tsukikage

Why the indirection? - why not have the jump table reference f0..fX directly rather than jumping to a jump?

The article makes no mention of optimisation use cases. In case anyone already knows off the top of their heads and before I start digging for the answer, is part of the intent here for these to inform the optimiser stages, perhaps similarly to what the if (!(some_condition)) { __builtin_unreachable(); } idiom can achieve?

Between this, the social media stuff and whatever further lockdown attempts are in the pipeline, that's an incredible amount of incentive to find hacks and workarounds. We'll end up with a generation that actually gives two hoots about owning and controlling their hardware and media, and has the engineering skills to achieve that.

A somewhat roundabout way of accomplishing that goal, but I'll take it!

"One body, three behaviours"

Much of the text is generated as well.

If your benchmark score beats the current incumbent, you get to wipe and replace them. There can only be one!

It's just like cookie banners: they shouldn't exist, but people's bonuses rely on never admitting that, so here we all are.

> Nobody wants this

"Oh. Well, then stop."

Have you ever considered that "well, it works on MY machine!" is perhaps not the most helpful response to people voicing frustration at something that is not working as they expect for them?

Realise that you are talking to some mathematics in a box. You can't get a rise out of it. It cannot feel guilt or remorse. Whatever emotional payoff one might want from "I asked for ONE button" cannot be had here. Meanwhile, not only are you being charged by the word, but every word you send that is not directly on the path to getting what you want done is just noise in the maths getting in your way.

So leave emotion at the door and make your words count. Voice your frustrations at the stress doll next to your monitor, sure, but spending tokens on them us just wasting time and money. Explicitly typing out your exasperation will not get you closer to whatever it is you are trying to accomplish. What can you type that will?

Laser focus: clearly, concisely explain what is wrong right now, and how you want it solved. Then do that again until all the moles have been whacked. If the AI is stuck in a sycophancy loop, start a new clean session. Do this frequently anyway: every turn in the same session charges for all preceding conversation again /and/ degrades the LLM's performance.

Traditional forum/wiki spam is intended for search engine crawlers, not humans. How this new stuff relates to human users of the resource does not seem to differ in any way that is significant.

“You could have LLM, LLM, chips and LLM. There’s not much LLM in that.”

We don't understand what those are or how they come about, so how does that help us objectively reason about whether a particular black box does or does not have them, especially when it behaves as though it does, beyond arbitrary I-know-it-when-I-see-it "things I don't like don't have the property" style declarations?

Meanwhile, entirely too frequently I see unit tests that have atol / rtol values which permit errors of one part in 200 or even higher, where people bother with unit tests at all. At some point someone in the pytorch ecosystem used rtol=5e-03 in a unit test and crowds of people started passing that around unquestioned.

With attention matrix sizes being what they are, that's high enough that you can literally zero out a row or two and still have the test suite pass. Guess how I know!

Check your numbers, folks.

> Or are they just using it for training

You’re making a distinction here, but training is something you repeat for every new point release, so you need to keep the data if you want to use it for training.