Nadav Timor (@NadavTimor) / X

Nadav Timor

301 posts

Nadav Timor

@NadavTimor

AI inference, speculative decoding, open source. Built novel decoding algorithms – default in Hugging Face Transformers (160+ ⭐). Making AI faster + cheaper

nyc

github.com/keyboardAnt

Joined December 2017

Pinned
Nadav Timor
@NadavTimor
Jul 18, 2025
Humbled that @TheRegister covered our oral presentation at #icml25! ⚡️
The Register
@TheRegister
Jul 17, 2025
Boffins detail new algorithms to losslessly boost AI perf by up to 2.8x dlvr.it/TLyjVv
Nadav Timor
@NadavTimor
Jul 1
reality has no reset button. that's why the next training paradigm might be dreaming. @dwarkesh_sp's new episode makes the case for training in imagination. think of teaching a robot to hold a glass. the outcome is verifiable (did it break?) but not replayable: once it shatters,
Ravid Shwartz Ziv
@ziv_ravid
Jul 1
1/ On Training in Imagination - Dwarkesh's episode has a segment on dreaming as one of the next training paradigms. The idea is that a model learns mostly inside its own, by imagining what would happen, instead of trying out for real. We have a recent paper on exactly this
Nadav Timor
@NadavTimor
Dec 24, 2025
@sgl_project released eagle3 checkpoints for sota models (incl. kimi-k2, gpt-oss, deepseek-v3.2) + the training recipe
LMSYS Org
@lmsysorg
Dec 23, 2025
Speculative decoding has shown a lot of promise, though broader adoption has taken time due to the complexity of building production-ready tooling and high-quality draft models. We’re releasing SpecBundle, a collection of large-scale EAGLE-3 draft models trained with SpecForge
Nadav Timor
@NadavTimor
Dec 17, 2025
Even w/o training, you can still use speculative decoding. No need to train a speculator per model. Our spec decoding algos for heterogeneous vocabs (open-sourced in HF Transformers; not yet in vLLM) let any off-the-shelf model serve as the speculator. ♻️ That means day-0
Red Hat AI
@RedHat_AI
Dec 16, 2025
Speculative decoding is a powerful way to improve inference performance, but in practice it has been hard to adopt. Training a unique draft model per LLM is time-consuming, and production-ready training utilities that work cleanly with vLLM have been limited. Speculators
Nadav Timor
@NadavTimor
Nov 18, 2025
Tons of high-impact opportunities! And btw, our NYC open-space inference hub is still welcoming active vLLM/SGLang contributors
Greg Brockman
@gdb
Nov 17, 2025
inference is perhaps the most valuable emerging software category. as models get smarter and more economically valuable, compute will increasingly be spent drawing samples from the models. if you'd like to work on inference at openai, reach out — [email protected]. include a

New to X?