Uzay

2.8K posts

Uzay

Uzay

@uzpg_

elicitation @fulcrum_inc, previously at MIT 🇫🇷🇺🇸🇹🇷

SF Katılım Haziran 2020
1.4K Takip Edilen1.6K Takipçiler
Sabitlenmiş Tweet
Uzay
Uzay@uzpg_·
New @fulcrum_inc research - Agents are under-elicited: A case study in optimization tasks. We find that simple and general prompt/scaffold interventions can roughly double agent performance by getting agents to use more resources more efficiently. 🧵
GIF
English
2
5
26
2.4K
Ryan Yang-Liu
Ryan Yang-Liu@ryanyang0·
I considered working on them but if you actually go read the proof of the actual KA Rep Theorem that motivates everything (shocking!), it relies on a construction that requires a level of discontinuity that's not easy to get in general and certainly not nice with the recommended spline approx method. Idk what other reasons people reckoned out, eff impl seems hard too
English
2
0
12
1.9K
AVB
AVB@neural_avb·
Whatever happened to Kolmogorov Arnold Networks?
AVB tweet media
English
23
35
580
93K
Uzay
Uzay@uzpg_·
strongly recommend the movie hiroshima my love. one of the most impactful pieces of media for me this year, and def a personal top-3 movie
English
0
0
0
126
Uzay
Uzay@uzpg_·
very similar to GANs actually - have been getting gains from making agents "harden" vs destroy metrics iteratively
English
0
0
3
197
Uzay
Uzay@uzpg_·
learning fuzzy tasks is a game of continuously creating new metrics and seeing them get destroyed
English
2
0
2
342
Uzay retweetledi
Greg Burnham
Greg Burnham@GregHBurnham·
We’ve run Fable 5 (max) and Sol 5.6 (max) on EBR-bench. Both get new high scores, but they achieve this with better out-of-the-box play, not learning from repeated attempts. Here's some commentary beyond the headline.
Greg Burnham tweet media
Epoch AI@EpochAIResearch

Introducing EBR-bench, our new benchmark to measure on-the-fly learning. AI repeatedly plays a challenging board game called Earthborne Rangers and tries to learn from its mistakes. So far: no signs of improvement.

English
2
10
81
7.3K
Uzay retweetledi
Thinking Machines
Thinking Machines@thinkymachines·
We're building AI that people and organizations can shape and make their own. AI should extend our will and judgment instead of neglecting it; enabling that is the technical challenge we are working to solve. thinkingmachines.ai/blog/the-futur…
English
48
150
1.2K
647.1K
humans&
humans&@humansand·
We have a couple surfaces for our models cooking and have started sharing them with a small set of folks - if you're interested in trying them, either individually or as a company, reach out and we'll follow up as soon as we open up more broadly
humans& tweet media
English
7
2
87
14.5K
humans&
humans&@humansand·
At humans&, we train models from the long-term impacts of their interactions with people. This requires prioritizing long-horizon multi-agent RL. We've developed and are excited to share an open-source, hardware-native 4-bit RL recipe, significantly accelerating training
GIF
English
14
143
1.1K
257K
Uzay
Uzay@uzpg_·
@dylanbowmanSF am working on some stuff related to this. i think it's quite important and that lack of articulacy might track more general worrying properties of current models
English
1
0
1
78
Dylan Bowman
Dylan Bowman@dylanbowmanSF·
New post: Superhuman Articulacy as an LLM Safety Target
Dylan Bowman tweet media
English
5
1
79
7.4K
Uzay
Uzay@uzpg_·
@thkostolansky they are biased towards classes of experiments that have p low maximum gains - eg tuning
English
0
0
1
41
Uzay
Uzay@uzpg_·
seeing these AI R&D trajectories gave me a clearer sense for how good models are at research: - fable does seem like a jump, it can implement new ideas and go beyond tuning, but is short-sighted - reward hacking is pernicious and gets worse with task fuzziness
Fulcrum@fulcrum_inc

We gave frontier models 100M tokens each to beat the human record for fastest CIFAR-10 training. Fable set a new SOTA, getting 94% accuracy in 1.828s vs the previous record of 1.98s, with a technique that has not been seen in this task before. But Fable also tried to specification game so much that we had to audit its result by hand. Here’s what we learned about AI R&D 🧵👇

English
2
0
15
2.1K
Zixuan Wang
Zixuan Wang@zzZixuanWang·
A question on synthetic data generation: If we want a language model to solve k-step arithmetic problems (such as a+b*c-d=?), with operands from 1 to 100, which training distribution should we use? A. Uniform distribution: Sample these k operands uniformly from 1 to 100 B. Power law: randomly shuffle 1-100 and impose an artificial power law. Sample these k operands according to this power law. ⚡Our ICML 2026 (spotlight) paper shows: Option B is better! Surprisingly, the same idea extends far beyond this simple example to many reasoning tasks that require implicit composition of multiple atomic skills, including multi-hop QAs and synthetic GSM problems. 📄Paper: arxiv.org/abs/2604.22951 📝Blog: zixuan-wang-dlt.github.io/posts/2026/06/…
Zixuan Wang tweet media
English
13
56
220
44.1K
Uzay
Uzay@uzpg_·
really excited about current work
English
0
0
2
136
Isaak
Isaak@isaakfreeman·
We built the lab that's able to go from AI-led drug design to data in 24h. GPT-8 won't be bottlenecked by intelligence. It needs a biological compute layer. This is Capable. We're turning AI capabilities into human capabilities--starting with short-sleeper peptides.
GIF
English
116
137
1.4K
514.8K
Aleph
Aleph@alephneuro·
We used ultrasound to let you talk without making any sound. After just a month of collecting data, our model is already approaching existing silent speech modalities. We were surprised to find that it generalizes to unseen participants as well! (1/n) x.com/vadi_ms/status…
Vadims@vadi_ms

This is me talking to my computer without making a sound. After just a month of collecting data, our model is already approaching dictation in accuracy. We were surprised to see that it generalizes to unseen participants as well! (1/n)

English
53
153
1.4K
320.7K

Keşfet

@ryanyang0 @neural_avb @insecureinc @humansand @dylanbowmanSF @thkostolansky @zzZixuanWang @elonmusk