SwellJoe

16 ms·

331.
by SwellJoe 3mo ago
That sounds cool, but it's still pretty meaningless without information about what your home lab looks like. A few DGX Sparks wired up with their fancy super fast network is much different than a few laptops on wifi.
332.
by SwellJoe 3mo ago
I note the lack of performance information. I can only imagine it's much, much, slower than any other way to run a larger model (including, e.g. using system RAM and streaming some stuff from disk). Consumer networks, even 10gbit ether
333.
by SwellJoe 3mo ago
Is it? I don't see anything on the website about splitting a model across multiple devices, only about putting local models on the internet, a wholly orthogonal problem (which is already easy with existing tools, since models use an ht
334.
by SwellJoe 3mo ago
They've got to be getting tired of all the hype going to Anthropic because of the big "I rewrote X in Rust in two weeks with Claude stories." Anthropic has their PR game on lock.
335.
by SwellJoe 3mo ago
The git repo on github is a publish-only copy. It doesn't accept pull requests, though apparently there is a mechanism to turn pull requests into a patch for the mailing list they may accept.
336.
by SwellJoe 3mo ago
My first thought on hearing about this was, "what's that boat like? I wanna see the boat." Not to take away from Pfendler's incredible achievement. She's amazing. But, I'm the kind of nerd that immediately went
337.
by SwellJoe 3mo ago
I don't think LLMs get us there, at least based on my understanding of them. You don't get the reasoning at a very high level without a similarly large parameter size. Though, I don't think we're really at the maximum de
338.
by SwellJoe 3mo ago
I've been saying efficiency is the next "frontier" in AI, at least for LLMs that people use daily. Companies have started to really balk at token costs from the major providers, and there's some evidence that the cheaper
339.
by SwellJoe 4mo ago
Er, actually, Pro is a big 1T model (which makes more sense given how well it does, sometimes). Regular MiMo is small, and I haven't tested it.
340.
by SwellJoe 4mo ago
I assume they're jealous of the Fable/Mythos hype. People talk about Fable like it's a whole new thing, rather than another incremental improvement over the existing best models (which has happened several times and continues
341.
by SwellJoe 4mo ago
Yeah, I pretty much had to switch to using GPT rather than Opus completely for all my security benchmarking and harness development. I was annoyed enough to blog about it: https://swelljoe.com/post/why-i-had-to-switch-t
342.
by SwellJoe 4mo ago
Gemma 4 is the first really small model that feels smart, to me. I mean, Qwen 3.6 is arguably better at some coding tasks. But, Gemma 4 has shockingly good reasoning for a small model. Even the tiny 12B, at 7GB on disk in the 4-bit QAT quan
343.
by SwellJoe 4mo ago
Nearly every model can be found on OpenRouter and used with a single key. Meta Spark is not among them, but Grok and almost every other model is. That's how I try models I don't already have an account for.
344.
by SwellJoe 4mo ago
MiMo v2.5 Pro is very spiky, in my experience. Sometimes excellent, sometimes mediocre. Weirdly high non-deterministic behavior. Run the same task three times, get three different results. I mean, they're all rolling dice for the next
345.
by SwellJoe 4mo ago
The Hy3 preview has been a mediocre performer in my benchmarks of security auditing with models, and yes, it is outperformed by Gemma 4 (31b soundly beats it, the MoE does slightly better, even at 4-bit quantization when using the QAT versi
346.
by SwellJoe 4mo ago
"One command installs 60+ community skills, 67 themes, MCP support, sub-agent support, Claude Code CLI provider support, memory, and more." What, no kitchen sink?
347.
by SwellJoe 4mo ago
I'm not talking about Anthropic.
348.
by SwellJoe 4mo ago
Yeah, I guess I could also write the code myself, if all the models refuse. But, it seems worrying to have a handful of the largest corporations and a few governments having access to the best models while the rest of us are using hobbled o
349.
by SwellJoe 4mo ago
I think the SpaceX deal was to goose the IPO. Google holds a significant stake in SpaceX, and made a fortune on the IPO.
350.
by SwellJoe 4mo ago
I mean, I'm sure there are people within Google who are behaving as though they can keep the dream of the 00s alive in Mountain View, but there's also a whole bunch of people doing work at the frontier in AI. Google has a large le
351.
by SwellJoe 4mo ago
And, yet. https://www.cnbc.com/2026/01/12/apple-google-ai-siri-gemini....
352.
by SwellJoe 4mo ago
I've found in my current work on a security auditing harness and benchmarks, both Fable and Opus are useless. I recently switched to using GPT for Nelson and the security benchmarks I've been doing because Opus started refusing
353.
by SwellJoe 4mo ago
Google is playing a different game. I don't really know what game they're playing, but they're not trying to beat Claude Code. They have coding capabilities and Antigravity, but I'd be surprised if it's much more th
354.
by SwellJoe 4mo ago
I have purchased every Civilization (and closely related) game since III. I have played more hours of any single Civ than all other games combined. I didn't buy Civ VII, because I don't buy malware. As long as it's got Denuvo
355.
by SwellJoe 4mo ago
Yeah, the stuff they preinstall is crazy. It's a ~16GB OS image, which is a huge pain in the ass (in addition to all the other things that are a pain in the ass about Nvidia's OS situation); I can't ship their 16GB pile of cr
356.
by SwellJoe 4mo ago
This particular situation just feels like Fable is able to figure out that it's in a simulation, so it's just playing the game it's been put in. That's not say I think a machine should ever be in a situation where it is
357.
by SwellJoe 4mo ago
This is couched in prepper nonsense, but it's got LLM, WikiPedia, maps, etc. A bunch of genuinely useful stuff to keep on a USB stick or whatever: https://www.projectnomad.us/ But, the current model you really want for
358.
by SwellJoe 4mo ago
Ah, that makes all the difference. Being normal has a lot of value. I mostly take it back. I still don't like the custom Ubuntu Nvidia ships, but if the system is normal any Linux distro for ARM works, it becomes an entirely viable thi
359.
by SwellJoe 4mo ago
Efficiency is the next frontier in LLMs, and I'm not confident the American companies are taking it seriously enough. DeepSeek, even in a naive API-calling loop, serves something like 80-90% cached tokens at an absurdly low price per t
360.
by SwellJoe 4mo ago
I said, "Like, DS4 (a 1-bit quantization of DeepSeek V4 Flash) runs at something like 9-13 tokens/second, with a loooong time to first token." Which almost exactly matches the benchmark you just linked. Looks like it's p

More ›