Thomas Scialom (@thomasscialom) 's Twitter Profile
Thomas Scialom

@thomasscialom

AGI Researcher @MetaAI -- I led Llama 2, built post-training from scratch. Also Toolformer, GAIA, Llama-3.0, CodeLlama, Galactica. Now working on Agents.

ID: 942694791707545600

linkhttps://www.linkedin.com/in/tscialom/ calendar_today18-12-2017 09:55:27

1,1K Tweet

7,7K Followers

224 Following

Thomas Scialom (@thomasscialom) 's Twitter Profile Photo

In fact there is on perplexity demo a specific system prompt that amplifes over safe responses. It has been removed from other demos like HF. Perplexity Denis Yarats could we deactivate it as well by default please?

Thomas Scialom (@thomasscialom) 's Twitter Profile Photo

It did in fact. RLHF is the technology behind chatgpt and probably dalle3. To panned out on real-world problems it needed nothing more than human feedback rewards.

Thomas Scialom (@thomasscialom) 's Twitter Profile Photo

I strongly disagree. There are many paths to success, and doing a PhD is never a suboptimal choice. Both professionally and personally.

Thomas Scialom (@thomasscialom) 's Twitter Profile Photo

At the AI-pulse today I talked about -- surprise -- LLMs. There short history, a deep dive into Llama 2, the magic behind RLHF, and my vision of where of the future of the field. Thanks Scaleway for the opportunity!

At the AI-pulse today I talked about -- surprise -- LLMs. There short history, a deep dive into Llama 2, the magic behind RLHF, and my vision of where of the future of the field.
Thanks <a href="/Scaleway/">Scaleway</a> for the opportunity!
AK (@_akhaliq) 's Twitter Profile Photo

GAIA: a benchmark for General AI Assistants paper page: huggingface.co/papers/2311.12… introduce GAIA, a benchmark for General AI Assistants that, if solved, would represent a milestone in AI research. GAIA proposes real-world questions that require a set of fundamental abilities such

GAIA: a benchmark for General AI Assistants

paper page: huggingface.co/papers/2311.12…

introduce GAIA, a benchmark for General AI Assistants that, if solved, would represent a milestone in AI research. GAIA proposes real-world questions that require a set of fundamental abilities such
Thomas Scialom (@thomasscialom) 's Twitter Profile Photo

Despite being an amazing paper, chinchilla did/could not be open-source. Llama-1 has now more than 10x citations than Chinchilla.

Thomas Scialom (@thomasscialom) 's Twitter Profile Photo

Delighted to finally introduce Llama 3: The most capable openly available LLM to date. Long jouney since Llama-2, a big shoutout to the incredible team effort that made this possible, and stay tuned, we will keep building🦙 ai.meta.com/blog/meta-llam…

Delighted to finally introduce Llama 3: The most capable openly available LLM to date. Long jouney since Llama-2, a big shoutout to the incredible team effort that made this possible, and stay tuned, we will keep building🦙
ai.meta.com/blog/meta-llam…
Thomas Scialom (@thomasscialom) 's Twitter Profile Photo

We had a small party to celebrate Llama-3 yesterday in Paris! The entire LLM OSS community joined us with Hugging Face, kyutai, Google DeepMind (Gemma), cohere As someone said: better that the building remains safe, or ciao the open source for AI 😆

We had a small party to celebrate Llama-3 yesterday in Paris! The entire LLM OSS community  joined us with <a href="/huggingface/">Hugging Face</a>, <a href="/kyutai_labs/">kyutai</a>, <a href="/GoogleDeepMind/">Google DeepMind</a> (Gemma), <a href="/cohere/">cohere</a>
As someone said: better that the building remains safe, or  ciao the open source for AI 😆
Thomas Scialom (@thomasscialom) 's Twitter Profile Photo

I am at ICLR.. 🦙 Llama-3: I ll be every morning at 11am at the AI at Meta for Llama-3 QA sessions 🤖 GAIA: General AI Assistant benchmark w/ Gregoire 🔭 NOUGAT: for Scientific OCR w/ Lukas And if you are interested in post-training, rlhf, agents i m down for ☕&🍺 ICLR 2026

Thomas Scialom (@thomasscialom) 's Twitter Profile Photo

The team worked really hard to make history, voila finally the Llama-3.1 herd of models...have fun with it! * open 405B, insane 70B * 128K context length, improved reasoning & coding capabilities * detailed paper ai.meta.com/research/publi…

The team worked really hard to make history, voila finally the Llama-3.1 herd of models...have fun with it!
  * open 405B, insane 70B
  * 128K context length, improved reasoning &amp; coding capabilities
  * detailed paper ai.meta.com/research/publi…
Latent.Space (@latentspacepod) 's Twitter Profile Photo

🆕 pod with Thomas Scialom of AI at Meta! Llama 2, 3 & 4: Synthetic Data, RLHF, Agents on the path to Open Source AGI latent.space/p/llama-3 shoutouts: - Why Yann LeCun's Galactica Instruct would have solved Lucas Beyer (bl16)'s Citations Generator - Beyond Chinchilla-Optimal: 100x

Miles Brundage (@miles_brundage) 's Twitter Profile Photo

Are we failing to grasp how big Internet-scale data is/how far interpolation on it goes? Are we underappreciating how fast GPUs are or how good backprop is? Are we overestimating the difference between the stuff we do vs what animals do + they’re similar in some deep sense? Etc.

Deedy (@deedydas) 's Twitter Profile Photo

All languages covey information at a similar rate when spoken (39bits/s). Languages that are spoken faster have less information density per syllable! One of the coolest results in linguistics.

All languages covey information at a similar rate when spoken (39bits/s).

Languages that are spoken faster have less information density per syllable!

One of the coolest results in linguistics.
Thomas Scialom (@thomasscialom) 's Twitter Profile Photo

Achieving AGI inherently leads to ASI because AI capability scales directly with compute and digital infrastructure. Think of Google Search already vastly outperforming humans at search. Now imagine applying the same scale to a truly intelligent system.