behohippy

It's pretty simple nowadays if you know conceptually how they work. Running the LLM calls in a loop with tools is an agent. You only need 10 or so basic tools to accomplish nearly anything, and you can build a dynamic skill system from that. Look at https://github.com/patw/pengy, ignore the app look at the spec.md file, feed that to your current agent of choice and make your own version. Use whatever tech stack or UI you're comfortable with. Change some of the choices in how it works, so it fits what you want to work.

You might have a business idea there. I wouldn't mind a twinscan plushie for sitting on top of the workstation.

Yeah 48g, sub 200W seems like a sweet spot for a single card setup. Then you can stack as deep as you want to get the size of model you want for whatever you want to pay for the power bill.

Sure, all the slop code projects I produce get MIT licensed on public repos. It wasn't mine to begin with, so I wouldn't prevent anyone from using it.

Used 3090s have been getting expensive in some markets. Another option is dual 5060ti 16 gig. Mine are lower powered, single 8 pin power, so they max out around 180W. With that I'm getting 80t/s on the new qwen 3 30b a3b models, and around 21t/s on Gemma 27b with vision. Cheap and cheerful setup if you can find the cards at MSRP.

About 768 gigs of ddr5 RAM in a dual socket server board with 12 channel memory and an extra 16 gig or better GPU for prompt processing. It's a few grand just to run this thing at 8-10 tokens/s

These articles are gold, thank you. I used your gemma one from a few weeks back to get gemma 3 performing properly. I know you guys are all GPU but do you do any testing on CPU/GPU mixes? I'd like to see the pp and t/s on pure 12 channel epyc and the same with using a 24 gig gpu to accelerate the pp.

I run the KV cache at Q8 even on that model. Is it not working well for you?

Qwen is a little fussy about the sampler settings, but it does run well quantized. If you were getting infinite repetition loops, try dropping the top_p a bit. I think qwen likes lower temps too

You probably won't be running fp16 anything locally. We typically run Q5 or Q6 quants to maximize the size of the model and context length we can run with the VRAM we have available. The quality loss is negligable at Q6.

Just this pic: https://imgur.com/ip8GWIh

I don't have a video but here's a pic of the output: https://imgur.com/ip8GWIh

It's a 3b model so the creativity is pretty limited. What helped for me was prompting for specific stories in specific styles. I have a python script that randomizes the prompt and the writing style, including asking for specific author styles.

I have a mini PC with an n100 CPU connected to a small 7" monitor sitting on my desk, under the regular PC. I have llama 3b (q4) generating endless stories in different genres and styles. It's fun to glance over at it and read whatever it's in the middle of making. I gave llama.cpp one CPU core and it generates slow enough to just read at a normal pace, and the CPU fans don't go nuts. Totally not productive or really useful but I like it.

I had this same issue with incomplete answers on longer summarization tasks. If you ask it to "go on" it will produce a better completion, but I haven't seen this behaviour in any other model.

It's probably an evolution of the phi-1/1.5 "Textbooks are all you Need" training method: https://arxiv.org/abs/2309.05463

No joke, that would be an awesome LLM project name!

Top_p and top_k are pretty important concepts for LLMs same as temperature so P,K,C and F are underutilized

Hey emad, thanks for SD and this! What's the plan if Meta does Apache 2.0 for LLaMA? Just keep going and making the 30b and 65b or build different models?

Vicuna-13b (4bit) got the answer right, the first time as well.

I've noticed the same with my Asus TUF laptops. I've had 2 generations of the 15" models with Ryzen processors and adding a second stick seemed to "wake" them up in a noticeable way. They Ryzen memory controller really seems to benefit from memory running in DDR mode (2 sticks).

My dad builds houses in northern Ontario and heat pumps seem to be getting more popular on new builds. This is a place that regularly gets below -30C in the winter. The heat pump (usually a heat pump/AC combo unit) by itself doesn't work in these conditions, you'll always pair it up with another system like a natural gas furnace, or resistive heat electric.

My personal heat pump is an older unit, it works down to -10C then then forced air electric (resistive) furnace kicks in. That thing is pricey to run, so I also light up the wood stove at that temp to reduce the costs.

Not a good fit for lifestyle. I'm rural, and I haul around ATVs and dirt bikes with the main destination being 300km away and even more rural. My pickup truck is a way better choice for that.

I used to tow a utility trailer with a small gas car. It worked fine, but this is just so much easier and more luxury. I don't believe any EV could handle this use case.

I love this idea. I've been desperate enough to reuse gaskets in a few repairs, and suffer oil leaks down the line. Usually it was because of back ordered gaskets that would be weeks or months out. Riding season in Canada is short, so... Take the risk and reuse. If you can ship in a week, I'm a customer.

Asus TUF series (A15). Inexpensive with really good hardware specs and upgradable ram, storage and wifi cards. I usually go AMD on them, and midrange video cards. They're shockingly good as desktop replacements but they suffer a bit on battery life. My 4900H/RTX2060 feels snappier than my work M1 Pro in everything I do and it can play pretty much every modern game smoothly.

I believe this is mostly due to declining sales. My local dealer had good stock levels during the car shortages and the sales people were complaining about lots of Chargers, but no Grand Caravans. They could sell every mini van they produced, if they could get more.

3 of the last 4 roles have been recruiters from LinkedIn (Solution Architecture), including the current one. The one exception was reaching out to a customer to let them know I was leaving and they counter offered to get me to join them. LinkedIn is probably the most valuable one to me, by far.

Give Rimworld a try. It's easy to get a few things done every break and pausing fits in naturally. You can even almost get away with backgrounding it where you set up a bunch of stuff to be done, and let the pawns do their own thing.

If I'm reading it right, it's more the acceptance of loss of what you had and more appreciation for what age brings you now. Also the start of removing things from your life, rather than constantly adding new things. To bring it to a HN context, maybe learning that new language won't be as satisfying as just getting deeper on the ones you already know? I'm just guessing at this point.

3 desktop class machines, one acting as a server. 3 laptops, one is a chromebook. 8 RPi's doing PiHole, OctoPrint and part of my temp/humidity sensor network. It gets fuzzy with the Pi's because I tend to load them up with multiple duties depending on location. With spare hardware I could probably do a few TV PCs.