@robdessi

Research Scientist at Not Diamond | Past: RS at Samaya, PhD student at UPF and Meta AI, intern at Amazon Alexa, Fondazione Bruno Kessler, Xerox Research

Joined August 2015
  • Tweets425
  • Following505
  • Followers474
41Media
Roberto Dessì retweeted
Vals AI
@ValsAIOct 14
Models are shipping faster than ever. In this past month alone, Vals tested 16 new LLMs and no single model leads on everything. That is why routers have become a core primitive in the AI stack. We partnered with @OpenRouter, @tryramp, @FireworksAI_HQ and @notdiamond_ai to share what’s missing in how routers get evaluated. Check out our latest blog:
7
8
3
75
5,636
Great work led by @LucaZh00 🎉 At ICML? Go find him to chat model merging, decoding, and all things LLM!
Luca Zhou@LucaZh00Jun 21
Are questions where your LLM fails actually harder❓ Not always. Some are just "unreached"‼️ Our paper "Hard or Just Unreached?" is accepted at the 3rd AI for Math Workshop @ ICML 2026! 🎉
1
3
8
1,432
Roberto Dessì retweeted
Huiru@huiruruApr 15
using it has made me more intentional about prompts. sometimes the answer is just: use the thing above my neck. free + open source → github.com/huiruru/ecoguilt

GitHub - huiruru/ecoguilt: Claude Code plugin that tracks environmental impact of AI coding...

Claude Code plugin that tracks environmental impact of AI coding sessions. - huiruru/ecoguilt

github.com
3
1
4
455
Given the lively debate on #moltbook, let me signal this pre-print we recently released on the need for an interactionist paradigm for studying multi-agent AI systems. We discuss 4 directions we consider crucial for studying LLM-based collectives. arxiv.org/abs/2601.10567

Generative AI collective behavior needs an interactionist paradigm

In this article, we argue that understanding the collective behavior of agents based on large language models (LLMs) is an essential area of inquiry, with important implications in terms of risks...

arxiv.org
1
4
1
10
697
Roberto Dessì retweeted
Tomas Hernando Kofman
@tomas_hkJan 20
Today we are releasing a new Prompt Optimization algorithm with significant accuracy and efficiency gains, generally available now through our API.
5
14
6
37
5,851
If you’ve ever spent time updating a prompt, you know the struggle. The Not Diamond crew just dropped something that'll level up your app/product/model and save you a ton of time. Huge shoutout to Tomás and the team!
Tomas Hernando Kofman
@tomas_hk20 May 2025
Today we’re launching Prompt Adaptation, a state-of-the-art agentic system that automatically adapts prompts across LLMs. Prompt Adaptation outperforms all other methods and significantly improves accuracy over manual prompt engineering, saving you thousands of hours per year.
1
1
8
595
✨ Meet #ResiDual, a novel perspective on the alignment of multimodal latent spaces! Think of it as a spectral "panning for gold" along the residual stream. It improves text-image alignment by simply amplifying task-related directions! 🌌🔍 arxiv.org/abs/2411.00246 [1/6]
2
11
30
3,073
I'm not a Queen 🎙️ fan and I'm not a fan of prisons 👎, but I'm a fan of interdisciplinary LLM projects that are not about SOTA-chasing 🤖 🎉 Check our recent preprint led by the great @CampedelliGian
Gian Maria Campedelli@CampedelliGian10 Oct 2024
💭Remember the Stanford Prison Experiment? 📣“I Want to Break Free! Anti-Social Behavior and Persuasion Ability of LLMs in Multi-Agent Settings with Social Hierarchy”, our new pre-print, is out! W/@penzo_nicolo, @itakello, @robdessi, @m_guerini, @brulepri, @stjaco Check this🧵
15
810
I believe that cracking model routing is *the* next big challenge with LLMs. @tomas_hk and the team have just released an amazing model for this. If you are developing any advance service on top of an LLM you probably need Not Diamond and should check it out
Tomas Hernando Kofman
@tomas_hk30 Jul 2024
Today we're releasing Not Diamond… The world’s most powerful AI model router. Not Diamond maximizes LLM output quality by automatically recommending the best LLM on every request at lower cost and latency. And it takes <5m to set up. Watch this to see how to start using it:
1
5
14
1,559
Roberto Dessì retweeted
AS Roma
@OfficialASRoma6 Apr 2024
✊
77
932
145
10,860
260,819
Roberto Dessì retweeted
AS Roma
@OfficialASRoma6 Apr 2024
🟨🟥🟨🟥
44
663
104
4,110
195,165
1/9 huggingface.co/notdiamond/no… We’re open-sourcing a lightweight preview of our router that sends queries to either GPT-3.5 or GPT-4, maximizing accuracy while drastically reducing costs and latency. Routing to Gemini, Mistral, Claude, and Llama coming soon. A few quick points:

notdiamond/notdiamond-0001 · Hugging Face

We’re on a journey to advance and democratize artificial intelligence through open source and open science.

huggingface.co
13
60
15
477
96,508
Liverpool’s reaction when they rob Roma TWICE in the 2018 Champions League semi final? “Who cares” ——————- Liverpool when they are robbed of a goal vs Tottenham in matchday 7 of a 38-game league season in 2023 “Replay the game”
Rousing The Kop@RousingTheKop3 May 2018
Reporter: Roma should have had two more penalties. Virgil van Dijk: Who cares?
402
4,308
339
19,766
3,556,568
Augmented Language Models: a Survey Grégoire Mialon, Roberto Dessi, Maria Lomeli et al.. Action editor: Yujia Li. openreview.net/forum?id=jh7w… #interpreter #alms #language
2
1
18
4,247
Today I'm (remotely 🥲) presenting our work at #CVPR2023 . It's on how multimodal LMs + RL improves captioning. We cast captioning as a discrimination game and tune a captioner using a frozen CLIP. You can find the paper here arxiv.org/abs/2304.01662

Cross-Domain Image Captioning with Discriminative Finetuning

Neural captioners are typically trained to mimic human-generated references without optimizing for any specific communication goal, leading to problems such as the generation of vague captions. In...

arxiv.org
16
850
GIANLUCA MANCINI 👋👋👋
12
58
10
733
30,705
Roberto Dessì retweeted
Andrea Santilli@teelinsan25 Mar 2023
I'm excited to introduce you Camoscio: an Italian instruction-tuned LLaMA, following Stanford Alpaca. The model should provide output of similar quality to GPT text-davinci-003 and has been finetuned by translating the Alpaca dataset to Italian. github.com/teelinsan/camosci… 1/3

GitHub - teelinsan/camoscio: Camoscio: An Italian instruction-tuned language model based on LLaMA

Camoscio: An Italian instruction-tuned language model based on LLaMA - teelinsan/camoscio

github.com
8
39
3
169
71,176
Investigating prompt robustness across LMs in our upcoming ICLR paper, great work led by @carraznathanael!
Nathanael Carraz@carraznathanael22 Feb 2023
We study the generalization of automatically-induced prompts across language models arxiv.org/abs/2302.09865. #ICLR2023
14
1,321
Load more