I think AI can be used to enhance stuff, like enlarging a small image without detail or quality loss, just that the details can’t be relied upon to be correct.
Like to enhance a license plate to make it readable. We don’t know if the AI will just put random numbers there.
To make good art, cultural work like poetry, the writer needs to have understanding. LLM doesn’t understand or know anything, all it does is fill in the blank using statistics.
what kind of AI companies are listed in the taiwan stock exchange?
You don’t know what you’re talking about man. It’s still a very new technology, and you can’t take the current situation as representative of what’s possible or will become possible.
The people developing these are obviously aware of the potential for errors and the need to reduce those. Go google “scalable oversight” or something (I know you won’t). I happen to be working on a project related to that at the moment — something about evaluating and improving the ability of AI systems to detect errors in complicated topics output by LLMs.
Yes you can. And a lot of humans aren’t writing poetry or creating art either.
Many humans have exactly the same problem. 
Nope, I wouldn’t say so. What “understanding” do you think went into something like the below, for example?
Is there anything you aren’t working on?
First you’re working in chemistry, now you’re working in AI? These are completely different fields.
I’m not sure if you’re calling me a liar here or just curious. 
But it’s a project related to developing chemistry-based tasks for testing AI systems — specifically evaluating LLM outputs about chemistry topics and trying to come up with difficult-to-detect errors on fairly complex subjects (part of the scalable oversight thing I mentioned). It’s probably more interesting than it sounds.
So it’s both. Just so you know, there are a ton of AI applications in chemistry, as with other fields. Not completely different. Computational chemistry has been a thing for like 40+ years, and the same with machine learning etc.
It’s an interesting comment for sure.
Scalable oversight is interesting. Has to be a bit more complicated than Asimov’s three rules. 
I’m not too worried as long as humans control physical resources. Go ahead and be as smart as you want. No power, no worry. 
Now when robots and AI are mining and processing and manufacturing stuff on Mars— well, wake me up gently.
I’m rather enjoying my tiny part in it. It’s interesting to be working on something new and different (to me), and quite good brain exercise to be trying to think up convincing explanations for why things that aren’t true are true while staying within the constraints of the question (and also remembering things learnt in university 20 years ago then forgotten after an exam).
A friend of mine owns a chemical company and they are in the process of “inventing new “ compounds and molecules and such things. I told him that AI is doing that stuff already— but surely by quantum driven big pharmaceutical computers. 
We are limited in our thinking. It’s cool that AI can be an additional resource. But it would make us feel better for sure if we knew that AI was our bitch.
The way we evaluate a neural network for complex tasks, at my work and pretty much everywhere, isn’t that AI has to get everything right, but that AI does at least as good as its human counterparts. As soon as that happens, the speed and consistency of AI already make it way more reliable and useful. In my experience, they usually do a bit better than humans anyway.
So for the ARC-AGI contest, the threshold for winning the contest isn’t set at 100% or even 90%, it’s set at 85% because on average that’s how well people do on these tests.
As for being creative, great human poets and artists don’t exist in a vacuum either. They too consumes tons of previous works before being able to create themselves. I don’t really see the difference between how people learn to create and how LLMs learn to create.
Hmm— I’d like to see the cognitive leaps comparison. The complexity of the human brain shouldn’t be downplayed. Surely there are things we can do but don’t as of yet.
Have you ever used AI to ‘time travel’ and talk to the great chemists of the past? That’s fun and eye-opening.
You can use Co-Pilot to traverse the entire corpus of online chemistry texts. I’ve done it a lot recently to learn brewing, longevity and bread-making, and it makes learning a lot of fun. I would be lost without it.
I would be worried that the hallucinatory nature of AI responses would make it more error prone in obscure, low-documented equations and reactions. I would also be terrified about asking it too many chemistry math questions… seeing as it might easily get a decimal point in the wrong place… but I guess you know this, hence the project you are contributing to.
If you ask an AI “please count the number of letters in the word ‘sparrow’ it can easily get it wrong.” At least, until recently.
I guess you are working on some way to double-check hallucinated answers against wikipedia, or something. I found Co-Pilot is far better at this than others… it seems to have good ‘web-coupling’ and a decent oversight layer, If those are the correct terms.
Anyway, if you have a working public site, please share…
Sure, but if what differentiates people and the current artificial neural networks comes down to model size, then that’s not really a difference in learning and understanding. It also means as we figure out how to make the computation cheaper and less power hungry, ANNs will eventually have more parameters than a human brain.
I tend to think that it’s not just learning or understanding but also experiencing that gives people an edge— for now.
AI begins its ominous split away from human thinking (newatlas.com)
Companies like Tesla, Figure and Sanctuary AI are working feverishly at building humanoids to a standard that’s commercially useful and cost-competitive with human labor. Once they achieve that - if they achieve that - they’ll be able to build enough robots to start working on that ground-up, trial-and-error understanding of the physical world, at scale and at speed.
Are these the droids we were looking for?
They might solve certain ahem-- logistical issues:
Longshoremen union’s demand for total ban on automation questioned as port strike looms (msn.com)
This kind of thing has been talked about for quite a few years now, but I think it’s still at the not-yet-possible stage for the time being. One example here:
https://www.science.org/content/blog-post/ai-drugs-so-far
You’re not too far off. It’s supposed to be questions that are difficult to answer for subject experts and very difficult to answer for non-experts, both with free access to whatever internet resources they want.
It’s something like this (but no affiliation to these particular authors)
Sure, this is just the beginning. We have to cure restless leg syndrome first!
That sounds interesting. Scalable oversight is a fascinating topic, I’m going to read up on it… as AI is getting real fun right now. My questions are like “Ummm, can you tell me about yeast and H2o2 again.” so I am in no danger of exhausting it. ha.
Should you need a data set on the early writings of Boethius, or the cultural impact of Batman, let me know.
google is giving out their model for free.
Here more, from Nature:
https://www.nature.com/articles/d42473-024-00248-3
Fascinating how they are learning from how kids learn new words.
Artificial intelligence (AI) offers the tantalizing promise of revealing new drugs by unveiling patterns lurking in the existing research literature. But efforts to unleash AI’s potential in this area are being hindered by inherent biases in the publications used for training AI models.
By adopting an approach that mimics the strategies children use to understand unfamiliar words when encountering1,
a Japanese company is seeking to bypass this limitation. FRONTEO Inc., an AI-solutions company, with its headquarters in Tokyo, has developed a natural language processing (NLP) model that adds a critical parameter — context — to the AI-powered analysis of research literature.
“When children encounter an unfamiliar word, they grasp its meaning by looking at the surrounding context,” says Hiroyoshi Toyoshiba, chief technology officer at FRONTEO. “Similarly, our engine automatically determines meanings based on context, without relying on pre-existing definitions.”
Promising results obtained by applying this approach hint that it could lead to ground-breaking health discoveries.
Isn’t that just a transformer encoder, like BERT?
hmm… what does that mean? No tokenizer?
I agree. I might go further and say I don’t necessarily see a difference between how LLMs regurgitate bits of their vast learning set and how humans think.