Ely Hahami @ NeurIPS
@ElyHahami
Followers
625
Following
2K
Media
16
Statuses
252
@Harvard Math / CS, deep learning
Joined November 2023
The Internet Contraceptive. In 2 hours, I built what people are calling protection against the algorithm™ A chrome extension using LLMs to filter out the slop on your feed. Soon I'll add more features, go multimodal, and improve latency. Comment 'cherri' for the source code.
140
40
780
why is Neurips 2025 just people following Jeff Dean and Richard Sutton around…
0
0
2
when prime intellect posts, it’s not just prime intellect, it’s a swarm of all their employees. Their swarm is stronger than the big labs.
0
2
25
I'll be at Neurips dec 2-7th, reach out / DM if you want to meet!!
0
0
3
With AI, we must rethink the role of a researcher. Zechen’s vision is to fundamentally change the abstraction layer, allowing a researcher to focus on ideas and intuitions, not wasting time setting up code infrastructure. Everyone should check out Orchestra, and this is just
Before Gemini drops today, I want to share something we've been building—Orchestra, the scientific Jarvis. The Iron Man dream is actually real this time. Not sci-fi. Just... talking to research agents about your ideas and they will handle everything else for you. Last week I
1
0
4
dope, ZZ!
So vibe coding is cool, but what about vibe all the way down to do vibe film-making using diffusion models & coding agents? So in the past few days, I fed the veo-3 docs to claude code, and brainstormed with it my idea of scenes and plots for the film, the result is here. This is
1
0
2
it appears xAI will be the first frontier lab to have a diffusion language model
0
0
1
an underrated part of working on LLMs is that the training data is actually interesting to read sometimes 😂
0
0
3
people care about "legacy." Then, they should get their work good enough to be on Grokipedia, as they will be etched by laser into the solar system, forever! @elonmusk
1
0
3
Overall, such knowledge injection by fine-tuning (ie weight changes!) will serve as a cornerstone for a self-evolving AI in the era of experience, and we are excited about tuning with a mask recovery objective! Github repo:
github.com
official repo for Closing the Data-Efficiency Gap Between Autoregressive and Masked Diffusion LLMs - xup5/masked_arLLM
0
0
5
The mask recovery objective uses a more flexible factorization, which can be seen as an implicit data augmentation, and leverages how diffusion language models solve the reversal curse!
1
0
4
Our "masked autoregressive LLM" tuning performs better than regular autoregressive tuning, on both forward-style and reversal-curse-style (backward) questions. Eg. Biography dataset: autoregressive QA accuracy: 12% (forward), 0.02% (backward) ours: 97% forward, 60% backward!
1
0
2
In our new paper, we show: 1) Diffusion language models are MORE data-efficient in post-training than AR models. They don't need data augmentation and can learn from a SINGLE sample! 2) Inspired by this, we propose Diffusion-style Autoregressive Fine-tuning Results are good!
1
0
1
New paper! https://t.co/5JNKMno5LD The problem: LLM weights need to update to new information. Traditional fine-tuning often fails at "reversal curse" style questions: if model learns "A→B", it struggles with "B→A". Then, data augmentation is needed.
arxiv.org
Despite autoregressive large language models (arLLMs) being the current dominant paradigm in language modeling, effectively updating these models to incorporate new factual knowledge still remains...
1
0
2
How do you inject new knowledge into LLMs? Our solution: *Diffusion-style Autoregressive Tuning* → Put a MASKED version of training sample in-context → Make the target the ORIGINAL (unmasked) sample → Keep it autoregressive! Our method improves QA acc by 65%! (Wiki Dataset)
3
2
12
me when I read another one of these "memory for AI agents" papers, command f "weights", and nothing pops up 🤔
0
0
7
The sukkah teaches that what protects you isn’t how solid your roof is, but how open your heart is. @ElyHahami
0
1
7
LoRAs, personalized for everyone, everywhere :)
Introducing Tinker: a flexible API for fine-tuning language models. Write training loops in Python on your laptop; we'll run them on distributed GPUs. Private beta starts today. We can't wait to see what researchers and developers build with cutting-edge open models!
0
0
6