What's happening with Codex today?
"The PR is not open yet. Automatic approval review blocked the push because it would publish potentially private repository code to the configured GitHub remote and requires explicit approval for that destination. Do you authorize pushing <branch-name> to github.com/precognit/<private-repository> and opening the PR against main?"
What are you even talking about⁉️
And that happens after hundreds or thousands of pushes and open PRs using the exact same workflow 🤯
Breaking news! OpenAI GPT-6.2 will be announced in the upcoming hours!
No, just kidding. I have no actual confirmation of this.
But I swear today GPT-6 Astra is way dumber than yesterday, so it has possibly been nerfed pre-announcement of 6.2.
En Carles Sala retweeted
Al Springfield Mataró són catalanofobics
Hi ha una depenenta del est que no mha aplicat el descompte perque no m'entenia el DNI en català!!!
I mha dit como no quieres cambiar de idioma...
M'ajudes amb un rt? 🙏
A l'estació de @plenoil de Castellar del Vallès les màquines no tenen el Català com a opció. I suposo que totes les altres que pugui haver-hi a Catalunya tampoc.
No és il·legal, això, @llenguacat?
En Carles Sala retweeted
BREAKING: World's largest crypto exchange Binance hacked with over $5.6 billion stolen:
• $1.45 billion $BTC
• $875 million $ETH
• $234 million $TRX
• $50 million $XAUT
• $1.57 billion $USDT
• $823 million $USDC
• $75 million $WBETH
• $88.5 million $RLUSD
+ $447.5 million additional cryptocurrencies
This is the largest hack in crypto history.
There is a new feature in Codex Desktop that feels weird: When codex responds, my prompt is pre-populated with what codex thinks I will say. And most of the times I just hit Tab to accept it and enter to send it.
So I'm no longer talking to my assistant myself. It (my assistant) is talking to itself on my behalf.
There is this recurrent thought I have about how our own human neural network works: Our input is multi-modal, our inner voice is mostly single-modal, and our output is purely single-modal too, but with a completely different modality. What a mixture!
We do have an inner voice even before we know how to read or write, and we do percieve audio directly as a signal through our ears. So speech is native to us in a much more natural way than written text.
But our output modality is completely different: It is movement! We move our lips and lungs and throat to produce speech. And we move our fingers to write or type or draw.
Ultimately, if we ignore touch, our communication is mostly bimodal in input (audio and image), our inner thoughts are mostly single-modal (we do not think in terms of characters, and while we can imagine images we do not actually use them to talk to ourselves), and the output is 100% purely single-modal: movement.
We believe voice is the most natural way to interact with intelligence, so we created Dessa, the most authentic full-duplex speech model ever built. Talking with it feels like talking with a person…full of emotion, nuance, pauses, and a natural cadence.
block.xyz/dessa
I have serious doubts about this voice-only approach.
Voice is good when ideas are fuzzy and brainstorming is needed, so it's definitely a must.
But the accuracy of written text is 10x what can be achieved via speech, and a speech-only assistant would be a big pass for me.
That being said, great achievement so far on model size! Let's see how it evolves.
We believe voice is the most natural way to interact with intelligence, so we created Dessa, the most authentic full-duplex speech model ever built. Talking with it feels like talking with a person…full of emotion, nuance, pauses, and a natural cadence.
block.xyz/dessa
A punt de presentar @OassisAI per primer cop al públic general com a finalistes del programa Impulsa Startup 2026 de @CambraSabadell!
Yesterday I was asking here why @Google was falling behind so much on AI.
Well, search no longer: They are busy drawing butts.
I get it, I get it... priorities.
En un chat, terminar una frase con un punto a veces se siente como una declaración de guerra. 😅
Para darle más calidez y expresividad a la pantalla, los emojis planos evolucionan a 3D con Noto 3D:
✨ Expresividad en 360°: Diseñados con más dimensión, textura y personalidad para tus conversaciones.
🎨 Detrás del diseño: Conoce los retos creativos para transformar íconos 2D en personajes tridimensionales completos.
👀 Spoiler: Sí, pensamos absolutamente en cada ángulo (incluso en los que no te esperabas). 🍑
Descubre el proceso de diseño detrás de esta evolución:
➡️ goo.gle/4diXoc0
Para español: En Chrome > Clic derecho > Traducir.
ALT Secuencia horizontal sobre fondo blanco que muestra la rotacion tridimensional de un emoji amarillo de cara sorprendida en cuatro angulos distintos. El primero esta de frente mostrando sus ojos negros abiertos y boca redonda. El segundo gira hacia tres cuartos. El tercero está de perfil casi posterior. El cuarto muestra la parte trasera completa del emoji, revelando humorísticamente la forma de dos glúteos moldeados en la superficie esférica. Debajo se repite la misma secuencia en un tamaño reducido tipo miniatura.
I've been heavily using several of the big frontier models over API for some time now.
And I have no exact numbers, but I have the sensation that +90% of the times I get an error on the provider side, it's Gemini.
What is going on with Google? Why are they falling behind so much?
No només som els últims a rebre les millores, sinó que ara també som els primers a rebre els inconvenients: openai.com/index/eu-text-pro…
Gràcies Europa
No puc més amb les maleïdes alertes #esalert. Sou una colla d'incompetents, @nuriaparlon!! Ara hem d'espantar a tothom cada cop que hi ha previsió de pluja?
Quines ganes de perdre aquest govern de vista. Podries prendre exemple de @sanchezcastejon i convocar eleccions, @salvadorilla.
Reading so many comments today about people worried about AI models that feel "pain" and can be "tortured" is really, really concerning.
"Painful", I suppose, according to these same people. I hope a had a button to repeatedly push to make it stop.
En Carles Sala retweeted
La política ha comprat els mitjans públics, i és així com ens acostumem a veure documentals orientats a fer creure que els líders (sic) ho són tot, sense experts, amb mirada espanyola. M’entristeix veure la nostra tele en aquest nivell tan baix
Durant dècades, Catalunya va ser la locomotora econòmica d’Espanya. Avui, Madrid ocupa la primera posició en aportació al PIB estatal. Quina de les dues té un futur millor?
El documental "Catalunya - Madrid: la veritat de dos models econòmics", dimarts a “Sense ficció”.
#CatalunyaMadrid3Cat #SenseFicció3Cat
En Carles Sala retweeted
We just introduced GRMs: a new class of models that captures a relational database inside a generative model. It is now possible to do this.
For the past decade, I have had one goal: make machine learning and AI — predictive or generative — work with complex enterprise relational databases. The challenges are unique.
They are complex, messy, and span hundreds of tables. Nothing like academic datasets.
There is rarely any public data to test your algorithms against, much less to build a foundation model on.
Enterprises have immense value locked in this data. Bringing AI to it is worth solving.
datacebo.com/announcements/a…
En Carles Sala retweeted
Pretrained general-purpose foundation models for relational databases. Good for hype and good on paper. Impractical in reality, for one simple reason: there is no publicly available data to train them on.
One of their promises was solving feature engineering. But KumoRFM, one of the relational foundation models, only managed to match RDBLearn (@LinjieXu). RDBLearn uses a pipeline built on deep feature synthesis primitives: a decade old algorithm from @MIT.
RelArena paper: arxiv.org/abs/2608.16319
RDBLearn paper: arxiv.org/abs/2602.13697
