This reminds me of the classic "Growing a Language" talk by Guy Steele: https://www.youtube.com/watch?v=_ahvzDzKdB0

That's nice if you're starting from scratch, but if you have existing code to deal with, you don't have the privilege of ignoring the insane subset.

Claude Code already does this, you can access it with /resume, /rewind and /fork. I'd imagine building a version that saves in the repo instead of in the home folder would take very minimal effort.

I'm using it on a large set of existing codebases full of extremely ugly legacy code, weird build systems, tons of business logic and shipping directly to prod at neckbreaking growth over the last two years, and it's delivering the same type of value that Karpathy writes about.

If you watch the demo video you can see how they would get this: the model is not aggressive enough. While it doesn't cut you off, which is nice, it also always waits an uncanny amount of time to chime in.

That should lead to a low recall: too many false negatives. I wonder how they are calculating it.

I hate to be the first one commenting to say this, but here it goes: the flashy LLM writing style, "Apple Event Dialect" in the README and in this comment is very recognizable and also quite irritating. If this is supposed to be boring then just state the facts and the benchmarks to prove them.

For a comment that goes on about not being flashy, the writing tries it’s very best to be flashy

Maybe rather an interactive explanation and exploration environment with embedded programming?

All life is basically refining energy - standing up to entropy and temporarily winning the fight.

It's all about putting the entropy somewhere else and keeping your own little area organised.

People of the earth, remember: unnecessary arm and leg movements increase the entropy! Fear of the heat death of the universe! Lie down when possible!

Yes, in a very local context it appears so, but net entropy across the system from life's activities is increased

"the purpose of life is to hydrogenate carbon dioxide"

-- Michael Russel

I can transfer money from Europe to Brazil in seconds with Wise. I press the button and the money is nearly instantly available in the Brazilian account via PIX. The same in the reverse direction is possible but only if you have a more modern bank in Europe, eg. N26 or Revolut.

I was thinking more of gp's comment "and the first user is an Argentinian bike importer that finds transacting with their suppliers to be challenging"

Wise isn't great for paying suppliers. Their business account limit for debit/credit is $2k, and for ACH is $50k. They have higher limits if you fund with wire, but then we're back at the starting problem again...

And still, you have no way of knowing that the receiving party actually got it. On a blockchain, the source-of-truth "database" is public.

My understanding is that Wise isn't a true international transfer. Wise has money already in a Brazilian account, and when they receive money in their European account then they send you money from their Brazilian account. If they don't have enough money in that Brazilian account then it can't be instant like it is today.

From my perspective, if it quacks like a duck...

and if wise bans you?

Not the full picture: Wise is that big that it has already lots of local accounts and/or correspondent banks; so basicly "you get the money from Wise" but from a "local payment way/scheme" (to which Wise is connected in the background through several layers)

This is a much more optimal solution than blockchain.

that we need one company to achieve such big scale that they literally are regulated in every single country and basically become monopolistic in terms of their influence?

a Blockchain based system can maintain similar effects but with a balance of power

think about it

No worries, if any DLT-system would become somewhere somehow relevant, any regulator will catch it sooner or later.

Though, its an issue for the company, as this allocates a lot of resources wrongly.