Pinned Tweet
I won 1st place in the @OpenBMB Backyard AI track for @huggingface Build Small Hackathon!!!
Thank you so much!!!! Small open models for the win!
É DO BRASIL!!! É NOSSO!!! 🇧🇷🇧🇷🇧🇷
Muito obrigada!!!
Finally, my Build Small Hackathon submission:
1 Person. 1 Small Business. 1B models.
Extracting valuable information from unstructured WhatsApp messages into a live dashboard, with Q&A in domain-specific brazilian portuguese.
Thank you all @OpenBMB @Gradio @huggingface
Marina Rosa retweeted
eu queria só lembrar que a pessoa aqui que não agrega nada não tá te vendendo nada também.
mas a galera ~DO BEM~ quer te empurrar qualquer ciosa vibecodada como produto que só tem audiência por relações parassociais.
passar bem pessoal
We shrank 1 GB of Wikipedia to 105.39 MB, losslessly.
Introducing tufazip: the best entry on the Large Text Compression Benchmark using a neural network as its sole predictor. Everything above it mixes a transformer with hand-built models.
Full post 👇
tufalabs.ai/research/tufazip
Marina Rosa retweeted
it’s punk now to be optimistic.
it’s punk now to be ambitious.
it’s punk now to be obsessed.
it’s punk now to be curious.
it’s punk now to be disciplined.
reject the mainstreaming of loserdom.
Thanks @ClementDelangue, I feel like sometimes you know me better than I know myself
Introducing Qwen3.8-27B-pi-adapted 🤗
We adapted Qwen's model specifically for the pi coding agent using reinforcement learning and a LoRA adapter.
The results, measured as pass@1:
🚀 SWE-bench Pro (Hard 51): 29.4% → 62.7%
💻 Terminal-Bench 2.1: 39.5% → 55.8%
⚡ 55–62% fewer tool calls per task
This is the power of open models: you can adapt them to YOUR tools and YOUR use cases. Let’s build! 💛
Open weights: huggingface.co/harness-hub/Q…
Methodology: huggingface.co/blog/harness-…
Marina Rosa retweeted
O pessoal às vezes me xinga aqui por causa de modelos pequenos como se estivessem fazendo um grande ensaio e me ensinando que Opus 5.5 é superior... só que eu sei disso, e não vou ficar mentindo pra ninguém e dizendo pra programarem só com modelos dentro da casa, sendo que os das big techs é que fazem o serviço completo de revisão arquitetural e entregas com poucos prompts.
A ideia nunca foi essa. Até porque empresas de SaaS são marias-vai-com-as-outras, elas pagam Opus pra você, fecham um contrato pro povo do compliance ficar feliz, e esperam que você seja hiper-produtivo com isso.
Acho que não entenderam ainda que o volume de pessoas usando IA pra programar é irrisório... As pessoas usam GPT pra pesquisa e pra perguntar horóscopo, gente. Modelo 4B faz isso e já é muita coisa.
Mas muito além disso, desde que eu me entendo por gente, TUDO o que eu falo vai especialmente em cima de cultivar o ESPÍRITO HACKER nas pessoas: a curiosidade, a vontade de pegar o negócio, virar de ponta-cabeça, entender, e usar por vontade e benefício próprio. O papo do mundo corporativo e da eficiência no seu serviço de programador eu deixo pro pessoal que tem paciência pra falar de IA no mercado de trabalho. Se vocês chutarem uma árvore, cai vinte.
Marina Rosa retweeted
Introducing Qwen3.8-27B-pi-adapted 🤗
We adapted Qwen's model specifically for the pi coding agent using reinforcement learning and a LoRA adapter.
The results, measured as pass@1:
🚀 SWE-bench Pro (Hard 51): 29.4% → 62.7%
💻 Terminal-Bench 2.1: 39.5% → 55.8%
⚡ 55–62% fewer tool calls per task
This is the power of open models: you can adapt them to YOUR tools and YOUR use cases. Let’s build! 💛
Open weights: huggingface.co/harness-hub/Q…
Methodology: huggingface.co/blog/harness-…
Honestly, we're not sure if we should keep spamming the timeline with reasons why you should be on Urbit instead of relying on megacorp identity systems and SaaS services.
Seriously. Urbit is your Ark. You don't get on the Ark because it is more comfy and compelling than the outside world. Than sitting by the fire in your cozy home.
You get on the Ark because a flood is coming.
JUST IN: Users report outages affecting multiple Apple $AAPL services.
• Apple Pay
• Apple Store
• Apple iCloud
• Apple Support
• Apple App Store
How open do you want your networks to be?
VPNs like @tailscale are great, but do you really want to be closed off into a private tailnet?
Maybe you say, "a tailnet is too closed off, I'm going to put everything behind @Cloudflare proxies and they are going to help me manage the bot traffic and digital trust issue."
But then you have handed all that power to a middle man--just like you did with email and @Microsoft.
How'd that work out for you? Oof. (Wouldn't want to be an exec on stage at Microsoft Ignite right now...)
Clearly a fully open network isn't what you want--this is why we have authentication and authorization tooling built into just about every 'app' you use--but a fully closed network is a blocker to collaboration and connection, and handing the gatekeeping to a megacorp is also a risk.
So what do you do?
There's Urbit, which we are obviously inclined to point you towards. But maybe you aren't ready for the full jump into Urbit OS. Maybe you like tailscale and just need a better solution for cross-subnet collaboration.
In that case, check out azimov.network from ~datder-sonnet and @TrueNorthD.
Think 'Tailscale, but with your pals'. Use your Urbit ID for device identification and networking, rather than... your Google login connected to Tailscale's servers.
General-purpose zkVM for Nock with the full pipeline checked in Lean 4.
Great work from @nockchain developers.
Marina Rosa retweeted
Introducing the Open Proof Machine (OPM), the largest technical milestone in Nockchain's history and the result of four years of research.
Our general-purpose zkVM is now open source, with soundness and completeness formally proved in Lean 4 for the arithmetization of the entire Nock language.






























