soham (@sohamstwt) header
soham (@sohamstwt) profile picture

inference • systems • EffectTS

Joined August 2024
Photo by soham (@sohamstwt)Photo by soham (@sohamstwt)Photo by soham (@sohamstwt)Photo by soham (@sohamstwt)Photo by soham (@sohamstwt)Photo by soham (@sohamstwt)Photo by soham (@sohamstwt)Photo by soham (@sohamstwt)Photo by soham (@sohamstwt)Photo by soham (@sohamstwt)
shipped RSC Inspector v0.1.5. visualizes Server and client component boundaries directly in your nextjs dev server. - less visual clutter in the UI - supports async nextjs config wrappers try it out with next 16
1
1
3
403
talking shit about github status with my medical friend is the kind of solidarity I never thought i would've had but love it. in a way, non devs now understand the developer frustration more because of AI
Replying to @GergelyOrosz
This is the type of “small businesses that would have never hired any developer now able to build the software they need” that seems to be happening
27
can't leave react tbh. hope there is plans for better reconciler agnostic dx at the router layer @MichaelArnaldi fwiw heard only good things about foldkit from the effect peeps.
Huge news! @devinjameson, creator of @foldkit , is joining us full-time as a core maintainer of Effect 🚀 Better yet @foldkit is coming along with him, and we have big plans for it…
1
1
8
1,310
Dr.K would love this model
Gemini 4 Neon has the best CoT
21
saying again this guy needs to be pussified by china
Robert Pattinson really walked around not knowing Hermione Granger was down bad for him some people are just built different.
22
China has already 5-10 honeypots ready for robert pattinson and I'm all for it. This man must be neutralized with pussifier
16
tibo and openAI are nice, so this is coming from a place of love. IF YOUR MODELS ARE GOOD, WHY IS YOUR SOFTWARE SO SHIT AND UNSTABLE AHHHHHH
One of the funny reasons for the difference in vibes between the influencers and the crowd today on 6.2 Sol/Astra is that we gave a few folks access a while ago and then made a lot of changes in the last 3-4 days and I'm pretty sure no re-testing was done within those days and the videos were produced last week. Don't get me wrong, while the models are a hefty jump compared to their GPT predecessors, they are by no means perfect and there are a few things the teams are excited for the models to get better at in future iterations in terms of taste and niceness of use. Love all the feedback and we hear you!
1
21
I'll believe they got to opus 5.5 level when dots is stable, codex app doesn't start my fan every time I open it(still doesn't open btw)
1
15
CODEX IS FUCKING UNUSABLE. IT IS NOT EVEN OPENING FOR A WEEK. THEIR INTERNAL MODEL CAN PROBABLY TRAVEL BACK IN TIME, BUT CAN'T BUILD SOFTWARE FFS. HOW CAN ASTRA BE SO GOOD AT EVERYTHING BUT CAN'T WRITE GOOD CODE
5
soham retweeted
do you understand now why i've been saying all these inference companies are incompetent
we're bored we're just gonna do phase 2 now $200 of GLM 5.3 usage for $10 have fun
191
92
11
5,523
301,784
I don't care about any further improvements anymore. I can live with Opus 5.5 my whole career and don't look back. shut all the trainings for all i care
11
keep saying this, astra is a REALLY smart model, but can't code for shit
13
Suffice to say I've churned out of codex
10
the latest @dwarkesh_sp pod with the OpenAI board member is pretty revealing how much the METR evaluation of anthropic affected OpenAI. they're nowhere near that kind of RSI optimization, their internal models are not aligned, as of now, they can't seem to figure out alignment.
My fear has mongered. They got me
1
1
1
47
and anthropic has years of head start in alignment research and RSI because it were their entire ethos from the start that they will make safe intelligence and they would make it good at coding so that it will get to that exponential self-improving loop. Dario won
1
1
10
and to dog food this more, astra is the more "intelligent" model than the competition, but not at coding, and that 'but' is so important cause it slows down their RSI loop, and producing good experiments for the alignment data too. nitter.cf/sohamstwt/status/21202…
One thing I'm realising slowly is that astra is THE BEST MODEL available right now, but just sucks at code quality. Because the largest pie of token consumers is still the dev world, astra is underrated.
7
the latest @dwarkesh_sp pod with the OpenAI board member is pretty revealing how much the METR evaluation of anthropic affected OpenAI. they're nowhere near that kind of RSI optimization, their internal models are not aligned, as of now, they can't seem to figure out alignment.
11
people from future will look back at opus 5.5 as the first step to the new world order
Do my non robotics friends know what’s happening rn? Before October 20 it took 1 week to train a whole body controller (ie robot walking model) on 128 GPUs. Now it takes 3 minutes on 1 GPU
15
My fear has mongered. They got me
1
47
opus 5.5 will be remembered as the first taste of AGI bliss
13