[–] 3 points 1 week ago

Yea they make the adapters for like $100 on eBay for smx2/3 cards. They're reference designs and loud as fuck. So you have to retro fit proper fans to them. So it's doable.

The L40s are expensive as fuck right now being a recent gen and all but shit if the bubble pops, then might see them getting sold for few hundred a piece. That'd be cool.

  • source
  • parent
  • context
  • [–] 4 points 1 week ago (1 child)

    You can't repurpose a DC/clippy 2.0 GPU for gaming

    If it's AMD hardware, you're SOL - they removed ROPs entirely ~4 gens ago.

    If it's nVidia, it depends on what it is- firstly you need to get PCIe versions not SXM (unless China makes an SXM to PCIe adapter) then If it's Hopper (2 gens ago) or Blackwell (Latest), it will perform badly for the specs, but it's possible, if it's Lovelace (last gen) or Ampere (3 gens ago), you could make it happen.

    Hopper and Blackwell both skip RT cores and only have 24 ROPs no matter what model. Lovelace and Ampere are full-fat. An L40 should actually be a better GPU than an RTX4090 for the most part.

  • source
  • parent
  • context
  • [–] -2 points 1 week ago (8 children)

    How long do they last though?

    Do they increase or decrease insurance costs? (hard to quantify, I think it'd likely overall raise insurance right now, but it will come down as trust increases)

    How many robots does it take to replace one person (or vice versa)?

    What are the employees being replaced going to do?

  • source
  • parent
  • context
  • [–] 1 point 1 week ago* (1 child)

    Yeah, webdev is one of those areas where it's really not that deep and LLMs have pretty much figured that out by this point so the distance between local and frontier is basically non-existant. I mean ask Fable 5 or a model on my B70 to make a website and it's pretty much going to look like the identical vibecoded slop site, that benchmark is old news. But to have it THIS close is honestly pretty funny, even in this benchmark I didn't expect the difference to be a rounding error.

  • source
  • parent
  • context
  • [–] 1 point 1 week ago* (last edited 1 week ago) (3 children)

    Which benchmarks? Are you sure you didn't misread Qwen3.8-Max for Qwen3.8-27B? That one's up there, but if a benchmark ever compares 27B to Fable then that benchmark is probably just trying to sell you their vibecoded harness or something lol

    Edit: Don't get me wrong, I've been especially impressed with Qwen3.8-Flash-Next on my PC and to say it's barely better than Opus 4.6 is a huge achievement in my eyes. Opus 4.6 was the point where people finally felt like that AI actually spits out mostly useful stuff and having models that can be ran on a realistic PC finally catching up is fantastic, but it's still far from the frontier stuff. No way in hell am I going to get anywhere with this for decompiling/reverse-engineering stuff unfortunately. I really like coding the important stuff by hand and having it spit out bash scripts in parallel or reviewing my code in the end for oversights, for that I really can't complain anymore and like Opus 4.6 any model better than that is just luxury. Would I let it vibecode the whole thing then it would end up being unmaintainable slop, but as an actual tool, not a crutch, it's fantastic. But yeah, Fable 5, not in a dream lol

  • source
  • parent
  • context
  • [–] 1 point 1 week ago (5 children)

    Nah, that's exaggerated. I have a 32GB Intel Arc Pro B70 and I would say Qwen3.8-27B is barely behind; Qwen3.8-Flash-Next is barely ahead of Opus 4.6, neither of them come remotely close to Fable 5. Some of the Chinese multi-trillion parameter models do compete with the frontier models tho, but nothing you could even remotely hope to run on local hardware.

  • source
  • parent
  • context
  • [–] 2 points 1 week ago (8 children)

    Naa they'd be doing this already. There are tons of AI slop GPUs on eBay for super cheap. The only thing they have done with them is retrofitting them for pcie and cooling so you can put them into your PC to run local LLMs.

  • source
  • parent
  • context
  • [–] 5 points 1 week ago* (9 children)

    All current datacentre hardware is being run at power and density levels that hover somewhere between "operational" and "pile of glowing silicon slag".

    I'm sure some enterprising Chinese companies will take all the surplus hardware and create underclocked "consumer-capable" versions for inference. I'd happily pick up a couple of cards that are 75% of cutting edge performance if they're a hundred bucks a pop.

  • source
  • parent
  • context
  • [–] [S] 7 points 1 week ago

    The models won't disapear. In the end, I think saving $5T-$10T in repercussions from not giving them $9b, ends with giving them $9b. The quiet limbo part is telling. Everyone who pumped AI bubble is already too overextended. So much other investment is also being requested.

  • source
  • parent
  • context
  • [–] 22 points 1 week ago (10 children)

    You're not the first to claim this, but it sounds pretty convincing… I'll have to see it with my own eyes before I believe it though. Not sure how excited I will be due to the economic repercussions such an event would have, and I might miss ChatGPT, but I am damn well aware that the longer this keeps on going the worse the repercussions will become.

    Either way, if hardware prices are coming down I will be buying.

  • source
  • [–] [S] 12 points 1 week ago*

    each datacenter gw, before nvidia price hikes and construction inflation, is $60b. They have promised to rent 33gw of future datacenters. A large chunk of dumbass bubble fuckers have been building these with upfront cash/debt thinking this is possible.

    The other problem for OpenAI is that they lose big money just running their existing models. They raised over $21b in March of this year, with even more in "compute credits" and now its gone. They have a 20% tax on revenue to pay MSFT, and in total outside of compute expenses, $2B/month cash burn.

  • source
  • parent
  • context
  • view more: ‹ prev next ›