Nick Oak

4.3K posts

Nick Oak banner
Nick Oak

Nick Oak

@buildoak_

AI engineer | building with Claude and for Claude | previously: quant, Revolut, 1inch, computer vision lab | https://t.co/JqVJIzjzrB

Dubai, United Arab Emirates Beigetreten Kasım 2018
688 Folgt251 Follower
Nick Oak
Nick Oak@buildoak_·
I think that issue is not just the default effort being medium (that if easily toggle-able though) but also the noticeable degradation of reasoning patterns on high and max which seems to be quite coherent with the peak hours I have adjusted my working day in a manner so that I can get reliable performance (off peak) otherwise doing computational math with CUDA end up being impossible due to the explosion of wrong assumptions and slight operational impotence of the model
English
0
0
1
376
Boris Cherny
Boris Cherny@bcherny·
This is false. We defaulted to medium as a result of user feedback about Claude using too many tokens. When we made the change, we (1) included it in the changelog and (2) showed a dialog when you opened Claude Code so you could choose to opt out. Literally nothing sneaky about it — this was us addressing user feedback in an obvious and explicit way.
English
269
68
2.3K
231.5K
Teng Yan
Teng Yan@tengyanAI·
basically: anthropic sneakily turned down how hard claude thinks before editing code, changed the default from "high" to "medium" effort, and hid the reasoning from session logs. all without telling users. an amd director had 7k sessions of telemetry to prove the degradation was real and measurable (not just vibes). anthropic admitted to the changes. there's a workaround (use "/effort max"). the uncomfortable part is most users had no data to notice it happened at all.
ℏεsam@Hesamation

AMD Senior AI Director confirms Claude has been nerfed. She analyzed Claude's session logs from Janurary to March: > median thinking dropped from ~2,200 to ~600 chars > API requests went up 80x from Feb to Mar. less thinking and failed attempts meaning more retries, burning more tokens, and spending more on tokens > reads-per-edit dropped from 6.6x → 2.0x. model stops researching code before touching it. > model tried to bail out or ask "should i continue" 173 times in 17 days (0 times before March 8). > self-contradiction in reasoning ("oh wait, actually...") tripled. > conventions like CLAUDE.md get ignored because there's less thinking budget to cross-check edits > 5pm and 7pm PST are the worst hours, late night is significantly better. this means the thinking allocation is most likely GPU-load-sensitive.

English
160
530
7.2K
1.3M
Nick Oak
Nick Oak@buildoak_·
@kotickmurochka ‘golubtsy’ funny enough in Russian: golub’ = pigeon goluboy = light blue color and also neutral-negative way to call someone a gay golubki = neutral sweet way to call a male-female pair, neutral-negative way to call a gay pair
English
0
0
1
102
ваня kiss
ваня kiss@kotickmurochka·
I cooked a Russian dish called faggots
ваня kiss tweet media
English
60
27
2K
48.6K
Nick Oak
Nick Oak@buildoak_·
@Nikiita_648 is Nikiita Japanese name as well? (my full name is Nikita - so I’m geniuenly interested)
English
0
0
4
404
Nikiita
Nikiita@Nikiita_648·
今日もロシア語をたくさん摂取したな。 やっぱり、私はロシア語やロシアの文化が好きだ。 ロシア人の友達とブランコを乗りながら馬鹿みたいな会話をして大笑いして、これが私の生きたい道だと気付けた。 私は日本の田舎なんかに嫁ぎにいかない。 良妻賢母なんてなるもんか。
日本語
46
12
957
18.9K
Nick Oak
Nick Oak@buildoak_·
actually it depends there are *generally* two types of mode collapse due to operational or conceptual complexity operational mode collapse could probably be caused by vibe coded mess - yes but conceptual collapse - it’s a whole different beast it’s when on the same succinct context model just fails to grasp on the essential abstractions I’ve been beating the same math problem for well over 2 months the problem could be defined in 2-3 sentences and the opus 4.6 quality has been quite volatile for me starting 21nd of March Though I had several days of brilliant opus 4.6 since then but the quality decrease is noticeable - especially though the peak hours
English
0
0
0
157
tmuxvim
tmuxvim@tmuxvim·
i have a theory that all the people complaining about Claude getting worse actually just vibe-coded themselves a massive ball of slop without paying any attention to architecture and they gradually start noticing poor Claude struggling with it more and more
English
231
38
862
46K
Nick Oak
Nick Oak@buildoak_·
nothing special here though observed concept in quantum physics isn’t something entitled to intelligence deterministic algorithm observing the environment will yield identical result to human observing it
English
0
0
0
18
Nick Oak
Nick Oak@buildoak_·
how does the “observer paradox” from quantum physics resolves when AI observes the environment where experiment takes place ?
English
1
0
0
34
teo
teo@teodorio·
I told my girlfriend I was talking with my Hermes agent on Whatsapp. She got really happy and a bit confused and asked me since when I have a Hermes agent? I told her since this morning, at which point she excitedly asked me if I am planning on buying her something.
English
71
59
2.4K
225.9K
Nick Oak
Nick Oak@buildoak_·
this sounds like not a lot but 400m primes / s give 7 min for full octant segment of annular ring at 850M with “width” of 8192 which makes 1M sweep to be only 14h at 4090 which makes not only the sqrt(40) moat possible - but also several consequent after it
English
0
0
0
13
Nick Oak
Nick Oak@buildoak_·
prior setup as per my experiments* in this field prior fastest been ~150-400k primes / s - Tsuchimura 2005 paper on a whole cluster
English
1
0
0
13
Nick Oak
Nick Oak@buildoak_·
CUDA optimizations and deep dive into registers, memory, cache and SMs have finally paid off prior fastest setup has been ~5M gaussian primes / s processed on 4090 with clean revamped approach I now have 400M gaussian primes per second on 4090! fells insane paper soon
English
1
0
0
25
Nick Oak
Nick Oak@buildoak_·
Once I’ll finish my math research I will dive all of the way into the LLM research there is incredibly lucrative edge on “latent thinking” going on right now I have some fun ideas of the proper regulrizarion that will make it work
English
0
0
0
23
Nick Oak
Nick Oak@buildoak_·
optimizing Cuda kernels on Jetson Orin Nano and then pushing them to 3090 for ~25 times performance bump feels like magic
English
0
0
0
20
Nick Oak
Nick Oak@buildoak_·
random fact but in Brunei there is a smell of flowers everywhere
English
0
0
0
47
Nick Oak
Nick Oak@buildoak_·
CUDA is a beast that is, well, hard to tame
English
0
0
0
28
Nick Oak
Nick Oak@buildoak_·
@Mayu_kosho a lot of Russians abroad using TG though and plenty of tech savvy inside the country using it as well
English
0
0
0
30
14ひきの刺身
14ひきの刺身@Mayu_kosho·
今ツイッターできてるロシア人ならTGも見れんじゃねと思ったけどロシアにいない人が大半やったりするんやろか
日本語
79
0
157
7.2K
14ひきの刺身
14ひきの刺身@Mayu_kosho·
ロシア人の皆様にはぜひ頑張ってロシア語で書いたりしているTGチャンネルのほうを登録してほしさがあるが、ここに出すと日本人に垢バレする恐れがあってめちゃめちゃ迷っている
日本語
276
17
1.4K
47.5K
Nick Oak
Nick Oak@buildoak_·
@kayintveen @enjojoyy true! and this is how you can make the big fella to think for the 10-15 min and come back with genuinely bright ideas
English
0
0
0
13
Kay
Kay@kayintveen·
@buildoak_ @enjojoyy this is the thing. low-level stuff where you're negotiating cache locality and branch prediction with the compiler, that's where claude becomes a genuine thought partner not just an autocomplete. the tight loop matters
English
1
0
1
35
albina
albina@enjojoyy·
So I do currently ship 100% of code using agents And I do miss coding manually a lot now Do you guys ever code by hands just not to forget how to do it? For the good time? I think I would schedule around one hour every day to code only by hand before I would start using AI Also in order to have a better architecture image in mind
English
20
0
24
1.3K
Nick Oak
Nick Oak@buildoak_·
depends on the task you wanna do throwing xhigh at some vanilla stuff will likely lead to over engineering though using it for audit & hard stuff - yields the best results for C++/CUDA code review there is no better model than gpt 5.4 xhigh - opus 4.6 and gemini 3.1 pro preview all miss critical moments
English
0
0
3
722
Minh-Phuc Tran
Minh-Phuc Tran@phuctm97·
Is it true that Codex GPT 5.4 with Extra High thinking effort is worse than with Medium/High thinking effort? If it’s true, that’d be very bad design. 😅
English
80
1
195
44.4K

Entdecken

@kotickmurochka @Nikiita_648 @teodorio @elonmusk @BarackObama @taylorswift13 @cristiano @BillGates