While this is true - and I see the same sentiment as i am about to make ad nauseam - it is kind of amazing how far the goal posts have moved in such a short space of time.

It’s not a goal post. It was literally a fucking benchmark of the very first LLMs.

sorry, can please explain this? which benchmark are you referring to?

Not sure why ds cache costs are confusing? Their kv cache is absurdly compressed compared to almost everyone else.

It is confusing because they are the only ones doing so whereas everyone has an incentive to apply similar techniques.

DS4's architecture was redesigned for cache efficiency. It's the model itself that is just way more cache efficient, in addition to a distributed KV cache (which they open sourced last year). You could run your own LLM provider just for DS4 models, recreate their setup, and get the same cache efficiency, but it would only work for that one model family.

Other models would also need to be rearchitected to get the same benefit, and I suspect they don't want to, because it seems to create both larger models and slightly less intelligence (DS4* has smaller active experts to get faster results).

Solve by collecting taxes properly and then using the funds to provide research grants. Then it's a shared risk, shared reward.

But no instead we have our governments pissing away money on useless bullshit and private enterprise bending over backwards to avoid paying corporate taxes, entire countries who exist to be tax havens. And an apathetic voterbase who don't understand anything but will tell you why they don't like it all, anyway.

That’s certainly the long term goal, but it seems foolish to abandon it without a solution to the problem.

>(I'd need a Lisp that runs in the browser, for starters!)

We would be in a very different world if Netscape’s exec team had let Brendan Eich just use scheme as he had planned to.

Block scoped let would have been there from the start, for one thing, rather than nearly two decades of function scoped var.

You could say the same about more or less any language feature - the thing is though that having a standard feature people can agree on is better than several people’s divergent opinions on which macros are sufficiently aesthetically pleasing.

You’re only considering the front half of the network though, classification is a helpful example for explaining the concept - but if it follows through to the output then you could quite easily imagine just shunting some activation from one branch of the representation tree to another to steer behavior. Which we already do with steering vectors - but if you could read off the actual topology you could do that with much higher fidelity.

So make one. Not everything is for you.

presumably if you are generating text for yourself to read prior to you writing something for a human.

>I see this quote being thrown around a lot as if it supports the article's point, but this quote is just some of the most famously smart, skilled, and effortful people ever admitting even they cannot condense their knowledge that much.

But that’s explicitly not what the quote says - it’s not that they can’t, it’s that it takes work.