Gamers Nexus

  • 🔧 Site instability resolved. You can report double-posts and broken attachments. For bigger issues, use the Technical Grievances thread.
    🇵🇦 Nuestro primer dominio localizado está en español en kiwifarms.pa. Our first localized domain is on Spanish on kiwifarms.pa.
  • Want to keep track of this thread?
    Accounts can bookmark posts, watch threads for updates, and jump back to where you stopped reading.
    Create account
Destacado
Sam Losco dijo:
Jensen sounds like a complete buffoon.
"It's ok to clap." No, Jensen, it isn't.
Presiona para expandir...
If there's one good thing about Steve's constant blackpilling of NVidia is that it shows just how fucking psychotic Jensen is.

This man should be nowhere near technology because he is stone cold fuck nuts. And the worst part is that his fucking ego has encompassed the entire globe.
 
Celebrate Nite dijo:
If there's one good thing about Steve's constant blackpilling of NVidia is that it shows just how fucking psychotic Jensen is.

This man should be nowhere near technology because he is stone cold fuck nuts. And the worst part is that his fucking ego has encompassed the entire globe.
Presiona para expandir...
That and that he controls the entire goddamn industry. Lip Bu Tan and Lisa Su are in cahoots with this egotistic fag, which is why Intel and AMD are never competition, but rather, collaborators.
 
WULULULULU dijo:
That and that he controls the entire goddamn industry. Lip Bu Tan and Lisa Su are in cahoots with this egotistic fag, which is why Intel and AMD are never competition, but rather, collaborators.
Presiona para expandir...
and now Nvidia has Intel shares too.
 
generic sauce dijo:
Middle aged rocker man realizes that corporations aren't friends.
Presiona para expandir...
He also realized that corporations give money to PACs, not just to those affiliated to the DNC.
One day he will find out that Nancy Pelosi was quite good at playing with the market via her Nvidia shares during Covid.
 
Dead Lion dijo:
Does anyone actually use AI agents? Chatbots are handy but agents sound retarded.
Presiona para expandir...

I tried to setup Volt last week, but their entire web-ui is hidden behind a login wall (Github/Google) even though it only connects to your local instance. No option to run the single-page app locally so they can just gather your personal info. Fuck them.

I tried Mastra and it won't work with my Claude OAuth token. I don't want to pay these companies, and I get the $100 Pro plan free from work. You can't use those for API. That requires a separate Anthrophic Console account that's pay-per-use. You can proxy requests using some tools, but it could lead to an account ban, and I can't risk that with my work account. I might try that with my work Copilot account since I don't care about that one at all.

I didn't realize this split between API and web existed this way until just last week. I knew the API access was more expensive and looked at it briefly months ago, and now I know why. I think a lot of the locking down probably happened when the vibe coded piece of shit OpenClaw got big.

So no, still haven't tried using an "agent," and kinda don't care at this point. I don't know what I'd get since even with Mastra/Volt I ran them in isolated containers. I don't even let Claude prompts run commands a lot of the time. I would never give them access to keys for a Grafana dashboard much less SSH or anything else.

Celebrate Nite dijo:
https://youtube.com/watch?v=IhWkBmFWU64
Presiona para expandir...

He talks about how people shouldn't pay for these models or these companies, running them locally if you can. Honestly, 90% of ChatGPT users on their $20 plan could probably do everything they want with ollama and a free model off hugging face. But that requires a minimal amount of learning and figuring things out, and people who use this shit would never be able to do that.

There are applications in computer vision, mostly for domestic spying companies like Flock. But I suspect they run their own, simpler models.

The big push is the software engineering space. I have been a lot more skeptical of these tools than some of my peers. I had an old manager who loved Claude and "spec driven development" and when he left I was pretty much put in charge of one of the biggest piles of AI slop you can imagine. I did my best to review his 10k line PRs, but often he'd merge them in as I was leaving comments.

There's another manager who straight up told me he has stuff in production he's made with Winsurf, and he doesn't fully understand how it works.

I use it like a scalpel for small discrete pieces. I test them to make sure they work. If it's my own work, I refactor over half of it. If it's for work .. I care a bit less but still try to make sure it's not shit. I have other friends in the industry who told me they haven't written actual code in months!

The big problem is that one of the very few places LLMs actually make sense is coding, and you can't get away from using these tools at this point. I've heard qwen3.6:35b-a3b has gotten really good and on par with some of the OpenAI models, but it won't run on my 32GB AMD 9700R. Not enough VRAM! That's literally the maximum you can get on any consumer level discrete card unless you jump up to an RTX 6000, a Mac Studio or one of the Spark devices .. and then you're getting into the $8k ~ $10k range. Same goes for the larger Deepseek models.

So even if you want to try one of these open weight models, you're very liklely still going to have to use them by purchasing them from some SaaS provider.

These are the dangers GN has been pointing out, but also since he's not on the development enthusist side like Wendel, he doesn't have a good grasp of the scope of the problem. (Wendel is a little to pro-AI for my taste).

The Fabel model is honestly ... surprsingly good. There are still a lot of issues I've found, but I was pretty amazed at its ability to actually structure code like your existing project. But I also realize Anthorpic is a drug dealer pushing heroin, and these costs are all going to jump. This chart has beeng going around lately:

1781366366098.png

These companies depend on a ton of customers pay $20 and using $1 of compute to subsudize people who pay $100 and use $800 of compute. That's not substainable, and it won't be long before Anthropic does some Copilot/OpenAI style rug pulls in pricing.

But people will pay it, because none of the local models get close.
 
HW News had a few tidbits in them

NVidia (as well as AMD and Intel) were dicks at computech (to the surprise of nobody)


Speaking of Intel, they were only relevant enough to get a small blurb instead of an actual video like NVidia/AMD did
 
Dead Lion dijo:
Does anyone actually use AI agents? Chatbots are handy but agents sound retarded.
Presiona para expandir...
Took some workshop class on it at work. It does seem somewhat useful in a fabricated and retarded context. One person has the rights to set up an agent; feed it thousands of instructions and examples and what not, then make it readily available for other employees to use. You know, the perfect blue-eyed aryan agent for pajeets to throw a mail into and it'll come out like proper dudebrah prose.

Unless of course "GPT agents" already changed meaning.
 
⋖ cørdion ⋗ dijo:
Took some workshop class on it at work. It does seem somewhat useful in a fabricated and retarded context. One person has the rights to set up an agent; feed it thousands of instructions and examples and what not, then make it readily available for other employees to use. You know, the perfect blue-eyed aryan agent for pajeets to throw a mail into and it'll come out like proper dudebrah prose.

Unless of course "GPT agents" already changed meaning.
Presiona para expandir...
To my understanding it's basically BonziBuddy but with an LLM serving as a back end. From responding to emails to vibecoding shellscripts on the fly.

Most people just use them to autorespond to emails though.
 
Destacado
Debes iniciar sesión o registrarte para responder aquí.
Atrás
Top Abajo