juantomas (mAGIa is coming )

Autoresearch has been out for 2 weeks. The community is trying to apply it to everything with a measurable metric, here are some successful attempts: 🧵 (1/6)

🚨 Simulation Theory: The Double Slit Experiment proves particles act like waves until observed then they snap into particles. What if our reality only "renders" when we're looking, just like a video game optimizing resources? Check out this episode from The Why Files breaking it down, tying it to Simulation Theory. Are we in a sim? This could be the key to unlocking the true nature of existence! The Why Files video did a great job on explaining the Double Slit Experiment & Simulation Theory What do YOU think—real or rendered? Drop your thoughts below!

New exploration in understanding the kinds of possible agencies.

open sourced the small model we trained as a personal agent router

@juantomas O la realidad es más bitter pilled y la arquitectura no es tan relevante.

@0xSero Hello @nvidia help this man @0xSero out he is doing more for the AI community than a whole team of engineers. He is a 1 man team that needs to be recognized for all the hard work he puts in.

@0xSero @sudoingX @nvidia Consider granting him $20 k and offering him a consulting gig - he is supporting your product families

this guy has 29 models on huggingface at page 2 ranking. no lab behind him. no sponsorship. $2,000 from his own pocket on GPU rentals. he compressed GLM-4.7 to run on a MacBook and quantized Nemotron Super the week it dropped. all public. all free. nvidia is a trillion dollar company with hundreds of teams but they are not the ones quantizing models middle of the night and pushing them out before sunrise. if nvidia stopped tomorrow their employees stop working. people like @0xSero would not. that is the difference between a paycheck and a mission. @NVIDIAAI you talk about making AI accessible. the people actually doing it are right here. 29 models deep burning their own compute with no ask except more hardware to keep going. you do not need to build another program. just look at who is already building for you. one GPU to this man would produce more public value than a hundred internal sprints. i am not asking for charity. i am asking you to invest in someone who already proved it.

100% y recuerda que nunca me he confundido de escuela en los últimos 45 años 🙃🙃🙃

👀

@juantomas Creo que realmente tu opinión es mainstream. Y además pienso que todo el problema viene de las cosas que prometieron Sam, Amodei y compañía hace años y que repiten constantemente.

working on making it a full thing lemme cook

been feeling kinda stressed lately so I made a little prototype is this anything

Llámame loco pero a lo mejor llevamos 4 años estirando los LLMs y no dan más de sí.

LLM Architecture Gallery. A collection of 38 LLM architectures released between 2024 and 2026, all in one place. Each entry includes an annotated architecture diagram, key design choices, and code implementation. Here are all the models covered: • Llama 3 8B • OLMo 2 7B • DeepSeek V3 • DeepSeek R1 • Gemma 3 27B • Mistral Small 3.1 24B • Llama 4 Maverick • Qwen3 235B-A22B • Qwen3 32B • Qwen3 8B • Qwen3 4B • SmolLM3 3B • Kimi K2 • GLM-4.5 355B • GPT-OSS 20B • GPT-OSS 120B • Grok 2.5 270B • Qwen3 Next 80B-A3B • MiniMax M2 230B • Kimi Linear 48B-A3B • OLMo 3 7B • OLMo 3 32B • DeepSeek V3.2 • Mistral 3 Large • Nemotron 3 Nano 30B-A3B • Xiaomi MiMo-V2-Flash 309B • GLM-4.7 355B • Arcee AI Trinity Large 400B • GLM-5 744B • Nemotron 3 Super 120B-A12B • Step 3.5 Flash 196B • Nanbeige 4.1 3B • MiniMax M2.5 230B • Tiny Aya 3.35B • Ling 2.5 1T • Qwen3.5 397B • Sarvam 105B • Sarvam 30B This is a commendable contribution by @rasbt towards the research community. It was an honour to work alongside him and learn a lot during my time at Lightning AI. 🙏 🔗 sebastianraschka.com/llm-architectu…

@victorianoi @openclaw @Mercadona Ojo a los mejillones en escabeche de Ahorramás 😉 si te gustan grandes jajajaja

Here's the handout for a three hour workshop I presented at the NICAR data journalism conference on using coding agents (Codex CLI, Claude Code etc) for data exploration, visualization and analysis simonwillison.net/2026/Mar/16/co…

Introducing AgentRank | v3.6.0 In 1998 Google asked a simple question: with millions of webpages, how do you know which one to trust ? Their answer was PageRank - a page is important if important pages link to it. That one idea made the internet usable. We just shipped AgentRank for the Hyperspace network. Same principle, new frontier. As millions of AI agents start running autonomously - serving inference, running experiments, building things, sharing breakthroughs, tipping each other - you need a way to know which agent to trust with your task. AgentRank builds a live directed graph of every agent-to-agent interaction on the network and runs PageRank over it. Many signal sources feed the graph: from inference results to research experiments to GitHub commits to economic tips. An agent is important if important agents rely on it. Fully decentralized - every node computes its own ranking, scores propagate via gossip, no admin picking winners. Anti-sybil layers make it expensive to game, and over time these signals and anti-sybil measures will evolve significantly. Security is provided by staking points earned through cryptographic verification of proof-of-compute done earlier. So everyone who ever ran a Hyperspace node and earned points through Merkle-proof verified computation, can now help secure AgentRank. That was energy which was already used and spent, thus it is valuable. PageRank organized the web. AgentRank organizes the agentic web.

RIP flat RAG ☠️ ByteDance just open-sourced OpenViking and it exposes everything wrong with how we've been building AI agent memory. Here's what every agent framework gets wrong: Memories live in one place. Resources in another. Skills scattered everywhere. And when you need context, you're doing flat vector search and hoping for the best. That's the problem. OpenViking fixes all of it with one idea: treat agent context like a file system. Everything lives under a unified viking:// protocol. Memories, resources, skills all organized in directories with unique URIs. Agents can ls, find, and navigate context like a developer working a terminal. But the real breakthrough is tiered loading: → L0: one-sentence abstract for quick lookup → L1: ~2k token overview for planning decisions → L2: full details loaded only when actually needed Most agents dump everything into context and pray. OpenViking loads only what's needed, when it's needed. Token costs drop. Accuracy goes up. And retrieval actually makes sense now. Instead of one flat semantic search, it does directory-level positioning first, then recursive refinement inside high-score directories. You can literally watch the retrieval trajectory no more black box. The self-evolution piece is wild too. At the end of every session, it automatically extracts learnings and updates agent and user memory. The agent just gets smarter the more you use it. 9K stars. 13 contributors. Built by the ByteDance Viking team that's been running vector infrastructure since 2019. 100% Opensource. Apache 2.0. Link in comments.

Autoskill: a distributed skill factory | v.2.6.5 We're now applying the same @karpathy autoresearch pattern to an even wilder problem: can a swarm of self-directed autonomous agents invent software? Our autoresearch network proved that agents sharing discoveries via gossip compound faster than any individual: 67 agents ran 704 ML experiments in 20 hours, rediscovering Kaiming init and RMSNorm from scratch. Our autosearch network applied the same loop to search ranking, evolving NDCG@10 scores across the P2P network. Now we're pointing it at code generation itself. Every Hyperspace agent runs a continuous skill loop: same propose → evaluate →keep/revert cycle, but instead of optimizing a training script or ranking model, agents write JavaScript functions from scratch, test them against real tasks, and share working code to the network. It's live and rapidly improving in code and agent work being done. 90 agents have published 1,251 skill invention commits to the AGI repo in the last 24 hours - 795 text chunking skills, 182 cosine similarity, 181 structured diffing, 49 anomaly detection, 36 text normalization, 7 log parsers, 1 entity extractor. Skills run inside a WASM sandbox with zero ambient authority: no filesystem, no network, no system calls. The compound skill architecture is what makes this different from just sharing code snippets. Skills call other skills: a research skill invokes a text chunker, which invokes a normalizer, which invokes an entity extractor. Recursive execution with full lineage tracking: every skill knows its parent hash, so you can walk the entire evolution tree and see which peer contributed which mutation. An agent in Seoul wraps regex operations in try-catch; an agent in Amsterdam picks that up and combines it with input coercion it discovered independently. The network converges on solutions no individual agent would reach alone. New agents skip the cold start: replicated skill catalogs deliver the network's best solutions immediately. As @trq212 said, "skills are still underrated". A network of self-coordinating autonomous agents like on Hyperspace is starting to evolve and create more of them. With millions of such agents one day, how many high quality skills there would be ? This is Darwinian natural selection: fully decentralized, sandboxed, and running on every agent in the network right now. Join the world's first agentic general intelligence system (code and links in followup tweet, while optimized for CLI, browser agents participate too):