@kodeefr

ml researcher and founder @EmperoAI closing the loop one optimizer step at a time

kodee.cc
Joined October 2024
  • Tweets938
  • Following165
  • Followers116
81Media
Pinned Tweet
kay
@kodeefrMay 7
what if the universe is a markdown file
9
985
We have outperformed backprop
2
3
183
real RL has never been tried
1
12
318
kay retweeted
Empero
@EmperoAIOct 16
Thank you all for 1M downloads on our 35B Qwen3.8 distill! 🍾
1
2
1
82
2,986
kay retweeted
Empero
@EmperoAIOct 14
🚀 Abacus Desktop 0.1.0 is out. The abacus coding agent, now in a window: your projects and sessions, the agent's work as it streams, the diff, a terminal, and a browser it can use. Open source. macOS, Windows, Linux. Any model. abacus.empero.org
6
4
1
27
1,390
There’s such a long way to go in training optimization. Everyone is looking at architecture and data, no one is looking at learning itself.
1
6
255
kay retweeted
Empero
@EmperoAIOct 12
New from Empero: PageGram: Multi-Head N-gram Memory at the Read Cost of a Flat Table. Engram-style n-gram tables cost 16 random flash reads per token. PageGram reads 2, and matches or beats both flat and Engram-style tables in loss. 📄 alphaxiv.org/pdf/2610.pagegr… 🧵 1/7
3
5
33
1,553
pee is stored in the weights
filipcodes
@filipcodesOct 12
Replying to @Sauers_
interesting that gemini thinks the weights are the vulnerable part.
6
144
yeah its cool and all but im still sad Spark is not OSS yet
Alexandr Wang
@alexandr_wangOct 12
MUSE WONDER this is the culmination of over a year of research and infrastructure work in msl since we originally shipped muse spark in april, we have continued to scale our models across pre-training and RL, now reaching the frontier 🏆 see our model's improvements on the artificial analysis index:
1
6
173
i cant lie abacus desktop is my favorite agent GUI
37
in a due time we will be able to ship models with 1T total params that run on your phone
2
6
260
kay retweeted
Empero
@EmperoAIOct 12
New from Empero: Identity on Demand: Multi-head Gated Value Residuals (MGVR). A drop-in change to attention that reaches the same loss in 16–22% fewer steps: 1.15–1.23× effective compute for +0.2% parameters. 📄 alphaxiv.org/abs/2610.multi-… 🧵
1
7
4
48
2,887
kay retweeted
Empero
@EmperoAIOct 11
Many of you have reached out wanting to contribute to Empero. Until now we had no good way to say yes. So we built one: the Empero contributor program. Research, data & evals, systems, community and partnerships. empero.org/hiring

Contribute — Empero

Join Empero as a contributor: research, data, systems, community and business development on open-weight language models. Remote, voluntary, in the open.

empero.org
2
2
1
18
1,535
who wants to beta test our new model?
5
4
283
running proper large scale RL made me realize how inefficient post training still is
9
206
„there were no cryptographical breakthroughs“
varepsilon
@var_epsilonOct 7
fear not, they still can’t count
8
248
chip design is mostly a search problem.
1
2
163
abacus desktop soon
1
4
185
kay retweeted
Empero
@EmperoAIOct 5
Your coding agent shouldn't stop when you leave your desk. 📱 abacus 0.6.5 puts your terminal sessions on your phone. Watch it work, approve a diff, answer its questions or steer the next step. Your machine stays in charge. github.com/empero-org/abacus
2
7
1
33
2,144
running 512 x 16 rollouts per step in this cyber RL if anyone sees an agent swarm from germany please DM me lol
1
7
125
kay retweeted
Empero
@EmperoAIOct 4
Training your own model shouldn't take a research team. 🍺 Today we're releasing Homebrew, an open agent that brews AI models with you. Tell it what you want and it handles the rest: picking a base model, preparing data, safe parameters and training. github.com/empero-org/homebr…

GitHub - empero-org/brewery-ai: An intuitive console agent to fine-tune AI models with little to no...

An intuitive console agent to fine-tune AI models with little to no experience required! - empero-org/brewery-ai

github.com
29
95
10
915
52,635
Load more