sumit

3.1K posts

sumit banner
sumit

sumit

@sumitdotml

learning how to think in matrices

日本 東京 Katılım Eylül 2022
549 Takip Edilen9.8K Takipçiler
sumit
sumit@sumitdotml·
just have to laugh at this point, i've exhausted everything i could possibly afford
sumit tweet mediasumit tweet mediasumit tweet media
English
5
0
5
2K
sumit retweetledi
Jarred Sumner
Jarred Sumner@jarredsumner·
Bun v1.3.14 releases tomorrow. If we do merge the Rust rewrite, this would be the last version in Zig
English
188
159
3.2K
595.6K
sumit
sumit@sumitdotml·
sumit tweet media
ZXX
0
0
9
381
kannav
kannav@Kannav02·
@sumitdotml Most of the usage is subsidised for cursor’s in house model, composer 2, its a really good model when it comes to non-trivial tasks, i’ve been using cloud agents with composer 2, gets a lot of work done
kannav tweet media
English
1
0
1
81
sumit
sumit@sumitdotml·
somebody needs to tell me this is rather a weekly usage limit because how on earth is one meant to live on a cursor pro subscription otherwise? and if this is indeed a monthly usage cap, how are cursor even surviving? I’m actually laughing so hard this is so bleak
sumit@sumitdotml

so I overworked both codex & cc so much this week that they ran out of weekly limits had no additional money, remembered I had free cursor credits to get me a pro subscription ran 5 gpt-5 medium prompts in the cursor cli, 8% of the (monthly?) pro model usage already wiped out..

English
7
0
11
2.7K
sumit
sumit@sumitdotml·
so I overworked both codex & cc so much this week that they ran out of weekly limits had no additional money, remembered I had free cursor credits to get me a pro subscription ran 5 gpt-5 medium prompts in the cursor cli, 8% of the (monthly?) pro model usage already wiped out..
sumit tweet media
English
1
0
6
2.7K
sumit
sumit@sumitdotml·
@jqlive I see, didn't know!
English
0
0
0
31
JQ
JQ@jqlive·
@sumitdotml No, they still have colossus 2, and his plans for Terafab and Space GPUs. This is Elon renting out older unused GPU capacity to a customer, akin to what Google does with its TPUs.
English
1
0
6
117
sumit
sumit@sumitdotml·
oh wow I did not expect this so xai is essentially dead?
sumit tweet media
Claude@claudeai

We’ve agreed to a partnership with @SpaceX that will substantially increase our compute capacity. This, along with our other recent compute deals, means that we’ve been able to increase our usage limits for Claude Code and the Claude API.

English
2
0
14
2.1K
sumit
sumit@sumitdotml·
I've been doing lora study + self-driven research in the process, I ran the gsm8k benchmark on qwen3-8b: relatively ok score (comparable to llama3.1b) + provides some room to improve when I finetune this, good stuff onto some lora finetunings + comparisons now
sumit tweet mediasumit tweet media
English
0
0
16
647
Yacine Mahdid
Yacine Mahdid@yacinelearning·
are you very eager to make a research contribution to the ai field? motivated perhaps? no idea how to do it? willing to commit crimes even? well this little 16min video is your perfect starting point to make a solid impact
Yacine Mahdid tweet media
English
16
18
366
49.3K
sumit
sumit@sumitdotml·
seems tinker does not expose lora alpha? I checked their sdk code + the repo and got nothing wonder if it's a hidden default owned in the backend that gets calculated based on the rank field & the model I pick hmmm
English
0
0
1
241
sumit
sumit@sumitdotml·
less than 24 hours since I last accessed this, cool
sumit tweet media
English
0
0
5
483
sumit
sumit@sumitdotml·
are you actually serious
sumit tweet media
English
2
0
9
715
sumit
sumit@sumitdotml·
funny how I still like to use cursor for its editor mode because their markdown previewer is just so nice
English
1
0
8
486
Rafael Mendiola
Rafael Mendiola@GroundControl·
@sumitdotml Pro-tip you can switch to 4.6 and you'll actually have a bigger token budget than what you had before 4.7. They bumped up the token budget since 4.7 eats so much
English
1
0
3
37
sumit
sumit@sumitdotml·
opus 4.7 burns too many tokens too fast, I exhaust my usage limits too fast it’s actually demoralizing to work with
English
5
0
12
1K
sumit
sumit@sumitdotml·
@mrAy0xu from opus 4.6 to 4.7, no all it does is deplete my tokens 2x faster
English
0
0
0
76
Tobi Williams
Tobi Williams@mrAy0xu·
@sumitdotml do you find a significant leap in performance from opus 4.6 or sonnet to opus 4.7 for yoir workflows?
English
1
0
1
54
sumit
sumit@sumitdotml·
@th1nkp0l I still use claude because I like crafting model debate threads between them for robust plans (been pretty effective for me) but other than that it's been completely codex for me these days
English
0
0
1
97
cory
cory@th1nkp0l·
@sumitdotml I bounce back and forth, too. I try to use them each for the tasks they’re best at, and many days I don’t even hit either limit as a result. It feels a little clunky, but you reach into the toolbox and sometimes the tools are only slightly, but nonetheless importantly, different
English
1
0
1
65
sumit
sumit@sumitdotml·
never thought there’d come a time I would barely even touch claude code except for when I hit the codex usage limits but here I am today
English
5
0
15
1K

Keşfet

@KamStaszewski @Kannav02 @jqlive @yacinelearning @elonmusk @BarackObama @taylorswift13 @cristiano