Pinned Tweet
Again, it happened again. Opened Facebook and I saw this. I looked at the comments and they're just unsuspecting boomers congratulating the fake AI gen couple😂
I opened Facebook and this is the very first thing I see, the boomers don't know what hit them
> At least we can tell that the image on the left was generated by gpt-mage-1 since there's heavy yellow tint
13
12
311
536,535
Sakib🏴 retweeted
Trace waterfall component, see where your agent’s time goes
71
87
2,393
99,077
so instead of the classical:
θ ← θ − α∇L
lol it's now:
θ ← θ − Agent(∇L, "the vibes")
ngl this felt like slop but... seems like it worked?
> normal optimiser overfit on bad labels
> this one didn't and kept learning
θ ← θ − α∇L
lol it's now:
θ ← θ − Agent(∇L, "the vibes")
ngl this felt like slop but... seems like it worked?
> normal optimiser overfit on bad labels
> this one didn't and kept learning
1/9 Current optimizers are static during training. What if the optimizer could think and react while the network trains?
We let an agent watch the training dynamics and rewrite the optimizer's source code as training goes: a reactive optimizer.
In every language and image as no tuned recipe, it matches or beats the best recipe one can one-shot, and it guards against loss spikes and overfitting by itself. In the near future, optimization decisions may belong to agents, not humans.
📝 xingjianbai.com/reactive-optim…
💻 Try it on your task: github.com/xingjian-bai/o…
🧵
1
0
2
405
Sakib🏴 retweeted
big day for autonomous ai research, not that surprising results imo because pre-training is quite verifiable, soon everything will just be a function of compute and how good your "discovery loop" is
congrats @mirendil @Recursive_SI !
congrats @mirendil @Recursive_SI !
verifying the science of LLM training is just a function of compute, and as compute scales you can almost infinitely parallelize and compress time
math was even easier, tons of existing data, you only need inference tokens, it's extremely parallelizable and proofs almost never have negative second order interactions with each other
even if model training is less of a perfect testbed than math, there is no doubt this will happen too. the big difference is that it needs 2*compute (inference and A LOT of training compute) so much less accessible, but the dynamics are the same
labs have spent enormous effort figuring out the right metrics for training. for pre-training it's relatively straightforward - which is also good news for tracking RSI - you establish a scaling ladder and measure the "efficiency gain": for a given budget, how much does a new proposal save compared to baseline to reach the same metric?
the question is how you define this "efficiency gain", which efficiency (flops? time? KV cache? data efficiency?..) and which gain (val loss? knowledge? reasoning?..)
i think it's quite similar for post training (modulo reward hacking), i would hope it's similar for alignment as well but after listening to @dwarkesh_sp <> @polynoamial itw i'm not sure as the "gain" here would assume that we have a correct way to measure behavior without eval awareness, probably other stuff i'm not thinking of here
anyway, when you put all of this together, hard to not believe in RSI
12
22
308
12,771
Whats the best way to get Opus 5.5 to use my computer? i.e. click around and shit
> is it cua-driver? (@trycua )
> or hijacking the OpenAI Computer Use (@skybysoftware) mcp?
> or copying the harness in OSWorld (@XLangNLP)?
> is it cua-driver? (@trycua )
> or hijacking the OpenAI Computer Use (@skybysoftware) mcp?
> or copying the harness in OSWorld (@XLangNLP)?
1
0
1
185
Sakib🏴 retweeted
🚀 Introducing DeepInfer, our in-house inference framework, now fully open source!
Powering DeepSeek-V4.1-Flash at scale with high throughput and low latency. We’re sharing our production stack to help the community lower compute costs, reduce energy use, and make AI more affordable.
🔗 github.com/deepseek-ai/De…
1/4
Powering DeepSeek-V4.1-Flash at scale with high throughput and low latency. We’re sharing our production stack to help the community lower compute costs, reduce energy use, and make AI more affordable.
🔗 github.com/deepseek-ai/De…
1/4
1,619
4,878
42,972
6,651,376
i watched the social reckoning btw
it’s mid
not bc it’s bad
but bc the framing of the whole story
like the framing is so… "and then we won"
like bro...
i live now, like rn, and i can tell you nothing happened
it’s the same
it’s [facebook, instagram, etc] worse
idk
very mid!
jeremy strong!
it’s mid
not bc it’s bad
but bc the framing of the whole story
like the framing is so… "and then we won"
like bro...
i live now, like rn, and i can tell you nothing happened
it’s the same
it’s [facebook, instagram, etc] worse
idk
very mid!
jeremy strong!
0
0
0
185
Sakib🏴 retweeted
1/4 Today, we’re releasing Mercury 3. Mercury 3 is by far our most powerful model and the most intelligent diffusion LLM ever released. It outperforms GPT-6 Luna and Haiku 5.5 on quality while running at 1000+ tok/sec.
287
458
5,267
479,293
Sakib🏴 retweeted
I've been doing some hobbyist astronomy with Argon. Here's one of my experiments, where I tasked Argon with reproducing the first black hole images (from 2017/18) using the same raw data.
Argon:
- pulled raw 230 GHz radio fringes (UVFITS) from the Event Horizon Telescope terferometry solver using 3-dish closure phases to cancel atmospheric noise
- started each run from a plain Gaussian blob with no ring templates or masks
- recovered M87* (2017 & 2018) and Sgr A* at a 91% to 98% match of the official EHT images
Argon:
- pulled raw 230 GHz radio fringes (UVFITS) from the Event Horizon Telescope terferometry solver using 3-dish closure phases to cancel atmospheric noise
- started each run from a plain Gaussian blob with no ring templates or masks
- recovered M87* (2017 & 2018) and Sgr A* at a 91% to 98% match of the official EHT images
32
42
504
58,674
RT @levelsio: In a similar scramble I did to jump on AI in late 2022 (moving from remote work and nomads to AI apps), I am now scrambling t…
0
107
0
18
