Because they suspect that, sometimes, different content will be served by HTML vs alternate APIs

French text is somewhere around 10-30% longer than the corresponding English text. I would guess much of what you save on smaller vocabulary is lost on the lengthened text.

What would matter more is the token length, depends how well your tokenizer was trained on french I guess.

A solution to this is apptainer: you configure it to not see any of the host files by default, and mount the repo you want to work on at runtime.

Question, out of curiosity, do you know how User and Permissions work?

A mitm does not redirect your ssh login to their machine

This could unlock a new chapter of Venus exploration!

On the other hand, any source code leak could be catastrophic

Well it won't be 'non-mainstream' for one

Sure it would. No matter how many people see, say, Shpongle on their timeline, they're never going to be mainstream. No one is going to be talking about 空夜coo:ya in the same conversation as Taylor Swift or Bruno Mars. Being slightly more visible in the algorithm for a brief period isn't going mainstream.

And even if that were the case, again, so what? If Youtube makes good non-mainstream music more popular, that's still a good thing. That's exactly what one should want a recommendation algorithm to do.

I'm trying to see the problem here and I can't.

So is the SIMD the magic piece here, or is it the interpolation search? If the data is evenly distributed, that is pretty optimal for the interpolation search..

In the Intel CPU + cold cache case, the quad search matters. In the other three cases, only the SIMD matters.

To put it another way: this is addressed in the article.

> Once the ads are injected directly into the main response is when things get interesting.

This would be where you post-process the LLM response with a second LLM to remove the ad..

I think it will be difficult to remove bias when you ask a model to compare alternative products. The model will simply lie, as with a biased human opinion and you will need to consult multiple models for a diversity of opinion and presumably use a "trusted" model to fuse the results. Anonymity will be a key tool in reducing the model's ability to engage in algorithmic pricing.

Super easy. Barely an inconvenience.

Not only that, but the underlying model may be tuned to omit mentions or data about competitors entirely, an absence which can't easily be filtered.

Extortionate economic shadowbanning, here we come.

> will simply lie, as with a biased human opinion

Is this really how bias works?

Writers have many options to deceive their audience without outright lying.

If a journalist is given an all-expenses-paid trip to an exotic location for the launch of a new product, and they review the product and say it's great - are they lying?

If a reviewer writes an article comparing certain types of product, but their review only includes products where affiliate links pay a 10% commission - are they lying?

If a journalist is vaguely aware of rumours about newsworthy, under-reported Event X but also that their publication has a big sponsorship deal with folks that Event X makes look bad, and they don't investigate the rumours or report on them - are they lying?

If a reviewer hears a claim from X, and they report the claim credulously, without adding the context that X has a history making false claims - are they lying?

Oh no. Definitely not. Humans would never just lie. They always lie only if they're biased. That is, after all, the definition of how a bias works.

/s

I'm using bias to mean hidden motivations to the benefit of other parties. Feel free to substitute a better word.

EDIT: actually I'm really not sure what hairs we're trying to split here. I see bias as a departure from objectivity. It can be conscious or unconscious, but when someone is selling something, it's frequently conscious and self-serving, and I believe that's referred to as a lie.

This is already how email works in the corporate world.

A writes email with chatgpt to B.

B sees big blob of text and summarizes email with chatgpt.

Adding an LLM in the middle is just the next step.

It's like one of those memes about the worst possible date picker, except for a communication system.

Then you just end up in an arms race that ultimately leads to photocopy-of-a-photocopy output.

... and replace it with two.

If sharing all of your code with the closed providers is OK then it works. If that is a blocker, open weights becomes much more compelling...