Vicente Reyes

👨‍💻 software/data engineer 😎 science&tech enthusiast

[Vicente Reyes @vreyespue]

@Bentham2024 @simongerman600 this is the best insight 🙌 skewed time zones should be taken into account, note that Spain should be in London TZ but it is for historical reasons in Berlin TZ

[Vicente Reyes @vreyespue]

great post! 😎🙌 exactly to the point 🎯

[Geoffrey Litt @geoffreylitt]

Hot take: I think it's still important to understand the code that our agents write!

In this mega thread (based on my AIE talk today), I will explain why that's the case, and show some ideas for how to efficiently understand code. Alright, let's dive in. 1/

[Vicente Reyes @vreyespue]

I made this indie retro game, a small homage to places and things I like 😎 https://t.co/i0yB4BKXyP 😎 in principle you should be able to play this short platformer (3 levels) in any browser and any device, it is made simply with html, plain js and love ❤️🎮 enjoy!

[Vicente Reyes @vreyespue]

agent skills is a messy term used for prompt-based instructions, tool wrappers and workflow compositions 🤔 many might be a transitional step towards something where agents have internalized the skills, have durable memory, can access tools natively, and instructions can be vague

[OSS Insight @OSSInsight]

Agent skills got 250K GitHub stars in 10 weeks. But they're not the endgame — they're a transitional layer.

We analyzed 57K+ AGENTS.md repos and found skills are just 1/4 of the agent stack.

The full breakdown 👇
https://t.co/mDnCzvz1Eh

[Vicente Reyes @vreyespue]

@nic_o_martin @huggingface @nic_o_martin I created a PR with a simple exponential retry schedule, it seems to be working more smoothly 😎https://t.co/AZIJcqHViL

[Vicente Reyes @vreyespue]

very interesting to see #Gemma4 running locally in the browser 🚀 @nic_o_martin @huggingface great extension helping for learning about #LLMs😎 sometimes flaky and losing connection to tabs, maybe some bug fixing still ongoing ✅ thanks for sharing! 🙌 https://t.co/HBTSxuCSzw

[🤷 Nico Martin @nicodotdev]

I built an AI agent that lives inside your browser🎉
tabs, history, page content, tool calling. All local, no server powered by #Gemma4
But the interesting part isn't that it works. It's how you architect Transformers.js inside a Chrome extension to make it work. 🧵

[Vicente Reyes @vreyespue]

@Swizec agree 🙌 it's about 17s for 100m 🤔 any normal person would need a decent sprint to make that 🏃 maintaining that pace for 42km is simply out of this world 🌎 absolutely amazing

[Vicente Reyes @vreyespue]

amazing! two german AI startups raised $150M+ in Q1 2026 😎🚀🙌 @parloa_ai @quantumdrones

[Ivan Landabaso @IvanLandabaso]

A list of 37 European AI startups that raised $10M+ in Q1 2026.

All likely hiring:

[Vicente Reyes @vreyespue]

almost 2 years later 🤔 overconfidence is still a major challenge in newer LLM based systems, which overestimate the probability that their answer is correct 20%-60% ☝️ humans have accuracy similar to the more advanced LLMs, but far lower overconfidence https://t.co/pNAuKi9H6P

[Philipp Schoenegger @SchoeneggerPhil]

@metaculus However, we also find that the LLM crowd is poorly calibrated and shows substantial overconfidence.

In Study 2, we also find that human forecasts can improve LLM predictions and reduce their uncertainty, though aggregating machine and human responses exceeds these improvements.

[Vicente Reyes @vreyespue]

@TFConsult it was a great pleasure to have you all here at #REWE Digital 😎🙌🚀

[Vicente Reyes @vreyespue]

@r0_lang @Lambda_World @lenguyenthanh it was really nice to meet you all in my hometown #Cadiz 🎉♥️😎🙌

[Vicente Reyes @vreyespue]

I enjoyed the panel discussion at @Lambda_World 😎 many interesting ideas! 🤔 in the future we could let AI coding systems apply automated theorem proving to generated programs 💡 making them very robust and performant ☝️ functional programming might be more relevant than ever!🚀

[Lambda World @Lambda_World]

We close out @Lambda_World this year with a panel discussion on
"Does FP still matter in the era of AI?"
with Raul Raja, Jack Higgs, Isaias Bartelborth & Anurag Mendhekar

What do you think?
check out their response to the question here
https://t.co/UeNeS5Ul38
#lambdaworld #functional_programming #debate

[Vicente Reyes @vreyespue]

@fchollet Totally agree 💯😎🙌

[Vicente Reyes @vreyespue]

@karpathy it is very difficult to define AGI 🤔 @GoogleDeepMind proposes six levels: https://t.co/JJOFCAdzvh ☝️ they claim that only level 1 (emerging AGI) has been achieved

[Vicente Reyes @vreyespue]

@Swizec They provide MCPs instead of classic APIs? 🤔

[Vicente Reyes @vreyespue]

@Xebia_Global we also have lots of good experience using GitHub workflows for automating backend/frontend pipelines 😎 including for instance integration tests 👍

[Vicente Reyes @vreyespue]

I had the pleasure to attend live this great talk about tokenizer-free transformer models 🚀 increasing performance using two abstraction layers: a word-level backbone, and byte-level encoder/decoder modules 🙌 amazing research and results! @Aleph__Alpha @CologneAIML

[Pablo Iyu Guerrero @pabloiyu]

First high-performance inference for hierarchical byte models.
@LukasBluebaum and I developed batched inference for tokenizer-free HAT (Hierarchical Autoregressive Transformers) models, developed by @Aleph__Alpha Research. In some settings, we outcompete the baseline Llama.🧵

[Vicente Reyes @vreyespue]

@VoylesGames @Esoteric_Though @VraserX best answer 🙌

[Vicente Reyes @vreyespue]

@omarsar0 very useful 🙌

[Vicente Reyes @vreyespue]

great talk by @amateurhuman at @aiDotEngineer 🙌 I felt very much identified by this sort of objective ideas 😎 not overhyped, not rudely criticised, just some fair facts 👍
https://t.co/HBRGAVDme6