zacmps

People would do the same if curl was publicly funded, that doesn't mean it would be a bad idea.

It's not something I've needed to deal with personally.

We have run into added content filters in Azure OpenAI on a different application, but we just put in a request to tune them down for us.

I've found the highest accuracy solution is to OCR with one of the dedicated models then feed that text and the original image into an LLM with a prompt like:

"Correct errors in this OCR transcription".

In some countries (NZ for example) this does not exempt you from reporting requirements unless your total spend falls under NZD$15,700.

For me the demo pulled up flights, but in USD, I'd need it to be localised to be useful.

Airlines also tend to present different prices for different locations/currencies so I doubt this would be as simple as a UI change (though that would still go a long way!).

You can also use a mutex that doesn't own data in Rust if you need to.

Correct

Sight correction, at least in my understanding (physics major but only undergrad) this explicitly isn't communication.

It's not possible to use entanglement to transmit information faster than light, even though the entanglement itself could be simultaneous.

It's been brought up elsewhere in the thread but GCP deleting an entire tenant's account rules them out entirely from any serious business IMO.

Super cool to see you here.

I've also looked at ray for running data pipelines before (at much much smaller scales) for the reasons you suggest (unstructured data, mixed CPU/GPU compute).

One thing I've wanted is an incremental computation framework (i.e., salsa [1]) built on ray so that I can write jobs that transparently reuse intermediate results from an object store if their dependents haven't changed.

Do you know if anyone has thought of building something like this?

[1] https://github.com/salsa-rs/salsa

Personally I think chemistry is going to trail a long way behind biology for a while (in terms of ML solutions). The data, supporting libraries, and funding doesn't seem to be on the same level.

One of the (many) problems with the market is that consumers have imperfect information. In this case specifically it's almost never possible to have an accurate picture of a company's security in advance of a mistake like this one.

I'm sure it's no where near good enough yet, but a legal model getting the answer right 83% of the time is still quite impressive imo.

Syngas can also be converted to other products through gas fermentation.

Seconding this, I'd love to add an interface like this to a RAG setup.

Nope, I tried it for graph and diagram understanding and it wasn't good enough. Planning to repeat the evaluation with 4o when I have time.

They might have meant a type like `Result<T,E>`

What?

Anonymous functions are amazing for: * Iterators (map, reduce, filter) * Callbacks (ui actions) * Threads/async tasks

The reason I don't name them isn't because I can't think of a name, it's because: * There's no point, they'll only be referred to once * Introducing a separate function block breaks the flow of the code and makes it harder to read (especially for iterators!)

It might not be obvious when you hit that point, bad regexes can be subtle, just see that old cloudflare postmortem.

On an SSD I'd drop random to 1-1.2.

It absolutely makes sense viewed through the lens of early 2000s antitrust cases (esp. those against Microsoft).

The landscape has obviously changed since, but not for any good reason in my opinion.

Ah true, I guess it's still 258 positions by 100,000 possible tokens though.

Remember, if it's using a similar tokeniser to GPT-4 (cl100k_base iirc), each token has a dimension of ~100,000.

So 258x100,000 is a space of 25,800,000 floats, using f16 (a total guess) that's 51.6kB, probably enough to represent the image at ok quality with JPG.

I think the same thing happened to 'actual.'

> The author (the human learner) has some formal experience in linguistics and has studied a variety of languages both formally and informally, though no Austronesian or Papuan languages

From the language benchmark (parentheses mine).

I've been using 'build' as in, 'build a dataset'. I think it gets the same idea across without being as easy to confuse with fine-tuning.

It's possible to build RAG pipelines that support answering complex questions over multiple documents, and I don't think I would say the whole architecture needs to change.

Doing it fast is another story, LLMs are pretty high latency at the moment.

LLMs also have huge potential for helping us understand genetics, see Facebook's ESM for example.

Because it causes preventable harm.

If they've done any research at all they almost certainly do know more.

I also live in NZ, did you read the CBA report on our feebate scheme a while back?

It concluded the policy had a net economic benefit due to correcting consumer's failing to take into account the cost of fuel in making ICE/hybrid/PEV purchasing decisions.