk is too high that something gets leaked.

The engineers bought the Kool-Aid.

The reality is: The company is bleeding money, and SW isn't their product (they're a HW company). They don't want to pay the bill.

it's being "borrowed" in the sense that there's an implicit promise as part of the IPO sales pitch, but shareholders are last in line to collect in case of something as extreme as a bankruptcy - and that's if there is anything left if that was to happen, which by definition is not frequent.