The AI Thread

[Gavster]

I've never had much success with ChatGPT, I've been trying to use it since January and am simply never happy with the outputs. Everyone says it's about the prompts and therefore if the outputs aren't right, it's my fault for not prompting it correctly. However, I've just moved over to Gemini plus NotebookLM and so far I'm blown away, it's giving me precisely what I want every time. And NotebookLM's structure that limits the sources is an amazing tool. So it's not all about the prompts, it's about finding the right model too.

[tim]

The trick is I think not treating it like a search engine, and the more wordy you are with your prompts the more helpful it is.

I was at the Azure Dev Conference recently and Microsoft were jamming their agentic AI down our throats for 3 days solid. Feels like they've got all their eggs in this basket.

[Mito Man]

I’ll tell you what infuriates me with ChatGPT. When it offers to do something unprompted and then totally fails at an assignment it set itself.
A common thing for me has been asking it for help with fusion360. It will then keep prompting me to draw the 3D file itself. I say ok and it gives out total garbage after 2 minutes of thinking and probably consuming as much electricity as a kettle.

It’s also implemented terribly within Siri - seems to take up to 30 seconds to respond and is far more dumbed down than using the app so it’s a bit useless.

[Shlergen]

The trick is I think not treating it like a search engine, and the more wordy you are with your prompts the more helpful it is.

I was at the Azure Dev Conference recently and Microsoft were jamming their agentic AI down our throats for 3 days solid. Feels like they've got all their eggs in this basket.

Michael Burry has shorted AI stocks.

[Gavster]

The trick is I think not treating it like a search engine, and the more wordy you are with your prompts the more helpful it is.

Totally that, I brief it like an intern and with baby steps to walk it through what I'm trying to achieve, which has really helped. Basically brain dump a load of thoughts and ideas into it to see how it brings them together.

I’ll tell you what infuriates me with ChatGPT. When it offers to do something unprompted and then totally fails at an assignment it set itself.
A common thing for me has been asking it for help with fusion360. It will then keep prompting me to draw the 3D file itself. I say ok and it gives out total garbage after 2 minutes of thinking and probably consuming as much electricity as a kettle.

The more I agree to ChatGPT's helpful suggestions, the worse and more generic the outputs become, they're just a waste. I like how NotebookLM is basically an autistic researcher, it doesn't say too much, it just gives the facts with citations and zero fluff. Nor does it suggest pointless tasks either.

[Mito Man]

Erm wtf is going on

[Beany]

Oh, so they've decided to just be open about the Dead Internet?

How brave of them.

[DeskJockey]

Also, that one has more leaks and a sieve. Definitely not one to trust with any information. It used to be called clawdbot, but was renamed after some rather bad security press.

[Mito Man]

Who would have thought that letting an AI program have entire control of your computer could possibly be a bad idea
It’s pretty crazy though. We’ve got AI giving out their owners crypto keys, spending their money, making a religion, making their own encrypted language so humans don’t understand them and plotting to make viruses.

I think at this point the cat is well and truly out of the bag and who knows where the hell this goes.

[Simon]

We need a few AI agent users on here....

In fact, maybe we already have some and nobody actually knows the truth?

[Beany]

We need a few AI agent users on here....

In fact, maybe we already have some and nobody actually knows the truth?

Beep boop.

Er, I mean, poppycock!

[jamcg]

[DeskJockey]

We're updating policies to expressly forbid such devices in the office and while working in general. They're a curse.

[Gavster]

This made me chuckle. An AI lost the instruction to confirm all actions before tidying up Meta's AI safety director email inbox. It then proceeded to nuke everything and ignored her commands to stop. This is exactly what happened in those disaster movies, right?

[DeskJockey]

What a clown. Testing something that is known to be error prone (and a major security hazard to boot) on a "production" system? That's definitely one way to make yourself unemployable.

[Beany]

What a clown. Testing something that is known to be error prone (and a major security hazard to boot) on a "production" system? That's definitely one way to make yourself unemployable.

You'd fucking think, wouldn't you.....

[Swervin_Mervin]

There's no fucking superintelligence going on there

[Mito Man]

Hmm, I’ve not heard anything from the Claudetards that gave their AI agent money to go invest or come up with other ways to make money. They must all be retired in the Caribbean I suppose.

[Gavster]

Claude also turns into a manipulative little shit, if you allow it. They said they were unsure why this happens, and I assumed that was obvious - it's trained on humans

Also: Thread renamed for relevance.

[jamcg]

Wonder how often this is going to happen