why? We know that we influence it, but the direction is unfavourable. why its bad to want to control it in the right direction?

General question: what is this model good for? I have mixed results with Laya on a pdf classifier (Jev was doing much better).

I am not sure how this process looks like. When they "discover" these, what are they actually doing?

    The agents ran quantum-mechanical simulations of each crystal with the standard method for this, density functional theory, at two levels of approximation: a faster one (PBE+U) and a slower, usually more accurate one (HSE06). The band gaps and spin windows below come from the more accurate one.

So the agent runs a classic simulation or I am missing something.

A lot of the public successes with agents is really LLM-driven local search against an objective function that is evaluated in more traditional ways. This one seems to fit the pattern.

This fits one of the patterns I wrote about ~1.5 years ago:

https://orlp.net/blog/bad-ai/#objective-p-mathrm-relevant-1-...

I think it still holds up.

I think a better way to put this is that success with LLMs is guaranteed by defining what success looks like without LLMs first.

Modeling superconductivity with DFT is tricky, there are plenty of DFT reports from reputable groups explaining why LK-99 should be superconducting. It’s a limitation of the theory, DFT can’t model correlated electron states well, and it’s not great at finite temperature, and both of those are important for superconductivity.

Edit: I somehow missed that this is about magnetic semiconductors (not superconductivity) so DFT is a bit on better footing here. I still think it’s a bit challenging predicting magnetic ordering at elevated temperature, but maybe not as difficult as superconductivity

I'm under the impression that this kind of modeling is one of the applications that quantum computers are likely to be good at.

I'd imagine there's a lot of documented research which has attempted to find such things using classical computers.

Seems like there would be a lot of well structured context for somebody to use while directing agents to repeat that research, now with updated models once quantum computing is ready for that kind of task.

They ran Quantum Espresso which is ok, but by no means the 'state of the art' for DFT. And in case, any DFT computation has to be taken with a few pounds of grains of salt before getting too excited about it.

No offense to the person writing this (assuming they did at all), but I'm not sure they really understand what they're doing..

not a classic simluation- a quantum simulation. This means they put a lot more work into representing the wave function of the simulation and modelling quantum effects.

They used quantum espresso.. undergrads usually run this in certain classes: https://www.quantum-espresso.org They didn't do any work there.

You misunderstood what I said. I mean that quantum calcs are more computationally expensive than classical simulations ("more work"). I am not saying the authors of this blog did anything special.

This is a strange attitude. When an agent is optimizing a piece of code, comes up with 2 variations, and runs benchmarks on them to figure out which one is faster, then selects one of them based on tradeoffs between performance and other things it reasons about, do you ignore its explanation and all experiment runs?

>Their thought process is effectively undecipherable by humans (it's essentially information arising from information

Are you trying to say that human brains are incapable of inference?

You're confusing the weights of a model and internal chain-of-thought with the output of the model. Yes, we don't know a lot about how the internal mechanisms work. But with the correct prompt, agents will produce a worklog that documents exactly what solutions were tried and how the result was obtained.

What are you on about? I have had Fable come up with new shit for me several times (I do research for a living, so actual new shit nobody knew before), and each time it was perfectly understandable.

Of course I don’t know how it got its ideas for what to try. But heck, I don’t even understand how I get my ideas half the time. But the process, like what code it wrote, simulations it ran etc can be understood by (some) humans just fine!

Yes I saw 3Blue1Brown say the same thing in his tutorial on how neural nets worked where he built a simple model to recognize a particular letter. Good reminder.

I've been dabbling with some of my own (tiny) models recently and it's actually shocking at what they can "learn" despite having _zero_ mention of it in it's training data.

It seems the subscription based plans these big vendors have has different quality on the service side based on time. Is this fair?

RIIR when? On a more serious note frameworks like Rails are getting less interesting because generating the boilerplate got significantly cheaper with LLMs.

I am not sure what you mean by profitable here?

Why is this Jev story hyped this much? Is there any business outcome that was achieved and justifies this?

This is probably why:

Jev was founded by TypeSafe AI, a San Francisco-based startup co-founded by Diogo Almeida, Erik Gafni, and Sasha Sheng.

Diogo Almeida (CEO) is a former OpenAI researcher and primary author of the InstructGPT paper, which laid the groundwork for ChatGPT and GPT-4; he is also credited as a co-inventor of RLHF.

Complex systems have complex operations and problems. I think isolating the read and write workloads and separating the read queries by departments or even users is a distinct possibility especially with technologies like DuckDB.

That is exactly where all the high scale stuff like Trino and StarRocks have gone (and ClickHouse can use S3 as storage too). However, the problem you run into is when you want to index stuff- secondary indices are tightly coupled to the query engine, so in practice you arent going to use StarRocks to write and index data to the object store, and then use ClickHouse or Trino to query it. I think it would be a useful development to decouple writes and reads, but this requires a common indexing scheme, and then once you have this standard, in principle lots more indexing plugins can be written and all compatible query engines can use them. Apache DataFusion is probably the right framework to build off of.

Valid. We need to innovate on this and re-think secondary indices. Storage is cheap.