8j1" dZddlZddlZddlmZmZmZddlmZddlm Z ddl m Z ddd d d d d dddd Z gdgdgdgdgdgdgdgdgdgdd Z gdZdeded ed!e fd"ZGd#d$e Zy)%uJProveedor de datos de EJEMPLO: coste 0, sin login, sin riesgo para tu cuenta. Genera tweets plausibles (temática IA/tech) para tus cuentas, con métricas y fechas realistas, de forma reproducible. Sirve para ver el dashboard y el ranking funcionando. Cuando quieras datos reales, crearemos otro proveedor con la misma interfaz. N)datetime timedeltatimezone)List)Tweet) TweetProvideri(#iipii iii( i@i) openaideepmindkarpathyylecun andrewyng lexfridmansimonw rowancheung_akhaliq techcrunch)zNew blog post: building an LLM from scratch in a single readable file. nanochat is now ~8k lines and trains a small but coherent model on one GPU overnight.zThe most underrated skill in 2026 is reading code slowly. We rush to generate, but understanding the existing 10k lines is where the real leverage is.zHot take: most 'agent frameworks' are 200 lines of glue around a while-loop and a tool registry. You probably don't need the framework.zSpent the weekend re-deriving backprop by hand for a small transformer. Still the best way to truly internalize how attention gradients flow.)zIntroducing our most capable model yet, with a 1M token context window and substantially improved tool use. Rolling out to the API and ChatGPT this week.zSafety update: we're publishing the full system card and red-teaming methodology alongside today's release. Transparency is part of the product.zDevelopers: function-calling latency is down 40% and structured outputs now support recursive JSON schemas. Full changelog in the thread.zWe're hiring researchers focused on long-horizon agent reliability. If you care about evals that actually predict real-world failure, come talk to us.)zNew short course: Building and Evaluating AI Agents. Free for the first week. Evaluation is the real bottleneck for teams shipping agents to production.zAI won't replace you, but someone using AI will. The cheapest investment a company can make right now is teaching every employee basic AI workflow skills.zData-centric AI again: a team went from 76% to 91% accuracy not by changing the model, but by fixing 300 mislabeled examples. Look at your data first.zExcited about small, specialized models running on-device. Not every task needs a frontier model; latency and privacy often matter more.)zLLMs are an off-ramp on the road to human-level AI. World models that can plan and reason about the physical world are the real path forward.zOpen source AI is not a risk, it's a necessity. The future of AI must not be controlled by a handful of companies on one coast.zA system that can't plan, can't reason, and has no persistent memory is not 'intelligent' in any meaningful sense, no matter how fluent it sounds.zNew paper from the lab on video world models: the model predicts in representation space, not pixels. This is the direction, folks.)zTIL you can run a 30B model on a laptop with the new quantization in llama.cpp. Wrote up the exact steps and benchmarks on my blog.zI've been using LLMs to write one-off bash scripts for two years now. The productivity gain on 'glue' tasks is the biggest shift in my workflow.zPrompt injection is still unsolved, and most 'agent' products are shipping it as a vulnerability. Untrusted text + tools = a problem.zNew release of my CLI tool: pipe any file into an LLM and get structured output back. `cat data.csv | llm 'extract the emails'`.)zNew podcast episode just dropped: 5 hours on the future of AGI, consciousness, and what it means to build machines that think. Link below.zThe hardest problems in AI are not technical, they're about love, meaning, and what kind of future we want to build. Engineering is the easy part.ztGrateful for this community. We're living through the most important decade in the history of intelligence on Earth.z~Question for the timeline: if you could ask a superintelligent AI one question, what would it be? Best answers in the replies.)zToday we're announcing a breakthrough in protein design: a new model generates novel enzymes with verified catalytic activity. Published in Nature.zOur latest agent solved 6/6 problems at competition-math-olympiad level, reasoning through full proofs step by step. Details in the blog.zWeather update: our model now produces 15-day forecasts more accurate than traditional methods, and runs in under a minute on a single TPU.zWe believe AI should benefit humanity. Today we're open-sourcing the weights and training recipe for our smaller research model.)zBREAKING: OpenAI just dropped a new model and it changes the game for developers. Here are the 7 most important things you need to know:zThis week in AI was insane. 1) A new frontier model 2) A major open-source release 3) A robotics demo that went viral. Full rundown below.ztI've now tested every major AI video tool. Here's the one that's actually production-ready in 2026 (and 3 to avoid):zzThe gap between people using AI daily and people who aren't is becoming the biggest productivity divide of our generation.)zNew paper: a 7B model matches frontier performance on math by training on high-quality synthetic chains of thought. Weights on Hugging Face.zlGradio demo of the day: real-time image-to-3D generation running in your browser. Link to try it free below.zoOpen models are catching up fast: three frontier-class releases this month alone, all with permissive licenses.z}New SOTA on long-context retrieval. The trick is a smarter positional encoding, not just more parameters. Paper + code below.)zsBreaking: AI startup raises $300M Series B at a $2B valuation to build autonomous coding agents for the enterprise.zzExclusive: a leaked memo reveals big tech's plan to put a personal AI assistant in front of a billion users by year's end.zrThe EU's new AI Act enforcement begins this month. Here's what every company deploying AI in Europe needs to know.zOpinion: the AI-bubble talk misses the point. The infrastructure being built now will outlast the hype cycle, just like fiber did in 2000.) r r r r rrr rrr)umNuevo modelo open source supera a varios cerrados en benchmarks de razonamiento. Pesos y código disponibles.uUHilo: las 5 herramientas de IA que de verdad cambiaron mi flujo de trabajo este año.u_La evaluación, no el entrenamiento, es el cuello de botella para llevar agentes a producción.u^Demo del día: generación de vídeo a partir de texto en tiempo real, gratis en el navegador.handleidxtextreturnc|j}tjtj|d|d|ddj }t j|d}t||jddz}t||jddz}t||jdd z}t||jd d z} t||jd d z} |jdd} tjtjt| z } dtj||j z} t|d|||| |||| | d|d|  S)N:ig333333?g@g?g?g{Gz?g333333?g{Gz?g{Gz?ZgR@)hourslTls<-zhttps://x.com/z/status/) idauthorr created_atlikesretweetsrepliesquotesviewsurl)lowerrandomRandomzlibcrc32encode INFLUENCEgetintuniformrnowrutcrr)rrrkeyrnd influencer!r"r#r$r% age_hourscreated status_ids agent/providers/mock.py _make_tweetr:^sh ,,.C -- cU!C5$s)#=#D#D#FG HC c4(I  CKKS11 2E53;;tT223H%#++dD112G T400 1F  B++ ,E C&Ill8<<(99+EEG#djjC51E1E1G&HHI U!C5> VHHYK 8  c ,eZdZddedededeefdZy) MockProviderrsincelimitrc (tj|jt}t |Dcgc]\}}t |||}}}|Dcgc]}|j |k\s|}}|jdd|d|Scc}}wcc}w)Nc|jS)N)r )ts r9z0MockProvider.fetch_user_tweets..s!,,r;T)r3reverse) SAMPLE_TWEETSr.r'GENERIC enumerater:r sort)selfrr>r?textsirBtweetss r9fetch_user_tweetszMockProvider.fetch_user_tweets|s!!&,,.':8A%8HI8H1+fa+8HI#=Vq||u'rXs 22.$$TT$D           o= ~ #SU:=r;