Diego Calanzone
@diegocalanzone
Followers
247
Following
4K
Media
69
Statuses
660
« artificia docuit fames » // phd at @Mila_Quebec, intelligence by agency + deep learning for science // AI grad @UniTrento
127.0.0.1
Joined April 2015
🫡 so long old friend, thanks for the journey from day 1 of masters and zero research, to PhD year 1. You’ve seen evolving decision trees, quantized models with handcrafted losses, chains of latex syntax errors. ps: not thanks for the nvidia driver crashes.
2
0
15
0
12
0
unfortunately, this is not playing Pokémon. The agent chats with the audience to get hints, plenty of tokens of reasoning before one major action, hence “steps”, no actual exploration, but handcrafted game APIs to recall pretraining knowledge. Good APIs, but IMO no agency.
0
0
0
RT @RL_Conference: Ending with our last RLC oral, @RichardSSutton with "The Oak Architecture: A Vision of SuperIntelligence from Experienc….
0
8
0
!!!.
Could we tell if gpt-oss was memorizing its training data? I.e., points where it’s reasoning vs reciting? We took a quick look at the curvature of the loss landscape of the 20B model to understand memorization and what’s happening internally during reasoning
0
0
0
RT @giffmana: Amazing! Truly open review, through which we all gained more insights, i love it!. Result: in multi epoch setting, making AR….
0
36
0
RT @sporadicalia: just remembered that time Noam Shazeer dropped the hardest line ever written in an ML paper
0
622
0
And I’ve been mentioning sinks for months.
Summary of GPT-OSS architectural innovations:. 1. sliding window attention (ref: .2. mixture of experts (ref: .3. RoPE w/ Yarn (ref: .4. attention sinks (ref: streaming llm .
0
0
1
> you get H100s.> nodes are isolated from internet.> you decide to copy envs and llama 4 weights over SSH 🫠. I’m wondering how accessible compute actually is for researchers without a CS background.
0
0
1
with pretty viz!!!.
Excited to share Flow Matching Policy Gradients: expressive RL policies trained from rewards using flow matching. It’s an easy, drop-in replacement for Gaussian PPO on control tasks.
0
0
1
RT @lavoiems: 🧵 Everyone is chasing new diffusion models—but what about the representations they model from?.We introduce Discrete Latent C….
0
45
0
good points. Inherent constraints of the competition are part of the outcome performance and they shall be considered. Though Terence didn’t mention that human competitors are running on a negligible fraction of energy ;).
0
0
0
RT @PontiEdoardo: We blend imitation (SFT) and exploration (RLVR) in post-training with a simple idea:. Sample a prefix of an SFT demonstra….
0
7
0
partially brewed at Mila.
Google DeepMind just dropped this new LLM model architecture called Mixture-of-Recursions. It gets 2x inference speed, reduced training FLOPs and ~50% reduced KV cache memory. Really interesting read. Has potential to be a Transformers killer.
0
0
2
RT @steveazzolin: This is an issue on multiple levels, and authors using those "shortcuts"👀 are equally responsible for this unethical beha….
0
2
0
A comprehensive article on ways to Hierarchical RL!.
As AI agents face increasingly long and complex tasks, decomposing them into subtasks becomes increasingly appealing. But how do we discover such temporal structure?. Hierarchical RL provides a natural formalism-yet many questions remain open. Here's our overview of the field🧵
0
0
2
RT @BlancheMinerva: A good warning lesson on using AIs to write papers: this alleged response to the (dubious) "Illusion of Thinking" paper….
arxiv.org
Shojaee et al. (2025) report that Large Reasoning Models (LRMs) exhibit "accuracy collapse" on planning puzzles beyond certain complexity thresholds. We demonstrate that their findings primarily...
0
47
0
RT @BlancheMinerva: Two years in the making, we finally have 8 TB of openly licensed data with document-level metadata for authorship attri….
0
69
0
