We trained miniclaus 1.5B with GRPO on GSM8k.. Increased GSM8k, GPQA, MMLU. Available in HF hub with GRPO code and reproduction details: https://t.co/TZYsA4VPVB

u know gpt5.2 or a new model is imminent.. gpt5.1 max-codex-ultra-mega became homer simpson, just as usual.. the model coming out aint noticeable better, otherwise why would u nerf your previous models?... ah?

#claude #Anthropic down?

GPT5.1 is always better than GPT5.0, and we always make sure of it. GPT5.1 indeed is better than GPT5.0 HomerSimpson Edition. Is GPT5.1 better than GPT5.0 NonHomerSimpson (1 week-ish ago)? Questionable.

the best way to sell AImodel 7.2 is by degrading AImodel 7.1 . i guess contaminating $self with fp32 answers.. is an effective way to maintain fp32 official scores.. but how to fix the fact that users feels talking to homer simpson: faith that lisa is coming out soon.

the hidden gem.. model selection aside, the best coding cli.. Amazon Q, lovely features

glm4.6 .. "count your mcp tools" 57, in reality 60.. half hour later back&forth... stills failing, cant count a simple list that is present in his context.. even when told him 60vs57.. he lies with fake sums 5+5=60 and stuff like that.. just like claude.. trashtalkers

Tired of wasting tokens on agents doing Input<>Output of tools? introducing Agentools for pydantic-ai, programatic approach for A2A. We added some freebies, like a UI to design and build agentoolkits, docs and tests :) https://t.co/AP4Y71zZB5

releasing ClaudeBench a tool to manage properly Claude Code work and perform more solid engineering work with it. We obtain better results using ClaudeBench with SpecKit than without it: consistency. #Anthropic #claudecode #claude @AnthropicAI https://t.co/dsdFqGvYOE

releasing ClaudeBench a tool to manage properly Claude Code work and perform more solid engineering work with it. We obtain better results using ClaudeBench with SpecKit than without it: consistency. https://t.co/dsdFqGvYOE

Model limits.. non accumulated.. i can use them all today and thats my limit.. if i dont use them tomorrow its my loss and their benefit.. So either you become "dependant", or "upgrade".. we missing github/fair-burn-tokens project to use all tokens 5m before they get restart..

There is a big difference between training models with your data and novelty classifiers that dispatches conversations to internal engineering teams to literally steal your project and idea. That's not model training, it's something else.. data thief bandits.

We crossed 1B+ tokens routed to inference providers partners on HF, that we released just a few days ago. Just getting started of course but early users seem to like it & always excited to be able to partner with cool startups in the ecosystem. Have you been using any

🚀 Can AI make $1M as a freelance software engineer? 💰 OpenAI’s SWE-Lancer puts AI to the test with 1,488 real-world freelance coding jobs from Upwork! 📊 But even the best model—Claude 3.5 Sonnet—only earned $400K. Is this the biggest AI freelancing opportunity? Or proof AI

1/3 🔹 How does SWE-Lancer work? ✅ AI tackles real Upwork coding jobs—worth $1M total ✅ Two categories: IC SWE tasks: AI writes & tests code. SWE Manager tasks: AI picks the best proposal. ✅ Graded with triple-verified end-to-end tests—no cherry-picked problems!

wonder why @perplexity_ai R1 is only half-way smart as the original https://t.co/7OYWYpy7b1 R1. Wonder wether this is a Quant version and not the good real R1.. because when contrasted R1 (#DeepSeek R1 web) vs O3 (og).. the metrics are very different vs R1 (perplexity) vs O3.

@deepseek_ai free chat could be hosted in @huggingface space with a $logs>/dev/null code that everyone can see.. or @Cloudflare could #savethetech and not just the #savetheweb by serving R1 on his AI fleet with some free calls per day.. so we can separate tech from politics

* unsubscribed claude, coz is truly bad at AI research. * unsubscribed openai 200$ super mega, coz he always inject bugs. * never paid gemini or G's coz they desperately vacuum users data. * deepseek always under attack * subscribed to https://t.co/8O8h1DmEUA to use R1

kudos to #DeepSeekR1 @deepseek_ai for the model, but that thing of "hobby project based on spare compute" .. LOL.. GGWP.. look the stocks, plenty of folks believed it XDDDDD

true fact of today: AI hype StocksForEx rules everything... evals gets contaminated, benchmarks with invented maj512, non-sense published arxiv papers, shit models that has 35 trillions follows, underperforming datasets, learn this true fact: AI IS NOT A POWERPOINT