La Era Atómica

🎧🇪🇸 by a 🙍🏽‍♀️🇨🇴-n historian

open.spotify.com/episode/2Ufd...

👏🏼💪🏼🇨🇴 Grande!

Source: X/Twitter

x.com/iclr_conf/st...

@iclr-conf.bsky.social "ICLR 2026 is almost there! We have 6 exciting keynotes covering a range of areas from machine learning to robotics, neuroscience and AI for science:

Maja Matarić, Max Welling, Percy Liang, Katie Bouman, Karen Adolph, Pablo Arbeláez

See you all soon! #ICLR2026"

The AI for Peace Workshop schedule is live! Full program: aiforpeaceworkshop.github.io/schedule/

Source: X/Twitter

x.com/egrefen/stat...

@egrefen.bsky.social "When looking at deep learning profiles, one of the most obvious tells between a mediocre and a great candidate is whether they prefer chocolate or vanilla ice cream."

What are your all time favorite textbooks? Here are a few of mine.

@ericxw.bsky.social "When looking at faculty profiles, one of the most obvious tells between a mediocre and great candidate is whether they use PowerPoint or Google Slides."

Echoing @fchollet (Keras' creator) take between Jax & PyTorch

Heading to @cvprconference.bsky.social?

Please reserve your hotel room within the official CVPR hotel block as soon as possible to secure the conference rate.

More info: cvpr.thecvf.com/Conferences/...

NeurIPS 2026 is soliciting competition proposals on topics of interest to the NeurIPS community.

Read the call for competitions for more information neurips.cc/Conferences/...

Strevens knowledge machine in action!

Unifying Popper's penalizing Pendulum; and Kuhn's Paradigmatic Pendulum aligner.

www.strevens.org/scientia/

With more people embracing AI agents, students and others will outsource part or all of their writing to LLMs.

Golden rule: do not use peers as unpaid labelers for your automation.

If AI touched it, proofread it before sharing, and leave edit notes where needed.

Cuál sera la siguiente simplificación?
(Una fuera de politica) NO fuimos a la luna porque no regresamos. Artemis es un teatrillo.

Política y diplomacia van de la mano. Hoy se pacta, mañana se construye, pasado mañana se compite.

NeurIPS encourages and benefits from a diversity of papers and ideas, which can be developed in many different ways. This year, Main Track submissions can select a Contribution Type, including General, Theory, Use-Inspired, Concept & Feasibility, and Negative Results.

can't today, baby 🤒

Is the who more important than what?

Source: X/Twitter

x.com/juliekallini...

8/ TL;DR: lots of evidence that tokenization choices can improve model performance without pre-training! Maybe the conclusion is that the glory days of tokenizer research are not over? 💁🏻‍♀️
"

7/ Even if Anthropic added new vocabulary items, that still wouldn't imply new pre-training. Vocabulary expansion with light fine-tuning is possible. But this is the opposite of the effect they've reported, since that would improve compression rather than reduce it.

6/ It's possible that Anthropic found more domains in which non-canonical tokenizations help performance. You can almost view this as a form of test-time scaling, since you're forcing more compute per string via finer-grained tokenizations.

Tokenization counts: the impact of tokenization on arithmetic in frontier LLMs Tokenization, the division of input text into input tokens, is an often overlooked aspect of the large language model (LLM) pipeline and could be the source of useful or harmful inductive biases. Hist...

5/ Singh & Strouse (2024) had earlier shown that the direction of digit tokenization (right-to-left or left-to-right) is important for arithmetic tasks for LLMs at that time, and that the effect persists across model scales.

arxiv.org/abs/2402.14903

Broken Tokens? Your Language Model can Secretly Handle Non-Canonical Tokenizations Modern tokenizers employ deterministic algorithms to map text into a single "canonical" token sequence, yet the same string can be encoded as many non-canonical tokenizations using the tokenizer vocab...

4/ Zheng et al. 2025 ("Broken Tokens") showed that non-canonical tokenizations at inference time can improve instruction-tuned model performance. E.g., character-level tokenization improves string manipulation and code understanding tasks.

arxiv.org/abs/2506.19004

3/ More importantly, changing tokenization can improve capabilities without changing the base model, and without any additional training at all. We have direct evidence of this from several tokenization studies.

2/ Anthropic says the updated tokenizer may increase sequence lengths by roughly 1–1.35×. That behavior doesn't even require training a new tokenizer from scratch. E.g., you could force finer segmentation on certain text using shorter tokens within the original vocabulary.

@juliekallini.bsky.social: "

1/ "New tokenizer" does not imply "new base model," and "new base model" is not the simplest explanation. There are much simpler explanations that fit Anthropic's public description of Opus 4.7 equally well.

Applies for vibe-coding 😂

Sometimes advisor/PM/coach helps steer the boat, not as the captain, but as instrument.

Other times is to bother the crew:
- remember the commitments,
- remind that DEI, broad participation, etc. are not grant-writing decorations,
- ask the uncomfortable questions:
why this, why now, and for whom?

I thought I would do a thread, but honestly the post is so good: kyutai.org/blog/2026-04...

It explains "One View Is Enough! Monocular Training for In-the-Wild Novel View Generation" arxiv.org/abs/2603.23488 done in colab with the smart people at kyutai

Source: X/Twitter

x.com/csprofkgd/st...