Thibault Castells (@thibcastells) 's Twitter Profile
Thibault Castells

@thibcastells

Research engineer (diffusion models & edge AI). I mostly tweet about things related to ML.

ID: 894532726308646912

calendar_today07-08-2017 12:16:36

618 Tweet

92 Followers

79 Following

Taelin (@victortaelin) 's Twitter Profile Photo

I apologize for the emotional vagueposts. I'll be direct now a major breakthrough took place today for a context, I'm working on the problem of synthesizing a program by examples. that is, given a set of tests, like: f(3, [0,1,2,0,2]) = [1,1,1,0,2] f(1, [4,2,3,7,7]) =

I apologize for the emotional vagueposts. I'll be direct now

a major breakthrough took place today

for a context, I'm working on the problem of synthesizing a program by examples. that is, given a set of tests, like:

f(3, [0,1,2,0,2]) = [1,1,1,0,2]
f(1, [4,2,3,7,7]) =
Taelin (@victortaelin) 's Twitter Profile Photo

Optimal Evaluation is so mind-blowing If I ask NeoGen to synthesize 'add : ℕ → ℕ' and 'sum : [ℕ] → ℕ' together, it takes ~0.28 seconds to find both. Now, if I perform the same search, except I don't print the source of 'add'... literally, if I just remove that 'print'

Optimal Evaluation is so mind-blowing

If I ask NeoGen to synthesize 'add : ℕ → ℕ' and 'sum : [ℕ] → ℕ' together, it takes ~0.28 seconds to find both. Now, if I perform the same search, except I don't print the source of 'add'... literally, if I just remove that 'print'
Thibault Castells (@thibcastells) 's Twitter Profile Photo

> March: first time people ask us to never stop offering a model with such passion > April: we are stopping offering this model in 3 months It’s not an issue to me, but I find it funny 😆 community.openai.com/t/gpt-4-5-prev…

Anton Osika – eu/acc (@antonosika) 's Twitter Profile Photo

1. Be Replit founder 2. Have a decade-long head start 3. Watch small EU competitor, Lovable, surpass you in usage and making vibe coding secure 4. Copy it 4 weeks later 5. Bash Lovable for not being secure Incredible leadership.

1. Be Replit founder
2. Have a decade-long head start
3. Watch small EU competitor, Lovable, surpass you in usage and making vibe coding secure
4. Copy it 4 weeks later
5. Bash Lovable for not being secure

Incredible leadership.
Chubby♨️ (@kimmonismus) 's Twitter Profile Photo

Europe has its first reasoning model: Magistral After OpenAI was known to be the first to introduce AI reasoning last year with o1, DeepSeek followed suit in January with DeepSeek r1. Almost 5 months later, Europe now also has its own reasoning model with Mistral's Magistral -

Europe has its first reasoning model: Magistral

After OpenAI was known to be the first to introduce AI reasoning last year with o1, DeepSeek followed suit in January with DeepSeek r1.

Almost 5 months later, Europe now also has its own reasoning model with Mistral's Magistral -
Thibault Castells (@thibcastells) 's Twitter Profile Photo

That’s really good news, but it makes me curious What kind of hardware update can reduce the inference cost by 80%? And why wasn’t it possible before? 🤔