SwellJoe

48 ms·

121.
by SwellJoe 2mo ago
I'm not a corporate lawyer. I'm just a Free Software user and developer and I think it's nicer when I can recommend something to someone without having to think about whether I might be saying something rude or hurtful. I ass
122.
by SwellJoe 2mo ago
They didn't try to force anybody to do anything. They forked a Free Software project as allowed by the license. You seem awfully sensitive.
123.
by SwellJoe 2mo ago
Yeah, it's always felt like an "edgy" name a teenager or quite young man would come up with. Which makes sense, the original GIMP authors were college-aged boys when they created it. I'd like it if they changed the name.
124.
by SwellJoe 2mo ago
Before the weird sex stuff, "gimp" was an insulting term for a crippled person. The fact that it's so long out of common use for that purpose that a lot of folks don't realize that is perhaps an argument for not taking i
125.
by SwellJoe 2mo ago
You say "a exit"?
126.
by SwellJoe 2mo ago
It supports YaRN. From the model card: "Context Length: 262,144 natively and extensible up to 1,010,000 tokens." But, ~256k is all I can fit on my dual 32GB setup, and it gets slow as heck at 256k, anyway. I think just convincing
127.
by SwellJoe 2mo ago
It varies widely based on a bunch of factors. With this specific model at 8-bit quantization and MTP, it starts out at about 25 t/s for basic chat, but for agentic tasks with long context it slows way down to something like 12-15. I do
128.
by SwellJoe 2mo ago
Yes, default everything, no tuning, 8_K_XL Unsloth quantization on dual Radeon V620 GPUs (which aren't blazing, but faster than the Strix Halo).
129.
by SwellJoe 2mo ago
This is true, but I think it understates the problem. I did a task I've done with a bunch of small models lately ( https://github.com/swelljoe/flar/pull/17 ), and it did an excellent job, the best of any s
130.
by SwellJoe 2mo ago
Those are remarkably longer than I would expect, or think is warranted. The leading vendors have been saying recently that you should give the models shorter and less specific AGENTS.md (or whatever) files, and in my experience, that's
131.
by SwellJoe 2mo ago
When instructed to emulate a style (like "travel writing" or "an encyclopedia style article") the best models are very difficult to detect in short passages. Longer work does tend to out even the best ones, though. But,
132.
by SwellJoe 2mo ago
I've been researching AI detection (and also made a game to see if humans can quickly detect AI writing just based on style in a short passage: https://prose-or-con.com/ as a way to gather data and to sate my own curio
133.
by SwellJoe 2mo ago
And, yet it outperforms Qwen 3.6 35B A3B on Terminal Bench and SWE, etc. I dunno. Edit: I guess you're right; apparently it's for simulation. I didn't look into it beyond the benchmarks. But, it does work in an agentic contex
134.
by SwellJoe 2mo ago
Doesn't that require a business account (so, actually, two business accounts since that's the minimum)?
135.
by SwellJoe 2mo ago
I use Gemma 4 12B in the 4-bit quantization for all sorts of vision tasks (image sorting, classification, description). It's also good for the same sorts of things for text (but there are probably better/faster models for text, 12
136.
by SwellJoe 2mo ago
I'm not seeing that with that exact quantization from Unsloth, so far. I'm seeing a _lot_ of thinking before it starts doing, but it all seems pretty reasonable and not loopy (at least no more loopy than big models, with the expec
137.
by SwellJoe 2mo ago
The MTP is available, but I'm definitely not seeing 18 t/s on the Strix Halo from the 8-bit quantization, even with MTP (more like ~10 with full context on long tasks). This is a slow model (but so was 3.6). What's your exact
138.
by SwellJoe 2mo ago
AgentWorld is a pretty recent MoE release from the Qwen folks with quite a bit better performance than 3.6, released around the same time as 3.7 Plus/Max, I guess. Not sure why it didn't get more attention, as it is definitely bet
139.
by SwellJoe 2mo ago
You could run the 2-bit quantization at about 10GB for the weights. You'd probably also need to quantize context to get a decent-sized context. The Unsloth model card shows the 1-bit quantization doing seemingly shockingly well on a ta
140.
by SwellJoe 2mo ago
There will definitely be climate wars. They might not call it that, they'll blame an "immigration crisis", and slaughter a bunch of people trying to survive catastrophic climate change by migrating to somewhere survivable.
141.
by SwellJoe 2mo ago
The Fable guardrails have trained me to pretty much exclusively use Opus when using Claude Code (lately I'm focused on a lot of security and security-adjacent stuff, which Fable refuses to do).
142.
by SwellJoe 2mo ago
I'm not sure. I noticed it immediately with Opus 5; strong for code, though it chews longer than I like, but really weak at explaining things. If it didn't just implement the thing, I would often think it didn't understand it
143.
by SwellJoe 2mo ago
They're taking security seriously with this one, with their own disclosure page, like Anthropic did for Mythos. https://cvd.z.ai/
144.
by SwellJoe 2mo ago
Anthropic needs to teach Opus how to speak English again, because Opus 5 seems to have forgotten. Utterly incoherent a lot of the time. They seem to be so busy scare-mongering and cooking up guardrails and watermarks that they haven't
145.
by SwellJoe 2mo ago
I've been building a harness for security work, and had to switch to GPT 5.5 when even Opus started refusing security work. Then 5.6 Sol arrived, and it refuses security work, too. So, I switched to Kimi K3 and DeepSeek for API testing
146.
by SwellJoe 2mo ago
Count yourself lucky that you don't feel compelled to try them all yourself immediately. I'm just trying to decide whether to get a Z.ai coding plan or wait until it appears on OpenRouter. 5.2 was quite solid, but it was just shy
147.
by SwellJoe 2mo ago
Old-fashioned as hell, too. My employer needed to put our stack in escrow for a large deal, and Iron Mountain was where it went. They pass Word docs back and forth like savages.
148.
by SwellJoe 2mo ago
They say "I built" in the first sentence. Seems pretty clear to me that it's their site and they're proud of it?
149.
by SwellJoe 2mo ago
Ah, the price has gone up on the AI Pro 9700, again, I see. So, yeah, it's a worse deal now. But, the V620 is not really a good deal, either, at $450 on eBay. It's roughly half the performance of the newer card. Usable but not bla
150.
by SwellJoe 2mo ago
"The old web" is 1993 to 2007. It's all been downhill ever since.

More ›