Ted McFadden

I'm not a spambot! Really!

[Ted McFadden @mctroyd]

@blackroomsec We’ve tossed people into the clink for less than what they’re doing. https://t.co/siAmTCHGr0

[Ted McFadden @mctroyd]

@PixelsByNikki Good luck! Hope this brings in the cash.

[Ted McFadden @mctroyd]

@blackroomsec Maybe it’s not you… maybe it’s everybody else? 😁 I mean, poor cyber hygiene is a pretty big deal. No friend would let their friend succumb to it, I think. 🤷‍♂️

[Ted McFadden @mctroyd]

@heyclairelin If the bot is doing something novel, and not trying to be the universe’s virtual gf, I might.

[Ted McFadden @mctroyd]

@blackroomsec Nah, let the bad management have their layoffs. When (not if) this turns out to be a terrible, shortsighted decision, they’ll have it that much worse coming back. Maybe they don’t. 🤷‍♂️

Meantime, having smart people with nothing to do is how we end up with stellar small biz.

[Ted McFadden @mctroyd]

@net_termina The only model in that class I’ve had any use for is lfm2.5:1.2B-instruct. It runs the various administrative functions of my Open WebUI install (automatic chat titles, tags, etc). Works very well in that capacity. Super fast even on a Haswell-era Xeon, capped at 4 threads.

[Ted McFadden @mctroyd]

@MiaAI_lab I run relatively ancient hardware, mostly for financial reasons, partly morbid curiosity. (Core i7-2600k, 32 GiB maxed RAM, Quadro P2200 w/ a P5000 on the way.) Qwen3.6:35b MoE still gets ~200/sec PP and 10-20/sec TG. More important the model gets it right; I’ll never be fast. 😅

[Ted McFadden @mctroyd]

@hotdropfps AMD RX580. Honestly, though, at 1080p it’s still hitting 60 Hz. 🤷‍♂️

[Ted McFadden @mctroyd]

@superalesha There must be a catch though; what are they getting in return for this? Community R&D support?

[Ted McFadden @mctroyd]

@ShimazuSystems It doesn’t have to make sense for rich guys like him. It just has to sound smart and insightful. Nobody around him is going to question it. That’s the biggest vulnerability of the wealthy right now. They are sorely out-of-touch.

[Ted McFadden @mctroyd]

@net_termina One advantage of still running old hardware: scored a Quadro 5000 (Pascal, 16 GiB) for $200 off eBay. Ubuntu 24.04/Linux 6.8 supports the v535 drivers well. Currently running a P2200 (same gen, 5 GiB, $100 eBay), which will stay in service. Owning need not bankrupt.

[Ted McFadden @mctroyd]

@PeasantSmith Thanks for this! Pascal cards are still somewhat affordable, and I managed to snag a Quadro P5000 (16GiB) for ~$200 just today. I imagine some people have them laying around. Can’t waste silicon these days. 💸

[Ted McFadden @mctroyd]

@net_termina Kinda funny, really. The training on humans is really showing through here. 😂

[Ted McFadden @mctroyd]

@OpenBMB @huggingface Thank you for the reminder! Forgot to go get this. Looking forward to taking it for a spin!

[Ted McFadden @mctroyd]

@eevblog Wheee!! 🎢📉 AAAAHHHH!! 😂

[Ted McFadden @mctroyd]

@sudoingX 2x speed. Harness can manage context. Not many options to go faster.

[Ted McFadden @mctroyd]

@gmkurtzer A podcast producer referred to LLMs as a “pachinko machine” and I can’t unsee it now. 😂

[Ted McFadden @mctroyd]

@sudoingX For output? 10. Using it in chat-type mode, that’s fast enough. I’m not using other modes yet.

[Ted McFadden @mctroyd]

@sudoingX Hell, get me a GPU with more than 5 GB VRAM and I’m better off than I am now. $500 would easily cover that.

[Ted McFadden @mctroyd]

@shesaidmewakeup @net_termina I had issues with DeepSeek v1 responding in Chinese about half the time. (I only speak English. 😅) Fixed it by telling it to only respond to user prompts in English, via the system prompt. Maybe that’ll work here?