@Carboaceto Maybe in the future many university courses that nowadays are about technology, will steer towards pure maths and algorithms to help develop these low-level thinking skills...
Very well said, another important aspect of open weight models is the educational possibility they open. In particular projects like ds4 there are highly focused on one architecture, are perfect playground to hack and learn how LLM serving works and the tradeoffs involved, as well as LLMs' architectures in general.
@antirez IMHO the idea of focussing on ds-v4 alone was a better long term bet. There'll always be a new better model but if we keep switching from one to the other we'll never get full optimization and stability, and in 3/6 months a new ds is released perhaps with similar architecture...
recommended reading. i love this type of forensics. an internet friend thought me 8087 fpu programming in the 90ies. first time i had contact with a stack. good times.
https://t.co/IuNiFXQ0vr
@antirez Cool, it would be interesting to have a standard corpus of prompts for optimization and one for testing, to make this measurements more systematic.
๐ 30% US exposure sounds like a lot, but I'm actually underweight the US versus the MSCI World Index, where it sits around 72% (as of May 2026).
My exact country split isn't deliberate. But I do think geographic diversification is worth having, as it spreads both currency and geopolitical risk.
Copy Trading is not investment advice | Capital at risk | Past performance does not guarantee future results
GOLDEN MOMENT ๐ฅ
History was waiting at the top of the wall! ๐ฎ๐ฉ
With a brilliant climb in the menโs Lead final, Putra Tri Ramadani secured his first-ever World Climbing Series gold medal and became the first Indonesian climber to win a medal outside of the Speed discipline. ๐
@_Suresh2@antirez With 32GB getting to 10t/s without logits drifts is though. There might be something to gain employing non-uniform caching. 1st and middle layers seem to have lower hit-rate. Exact logits are not a big deal to me. You are not running ds4f no more, but as long as it is useful...
Running Deepseek-v4-flash with @antirez ds4 ssd-streaming feature and a few more tricks, on a MacBook pro M2 Max with 32GB of RAM.
10 t/s at decoding is approaching the usability threshold.
I must say that logits correctness is not guaranteed with this implementation.
Yes, since with 32GB tradeoffs is pretty much the only thing one can do, that is want I have done.
ds4-eval tests are passing. Overall I'm finding it more capable than qwen3.6 in my personal use.
Perhaps logits correctness is not what one should aim for when running on very limited hw.
@antirez@lollipop_stat We might call it a bug , but I made it that way intentionally. I understand that doesn't follow the ds4 philosophy of logits correctness, which I respect. Still I find this kind of exploration interesting because it could enable useful local AI with cheeper machines.
@lollipop_stat@antirez No, it is an intentional architectural change under the assumption that there are different "internal trajectories" to get to the same macro qualitative result.
I need to run more tests, but so far the thesis is holding.
@antirez The underlining idea is that there are a thousand ways to come up with the same idea! ๐
It seems to work well, but I haven't tested logits correctness.
More then an optimization is a tradeoff ai would say...
@antirez I am picking 6 among the top-k (k>6) experts selected by the router, if they are cache residents already.
The result is an increase in hit-rate but the forward pass, follows a different international trajectory.
DeepSeek v4 Flash is able to use an EDIT tool I re-designed compared to what people normally use. The READ tool returns lines+tags, like: 1:f3_c int main(...), where the four chars are base64 crc of the line. So when there is to edit, there is no need to repeat the old line, just the tag. This saves tokens and produces less failures. The wager was if DS4F could really use it. Apparently yes.
Did you know you can get a 4% cashback on every purchase you make with your eToro Card?
Don't miss out on this new feature, exclusively from eToro.
Want to know how to get your card and start earning stock with your daily spend?
Comment MORE๐ธ, and I'll send you all the information directly via DM!