Hacker News Board Hacker News, browsed like a classic message board
Live data from Hacker News

A conversation about AI for science with Jason Pruet

lanl.gov

Post reply on HN
Print view View original

141–150 of 163 posts

Re: A conversation about AI for science with Jason Pruet

#141

Post by mallardgryph »

__loam wrote: post #134
KolibriFly wrote: post #128

Earlier quoted context omitted.

The gap between demo impressive and reliably useful in real-world scenarios is still huge in a lot of cases. That said, I think the concern isn't that AI is flawless now, but that it's improving at a pace we haven't really seen with other technologies

It's really funny to see someone say this technology, which seems to be plateauing with marginal improvements after a few years, is improving at a pace we have never seen before, in an industry that is built on the microprocessor. Astoundingly ignorant.

The marketing for it all has been something else. It's a technology that demos extremely well, and is surface level very impressive, and to those who wern't paying any attention to its evolution, appeared to come out of nowhere.

But this line of "its only going to get even better" is a mantra that's endlessly expounded on, a brain worm even. It's never backed up with any observable evidence. It's marketing that they're tricking people into repeating.

Re: A conversation about AI for science with Jason Pruet

#142

Post by sampullman »

fsflover wrote: post #136
sampullman wrote: post #120

Earlier quoted context omitted.

I'm not sure if the analogy fits. I would say AI is more like a fast car that turns the wrong way every couple of blocks, or a car that randomly refuses to go on certain roads. Is it still a useful mode of transportation? Sometimes, but, it's not reliable. The sibling comment says it better - it's about predicability of mistakes, and the effort and knowledge necessary to correct them. Maybe the next generation of mod…

> AI is more like a fast car that turns the wrong way every couple of blocks So if you drive well and can control this "car", you are able to fix that and benefit from the fast "car" a lot. > or a car that randomly refuses to go on certain roads This doesn't seem like a good comparison. AI will do anything, it may just be wrong.

The AIs that I use will definitely not do anything, and it the limits are not always obvious without a lot of prodding.

I'm fine with ditching the car analogy though.

Re: A conversation about AI for science with Jason Pruet

#143

Post by amelius »

saretup wrote: post #117
sampullman wrote: post #115

Earlier quoted context omitted.

I appreciate the value of AI in my own work, but "they’re very good coders" and "they make trivial mistakes or hallucinate" seems incongruous. For me, AI tools shine when I know enough about a topic to quickly error check, but not enough where I can code fluently without documentation. I'm sure it will get more useful over time, but that's where it's been for me for the last year or so.

> "they’re very good coders" and "they make trivial mistakes or hallucinate" seems incongruous. It seems that way if you judge them the same way you would human coders, but they’re different. They might be able to do things that veteran coders can’t without spending days on it, and fail at things that beginners can do in half an hour. A car might not be able to traverse difficult terrain as well as a horse, but it do…

The problem I have with AI is that you cannot reason with them after they make mistakes. They always reply with "Good point!" and then they give me another iteration of bad fixes. You really have to know when you should give up and do things manually.

Re: A conversation about AI for science with Jason Pruet

#145

Post by chairhairair »

amelius wrote: post #143
saretup wrote: post #117

Earlier quoted context omitted.

> "they’re very good coders" and "they make trivial mistakes or hallucinate" seems incongruous. It seems that way if you judge them the same way you would human coders, but they’re different. They might be able to do things that veteran coders can’t without spending days on it, and fail at things that beginners can do in half an hour. A car might not be able to traverse difficult terrain as well as a horse, but it do…

The problem I have with AI is that you cannot reason with them after they make mistakes. They always reply with "Good point!" and then they give me another iteration of bad fixes. You really have to know when you should give up and do things manually.

Gemini 2.5 is much more confident.

Anecdotally, I think this behavior is undesirable for most commercial LLM use cases. I have several friends that have complained about Gemini’s “back talking” and prefer ChatGPT’s relative sycophancy.

Re: A conversation about AI for science with Jason Pruet

#146

Post by hbartab »

hdivider wrote: post #13
hbartab wrote: post #6

> We certainly need to partner with industry. Because they are so far ahead and are making such giant investments, that is the only possible path. And therein lies the risk: research labs may become wholly dependent on companies whose agendas are fundamentally commercial. In exchange for access to compute and frontier models, labs may cede control over data, methods, and IP—letting private firms quietly extract value…

I fail to understand the sentiment here. This is the intention of tech transfer. To have private-sector entities commercialize the R&D. What is the alternative? National labs and universities can't commercialize in the same way, including due to legal restrictions at the state and sometimes federal level. As long as the process and tech transfer agreements are fair and transparent -- and not concentrated in say OpenA…

If this were just about tech transfer, in which private firms commercialize public research, I agree. But that's not what Jason Pruet is saying. In the Q&A he notes:

> “Why don’t we just let private industry build these giant engines for progress and science, and we’ll all reap the benefits?” The problem is that if we’re not careful, it could lead us to a very different country than the one we’ve been in.

This isn't about commercialization, it's about control. When access to frontier models and SOTA compute is gated by private interests, academics (and the public) risk getting locked out. Not because of merit, but because their work doesn't align with corporate priorities.

Re: A conversation about AI for science with Jason Pruet

#147

Post by bobxmax »

fumeux_fume wrote: post #113

> If you’ve played with the most recent AI tools, you know: They’re very good coders, very good legal analysts, very good first drafters of writing, very good image generators. They’re only going to get better. Most of the bullshitters will tip their hand pretty early that they're just hype men for AI. Right off the bat, the fact that AI is disruptive and transforming society is apparently self-evident because they n…

I'm sick of this forum's lame performative cynicism masquerading as depth—it's a lazy cop-out that spares you people from the uncomfortable, demanding work of actually building something better.

Technology that any of us 5 years ago would've thought was a hundred years away and all you read is moving goalposts from ornery developers in denial.

Silicon Valley is cooked.

Re: A conversation about AI for science with Jason Pruet

#148

Post by bobxmax »

KolibriFly wrote: post #128
fumeux_fume wrote: post #113

> If you’ve played with the most recent AI tools, you know: They’re very good coders, very good legal analysts, very good first drafters of writing, very good image generators. They’re only going to get better. Most of the bullshitters will tip their hand pretty early that they're just hype men for AI. Right off the bat, the fact that AI is disruptive and transforming society is apparently self-evident because they n…

The gap between demo impressive and reliably useful in real-world scenarios is still huge in a lot of cases. That said, I think the concern isn't that AI is flawless now, but that it's improving at a pace we haven't really seen with other technologies

there are hundreds of thousands of developers, lawyers, artists, musicians, etc who have transformed what they're doing with AI

Re: A conversation about AI for science with Jason Pruet

#150

Post by amelius »

chairhairair wrote: post #145
amelius wrote: post #143

Earlier quoted context omitted.

The problem I have with AI is that you cannot reason with them after they make mistakes. They always reply with "Good point!" and then they give me another iteration of bad fixes. You really have to know when you should give up and do things manually.

Gemini 2.5 is much more confident. Anecdotally, I think this behavior is undesirable for most commercial LLM use cases. I have several friends that have complained about Gemini’s “back talking” and prefer ChatGPT’s relative sycophancy.

I was talking about ChatGpt, but other llms also have this problem. I don't really care how polite they are, etc., I just want to get the job done, but the llm often gets stuck at some point.
Post reply on HN