Pinned Tweet
*Alice's book got a (minor) upgrade!*
Thanks to the dozens of people who gave feedback, now with 1000% less typos and errors, a novel set of Colab lab sessions, and a brand-new CC-BY-SA license. 🙃
sscardapane.it/alice-book/
Thanks to the dozens of people who gave feedback, now with 1000% less typos and errors, a novel set of Colab lab sessions, and a brand-new CC-BY-SA license. 🙃
sscardapane.it/alice-book/
9
196
1,335
60,943
Simone Scardapane retweeted
Our book "The Elements of Differentiable Programming" can now be read directly in the browser:
diffprog.github.io
diffprog.github.io
11
92
788
58,765
Hey Twitter, long time no see!
(Like everyone?) I am wasting my time building a small personal agentic system; as all my friends have asked me to stop talking about it, here's a Claude write-up about the current system if anyone here is interested:
sscardapane.it/projects/how-m…
(Like everyone?) I am wasting my time building a small personal agentic system; as all my friends have asked me to stop talking about it, here's a Claude write-up about the current system if anyone here is interested:
sscardapane.it/projects/how-m…
0
0
21
1,499
Simone Scardapane retweeted
Check our last paper: Anomaly detection at the HL-LHC in learned latent spaces: we learn a representation of SM collisions, flag anomalies with an autoencoder, then explain locally what makes them anomalous.
@s_scardapane @Stefano
@Thea_kaa
@s_scardapane @Stefano
@Thea_kaa
5
1
4
288
*Retrieval-Centric Deep Learning in
Growing Nonparametric NNs*
by @mtavitschlegel @oswaldjoh et al.
Really cool work pushing forward the "dual view" of a linear layer that turns it into an equivalent linear attention layer over the training points.
arxiv.org/abs/2610.03858
Growing Nonparametric NNs*
by @mtavitschlegel @oswaldjoh et al.
Really cool work pushing forward the "dual view" of a linear layer that turns it into an equivalent linear attention layer over the training points.
arxiv.org/abs/2610.03858
2
5
19
1,423
Simone Scardapane retweeted
New paper from Japan! Our work “Authorship Attribution and Aesthetic Evaluation of AI Poetry: A Case Study with Haiku” is now on arXiv and accepted as a full paper at SCIS&ISIS 2026 🎉
Work done by Livia Oddi with me, @s_scardapane and Toru Sugimoto
arxiv.org/abs/2609.15511…
Work done by Livia Oddi with me, @s_scardapane and Toru Sugimoto
arxiv.org/abs/2609.15511…
0
2
7
398
*Sliding-window beats linear attention*
by @jm_alexia @RheaSukthanker @CameronPashmina @Emy_Aze
Static attention masks (local + sinks) can outperform replacing the attention layer with a linear attention variant in terms of accuracy and efficiency.
arxiv.org/abs/2608.28444
by @jm_alexia @RheaSukthanker @CameronPashmina @Emy_Aze
Static attention masks (local + sinks) can outperform replacing the attention layer with a linear attention variant in terms of accuracy and efficiency.
arxiv.org/abs/2608.28444
0
3
29
1,455
As a programmer I can imagine a world where all code is written by AI and old style programmers wither away, but as a scientist I have zero idea what a world where "AI does all the math" even looks like.
4
0
43
4,207
**Recirculation**
by @mc_mozer @ShoaibASiddiqui @savvyRL et al.
Very simple test-time intervention where the activation of a late layer is "leaked" into an early layer to steer it towards a more effective answer.
arxiv.org/abs/2608.17981
by @mc_mozer @ShoaibASiddiqui @savvyRL et al.
Very simple test-time intervention where the activation of a late layer is "leaked" into an early layer to steer it towards a more effective answer.
arxiv.org/abs/2608.17981
0
1
7
901
I really can't help but feel we are destroying the confidence of generations of mathematicians partly for financial gain, and we'll regret this more than scaling test time compute too much.
9
7
106
4,181
Simone Scardapane retweeted
Things you can do with @NVIDIAAI 's Object Oriented Agents (in simple Python🐍)
1. Subclass an Agent !
The PhD Agent inherits from the Researcher Agent class🎓
1. Subclass an Agent !
The PhD Agent inherits from the Researcher Agent class🎓
2
5
21
1,530
Simone Scardapane retweeted
Un Bilancio positivo ma solo per #ATAC
Il servizio sotto l’attuale Amministrazione è stato INFERIORE in media a quella precedente
Le auto private a #Roma sono aumentate del +11% in soli 4 anni
Al link sotto un'analisi di quanto svolto da ATAC in base alla tipologia di servizio
Il servizio sotto l’attuale Amministrazione è stato INFERIORE in media a quella precedente
Le auto private a #Roma sono aumentate del +11% in soli 4 anni
Al link sotto un'analisi di quanto svolto da ATAC in base alla tipologia di servizio
2
3
17
844
*BERT is just a Single Text Diffusion Step*
by @nathanrs
Easiest way to understand diffusion LMs: take a BERT model and fine-tune it to work with multiple masking rates - you now have a working DLM. 😄
nathan.rs/posts/roberta-…
by @nathanrs
Easiest way to understand diffusion LMs: take a BERT model and fine-tune it to work with multiple masking rates - you now have a working DLM. 😄
nathan.rs/posts/roberta-…
0
5
40
1,950
*Local MoE: Essentially Free Test-Time Training via Model Merging*
by @jonashubotter et al.
They train thousands of LoRAs on very small data clusters and combine them at inference-time dynamically via weight merging.
arxiv.org/abs/2505.14136
by @jonashubotter et al.
They train thousands of LoRAs on very small data clusters and combine them at inference-time dynamically via weight merging.
arxiv.org/abs/2505.14136
0
18
113
7,062
*Language Models are Injective and Hence Invertible*
by @GiorgosNik02 @tommaso_mncttn @DonatoCrisosto1 @teelinsan @EmanueleRodola
They show that decoder transformers are injective, meaning you can easily recover an input sequence from the hidden states.
arxiv.org/abs/2510.15511
by @GiorgosNik02 @tommaso_mncttn @DonatoCrisosto1 @teelinsan @EmanueleRodola
They show that decoder transformers are injective, meaning you can easily recover an input sequence from the hidden states.
arxiv.org/abs/2510.15511
8
57
431
20,387
Simone Scardapane retweeted
Introducing NVIDIA Object Oriented Agents (NOOA), probably the most Pythonic way to build AI agents 🐍
In NOOA, an agent is just a Python class.
If you know Python, you already know NOOA. And yes, building agents can be this simple ↓
In NOOA, an agent is just a Python class.
If you know Python, you already know NOOA. And yes, building agents can be this simple ↓
7
15
45
3,267
*Encode, Think, Decode: Scaling test-time reasoning with recursive latent thoughts*
by @yeskendir_k @nicola_cancedda @AldoLipani
They modify the mid-training phase of an LLM with looping over the middle layers, which are assumed crucial for reasoning.
arxiv.org/abs/2510.07358
by @yeskendir_k @nicola_cancedda @AldoLipani
They modify the mid-training phase of an LLM with looping over the middle layers, which are assumed crucial for reasoning.
arxiv.org/abs/2510.07358
2
23
167
14,630
Simone Scardapane retweeted
Subliminal Clocks
Diffusion language models secretly encode their denoising timestep in internal activations, and we can steer them by manipulating this latent signal.
Diffusion language models secretly encode their denoising timestep in internal activations, and we can steer them by manipulating this latent signal.
2
17
125
7,699
Work lead by the super talented @MaximoRulli - who is looking for PhD positions, by the way! 👇
Diffusion language models develop 𝐬𝐮𝐛𝐥𝐢𝐦𝐢𝐧𝐚𝐥⏱️
Despite receiving no explicit timestep, mask-based DLMs encode denoising progress in a low-dimensional latent structure. We can steer this clock predictably and watch the model correct perturbations across layers.
(1/6)
0
2
12
1,862
Twitter friends, after @alfcnz introduced me to "second brains" I spent a lot of time trying to turn my Obsidian vault into an always-on personal agent à-la-Hermes. Anyone interested in reading something about this or working in a similar vein?
6
3
18
1,626
*A History of Large Language Models*
by Gregory Gundersen
Very well written blog post on the evolution of language models from simple n-grams to current LLMs via an overview of the major papers.
gregorygundersen.com/blog/2025/10/0…
by Gregory Gundersen
Very well written blog post on the evolution of language models from simple n-grams to current LLMs via an overview of the major papers.
gregorygundersen.com/blog/2025/10/0…
1
41
294
17,495