girvo

9 ms·

181.
by girvo 4mo ago
I think that’s an acceptable trade off, personally… but! You can implement persistent caching if you’d like :) you create the record store when you stand up the environment, so it’s possible to have that hydrated with data from somewhere el
182.
by girvo 4mo ago
I can say that its happening right now (not where I work, at a place I used to). Even being shipped to production as-is, with data loss and security issues in tow. We're cooked haha
183.
by girvo 4mo ago
> unless something is done about the industry to change course There will be no change of course, sadly.
184.
by girvo 4mo ago
AKA what Relay does out of the box haha Though its depressing how few actually use it to its full extent. My team is one of the few where I work; heavy declarative mutation directives with optimisticResponse (and optimisticUpdaters because
185.
by girvo 4mo ago
Yeah I cannot see how "we shipped 1 million lines of code in three weeks" is... something to be proud of haha
186.
by girvo 4mo ago
Just chiming in to say that I've seen the exact same that you have. Tests are better used to help validate that was was generated worked after the fact. That, and even the absolute SOTA models still suck at writing tests. Which shoul
187.
by girvo 4mo ago
The “-assistant” models released by Google are specialised tiny MTP draft models :) 31b-it-assistant is what enables MTP
188.
by girvo 5mo ago
Step 3.7 Flash on my Asus GB10 based mini pc is incredibly close to that today. I’m very impressed, and that’s without MTP to boost performance
189.
by girvo 5mo ago
Yes. I’ve got so many greyed out songs in my library due to this :(
190.
by girvo 5mo ago
You top out at 20 tokens per second on hardware with memory bandwidth this low for any local model actually worth using. Doing the maths, it’s not financially worth it. Only worth it for privacy and control reasons. I do love my GB10 Asus S
191.
by girvo 5mo ago
> inputs and outputs If the inputs and outputs are only and exactly those of the domain, sure. But software is more than that. And your logistics operator (or my actual work example: our extremely talented designers with deep understan
192.
by girvo 5mo ago
Same for me: download was okay, upload was completely wrong
193.
by girvo 5mo ago
They adopt different alignment, not no alignment.
194.
by girvo 5mo ago
> I suspect nobody is doing real student teacher distillation It gets used for quantisation, basically recovering accuracy for lower quants (Nvidia calls it QAD). Can’t speak to how widespread it is though
195.
by girvo 5mo ago
> takes our jobs and makes RAM expensive I mean it is doing both of those, so thats fair to be honest.
196.
by girvo 5mo ago
True, I might be thinking of some of the communities four-Spark clusters for it; it’s already int4 right?
197.
by girvo 5mo ago
DeepSeek v4 Flash, various quantised versions of Kimi K2.6, MiniMax 2.7, Qwen 3.5 “full sized, with a dual spark setup you can fit some decent setups on here My single spark has me running Qwen 3.6 27B and antirez’s specially quantised Deep
198.
by girvo 5mo ago
Australia! Catholic high school, though in practice amusingly secular
199.
by girvo 5mo ago
It’s a whole thing. The priest at my high school was way into Tolkien and, more interestingly to me as a teenager: the Wolfenstein series of games!
200.
by girvo 5mo ago
To me an all powerful government is one that has all the information the user just listed, because governments are by definition the most powerful force in our countries. Private companies should also be regulated with regards to data sales
201.
by girvo 5mo ago
Other jobs don’t require this kind of shield. Instead, they require insurance. Qualified immunity isn’t qualified, and it’s a horrific distorting function on your society, as officers get to act with impunity. They’re given more and more po
202.
by girvo 5mo ago
People are arguing with you as if you’re not making the same point they are, amusing
203.
by girvo 5mo ago
Wait, I went and looked far and wide today and I can’t find anything with ~100GB of VRAM that isn’t $20k a year, what am I missing
204.
by girvo 5mo ago
> This suggests that Gemini tokens are actually profitable, or at least not extremely unprofitable. Out of all the companies, considering their own silicon etc. I wouldn't be surprised. Though I do wonder in terms of total CapEx and
205.
by girvo 5mo ago
> renting really does beat owning, and cloud beats both
206.
by girvo 5mo ago
Fine tuning, and yeah training my own, experimenting with architectures and learning how it all works. Been a lot of fun
207.
by girvo 5mo ago
Less than I expected, though! And I get to run this all through the night I do still use Vast and Runpod for things too, but it’s much nicer to test a fine tuning run here to make sure I’m in the ballpark I also did literally say “It's
208.
by girvo 5mo ago
Yes I could, but that is annoying because of spot pricing and having my instance shut down, and it has fluctuating prices It’s also annoying because then I need to make sure my little “lab” setup is well automated, and I’m lazy :) Also, I l
209.
by girvo 5mo ago
I didn't spend that much, only $6500 AUD for a GB10 based Asus GX10 which is even slower than OPs, but I spent that because it makes for a great learning platform. Theres not much else that lets me fiddle with 128GB of RAM for my g
210.
by girvo 5mo ago
Right and all of my own evals back this up for Gemma 4... ...except its notably worse at coding in an agent context even with a harness setup to do exactly what Google says it should do (wrt. to sending summarised thinking back and so on) S

More ›