Cristian Georgescu

[Cristian Georgescu @c_georges]

@plainionist If you walk in the dark and keep walking head first into trees, it will feel like progress when you bump in a tree and then find your way around it.

[Cristian Georgescu @c_georges]

@asaio87 Some people can't imagine going to a nearby grocery store without taking the car, but then there are also lots of people who run ultra marathons.

[Cristian Georgescu @c_georges]

@decisionleader In a couple of years it will be 125%

[Cristian Georgescu @c_georges]

@decisionleader We don't have this philosophical discussion about a real and a counterfeit Monet, so why an AI Monet would somehow be in a different category than a counterfeit Monet? Other than the illegal part of trying to sell the counterfeit as a real one.

[Cristian Georgescu @c_georges]

@N_and_ni That's obvious: you'll need to be smarter than the AI l, and that shouldn't be that hard under your assumptions because "smarter than most humans" means smarter than 51% of humans, or smarter than the median human.

[Cristian Georgescu @c_georges]

@unclebobmartin Maybe Codex is more hungry for tokens than Copilot

[Cristian Georgescu @c_georges]

@jeffreyleefunk Email is a medium where form is not important - the informational content is. So writing an email with AI means you need help with the ideas.

[Cristian Georgescu @c_georges]

@asaio87 Maybe that's true also for human intelligence: humans are just primates with bigger brains.

[Cristian Georgescu @c_georges]

@fchollet Words are for communicating thoughts, not for thinking.

[Cristian Georgescu @c_georges]

@Techmeme @rafeuddin_ Instead of making token usage a metric why not use the coffee consumption instead?

[Cristian Georgescu @c_georges]

@unclebobmartin To make the AI vs human test more accurate I would just take the specs for the "Hunt the Wumpus" game, remove all references to the game name, change the rules, and replace all game artifacts like caves, smells, drafts, bats with different concepts. Then feed this to the agents.

[Cristian Georgescu @c_georges]

@unclebobmartin For LLMs there is a huge difference between in training data and out of training. It can play reasonably well chess but if you change the rules just a little bit then the quality decreases significantly. AI basically has a highly compressed lossy version of the all the internet.

[Cristian Georgescu @c_georges]

@skdh What matters is the intent

[Cristian Georgescu @c_georges]

@skdh If they can send messages in the past, then the paper should have already been written.

[Cristian Georgescu @c_georges]

@unclebobmartin @karrisaarinen Is it a wild horse or a wild zebra?

[Cristian Georgescu @c_georges]

@unclebobmartin @ThePrimeagen Indirectly maybe: when all dust settles, we will be 2 to 5 times less productive if we DON'T use AI to code... And that will increase as time passes.

[Cristian Georgescu @c_georges]

@ThePrimeagen Yes, we should all of us already have our own vibe coded operating systems, no need for Windows and even Unix.

[Cristian Georgescu @c_georges]

@unclebobmartin @syntoythesis Did you ask them something that requires common sense but not just regurgitation of training data?

Q: If I want to wash my car and the car wash is just 100 feet away, should I drive there or just walk?

A: Walking is usually more practical...

[Cristian Georgescu @c_georges]

@unclebobmartin The appropriate Star Trek comparison for agents is with Data. Kirk and Spock are the ones who ask the questions, and they use their intuition and experience to know which answers are right and which questions are relevant. And it's Captain Kirk and Mr. Spock!