Achievements of Artificial Intelligence (AI)

Not trying to derail the topic, but anyone else getting incredibly concerned at the recent developments?
I used to just laugh and shake my head at stuff like this, not anymore
I feel the same way. The increase in acceleration has been very extreme as policy, society and everything else seems to be left lagging behind wondering about possible implications, whilst the implications are already shaping their future. It seems these are extremely uncertain times with little vision on how this ship can be steered to a harbour that is enjoyable for all. One can only hope for the best.
 
Last edited:
you could get RNA-seq data for 10 million cells exposed to hundreds of different drugs, and that model will not be able to generate relevant predictions for what happens to cells exposed to e.g. a specific virus because that involves new gene programs and interactions therein.
Not so sure about this because, in essence, it comes down to how much one can extrapolate from the data we collect. How many more scientific rules and principles can an AI distill from those than us? I guess that it's likely a lot.

Perhaps the AI will be able to predict those novel interactions quite accurately based on the data that it does have. It only needed a few protein folding examples to predict almost all of them, and no human figured this out.

I see the same thing happening in the next couple of years for genetics. Nobody can tell me what the DecodeME results mean biologically except for some general conclusions. I suspect we already have sufficient data on the genome to distill general principles that will allow AI to accurately predict what specific DNA mutations do exactly.

The same will likely apply to many other biological systems, and AI might also help in design/creating new measurement tools for biology.

Perhaps a bit of a weird thought experiment, but one could have made similar reservations to the human enterprise of understanding the world several centuries ago: there's too much complexity, too little data; you won't be able to predict anything, etc. But one of the surprising things is how much progress we have been able to make in science by simplifying things and just focusing on recognizing general principles and extrapolating from them. For this reason, I think that AI will be able to make a lot of progress in biology and physics as well (but not as fast as in math or coding).
 
The increase in acceleration has been very extreme as policy, society and everything else seems to be left lagging behind wondering about possible implications, whilst the implications are already shaping their future.
I think most of the world doesn't see this yet, because you have to have one foot inside programming or math to really understand how powerful AI is at the moment.

Most people's experience of AI is asking it a question instead of googling it - so a single agent, on a free version, not the latest model, and a simple question instead of a complex task. In such scenarios, AI sometimes makes stupid mistakes. Same with AI images or videos, where you can still often see that it's 'fake'.

People might have heard news stories about AI breakthroughs, but these math problems are too complex and theoretical to understand how significant the solutions are. And when you look up info about AI, there are thousands of bots or influencer accounts trying to hype it. So it very much feels like crypto or NFTs a couple of years ago. I suspect a lot of people think it's just a new hype, some useful technology that is currently overvalued and in a bubble. And all those predictions and warnings from AI people are ignored because it's just a way to increase their stock value.

It's difficult to communicate all of this to the general public because if you explain how powerful AI is and might become, you just seem in on the hype. I think the only way is for top experts in the fields it disrupts, like math, coding, design, etc. to speak out about it.
 
Anthropic researcher says more than 10% chance AI 'could kill all humans'.

London — A lead researcher at Anthropic, one of the world's leading artificial intelligence firms, said Wednesday that he believes there is a more than 10% chance AI "could kill all humans" within the next decade.

"We really do earnestly believe AI could kill all humans! I personally think it is >10% within the next decade," Evan Hubinger, the San Francisco-based company's Alignment Science Lead, said in a post on X. "I believe Anthropic is trying its best, but we do not yet have a plan to solve alignment for superintelligence and are not clearly on track to."
 
Anthropic researcher says more than 10% chance AI 'could kill all humans'.

London — A lead researcher at Anthropic, one of the world's leading artificial intelligence firms, said Wednesday that he believes there is a more than 10% chance AI "could kill all humans" within the next decade.

"We really do earnestly believe AI could kill all humans! I personally think it is >10% within the next decade," Evan Hubinger, the San Francisco-based company's Alignment Science Lead, said in a post on X. "I believe Anthropic is trying its best, but we do not yet have a plan to solve alignment for superintelligence and are not clearly on track to."
 
Jacob Coxon researcher thread on X:

"I resigned from Anthropic today. I spent the last three years doing pretraining research at both OpenAI and Anthropic. Neither company is acting responsibly. They are racing straight to self-improving superintelligence and gambling with our lives. More thoughts below."
 
Some statsmfrom OpenAI’s own research papers (or press releases if you prefer) that seemed noteworthy on scale and cost here and in wider use. This is not to me a sign of cost reduction.

the group that produced the Navier–Stokes resolution involved on the order of 10,000 concurrent agents.
Across all attempted problems, the agents sent 4.9 million messages and used about 300 billion output tokens. In the process of resolving the Navier–Stokes problem, the agents sent 2.7 million messages and used approximately 130 billion output tokens.
Source: https://openai.com/index/navier-stokes-solution/

At the start of this year, the median researcher ranked by agent usage at OpenAI was using coding agents only in modest amounts. By mid-August, the median researcher was integrating agents daily into their work, using more than $600 per day of inference at API prices. The 90th percentile user in our research organization now uses more than $7,000 of tokens per day.
Source: https://openai.com/index/research-acceleration-view-inside-openai/
 
The same will likely apply to many other biological systems, and AI might also help in design/creating new measurement tools for biology.
I agree wih a lot of what you’ve laid out. I also think I see the potential limits others talk about. It feels like we may end up in a place where we can say a great deal about say, what genetic variation or protein interactions do, so be able to make use of that. But still be nowhere near say, modelling a cell. They seem different sort of problems and maybe different layers of abstraction.

And it’s a bit like weather prediction now, where AI systems are doing amazing things (with less compute than traditional forecasts interestingly). So they are obviously understanding something in patterns of complex systems but we don’t seem to really understand what. And there’s oddities like them sometime being worse short term but better long term forecasters. They can do that without solving all the underlying equations or all the physics problems we would need to run a full scale simulation of the weather system of the planet. So maybe we’ll be able to go a long way with some biological problems and treatments without really understanding all the things going on in a cell. It seems what we already do with a lot of medicine.


If you asked me months ago I would say the AI researchers were trying to juice their IPOs
There are still an awful lot of motivations here to do some odd things. It’s difficult to comprehend everyone’s reasons. Maybe an AI can tell us ;)
 
Last edited:
Some statsmfrom OpenAI’s own research papers (or press releases if you prefer) that seemed noteworthy on scale and cost here and in wider use. This is not to me a sign of cost reduction.

My knowledge may be outdated, as I haven't tracked recent developments, but I was under the impression that "scaling by data is dead" because all the data has been used, so it has been scaling by compute for a while now. And I think that's how you get these numbers.

The price of compute drops steadily but the rate is slowing down (see, e.g., Here) as we're nearing the physical limits. I think the AI companies are far from profitable yet, but they will need to turn profitable eventually. I don't dare to make a guess how it will turn out. Will the development stall? Will there be a crash? Or will they achieve their AGI?

For me, the latter seems to imply smarter algorithms but not sure how much progress is made on that front. It seems like scaling what we have was the way forward so far. To me, it seems the fundamentals haven't changed, as we still see LLMs making the same silly mistakes (though certainly less often and maybe even less silly).
 
Sam Altman tweeted last night they will try to point their new model at finding a room temperature super conductor. I imagine any of these break throughs in physics will pay for itself and will be a government asset so to speak. Cold War 2.0 except unlocking the secrets of the universe
 
Sam Altman tweeted last night they will try to point their new model at finding a room temperature super conductor. I imagine any of these break throughs in physics will pay for itself and will be a government asset so to speak. Cold War 2.0 except unlocking the secrets of the universe
I feel like this will have a far harder time than brute forcing hunches on lean. Since they can’t actually verify if a hunch is correct without a lab, funding, and lots of money.
 
without a lab, funding, and lots of money.
I think the only thing they are missing is a lab and you can pay for that pretty easy when you have seemingly unlimited money as they do

They are already doing that for bio, the Claude mini protein paper they just paid a wet lab to spin up their proteins after they designed them. No need to build infrastructure if you’re rich rich

Imagine the guys who just solved the NS millennium prize just goes to the government and says we have a super conductor at room temp, I’d wager they’d give them the resources even if there wasn’t private interest
 
Imagine the guys who just solved the NS millennium prize just goes to the government and says we have a super conductor at room temp, I’d wager they’d give them the resources even if there wasn’t private interest
Rather, we have 10.000 candidates for a super conductor, one may prove useful.

Obviously not deep enough into solid state physics to judge feasibility here, but everything from Altman I treat as marketing.
 
According to the blog it seems like the researches recognized the progress in an important subproblem for Navier-Stokes and then primed the AI groups accordingly (of which a massive amount was working in parallel on the problem). That probably reduced the search space (and cost and time) drastically.

It is unclear to me, how good those models are at recognizing important intermediate results themselves. If they are not good at this, then cost and time explode because it ends up being an (somewhat) unguided exploration of an exponentially growing space of possibilities.

In that sense Navier-Stokes was maybe tractable, because there has been a lot of progress already. Other problems are further away and may still be prohibitively expensive (to cite my computer science professor: "we have no idea how a method to solve P-NP would even look like").

Of course it is still extremely impressive they solved a millenium problem. Not trying to belittle this in any way.

Guess I'm just saying this does not mean that any specific problem is suddenly solvable (yet). It's still shotgun principle. Do a lot and see what sticks.
 
Cold War 2.0 except unlocking the secrets of the universe
Yes, it also feels like a pissing contest between OpenAI and Anthropic. Solving scientific problems as fast as possible to show investors how advanced their model is.

At least they're throwing talent and money at things that might actually be useful - probably better than creating credit default swaps or social media algorithms.
 
Except CoT doesn’t necessarily tell us what a model is doing. And some models are apparently increasingly manipulating their CoT to mislead. It’s wild that these ideas of observability and ‘what if a model did this’ were just ideas a few months ago and now seem to be happening. Another idea is ‘well are they only doing this because we discussed the idea of them doing this and that’s where they got the idea’. It’s all very….recursive!
That’s what you’d expect from any model. As soon as you target a measurement, it will throw everything you do not measure out the window in order to maximise that measurement.
 
Back
Top Bottom