Arrow in the knee

[Dr_Gingerballs @Dr_Gingerballs]

To be fair the primary event that occurred that few saw coming was Masayoshi Son going full blown AI delusion and betting the farm on them. Without that injection they probably wouldn’t have survived this long. Not sure about y’all, but going long delusion makes me uncomfortable.

[Arrow in the knee @arr_kni]

@nicbstme > it shows how early we are

It shows that they’re useless for general public. To use agents to book your flights and hotels you have to be actually insane given how unreliable they are. Many use cases are just invented by AI bros and have no sense IRL

https://t.co/x9qdmV1dxQ

[NEXTA @nexta_tv]

A Russian man was left without a trip to North Macedonia after trusting ChatGPT

The family relied on the AI and didn’t double-check the country’s entry rules.

The man, along with his wife and child, bought tickets from Istanbul to Skopje and booked hotels and a car rental.

Before the trip, he asked ChatGPT about visa requirements and was told that no visa was needed. In reality, a visa was required.

At the airport, they were simply denied boarding. Showing a screenshot of the AI answer didn’t help.

[Arrow in the knee @arr_kni]

@bagofdoom @Krubslaw @JeffLadish @edzitron I skip checking libs update

Human hallucinations are solved by firing humans, so they have an incentive and they take responsibility. Also there are PR reviews on much less code

> do you use agents to code review yet

Bro, of course I do it. But valid for an agent != valid

[Arrow in the knee @arr_kni]

@bagofdoom @Krubslaw @JeffLadish @edzitron I’m talking about file access for destructive operations

You can sandbox it to work with a repo copy. Another hassle, many will skip it, just like everyone uses auto mode now

But some tasks are destructive by nature, like email cleanup, organizing photos, you can’t sandbox it

[Arrow in the knee @arr_kni]

@8hipulin > крутит опенкло

Походу, кроме него больше никто и не крутит, нет такого ощущения?

[Arrow in the knee @arr_kni]

@bagofdoom @Krubslaw @JeffLadish @edzitron Also yes, humans hallucinate, but:
- you won’t give another human access to your files
- you didn’t have an apprentice to oversee, now you have and it’s forced on you
- this appeentice can’t be held accountable, you are
- but you’re expected to go 10x and take all responsibility

[Arrow in the knee @arr_kni]

@unixpickle @mishakgg I use Claude code, and I know that remembering and adapting happens in memory files, the weights of the model are not changed. Your local experience doesn’t translate across the world

[Arrow in the knee @arr_kni]

@entropia_acc @nic_carter Ok, that’s fair, although I’m sceptical about ads (some practices may be deemed barely legal given how much oAI knows about the user,their weak points; they can shift the conversation in a direction needed for ads so I bet some regulation will follow), so they can be the next AOL

[Arrow in the knee @arr_kni]

@bagofdoom @Krubslaw @JeffLadish @edzitron They are in fact hallucination machines (I use them a lot and I can tell, although I find them useful) and it’s oAI that says that the hallucinations can’t be fixed

Most benefits come from harnesses, eg Claude Code is only 2% AI. Deterministic code calling a model

[Arrow in the knee @arr_kni]

@Alien_AV @unixpickle OpenAI has been working on some stuff since 2015, yet in 2018 they used a Google’s technology developed in 2017

What is exactly your point? “Oh he didn’t do anything new in a <year since he departed from Meta”. Ok?

[Arrow in the knee @arr_kni]

@lemire Because you have no means to verify it (no, tests won’t cut it), good luck debugging an assembly for a microservice

It’s essentially a vendor lock, and that vendor has no intent on making a quality job. It may even introduce bugs so you pay it to solve them

“So what”

[Sauers @Sauers_]

UPDATE

[Arrow in the knee @arr_kni]

@mishakgg @unixpickle That’s right, AI researchers don’t give such estimates about LLMs

[Arrow in the knee @arr_kni]

@entropia_acc @nic_carter So I have to look up the information to disprove your claims based on nothing? Nah I’ll pass

https://t.co/Tl5rKVgEQm

[Rebecca Valentine @defnotbeka]

if you ever wonder why modern computers don't feel faster and snappier than past computers, it's because there are a lot of brain dead people making brain dead software

[Arrow in the knee @arr_kni]

@entropia_acc @nic_carter How many of these 1B users pay for inference (which is not free)? How many of those paying don’t cost 25-500x for oAI due to subsidies?

[Arrow in the knee @arr_kni]

@RodSquadWells @nic_carter Because the circular investment schemes and no ROI for users make people think something is wrong here

[Arrow in the knee @arr_kni]

@nic_carter @RobertHarrisonv Do you know how companies raise capital? They promise returns. There’s even a measurement for it, it’s called ROI

How will an unprofitable company return the investment? They just slashed costs for 5.6, burning even more VC money

[doomslide @doomslide]

Oh you've automated math have you? What are your foundations?... Oh and this is your machine checker? Is it decidable?... Curious, how about transitive?... Hmmm, did you formalize its metatheory?... Oh dear.