girvo

6 ms·

91.
by girvo 4mo ago
Depends on the model. Step (from StepFun) will happily yap about Tiannemen to you, if you're running it locally. Quite a lot of these models have "safety" (lol) filters in front of them, vs it being heavily encoded into the w
92.
by girvo 4mo ago
Truthfully? There isn't one. They all have flaws. Your best bet is to look at all of them, and then run a suite of evals yourself. Its rough out here!
93.
by girvo 4mo ago
The specific market positioning is... for me to use at my big tech company job, where we aren't allowed to use GLM and similar, but have fixed caps on how much token usage we're allowed to rack up a month.
94.
by girvo 4mo ago
Which basically only Nvidia does, because it’s very expensive. Though I’m currently working on QADing the smaller Qwen 3.5 models from FP16 teacher to NVFP4 student, to hopefully eventually apply it to 3.6 27B… harder to get right than I ex
95.
by girvo 4mo ago
But prompt processing is terrible
96.
by girvo 4mo ago
It’s half a sentence to be fair, easy to miss!
97.
by girvo 4mo ago
Not Llama 3.1, but Step 3.7 Flash is one of the few new high quality models in this size bracket. DeepSeek v4 Flash too
98.
by girvo 4mo ago
My GB10 Spark-alike is absolutely amazingly fun… but it is not cost effective. Step 3.7 Flash is shockingly capable (IQ4_XS and used for web dev mainly), but it cost me $6800 AUD. They’re even more expensive now. The numbers just don’t ma
99.
by girvo 4mo ago
Gemma is better than Qwen at everything except coding, in all my evaluations. Which is a shame because that is what I use them for!
100.
by girvo 4mo ago
The Super-R run-command dialog on Windows, and Notepad make it look crazy
101.
by girvo 4mo ago
Yes, it does: > a genie-generated description of YAGNI intended for the improvement of future generations of genies
102.
by girvo 4mo ago
Why would I care about Christian morals? In fact from what I can see of the US, you don’t have them either.
103.
by girvo 4mo ago
The article said it was indeed AI generated babble right at the start, which is why I skipped reading it the moment I saw that. A shame, really.
104.
by girvo 4mo ago
Which is itself a problem as (in my partners evaluations as an optometrist), LLMs used for clinical notes has a bad habit of dropping clinically important information, and the biggest providers don’t give you a copy of the raw transcript or
105.
by girvo 4mo ago
Aw that’s a shame; I’m running the official llama.cpp on my Spark-alike, and it works great now. Proper triple head too which is what it is trained on, gets me up to 35-40tk/s decode
106.
by girvo 4mo ago
That’s what I did with Pi, super simple :)
107.
by girvo 4mo ago
Being grafted onto the main model reduces layer duplication that you’d otherwise have: at least for Step and Qwen 3.6
108.
by girvo 4mo ago
You’d be quite surprised, I think. Fine tuning a model on one axis can have drastic impacts on another that as a human we would expect to be completely unrelated.
109.
by girvo 4mo ago
Hah the edit window here is the bane of my existence sometimes :) They’re a super interesting company, and their innovations are fascinating
110.
by girvo 4mo ago
Then you have no understanding of what DeepSeek has actually done. They publish their work openly: go have a look! Their architectural improvements are fascinating.
111.
by girvo 4mo ago
> Download DeepSeek-R1 671B Dunno why you'd want to though, considering v4 Pro (and even Flash) outpace it drastically
112.
by girvo 4mo ago
> I found Flash to be a bit shaky as well until I started using it in xhigh/max thinking effort This is true for most of the open weight chinese models, to be fair. They're really built around long reasoning chains. Also you&#x
113.
by girvo 4mo ago
Yeah! So at a much smaller scale, being able to boost Step 3.7 Flash up to 40tk/s on my Spark-alike with proper triple head MTP was the thing that made it superior to Qwen 3.6 27B in wall clock time despite Step reasoning more A lot of
114.
by girvo 4mo ago
Z.ai and MoonShot (and StepFun and some others who are another six months behind or targeting smaller use cases) are still open, surprisingly enough. Alibaba making Qwen close up shop for its best models isn’t that surprising, though sad. T
115.
by girvo 4mo ago
> If your web browser hacks a bank, but you didn't know and didn't expect it to, have you hacked a bank? Depends, as usual. Intent can matter, but depends on the statute (and jurisdiction) in question.
116.
by girvo 4mo ago
I still have mine :)
117.
by girvo 4mo ago
> The overwhelming majority of the population cannot perceive anything over 90 Hz This isn’t true. Unless you’re saying I’m some kind of fighter pilot at 35!
118.
by girvo 4mo ago
I really wish I had followed through when I was ask by some of the guys at IBM Research to apply when we had worked together on a partner project. Though I didn’t have a degree which I seem to remember was a sticking point, this is in the m
119.
by girvo 4mo ago
Not that it matters but most of the open weight models aren’t actually censored that way: they run another layer on top of to do that. At least some of them do, Step 3.7 Flash locally happily tells me about the Tiananmen Square massacre
120.
by girvo 4mo ago
Simple: a lot of the people claiming they’re reviewing the output of these models are lying. Also if you run the “loops” they’re now yapping about, it will burn through enormous amounts of usage as well.

More ›