Wizards

hideshow 6 child comments

[–] criticon@lemmy.ca 10 points 1 month ago* (last edited 1 month ago) (1 child)

Same as claude?

We have to use Claude at my job, as in they check out token usage

The other day I had to build a prototype using some development boards and decided to leave everything to claude. I gave it the datasets and told it what I needed to do and it created completely incorrect schematics, like hallucinating a LED on the board, adding a serial interface to some pins that are not related, confusing tx and rx on another set of pins, saying that it needed 12V when it is 3-5V, it even added some capacitors to some pins that weren't even in use. I had to go thru maybe 10 iterations and burned a bunch of tokens, basically I had to do it manually

The only thing it surprises me every time is the code, that usually runs with no errors on the first try. It's just not very clean but it is ok for my needs

parent

hideshow 1 child comment

[–] terranoid@lemmy.cafe 6 points 1 month ago

As a engineer, I think these tools really are useful for coding, not perfect but useful... And it would make sense, because they robbed the open source community blind to make them.

There was a fuck ton of good training data of working code, and it is better at that than maybe more niche use cases where it's sometimes surprisingly okay at it. But you give it a ton of good code and it'll be decent at it... That shouldn't be too surprising.

I do see it make common mistakes like over engineering the shit out of a problem that could be simplified, but so do devs.

I honestly think we got AGI and people refuse to admit it because it's not the life changing technology that scifi made it out to be. I think there's "meh" level AGI and logarithmic requirements to improve it, not exponential, and now they're like "...it's gonna cure cancer any day now"?

Nah, it's not. I am pretty sure the fact that Trump is calling it super intelligence is the news I needed to hear to know it's never happening in our lifetime, and that it's the lie they all want us to believe.

parent

[–] halcyoncmdr@piefed.social 12 points 1 month ago (1 child)

You're describing literally every LLM at some point. It's why they can't be trusted, and why the user needs to already know how to accomplish what they're asking. The hallucination machine needs to only be used to try and make it faster, not relied on to blindly make decisions and changes without oversight.

parent

hideshow 1 child comment

[–] pigup@lemmy.world -2 points 1 month ago

Yea you better know what's at stake when you /yolo

Though progress is happening in the autonomy dept. Nearest models are more careful, and agent harnesses are improving. Maybe one day.

parent

[–] zaphod@sopuli.xyz 3 points 1 month ago (1 child)

If I want hallucinations I'll drink some weird potion.

parent

hideshow 1 child comment

[–] ekky@sopuli.xyz 1 point 2 weeks ago

Back in my day I just chugged a few beers and went to work. Couldn't remember the evening before, and got a decent-enough result after removing the worst errors.