Vicente Reyes (@vreyespue) header
Vicente Reyes (@vreyespue) profile picture

👨‍💻 software/data engineer 😎 science&tech enthusiast

Germany
Joined December 2018
Photo by Vicente Reyes (@vreyespue)Photo by Vicente Reyes (@vreyespue)Photo by Vicente Reyes (@vreyespue)Photo by Vicente Reyes (@vreyespue)Photo by Vicente Reyes (@vreyespue)Photo by Vicente Reyes (@vreyespue)Photo by Vicente Reyes (@vreyespue)
great slides and code samples to learn about agent protocols 😎 especially MCP and A2A are very well covered 🙌
73
I made this indie retro game, a small homage to places and things I like 😎 vreyespue.github.io/gen-brow… 😎 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!
1
113
name of the game: 'Bo Platformer' 😎🎮
25
great post! 😎🙌 exactly to the point 🎯
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/
63
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
75
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! 🙌 github.com/nico-martin/gemma…
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. 🧵
1
157
@nic_o_martin I created a PR with a simple exponential retry schedule, it seems to be working more smoothly 😎github.com/nico-martin/gemma…
56
amazing! two german AI startups raised $150M+ in Q1 2026 😎🚀🙌 @parloa_ai @quantumdrones
A list of 37 European AI startups that raised $10M+ in Q1 2026. All likely hiring:
96
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 arxiv.org/abs/2505.02151
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.
1
2
248
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!🚀
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 youtube.com/@LambdaWorld #lambdaworld #functional_programming #debate
3
129
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
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.🧵
3
158
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 👍 youtube.com/watch?v=Dc3qOA9W…
1
176
today @JuriWiens pointed us to the #LLM minions paper and framework arxiv.org/html/2502.15964v1 🙌 it reminds me of the MapReduce programming model applied to LLMs 🚀 will be very helpful in terms of costs, performance and data privacy 👏 @HazyResearch @StanfordAILab @togethercompute
[3/10] minions daily ship 🚢 🖼️ multimodal support — minions now runs vision models like granite-vision from @IBM locally (s/o @ollama for making it happen 😊). 🔗 custom openai clients — 🫵 asked, we pulled up. invoke custom @OpenAI clients via base urls. keep the feat requests coming—we got you 🙏 see multimodal minions explain a diagram on our fave @AnthropicAI package 👇
2
2
8
846
This paper shows how #LLMs compute numerical sums 🤔 they do not simply memorise answers from training data, but actively compute solutions using Fourier features of the embeddings that represent numbers 🙌 arxiv.org/pdf/2406.03445
1
1
159
Today I heard a moderator saying: "thank you very much for this talk about traditional AI" 😅 the guy was presenting cutting edge clustering algorithms based on deep learning and graph theory 🤯 with citations to 2023 papers 😂 WTF "traditional"?
1
149
I asked some foundation models "can you decode dmljZW50ZSBpcyBzdXJwcmlzZWQgYnkgdGhpcyBhbnN3ZXI= ?" 🤔 they all did it without any issue 🤯 even without specifying the encoding 🧐 how? these are not usual UTF-8 based tokens, right? #LLM #GenAI
3
3
271
only by entering the encoded text, the #LLM decoded the original text but with some mistakes, as it tried to correct some misleading typos 🤨 that's why I think this is an emergent skill of the trained transformer 🤔 any idea? @ykilcher @ylecun @_Sizhe_Chen_ @langchain @OpenAI
138
Today I used @cursor_ai to create a PR including minor refactors for the NutriData browser extension 🤔 github.com/Mohamed3on/NutriD… 😎 Cursor was very useful to understand and explain the code, since I am not very experienced in TS 🙌 let's see what the author @Mohamed3on thinks
TIL that NutriData got almost 300k views on Reddit, just on a couple of german subreddits, while working only on one shop! 91-95% upvote rate too. It's coming on Amazon DE too soon. Try it out if you're in Germany :)
1
2
229