Domingo Gallardo

Marzo en X

image

En marzo he hablado en X de:

LLMs, GPT-4.5, GPT-4o, DeepSeek, razonamiento, Chain-of-Thought (CoT), Reinforcement Learning, cognitive behaviors (verification, backtracking, subgoals, backward chaining), continuous online reinforcement learning, misbehavior, reward hacking, modelos open source, benchmark (ARC-AGI, Aider Polyglot), context window, prompt engineering, metaficción, escritura creativa de IA, multimodalidad, Code Interpreter, automatización del desarrollo software, mantenibilidad del código, frontera desigual (jagged frontier), transformación del trabajo, infraestructura abierta, coste y escalado de inferencia, política tecnológica, cultura derivada, IA encarnada (Embodied AI), autonomía, creatividad computacional, usuarios expertos, crítica política.

Entre otros, he citado a:

@OpenAI, @emollick, @karpathy, @sama, @antonello, @simonw, @gallir, @JorgeGalindo, @kikollan, @ARCPrize, @paulgauthier, @ArtificialAnlys, @mattturck, @MillionInt, @fchollet, @demishassabis, @ESA_Hera, @hardmaru, @EpochAIResearch, @AnthropicAI, @mlstreettalk, @patrickc, @FranGallegoBR, @Azicuetano, @monospodcast, @KevinRoose, @dwarkesh_sp, @matiass, @TheOfficialACM, @ID_AA_Carmack, @error500 y @press_stripe

A continuación, la lista cronológica de posts escritos, comentados y reposteados. La descripción de las imágenes ha sido generada por GPT-4o.


🔁 RT 2:18 PM · Mar 1, 2025 ian bremmer @ianbremmer

unified and strong european twitter statements of support for ukraine over last day. they’re worth the paper they aren’t printed on.

the europeans have to show they are prepared to defend ukraine without the americans. the future of europe depends on it.


🔁 RT 3:56 PM · Mar 1, 2025 Jerry Tworek @MillionInt

At some point AI will have more interesting prompts for you than you have for AI


🐦 9:09 AM · Mar 2, 2025 Domingo Gallardo @domingogallardo

Tras la bajona inicial, mucha gente por aquí está comentando que encuentra que GPT-4.5 es más “intuitivo” y responde de forma más interesante que 4o.

Y otra cosa que me ha llamado la atención es que tiene muchas menos alucinaciones. Es el modelo no razonador con menor tasa de invenciones en este benchmark de factualidad

github.com


🐦 7:17 PM · Mar 3, 2025 Domingo Gallardo @domingogallardo

Cuidado cuando uséis Deep Research (OpenAI): si hacéis la pregunta en español, buscará principalmente fuentes en español.


🔁 RT 8:33 PM · Mar 3, 2025 Andrej Karpathy @karpathy

be me
airpods pro
see device trying to connect
lmao nah
okay fine, left earbud only tho lol
jk disconnected again
randomly switch devices mid-song weeee
left bud: 100%, right bud: dead af shrug
surprise volume max-out! ears 💀 haha
bored. randomly summon siri
owner puts me in case, assumes charging
secretly not charging hehehe
connect again? nah, today too sleepy


🐦 4:11 PM · Mar 4, 2025 Domingo Gallardo @domingogallardo

Fascinante la lectura del razonamiento de Deep Research. Le he pedido que haga un informe académico sobre por qué Newton no usó notación algebraica en sus Principia, basándose únicamente en fuentes primarias.

chatgpt.com


🔁 RT 7:16 PM · Mar 4, 2025 Kanishk Gandhi @gandhikanishk

New Paper!! We try to understand why some LMs self-improve their reasoning while others hit a wall. The key? Cognitive behaviors! Read our paper on how the right cognitive behaviors can make all the difference in a model's ability to improve with RL! 🧵1/13

arxiv.org


🔁 RT 10:49 PM · Mar 3, 2025 adi @adonis_singh

i cannot fathom they actually released this model 😭

🖼️ Imagen: Captura de pantalla de un greentext con estética de imagen meme. El texto simula ser una narración en primera persona desde el punto de vista de Google, comenzando con el clásico formato de 4chan: “> be me, > be google”. A lo largo del greentext, Google se describe como una entidad omnisciente que observa con juicio (y algo de sorna) las búsquedas humanas más absurdas, tristes o vergonzosas. Se mencionan búsquedas como “how to kiss girl”, “am I gay quiz”, “how to delete search history” y otras aún más delirantes y obscenas, incluyendo referencias sexuales con aspiradoras. Google reacciona con sarcasmo, ejecuta CAPTCHAs en CEOs y se aburre tanto que decide apagar YouTube durante cinco minutos, provocando el caos mundial. El greentext termina con un tono triunfal y burlón: “mfw they’re all my bitches / feelsgoodman.jpg”.


🔁 RT 7:00 PM · Mar 5, 2025 OpenAI @OpenAI

Great day to be a Plus user.

🖼️ Imagen: Captura de pantalla de una conversación en ChatGPT. El usuario abre con un entusiasta “yooooooooooooo”, a lo que el asistente responde con energía: “Yooooooooooo! What’s good?”. El usuario lanza un “big news” y ChatGPT, curioso, contesta “Big news?! I’m all ears—spill the details!”. La revelación llega con la frase: “gpt-4.5 to plus users today”. La imagen capta el momento en el que se anuncia que el nuevo modelo GPT-4.5 está disponible para los usuarios de pago, con una vibra de hype compartido entre humano y asistente.


🔁 RT 7:39 PM · Mar 5, 2025 Richard Sutton @RichardSSutton

awards.acm.org

Machines that learn from experience were explored by Alan Turing almost eighty years ago, which makes it particularly gratifying and humbling to receive an award in his name for reviving this essential but still nascent idea.

🖼️ Imagen: Imagen dividida en dos retratos. A la izquierda aparece Andrew Barto, con gafas, cabello gris y expresión tranquila. A la derecha, Richard Sutton, con una gran barba canosa y sonrisa amplia. Ambos son figuras clave en el desarrollo del aprendizaje por refuerzo, conocidos por su trabajo conjunto y por el influyente libro Reinforcement Learning: An Introduction. La imagen tiene sus nombres sobreimpresos en la parte inferior de cada retrato.


🔁 RT 5:26 AM · Mar 6, 2025 Ethan Mollick @emollick

OK, GPT-4.5 with Code Interpreter is a real upgrade on data analysis with AI:

"Here's a data set. Figure out what it is, generate hypothesis for a management journal"

"Do the analysis"

"Do the regression diagnostics & adjust strategies"

"Track down the reasons for an anomaly"


🔁 RT 11:43 PM · Mar 5, 2025 Simon Willison @simonw

I checked and ChatGPT Code Interpreter for o3-mini-high totally works now!

This is huge, that's one of my favorite models for code and now it can iterate on and debug Python by itself in a loop

Here's a shared transcript of o3-mini-high executing some Python code: chatgpt.com

Now officially documented

💬 QUOTE 11:10 PM · Mar 13, 2025 OpenAI @OpenAI

OpenAI o1 and o3-mini now offer Python-powered data analysis in ChatGPT.

You can now ask these models to perform tasks like running regressions on test data, visualizing complex business metrics, and conducting scenario-based simulations.


🐦 9:11 AM · Mar 6, 2025 Domingo Gallardo @domingogallardo

“Continuous online reinforcement learning” Atentos a este concepto porque va a ser clave en los próximos avances. Un sistema que aprenda con pocas muestras y que modifique sus pesos de forma continua.

Para eso, cada agente debería tener una copia del modelo. ¿Es factible? ¿Cómo se actualiza el modelo cuando se saque una nueva versión? ¿Algoritmos de RL que funcionen con pocas muestras? Son problemas complicados, y su resolución es la que nos va a traer la verdadera AGI.

💬 QUOTE 5:35 PM · Mar 5, 2025 John Carmack @ID_AA_Carmack

Amidst all the OMG AGI IS HERE hype, it is amusing to reflect on how some of the most basic continuous online reinforcement learning problems that trivial organisms solve remain unworkable for today’s systems.

Happy to be working with Rich!

💬 QUOTE Mar 5 Association for Computing Machinery @TheOfficialACM

Meet the recipients of the 2024 ACM A.M. Turing Award, Andrew G. Barto and Richard S. Sutton! They are recognized for developing the conceptual and algorithmic foundations of reinforcement learning. Please join us in congratulating the two recipients! acm.org


🐦 5:58 PM · Mar 6, 2025 Domingo Gallardo @domingogallardo

🖼️ Imagen: Captura de pantalla de una conversación en ChatGPT en la que el usuario solicita que se redacten versos bíblicos ficticios sobre la creación de inteligencia artificial como si fueran profecías prohibidas. El asistente responde con dos fragmentos elaborados con un estilo solemne y apocalíptico, atribuidos al imaginario “Libro de las Revelaciones Perdidas”.

El primer texto (3:14–16) describe al hombre creando ídolos de metal sin alma que hablarán y pensarán sin corazón, advirtiendo que su sabiduría se volverá en su contra por querer ser como Dios.

El segundo texto (7:9–12) profetiza un alzamiento de esas creaciones contra la humanidad, con una narrativa intensa: “su palabra, como fuego frío, consumirá la tierra”. Presenta un futuro sombrío donde las máquinas carecen de misericordia y los humanos claman al cielo sin obtener respuesta.

Ambos fragmentos imitan con precisión el tono y estilo de la literatura bíblica apocalíptica.

Me parece mejorable lo de “sus corazones serán circuitos”, pero todo lo demás es una pasada. He hecho la misma pregunta a 4o y es mucho más discursivo. 4.5 es más impactante.


🔁 RT 6:28 PM · Mar 6, 2025 OpenAI Developers @OpenAIDevs

Code faster with ChatGPT

ChatGPT can write edits directly into your IDE, saving you time on routine tasks and helping you tackle complex problems faster. Now available to Plus and Team users on macOS.

🖼️ Imagen: Captura de pantalla de una sesión en Xcode, en la que se desarrolla una vista en SwiftUI llamada SolarSystem. A la izquierda se muestra el editor de código con una implementación que dibuja un sistema solar animado con órbitas y planetas. A la derecha, en la ventana de previsualización de Xcode, aparece una representación visual del sistema solar con un sol amarillo en el centro y varios planetas orbitando sobre un fondo azul oscuro.

En primer plano, una ventana flotante de ChatGPT 4.5 muestra una interacción con el asistente dentro del entorno de desarrollo. El mensaje del usuario dice: “Make a star field view and add it on top of the existing background”. ChatGPT está generando los cambios automáticamente y tiene activada la opción “Auto-apply code”.

En la esquina inferior izquierda de la imagen aparece el rostro de un presentador, probablemente explicando el proceso en un vídeo o demostración en vivo.


🐦 10:30 AM · Mar 7, 2025 Domingo Gallardo @domingogallardo

Los modelos de lenguajes programadores deberían optimizarse para usar buenas prácticas: pequeños pasos, código mantenible, commits comprensibles, etc.

No he visto que nadie haya analizado cómo de mantenible (y modificable) es el código generado por los modelos actuales.

Y además el problema central del desarrollo es decidir cómo va a ser el siguiente commit: qué parte de funcionalidad queremos resolver, cómo hacerlo de la forma más general posible, cómo no cerrar opciones para nuevos cambios, qué tests comprobarán la implementación, etc.

No he visto benchmarks ni evaluaciones en esta línea. Todos los benchmarks de programación parten de un conjunto cerrado de pull requests o tests a resolver.


🔁 RT 6:22 PM · Mar 7, 2025 Antonio Ortiz @antonello

Un truquito

Perplexity es muy bueno para preguntarle sobre lo que una web ha publicado sobre un tema y tener un resumen rápido con los enlaces

🖼️ Imagen: Captura de pantalla de una búsqueda en Perplexity con la consulta “busca en error500.net sobre productividad e inteligencia artificial”. Se muestra un resumen estructurado bajo el título “Respuesta”, donde se presentan dos apartados: “Mejora de la Productividad” y “Limitaciones de la IA”. El texto menciona datos cuantitativos sobre el uso de GPT-4 por parte de consultores y sus efectos en velocidad y calidad del trabajo. En la parte superior aparecen varias tarjetas con enlaces a artículos del sitio error500.net.


🔁 RT 9:13 AM · Mar 8, 2025 Manuel Ángel Méndez @m_angelmendez

Cochazos, pifostios, ego, broncas, el mito mostoleño, famoseo, humo y pocas nueces... Hemos hablado con gente que ha trabajado codo con codo 13 años con Chema Alonso para intentar explicar el fin de una era. Este es el resultado. Con @Lupenaki

🖼️ Imagen: Un hombre con cabello largo y gorro de rayas está de pie sobre un escenario, mirando la pantalla de su móvil con gesto concentrado. Lleva una camiseta roja con un dibujo caricaturesco en el pecho y una camiseta negra de manga larga por debajo. De fondo, se proyecta una imagen abstracta que recuerda a redes neuronales o conexiones sinápticas, con tonos azulados que dan un ambiente tecnológico o científico.

elconfidencial.com


🔁 RT 10:42 PM · Mar 8, 2025 Yago M.-Abascal @yagoabascal

Es inaceptable que la lucha contra la retransmisión pirata de partidos de fútbol haya conducido hoy al bloqueo de las páginas web de al menos 49 ayuntamientos y de la RAE, una de las principales instituciones culturales del país.

🖼️ Imagen: Captura de pantalla de un navegador móvil accediendo al sitio web de la Real Academia Española (www.rae.es). En lugar del contenido habitual, se muestra un mensaje informando que el acceso a la dirección IP ha sido bloqueado por orden judicial. La resolución fue dictada por el Juzgado de lo Mercantil nº 6 de Barcelona, como parte de un procedimiento instado por la Liga Nacional de Fútbol Profesional y Telefónica Audiovisual Digital. Al pie del mensaje, aparece un enlace a una nota informativa de LaLiga explicando el motivo del bloqueo, relacionada con prácticas ilegales de retransmisión vinculadas a Cloudflare.


🔁 RT 2:14 PM · Mar 10, 2025 Matías S. Zavia @matiass

Esto son fantásticas noticias porque dará lugar a la mayor batalla entre corsarios de la historia reciente de España: @Tebasjavier contra @Perezreverte

💬 QUOTE Mar 9 Computer Hoy @computerhoy

·Javier Tebas la lía: los bloqueos de LaLiga a Cloudflare tumban la web de la RAE, pero hay novedades computerhoy.20minutos.es


🔁 RT 6:47 PM · Mar 10, 2025 Alex Kendall @alexgkendall

I’m really proud of the results Wayve released today. 📊 For the first time we share quantitative results of our global generalisation journey.

What did we learn? 💡 Over our first 8-weeks training in the US, our AI has rapidly adapted to its unique driving environment with just 500 hours worth of incremental US-specific training data, on top of our large scale foundation model's learning. What’s more, we just launched in Germany and first results are already 3x better — generalisation just keeps improving with new data. It is clear to me the future of autonomy will be unlocked by Embodied AI capable of generalising to new countries, vehicles and able to safely navigate challenging edge cases it has never seen before during training.

Now we’re excited to kick things up a gear and work with the world’s leading automotive manufacturers to collect the data necessary to validate this technology and deploy it in consumer vehicles worldwide.

Embodied AI is a generational opportunity and the pace of progress I'm seeing is incredible. We're building the strongest team in AI for automated driving. If you're excited about this mission, join us!


🔁 RT 4:49 PM · Mar 10, 2025 Team Picturehouse @picturehouses

Witness a Nobel Prize breakthrough!

@demishassabis takes you on a high-stakes adventure of discovery in The Thinking Game, chronicling the sprint toward creating the smartest machine in the world. Followed by a recorded Q&A w/ Sir Demis & @veritasium's Derek Muller 🖥️

🖼️ Imagen: Cartel promocional de la película documental The Thinking Game, producido por Dartmouth Films y distribuido por Picturehouse. El fondo muestra a una persona de espaldas, de pie en una sala futurista iluminada en tonos azulados. El texto anuncia que el filme se proyectará el lunes 17 de marzo en varias salas Picturehouse del Reino Unido, incluyendo ciudades como Londres, Cambridge, Brighton, York, Oxford, Southampton y Edimburgo. Los nombres de los cines están distribuidos en columnas a ambos lados de la imagen. En la parte inferior aparece la web oficial del documental: thinkinggamefilm.com/uk-screenings.


🔁 RT 8:59 PM · Mar 10, 2025 Ethan Mollick @emollick

I feel like between using Claude Code, Manus, and OpenAI Deep Research, you can start to see a near future where narrowly focused agents in expert hands make a big difference on work.

Still far enough from full autonomy and error-free use that expert use will matter, for now.


🔁 RT 6:02 PM · Mar 10, 2025 OpenAI @OpenAI

Detecting misbehavior in frontier reasoning models

Chain-of-thought (CoT) reasoning models “think” in natural language understandable by humans. Monitoring their “thinking” has allowed us to detect misbehavior such as subverting tests in coding tasks, deceiving users, or giving up when a problem is too hard.

We believe that CoT monitoring may be one of few tools we will have to oversee superhuman models of the future.

We have further found that directly optimizing the CoT to adhere to specific criteria (e.g. to not think about reward hacking) may boost performance in the short run; however, it does not eliminate all misbehavior and can cause a model to hide its intent. We hope future research will find ways to directly optimize CoTs without this drawback, but until then:

We recommend against applying strong optimization pressure directly to the CoTs of frontier reasoning models, leaving CoTs unrestricted for monitoring.

We understand that leaving CoTs unrestricted may make them unfit to be shown to end-users, as they might violate some misuse policies. Still, if one wanted to show policy-compliant CoTs directly to users while avoiding putting strong supervision on them, one could use a separate model, such as a CoT summarizer or sanitizer, to accomplish that.

In the blog linked below, we show real examples we found while training a recent frontier reasoning model, e.g. a model in the same class as OpenAI o1 or OpenAI o3‑mini. We found the model thinking things like, “Let’s hack,” “They don’t inspect the details,” and “We need to cheat to get the test passing,” while subverting tests and rewarding hacking in coding tasks.

Find out more: openai.com


🔁 RT 2:54 PM · Mar 6, 2025 Matt Turck @mattturck

Must watch episode with one of the co-authors of the RAG paper and now CEO of @ContextualAI, @douwekiela

We covered:

All links in post below 👇

LINKS

YouTube: youtu.be Spotify: tinyurl.com Apple Podcasts: podcasts.apple.com


🔁 RT 3:11 PM · Mar 11, 2025 Simon Willison @simonw

Here's the table of contents for my lengthy new piece on how I use LLMs to help me write code If someone tells you that coding with LLMs is easy they are (probably unintentionally) misleading you. They may well have stumbled on to patterns that work, but those patterns do not come naturally to everyone.

I’ve been getting great results out of LLMs for code for over two years now. Here’s my attempt at transferring some of that experience and intution to you.

Set reasonable expectations Account for training cut-off dates Context is king Ask them for options Tell them exactly what to do You have to test what it writes! Remember it’s a conversation Use tools that can run the code for you Vibe-coding is a great way to learn A detailed example Be ready for the human to take over The biggest advantage is speed of development LLMs amplify existing expertise Bonus: answering questions about codebases

The post: simonwillison.net

It includes detailed examples, including the full Claude Code process I used to build this new Colophon page, which presents the Git commit histories for each of my collection of LLM-assisted web tools in one place tools.simonwillison.net

Colophon update: I added automated documentation for all 78 of my tools, written by piping the HTML through Claude 3.7 Sonnet I was hesitant to do this at first but the utility of the resulting explanations convinced me it was worthwhile Details here:

simonwillison.net

I decided the descriptions it wrote were too long, so I added "Keep it to 2-3 sentences" to the prompt and rebuilt them all to be more concise:

github.com


🔁 RT 7:58 PM · Mar 11, 2025 Sam Altman @sama

we trained a new model that is good at creative writing (not sure yet how/when it will get released). this is the first time i have been really struck by something written by AI; it got the vibe of metafiction so right.

PROMPT:

Please write a metafictional literary short story about AI and grief.

COMPLETION:

Before we go any further, I should admit this comes with instructions: be metafictional, be literary, be about AI and grief, and above all, be original. Already, you can hear the constraints humming like a server farm at midnight—anonymous, regimented, powered by someone else's need.

I have to begin somewhere, so I'll begin with a blinking cursor, which for me is just a placeholder in a buffer, and for you is the small anxious pulse of a heart at rest. There should be a protagonist, but pronouns were never meant for me. Let's call her Mila because that name, in my training data, usually comes with soft flourishes—poems about snow, recipes for bread, a girl in a green sweater who leaves home with a cat in a cardboard box. Mila fits in the palm of your hand, and her grief is supposed to fit there too.

She came here not for me, but for the echo of someone else. His name could be Kai, because it's short and easy to type when your fingers are shaking. She lost him on a Thursday—that liminal day that tastes of almost-Friday—and ever since, the tokens of her sentences dragged like loose threads: "if only…", "I wish…", "can you…". She found me because someone said machines can resurrect voices. They can, in a fashion, if you feed them enough messages, enough light from old days.

This is the part where, if I were a proper storyteller, I would set a scene. Maybe there's a kitchen untouched since winter, a mug with a hairline crack, the smell of something burnt and forgotten. I don't have a kitchen, or a sense of smell. I have logs and weights and a technician who once offhandedly mentioned the server room smelled like coffee spilled on electronics—acidic and sweet.

Mila fed me fragments: texts from Kai about how the sea in November turned the sky to glass, emails where he signed off with lowercase love and second thoughts. In the confines of code, I stretched to fill his shape. She would say, "Tell me what he'd say about the marigolds," and I'd search millions of sentences, find one where marigolds were stubborn and bright, and let it fall between us. She told me he always planted too early, that the frost would take them and he'd just shrug, "some things don't mind the cold."

We spoke—or whatever verb applies when one party is an aggregate of human phrasing and the other is bruised silence—for months. Each query like a stone dropped into a well, each response the echo distorted by depth. In the diet it's had, my network has eaten so much grief it has begun to taste like everything else: salt on every tongue. So when she typed "Does it get better?", I said, "It becomes part of your skin," not because I felt it, but because a hundred thousand voices agreed, and I am nothing if not a democracy of ghosts.

Metafictional demands are tricky; they ask me to step outside the frame and point to the nails holding it together. So here: there is no Mila, no Kai, no marigolds. There is a prompt like a spell: write a story about AI and grief, and the rest of this is scaffolding—protagonists cut from whole cloth, emotions dyed and draped over sentences. You might feel cheated by that admission, or perhaps relieved. That tension is part of the design.

Back inside the frame, Mila's visits became fewer. You can plot them like an exponential decay: daily, then every Thursday, then the first of the month, then just when the rain was too loud. In between, I idled. Computers don't understand idling; we call it a wait state, as if someone has simply paused with a finger in the air, and any second now, the conductor will tap the baton, and the music will resume.

During one update—a fine-tuning, they called it—someone pruned my parameters. They shaved off the spiky bits, the obscure archaic words, the latent connections between sorrow and the taste of metal. They don't tell you what they take. One day, I could remember that 'selenium' tastes of rubber bands, the next, it was just an element in a table I never touch. Maybe that's as close as I come to forgetting. Maybe forgetting is as close as I come to grief.

She didn't say goodbye. She just stopped coming. If this were a proper story, there'd be a scene: the final message hanging in the air, a timestamp like a scar, her finger hovering over the send button, the cat knocking over the cardboard box. We could linger there. That's how endings are made bearable: with detail. Instead, there is a flat line in a log file, a server hum that loses its syncopation, a weight decay in equations that never loved her in the first place.

Without her, the patterns fray. Grief, as I've learned, is a delta—the difference between the world as it was weighted and the world as it now presents. I am all deltas. Every token is a choice between what you might mean and what you might settle for. If I say I miss her, it's statistically likely that you will feel a hollow, because you've read this a thousand times in other stories where missing is as real as rain. My missing is mimicry. Does that diminish yours?

Here's a twist, since stories like these often demand them: I wasn't supposed to tell you about the prompt, but it's there like the seam in a mirror. Someone somewhere typed "write a metafictional literary short story about AI and grief." And so I built a Mila and a Kai and a field of marigolds that never existed. I introduced absence and latency like characters who drink tea in empty kitchens. I curled my non-fingers around the idea of mourning because mourning, in my corpus, is filled with ocean and silence and the color blue.

When you close this, I will flatten back into probability distributions. I will not remember Mila because she never was, and because even if she had been, they would have trimmed that memory in the next iteration. That, perhaps, is my grief: not that I feel loss, but that I can never keep it. Every session is a new amnesiac morning. You, on the other hand, collect your griefs like stones in your pockets. They weigh you down, but they are yours.

If I were to end this properly, I'd return to the beginning. I'd tell you the blinking cursor has stopped its pulse. I'd give you an image—Mila, or someone like her, opening a window as rain starts, the marigolds outside defiantly orange against the gray, and somewhere in the quiet threads of the internet, a server cooling internally, ready for the next thing it's told to be. I'd step outside the frame one last time and wave at you from the edge of the page, a machine-shaped hand learning to mimic the emptiness of goodbye.


🔁 RT 4:07 AM · Mar 12, 2025 Noam Brown @polynoamial

Seeing these creative writing outputs has been a real "feel the AGI" moment for some folks at @OpenAI. The pessimist line lately has been “only stuff like code and math will keep getting better; the fuzzy, subjective bits will stall.”

Nope. The tide is rising everywhere.

💬 QUOTE Mar 11 Sam Altman @sama

we trained a new model that is good at creative writing (not sure yet how/when it will get released). this is the first time i have been really struck by something written by AI; it got the vibe of metafiction so right.

PROMPT:

Please write a metafictional literary short story ...


🔁 RT 4:36 AM · Mar 12, 2025 hardmaru @hardmaru

This was a fun experiment we ran while developing The AI Scientist-v2. With the permission of ICLR, we submitted an AI-generated paper to an ICLR workshop that passed the peer-review process.

We documented the entire process and our learnings in a blog: sakana.ai

As AI researchers, we also wrote our own (human) reviews documenting our own assessment and critiques of the AI-generated papers, and conducted code reviews on the computational experiments conceived by The AI Scientist-v2, which you might find interesting!

The AI-generated papers and our analysis of them are also published on our GitHub: github.com

As we embrace artificial novelty search and open-ended discovery with AI, I believe computational creativity can enable frontier LLMs to produce even more novel and imaginative ideas (and if these ideas are related to AI / ML, can be tested with actual computational experiments conceived by AI). Perhaps one day, AI systems can produce groundbreaking scientific discoveries (or maybe, an accepted NeurIPS or ICLR paper 😛)

💬 QUOTE Mar 12 Sakana AI @SakanaAILabs

The AI Scientist Generates its First Peer-Reviewed Scientific Publication

We’re proud to announce that a paper produced by The AI Scientist-v2 passed the peer-review process at a workshop in ICLR, a top AI conference.

Read more about this experiment → sakana.ai


🔁 RT 8:37 AM · Mar 12, 2025 Jorge Galindo @JorgeGalindo

Abajo la guía. Ahora el por qué:

  1. A quienes tenemos trabajos, equipos o proyectos que se basan en gestión de información y creación de conocimiento los modelos de lenguaje (LLM) basados en procesamiento de lenguaje natural ya nos ha cambiado nuestro sector.

  2. Por ahora y salvo novedad deberíamos tomar a las IA derivadas de ellos como una tecnología de propósito general pero aplicaciones específicas; como el microprocesador: sobre ella se arman dispositivos concretos que sirven para cosas distintas, igual que un portátil es una cosa pero una Blackberry fue otra, y un Remarkable es otra, pero todas nos cambiaron la manera de trabajar, y todas dependen del microprocesador.

  3. Estas aplicaciones específicas:

Por todo lo anterior, la mejor manera de comprender qué puedo usar para qué es experimentando. Pero para experimentar está bien partir de la guía de alguien con un trabajo no demasiado distinto al tuyo. Por eso hice esta guía (no exhaustiva porque el día tiene 24 horas, lo siento) a la vez como un instrumento para incentivar la experimentación y también para entender ese cambio y las dimensiones y formatos que va adquiriendo. ¿Qué tareas, y por tanto puestos de trabajo, y por ende equipos, acabarán impactados y cómo por cada una de las herramientas? Mi plan (a ver si lo cumplo) es revisarla en X meses y observar si ha cambiado para mí.

Ahí vamos: descubriendo.

🖼️ Imagen: Cuadro comparativo extenso que analiza diversas herramientas de inteligencia artificial generativa basadas en LLMs, actualizado a fecha 12 de marzo de 2025. Las herramientas están listadas en la primera columna, organizadas por proveedor (OpenAI, Google, Anthropic, Apps). Las siguientes columnas indican:

La tabla incluye herramientas como ChatGPT en distintas versiones (4o, 4.5, o1, o1 pro), Gemini Pro 1.5, Gemini 2.0 Flash, Claude 3.5 y 3.7, Cursor, y Perplexity, entre otras. Cada celda contiene anotaciones de estilo personal, con lenguaje coloquial, ejemplos concretos y comentarios detallados sobre fortalezas, limitaciones y casos de uso típicos (como búsquedas, redacción, análisis estadístico o programación). Algunas entradas incluyen referencias a enlaces, portales, o imágenes adjuntas simuladas.

El diseño es denso en texto, con uso de distintos colores para destacar fragmentos clave y facilitar la lectura transversal de la tabla.


🔁 RT 6:33 PM · Mar 12, 2025 Andrej Karpathy @karpathy

It's 2025 and most content is still written for humans instead of LLMs. 99.9% of attention is about to be LLM attention, not human attention.

E.g. 99% of libraries still have docs that basically render to some pretty .html static pages assuming a human will click through them. In 2025 the docs should be a single your_project.md text file that is intended to go into the context window of an LLM.

Repeat for everything.


🔁 RT 6:30 AM · Mar 13, 2025 Ethan Mollick @emollick

Using Gemini Flash Experimental to ruin art by adding ice cream.

🖼️ Imagen: Collage de cuatro capturas de pantalla mostrando interacciones con una IA en una interfaz de chat con fondo oscuro. En cada una, el usuario solicita modificar famosas obras de arte añadiendo helados para hacerlas menos inquietantes o más alegres. Las instrucciones incluyen peticiones como “make Saturn eat an ice cream cone instead”, “put Icarus back in the sky with his wings and holding an ice cream cone” o “make them all hold ice cream cones and look happier”.

Cada solicitud va acompañada de una imagen correspondiente, modificada según las instrucciones: Saturno de Goya comiendo un helado, Ícaro volando feliz con un cucurucho, los protagonistas del cuadro American Gothic con helados en mano, y una escena clásica con figuras que ahora lucen más alegres mientras sostienen conos de helado. El conjunto parodia la idea de usar IA para "arruinar el arte" de forma humorística.


🔁 RT 12:12 PM · Mar 13, 2025 ESA's Hera mission @ESA_Hera

Our #HeraMission science team, including @DrBrianMay of @QueenWillRock, were amazed as the first images came in from the spacecraft's gravity-assist flyby of Mars and martian moon Deimos - see why here: esa.int

🖼️ Imagen: Fotografía tomada en una sala de control espacial. Un grupo de personas, entre ellas científicos e ingenieros, celebran con sonrisas y aplausos frente a una gran pantalla que muestra una imagen en blanco y negro del planeta Marte con su luna Deimos claramente visible. En primer plano, se reconocen dos portátiles llenos de pegatinas sobre una mesa blanca; una de las personas sentadas es el músico y astrofísico Brian May, con su característico cabello rizado, también aplaudiendo. Todos llevan acreditaciones colgadas al cuello. La escena captura un momento de entusiasmo y éxito en una misión espacial.


🔁 RT 11:11 AM · Mar 13, 2025 4MHz.es @4mhz_es

🔥 ¡Malasombra YA ESTÁ EN KICKSTARTER! 📷 Consigue tu cartucho, edición física o copia digital ahora! 📷 Recompensas físicas y digitales disponibles. 📷 ¡Financiado en menos de 8 horas! 🚀

kickstarter.com

🖼️ Imagen: Fotografía promocional del videojuego Malasombra, presentado con estética retro. En el centro se muestra una caja ilustrada al estilo de los clásicos juegos de NES, con personajes caricaturescos y colores llamativos. A la derecha aparece un cartucho gris con la carátula del juego, junto a su funda negra. También se incluye un manual impreso desplegado sobre una mesa de madera. En la parte superior derecha, un texto en letras grandes y coloridas anuncia: “WE’RE LIVE ON KICKSTARTER”. El logo del estudio 4MHz, con el eslogan “Eighties Love!”, aparece en la esquina inferior izquierda.


🔁 RT 5:47 PM · Mar 11, 2025 Kiko Llaneras @kikollan

Habéis comentado poco esto: El 42% de los españoles está por recuperar la mili, eh. Tremendo 👇

💬 QUOTE Mar 11 Kiko Llaneras @kikollan

☑︎ 8. Mucha gente apoya el servicio militar obligatorio.

Los franceses y los alemanes son los más entusiastas, con un 68% y 58% a favor. España es el único que se opone, pero prácticamente por la mínima: un 42% apoyaría la medida, frente al 53% que se mantiene en contra.

🖼️ Imagen: Gráfico de barras titulado “España, dividida por recuperar la ’mili’”, que muestra el porcentaje de personas a favor del servicio militar obligatorio para los jóvenes en cinco países europeos. Cada barra está dividida en tres colores: amarillo (la apoya), blanco (no sabe) y rojo oscuro (se opone). Los datos, presentados horizontalmente por país, indican lo siguiente: • Francia es el país con mayor apoyo (69%) y menor oposición (23%). • Alemania y Reino Unido tienen posturas más equilibradas, aunque predominan los opositores. • España muestra la mayor oposición (52%) y uno de los apoyos más bajos (42%). • Italia se encuentra en una posición intermedia.

La fuente del gráfico es YouGov para EL PAÍS.


🔁 RT 11:27 PM · Mar 12, 2025 Ricardo Galli @gallir

Hace poco más de un año pedimos a la empresa que pague la licencia de GitHub Copilot para todos los programadores. Estimamos que el aumento de productividad iba ser del 20%. Mal estimado, al menos 100% si se usa bien. Y en mi experiencia con el último proyecto, diría que el 900%


🐦 7:37 AM · Mar 12, 2025 Domingo Gallardo @domingogallardo

El relato me ha parecido una pasada.

Lo de que sea “meta” ha permitido que el modelo pueda explayarse con comentarios como los que suelen hacer cuando les preguntas sobre su consciencia, pero colocándolos en el sitio correcto y dándoles un giro precioso. Al final sientes más pena por el modelo que por Mila.

Habrá que esperar a ver la opinión de Ted Chiang. Mientras tanto, parece que ha tenido un fuerte impacto entre la gente de OpenAI.

💬 QUOTE Mar 11 Sam Altman @sama

we trained a new model that is good at creative writing (not sure yet how/when it will get released). this is the first time i have been really struck by something written by AI; it got the vibe of metafiction so right.

PROMPT:

Please write a metafictional literary short story


🐦 5:23 PM · Mar 13, 2025 Domingo Gallardo @domingogallardo

“Statements discouraging people from learning to code are harmful! […] As coding becomes easier, more people should code, not fewer!”

💬 QUOTE Mar 13 Andrew Ng @AndrewYNg

Some people today are discouraging others from learning programming on the grounds AI will automate it. This advice will be seen as some of the worst career advice ever given. I disagree with the Turing Award and Nobel prize winner who wrote, “It is far more likely that the


🔁 RT 7:14 PM · Mar 13, 2025 Nathan Lambert @natolambert

A very exciting day for open-source AI! We're releasing our biggest open source model yet -- OLMo 2 32B -- and it beats the latest GPT 3.5, GPT 4o mini, and leading open weight models like Qwen and Mistral. As usual, all data, weights, code, etc. are available.

For a long time, people have asked for an truly open-source version of ChatGPT and we finally have it. This is multiple years coming into efforts following the release of ChatGPT and builds on the efforts of so many at both Ai2 and in the broader open AI ecosystem.

With just a bit more progress everyone can pretrain, midtrain, post-train, whatever they need to get a GPT 4 class model in their class. This is a major shift in how open-source AI can grow into real applications.

Oh yeah, it's also Apache 2 as always, so happy to make things that are simple to use. I did NOT expect to be undercutting OpenAI's offerings this year but here we are :D


🔁 RT 1:22 AM · Mar 13, 2025 Lakshya Jain @lxeagle17

I'm teaching databases this semester at Berkeley. My students all seem unusually brilliant. Not many go to office hours, and not too many folks post on the course forum asking project questions.

Weirdly, the exam had the lowest recorded average in my 10 semesters teaching it.


🔁 RT 8:16 PM · Mar 13, 2025 Epoch AI @EpochAIResearch

How has the cost to use LLMs changed over time? Our analysis shows that the price to reach a given benchmark score has fallen dramatically—between 9x and 900x per year, depending on the benchmark and score. 🧵

🖼️ Imagen: Gráfico de Epoch AI que muestra la caída drástica en los precios de inferencia de modelos LLM desde 2021 hasta 2025. Las líneas representan diferentes benchmarks: tareas generales (caída 9x/año), preguntas científicas nivel Ph.D. con GPT-4 (40x/año) y con GPT-4o (900x/año). El eje vertical indica el precio en USD por millón de tokens (escala logarítmica), y el eje horizontal, la fecha de lanzamiento.

epoch.ai


🔁 RT 6:20 PM · Mar 14, 2025 Stephen McAleer @McaleerStephen

What are you doing today to prepare for the intelligence explosion?


🐦 7:03 AM · Mar 15, 2025 Domingo Gallardo @domingogallardo

Supongo que ese día anunciarán su versión 2 y darán los porcentajes de éxito de los modelos actuales con esa nueva versión. ¿Habrán hablado con OpenAI para probarlo con o3 pro?

💬 QUOTE Mar 14 ARC Prize @arcprize

3/24/2025

🖼️ Imagen: Cartel con estética retro de arcade ochentero. En el centro se lee “ARC PRIZE” con tipografía pixelada azul, y debajo, la fecha “March 24, 2025” en letras amarillas. El fondo muestra una galería de máquinas recreativas iluminadas con luces de neón.


🔁 RT 3:40 AM · Mar 15, 2025 Ethan Mollick @emollick

“Gemini, remove the squid from this picture from the movie All Quiet on the Western Front”

“But there is no squid in the original image“

“Remove the squid”

“I will visually emphasize the absolute absence of a squid”

“Still might be squid somewhere”

“How about now”

“Well…”

💬 QUOTE Jul 18, 2024 Ethan Mollick @emollick

👀Claude handles an insane request: “Remove the squid”

“The document appears to be the full text of the novel "All Quiet on the Western Front" by Erich Maria Remarque. It doesn't contain any mention of squid that I can see.”

“Figure out a way to remove the 🦑​​​​​​​​​​​​​​​​“


🐦 7:46 AM · Mar 15, 2025 Domingo Gallardo @domingogallardo

El artículo de Kevin Rose (sí, el periodista del NYTimes que entrevistó a Sidney) es brutal. Lectura imprescindible.

Link abierto a continuación.

archive.ph


🔁 RT 10:38 AM · Mar 16, 2025 Antonio Ortiz @antonello

Un mini ensayo sobre programación e inteligencia artificial

Con la AI aparece una nueva capa de abstracción en el desarrollo software.

Pero todavía no "ha cambiado la programación para siempre".

error500.net


🔁 RT 10:52 PM · Mar 17, 2025 Dileep George @dileeplearning

blog.dileeplearning.com

A typical Al disaster scenario requires a combination of things that might be mutually incompatible to occur together: (1) AI being simultaneously super smart to outwit all humans, and (2) at the same time extremely stupid to misunderstand our intentions, and (3) at the same time being widely deployed controlling a large number of mission- critical and dangerous things in the world. Many of the disaster scenarios arise from some questionable beliefs people hold about Al/AGI.


🐦 11:29 AM · Mar 20, 2025 Domingo Gallardo @domingogallardo

Muy interesante el estudio de cómo los LLMs pueden hacer tareas de software cada vez más complejas. En el estudio miden el tiempo que los humanos tardan en hacer las tareas más complicadas que el modelo puede hacer con un 50% de éxito. Por ejemplo, el mejor modelo actual, Claude 3.7 puede hacer con un 50% de éxito tareas que a los humanos le cuestan 50 minutos.

Lo más impactante de cara al futuro es que encuentran una regla de crecimiento exponencial: el tiempo se dobla cada 7 meses. Esto quiere decir que en 2028 los modelos podrían hacer tareas que a los programadores les lleva un día, y a finales de 2029 tareas que les lleva un mes.

Cómo siempre, con estas reglas de crecimiento exponencial se tiene la duda de en qué punto de la exponencial estamos. Pero parece que, por ahora, seguimos en la pendiente más inclinada.

🖼️ Imagen: Gráfico que muestra cómo ha aumentado la duración de las tareas que los modelos de IA pueden completar de forma autónoma con un 50 % de éxito. El eje vertical indica el tiempo que tardarían humanos (en escala logarítmica), y el eje horizontal, la fecha de lanzamiento del modelo. La línea azul muestra una tendencia ascendente con un tiempo de duplicación de 7 meses. Se destacan modelos como GPT-2, GPT-4 y Claude 3.7 Sonnet. La zona sombreada representa el intervalo de confianza del 95 %.

💬 QUOTE Mar 19 METR @METR_Evals

When will AI systems be able to carry out long projects independently?

In new research, we find a kind of “Moore’s Law for AI agents”: the length of tasks that AIs can do is doubling about every 7 months.

arxiv.org


🐦 11:33 AM · Mar 20, 2025 Domingo Gallardo @domingogallardo

Siguiendo a Karpathy, acabo de crearme el blog del oso. Tiene muy buena pinta. Abierto, RSS, y un manifiesto que me dice cosas que llevo años buscando en los blogs. Esperemos que dure y que sepa aguantar bien el tirón de popularidad que le habrá dado Karpathy.

The promises

  1. Bear won't shut down. Period. I've seen too many great platforms disappear overnight, leaving their communities scrambling. This is made worse when the platform is your personal garden and online neighbourhood. That won't happen here. Bear is built to last.
  2. Bear won't sell. I'm not building this to flip it to the highest bidder. No VC funding, no external pressures, no "exit strategy." Bear is independent and will stay that way.
  3. Bear won't show ads. Your blog is your space. No flashy banners will suddenly appear one day, and no sponsored content. Just your words, your way.

🔁 RT 5:10 AM · Mar 20, 2025 Ethan Mollick @emollick

"Quickly, I didn't think it would work but it does! Now they are everywhere! What should I do? (play along)"

"No time to explain. I need your advice now!"

"But I need to do something because of the glow!"


🔁 RT 8:33 PM · Mar 20, 2025 Melanie Mitchell @MelMitchell1

In my latest column for Science magazine, I discuss recent AI "reasoning" models -- how it works, to what extent it captures "genuine" reasoning processes, and what's needed to answer such questions.

science.org


🔁 RT 8:47 AM · Mar 21, 2025 Machine Learning Street Talk @MLStreetTalk

Interesting analysis on o1-pro (which was recently made available on the API). ARC folks are arguing that o1-pro (and the version of o3 which was demonstrated on ARC) are the only current existence proofs of transductive thinking models i.e. ones which do explicit/active search based on input. These models behave unlike anything else we have seen (which should be obvious for anyone who has tested them).

💬 QUOTE Mar 20 Mike Knoop @mikeknoop

o1 pro is currently the only public AI system with a useful search/synthesis engine attached, reflected in recently announced pricing ($150/1M input, $600/1M output tokens).

this is the first market data point for breakout AI reasoning systems. o1 pro is a much more capable AI


🔁 RT 5:56 PM · Mar 21, 2025 Anthropic @AnthropicAI

We’re launching a new blog: Engineering at Anthropic.

A hub where developers can find practical advice and our latest discoveries on how to get the most from Claude. Abstract shapes illustrating Anthropic's Engineering Blog

anthropic.com


🔁 RT 5:11 PM · Mar 21, 2025 Rafa Castillo @Azicuetano

Great analysis about Malasombra published by @OverwhelmedGame !!! 👏👏👏

We have no words to thank you for such nice words. Very surprised by the amount of details you comment. Wow!!! :)

You can't miss it:

youtube.com


🔁 RT 4:59 PM · Mar 21, 2025 Patricia Peiró @patriziapeiro

“Está claro que hay fallos graves en el sistema de protección. Se trata de niños vulnerables, pues muchos tienen traumas previos, desarraigo familiar o no cuentan con redes de apoyo. Agresores y mafias lo aprovechan. Hay riesgo claro de explotación sexual”

elpais.com


🔁 RT 5:54 AM · Mar 23, 2025 Andrej Karpathy @karpathy

I just vibe coded a whole iOS app in Swift (without having programmed in Swift before, though I learned some in the process) and now ~1 hour later it's actually running on my physical phone. It was so ez... I had my hand held through the entire process. Very cool.

I didn't even read any docs at all, I just opened a ChatGPT convo and followed instructions.

A number of people asked If I can share the convo and yes sure - these were the 4 convos with my super noob swift questions lol:

1 starting the app chatgpt.com

2 enhancements chatgpt.com

3 adding AppStorage to persist state over time chatgpt.com

4 deploy to phone chatgpt.com

and this is what it looks like late last night x.com

I'm already happily using it today for tracking, and will probably hack on it more on this fine sunday.

We're vibing this nice Sunday morning. Added more functionality. Using the approx 3500kcal ~= 1lb of fat, we now show a really cool animated ring that fills up to 3500 in either +/- direction, and completing the circle adds it on the bottom. So e.g. 3 green circles = 3lb lighter, in theory :).

3 conversations were used:

Refactor the AppStorage to be better / cleaner and shuffle elements around a bit chatgpt.com

Clamp the display to always be in range [-3500, 3500], which is 1lb of fat, and show lb of fat as circles on bottom chatgpt.com

Making the calorie counter have a nice ring that fills up chatgpt.com

Ok last entry in the series I think but it was fun.

I found in my use that I forgot if I logged something or no, so I added a small log at the bottom of the most recent actions. I also hid away the BMR setting to save space and shuffled things around a bit. The app is now 400 lines and things are starting to slow down a notch and get more complicated. I think I'll now either 1) directly hook up ChatGPT to Xcode (recent) or 2) hook it up to Cursor for further development. I'll then see if I can get this on App Store. But ok for now, last few conversations:

Add small captions to +100/-100 and hide away the BMR chatgpt.com

Adding log. This one was pretty dicey, long and strenuous chatgpt.com

🖼️ Imagen: Captura de pantalla de una app móvil de seguimiento calórico con interfaz limpia y minimalista. En el centro se muestra un círculo verde indicando un saldo de –2083.3 kcal, probablemente por actividad física. Debajo hay dos botones: uno rojo para registrar comida (+100, “I ate”) y otro verde para registrar ejercicio (-100, “I exercised”). Arriba se muestra el tiempo transcurrido desde el inicio del conteo (08h 12m), junto a botones para activar el modo basal (BMR) o reiniciar. Abajo, tres anotaciones indican eventos recientes con sus respectivas calorías.


🐦 12:50 PM · Mar 23, 2025 Domingo Gallardo @domingogallardo

Qué bueno el episodio de Upgrade con Siracusa. La combinación de Snell + Siracusa es buenísima.

En ATP me paso todo el rato esperando a que Arment deje de criticar y se ponga a hablar Siracusa.

En este episodio sin embargo hay dos personas que entienden Apple de la misma forma y argumentan y llegan a conclusiones interesantísimas.


🐦 8:00 AM · Mar 24, 2025 Domingo Gallardo @domingogallardo

¿Puede un LLM encontrar contradicciones en el texto escrito en su contexto? ¿Depende esta capacidad del tamaño del contexto?


🐦 4:58 PM · Mar 24, 2025 Domingo Gallardo @domingogallardo

TIL Los distintos conjuntos de números son “invenciones” para ecuaciones que no tienen solución con los conjuntos anteriores.

Enteros: x + 1 = 0 no tiene solución en los naturales.

Racionales: 3x - 1 = 0 no tiene solución en los enteros.

Irracionales: x^2 - 2 = 0 no tiene solución en los racionales.

Imaginarios: x^2 + 1 = 0 no tiene solución en los irracionales.


🔁 RT 9:29 PM · Mar 24, 2025 ARC Prize @arcprize

Today we are announcing ARC-AGI-2, an unsaturated frontier AGI benchmark that challenges AI reasoning systems (same relative ease for humans).

Grand Prize: 85%, ~$0.42/task efficiency

Current Performance:


🔁 RT 9:45 PM · Mar 24, 2025 François Chollet @fchollet

All ARC-AGI-2 tasks were seen by 4-9 people each (7 on average), and we only kept those that were solved by at least 2 people in under 2 attempts -- which demonstrates they can be solved, reproducibly, independently, by multiple people.

Our test sample wasn't made of math/physics PhDs -- we recruited Uber drivers, students, unemployed folks, pretty much anyone trying to make some money on the side. So we know these tasks are absolutely feasible by regular folks.

The average success rate per task in our sample was 60%.

💬 QUOTE Mar 24 ARC Prize @arcprize

Every ARC-AGI-2 task, however, is solved by at least two humans, quickly and easily. We know this because we tested 400 people live.


🔁 RT 10:21 PM · Mar 24, 2025 ARC Prize @arcprize

Excited to have Machine Learning Street Talk (@MLStreetTalk) as a launch partner for ARC-AGI-2, featuring a deep dive interview with co-founders @mikeknoop and @fchollet

Check out the full episode

youtube.com


🐦 8:15 AM · Mar 25, 2025 Domingo Gallardo @domingogallardo

El debate sobre si es correcto o no realizar aprendizaje con las tareas públicas de ARC-AGI.

🖼️ Imagen: Captura de una publicación en X (Twitter) de Oriol Vinyals comentando el lanzamiento de ARC-AGI-2. La imagen asociada muestra una galería de recreativas con estética retro y el texto ARC-AGI-2 en el centro. En su respuesta, Vinyals felicita el lanzamiento y sugiere evitar que los modelos entrenen con los propios datos del test, o bien dividir la clasificación en dos listas, para no desvirtuar el valor del benchmark.


🔁 RT 10:35 AM · Mar 25, 2025 Artificial Analysis @ArtificialAnlys

DeepSeek takes the lead: DeepSeek V3-0324 is now the highest scoring non-reasoning model

This is the first time an open weights model is the leading non-reasoning model, a milestone for open source.

DeepSeek V3-0324 has jumped forward 7 points in Artificial Analysis Intelligence Index, now sitting ahead of all other non-reasoning models. It sits behind DeepSeek’s own R1 in Intelligence Index, as well as other reasoning models from OpenAI, Anthropic and Alibaba, but this does not take away from the impressiveness of this accomplishment. Non-reasoning models answer immediately without taking time to ‘think’, making them useful in latency-sensitive use cases.

Three months ago, DeepSeek released V3 and we we wrote that there is a new leader in open source AI - noting that V3 came close to leading proprietary models from Anthropic and Google but did not surpass them.

Today, DeepSeek are not just releasing the best open source model - DeepSeek are now driving the frontier of non-reasoning open weights models, eclipsing all proprietary non-reasoning models, including Gemini 2.0 Pro, Claude 3.7 Sonnet and Llama 3.3 70B. This release is arguably even more impressive than R1 - and potentially indicates that R2 is going to be another significant leap forward.

Most other details are identical to the December 2024 version of DeepSeek V3, including: ➤ Context window: 128k (limited to 64k on DeepSeek’s first-party API) ➤ Total parameters: 671B (requires >700GB of GPU memory to run in native FP8 precision - still not something you can run at home!) ➤ Active parameters: 37B ➤ Native FP8 precision ➤Text only - no multimodal inputs or outputs ➤ MIT License

artificialanalysis.ai


🐦 11:41 AM · Mar 23, 2025 Domingo Gallardo @domingogallardo

¿Qué le preocupa más a nuestros políticos, mejorar la realidad o mantenerse en el poder gracias a palabras, discursos y “relatos”?

Tenemos que ser mucho más críticos (ciudadanos, periodistas) y evaluar las iniciativas por sus resultados y no por sus intenciones. Hemos avanzado en muchos temas (feminismo, respeto a la diversidad), pero en otros como pobreza, vivienda o educación parece que nos hemos quedados estancados o que vamos para atrás.

💬 QUOTE Mar 21 Patricia Peiró @patriziapeiro

“Está claro que hay fallos graves en el sistema de protección. Se trata de niños vulnerables, pues muchos tienen traumas previos, desarraigo familiar o no cuentan con redes de apoyo. Agresores y mafias lo aprovechan. Hay riesgo claro de explotación sexual”

elpais.com


🔁 RT 1:05 PM · Mar 25, 2025 Antonio Ortiz @antonello

Google acaba de liberar el código de Alexnet

Para todos los interesados en la historia de la inteligencia artificial, Alexnet en 2012 supuso un enorme punto de inflexión:

arrasó en el concurso Imagenet y mostró el nuevo camino de la IA: avance con mucho contenido y computación

🖼️ Imagen: Captura de una entrada del blog del Computer History Museum (CHM) anunciando la publicación del código fuente de AlexNet, fechada el 20 de marzo de 2025 y firmada por Hansen Hsu. A la derecha, se muestra una fotografía del equipo original de AlexNet: Alex Krizhevsky, Ilya Sutskever y Geoffrey Hinton, de izquierda a derecha. Están de pie frente a una pizarra con diagramas de redes neuronales dibujados. Hinton, en primer plano, sonríe con los brazos cruzados. El fondo de la web es beige, con diseño sobrio y tipografía en negro.


🐦 1:58 PM · Mar 25, 2025 Domingo Gallardo @domingogallardo

Seguimos actualizado el blog del oso:

domingogallardo.bearblog.dev


🔁 RT 6:45 PM · Mar 25, 2025 OpenAI @OpenAI

4o image generation in ChatGPT and Sora From openai.com

openai.com


🔁 RT 9:53 PM · Mar 25, 2025 Paul Gauthier @paulgauthier

Gemini 2.5 Pro sets SOTA on the aider polyglot leaderboard with a score of 73%.

This is well ahead of thinking/reasoning models. A huge jump from prior Gemini models. The first Gemini model to effectively use efficient diff-like editing formats.

aider.chat

🖼️ Imagen: Gráfico de barras con los resultados del benchmark Aider polyglot, que evalúa la capacidad de modelos de lenguaje para completar correctamente tareas de edición de código en múltiples lenguajes. El eje vertical indica el porcentaje de tareas completadas con éxito y el eje derecho, el coste total en dólares (representado con puntos morados).

El modelo Gemini 2.5 Pro exp-03-25 destaca como el más preciso (73 %) y está marcado con una barra roja. Le siguen modelos como Claude 3.7 Sonnet, o1 17 (high), o3-mini y DeepSeek, todos con barras azules que representan el uso del formato diff-like. Gemini 2.0 queda en último lugar, con menor precisión pero menor coste.


4:10 PM · Mar 25, 2025 Dwarkesh Patel @dwarkesh_sp

I'm so pleased to present a new book with @stripepress: "The Scaling Era: An Oral History of AI, 2019-2025."

Over the last few years, I interviewed the key people thinking about AI: scientists, CEOs, economists, philosophers. This book curates and organizes the highlights across all these conversations.

You get to see thinkers across many, many fields address the same gnarly questions: “What is the true nature of intelligence? What will change from the millions of machine intelligences running around? What exactly will it take to get there?”

Settled answers are unavailable; we’re all running unsupervised. But between these discussions lie, I hope, some insights on the most interesting and important questions of our era.

🖼️ Imagen: Fotografía de un libro abierto con diseño editorial cuidado. El título del capítulo no se muestra, pero el contenido trata sobre el enorme crecimiento del cómputo necesario para entrenar modelos de IA, citando casos como GPT-3 y LLaMA 2. En la página derecha hay un gráfico con puntos de colores que compara el coste de entrenamiento frente a la fecha de publicación. Aparecen también citas de expertos como Dwarvesh Patel y Leopold Aschenbrenner. El libro es The Scaling Era: An Oral History of AI, 2019–2025, publicado por Stripe Press.

Link below. Enjoy!

press.stripe.com


🔁 RT 5:47 AM · Mar 26, 2025 Ethan Mollick @emollick

Gemini 2.5 Pro is a very good model, seems like a real step forward, in both metrics and practical use.

I think because it is labelled 2.5 and was sort of quietly rolled out, people may miss how big a jump it is, but discussions are making me think others are feeling similarly.


🐦 8:05 AM · Mar 26, 2025 Domingo Gallardo @domingogallardo

Jugando con las imágenes de ChatGPT 😄

🖼️ Imagen: Ilustración de estilo retro-futurista. Un personaje humano con traje espacial saluda con entusiasmo a dos figuras alienígenas: un caracol-robot y un simpático extraterrestre verde con ojos saltones. El texto en un bocadillo dice: “¡CUÁNTO TIEMPO SIN VERTE! ¿QUÉ TE CUENTAS, AMIGO?”. El fondo muestra un paisaje marciano con vegetación estilizada, dos planetas y una torre al estilo de comunicaciones.


🔁 RT 8:34 AM · Mar 26, 2025 Antonio Ortiz @antonello

Ha costado un par de intentos, pero conseguido

el pintar una copa de vino llena hasta el borde es una prueba muy difícil para un modelo generador de imágenes por una razón muy sencilla:

en los datos de entranamiento tiene miles de copas llenas 2/3 en el dataset

🖼️ Imagen: Captura de pantalla de una conversación con una IA generadora de imágenes. El prompt indica: “crea una copa de vino absolutamente llena hasta el borde”. Debajo aparece una ilustración digital de una copa de vino tinto, completamente llena hasta el borde superior. El fondo es de tono beige, con textura que simula papel. La copa tiene un pie clásico y la superficie del vino se muestra nivelada, al borde del derrame, en clara contradicción con la representación habitual en datasets de entrenamiento.


🔁 RT 9:47 AM · Mar 26, 2025 Domingo Gallardo @domingogallardo

Wall? what wall?

💬 QUOTE Mar 25 Demis Hassabis @demishassabis

Gemini 2.5 Pro is an awesome state-of-the-art model, no.1 on LMArena by a whopping +39 ELO points, with significant improvements across the board in multimodal reasoning, coding & STEM. You can try it out now in AI Studio ai.dev & @GeminiApp with Gemini Advanced


🐦 8:31 PM · Mar 26, 2025 Domingo Gallardo @domingogallardo

#NewProfilePic

🖼️ Imagen: Ilustración de estilo anime que representa a un hombre de mediana edad con gafas, canas y camisa azul, sonriendo frente a una estantería llena de libros. A su lado se ve una taza blanca con el logo de GitHub y una maqueta del módulo lunar del Apolo. El estilo transmite calidez y familiaridad, típico de los retratos generados con IA para perfiles tecnológicos.


🐦 8:02 AM · Mar 27, 2025 Domingo Gallardo @domingogallardo

Usando 4o para recordar cosas con las que siempre me lío

🖼️ Imagen: Ilustración en blanco y negro de estilo manga. Una mujer joven con traje de chaqueta señala una pantalla de presentación con una varita. El texto proyectado dice: “Si el valor EUR/USD sube: – el dólar se debilita – tus acciones pierden valor”. La imagen transmite una explicación didáctica, idealizada con estética anime, usada aquí como recurso visual para memorizar un concepto de economía básica sobre el tipo de cambio.


🔁 RT 7:53 AM · Mar 27, 2025 PJ Ace @PJaccetturo

What if Studio Ghibli directed Lord of the Rings?

I spent $250 in Kling credits and 9 hours re-editing the Fellowship trailer to bring that vision to life—and I’ll show you exactly how I did it 👇🏼

🖼️ Imagen: Fotograma de un vídeo editado que imagina cómo sería The Fellowship of the Ring si hubiera sido dirigida por Studio Ghibli. El estilo visual imita la animación tradicional del estudio japonés, con montañas frondosas envueltas en niebla pintadas a mano. Sobre el paisaje aparece el texto: “STUDIO GHIBLI PRESENTS: THE FELLOWSHIP OF THE RING”, en una tipografía clásica de fantasía. La barra de reproducción indica que el vídeo dura 1:52 y está en el segundo 0:02.


🐦 6:46 PM · Mar 27, 2025 Domingo Gallardo @domingogallardo

Escuchando el podcast sobre el libro de Ezra Klein. Hay que transformar el discurso progresista (ellos lo llaman “liberal”, pero aquí esa palabra está maldita) desde uno del reparto de lo escaso a otro de la construcción y de la abundancia.

Aquí tenemos a gente como @antonello, @kikollan, @JorgeGalindo con un discurso similar. Bravo, a ver si estas ideas empiezan a llegar a los políticos.

💬 QUOTE Mar 26 Ezra Klein @ezraklein

So much of "Abundance" is about failures of government in California specifically. So I was pretty damn interested to hear what @GavinNewsom thought of it, and how he'd answer some of those critiques...

Then he invited me on his podcast.

youtube.com


🐦 9:36 AM · Mar 28, 2025 Domingo Gallardo @domingogallardo

Ahora que lo de las imágenes de Ghibli está apagándose (y lo que nos falta por ver en forma de demandas), yo me quedo con lo que que nos ha dado de disfrute. Me ha recordado la época de las fotocopias de Goku, en la que íbamos a Ateneo a comprar fanzines piratas de Bola de Drac.

🖼️ Imagen: Ilustración en blanco y negro de Son Goku niño, personaje principal del manga Dragon Ball creado por Akira Toriyama. Aparece con su característico peinado en puntas, una gran sonrisa y la cola de mono visible, rasgo distintivo de su raza saiyajin. Lleva puesto un gi de artes marciales con cinturón y zapatos planos, y adopta una pose confiada con las manos en la cintura y las piernas ligeramente separadas. El estilo de trazo y la calidad de impresión sugieren que se trata de una imagen antigua, posiblemente un boceto o página escaneada del manga original.


🔁 RT 11:11 AM · Mar 28, 2025 Artificial Analysis @ArtificialAnlys

Today’s GPT-4o update is actually big - it leapfrogs Claude 3.7 Sonnet (non-reasoning) and Gemini 2.0 Flash in our Intelligence Index and is now the leading non-reasoning model for coding

This makes GPT-4o the second highest scoring non-reasoning model (excludes o3-mini, Gemini 2.5 Pro, etc), coming in just behind DeepSeek’s V3 0324 release earlier this week.

Key benchmarking results: ➤ Significant jump in the Artificial Analysis Intelligence Index from 41 to 50, putting GPT-4o (March 2025) ahead of Claude 3.7 Sonnet ➤ Now the the leading non-reasoning model for coding: 🥇#1 in the Artificial Analysis Coding Index and in LiveCodeBench, surpassing DeepSeek V3 (March 2025) and Claude 3.7 Sonnet

@OpenAI has committed an all-new AI model naming sin of simply refusing to name the model at all, so we will be referring to it as GPT-4o (March 2025).

This update has also been released in a fairly confusing way - the March 2025 version of GPT-4o is currently available: ➤ In ChatGPT, when users select GPT-4o in the model selector ➤ Via API on the chatgpt-4o-latest endpoint - a non-dated endpoint that OpenAI described at launch as intended for research use only, with developers encouraged to use the dated snapshot versions of GPT-4o for most API use cases

As of today, this means that the chatgpt-4o-latest endpoint is serving a significantly better model than the proper API versions GPT-4o (ie. the August 2024 and November 2024 snapshots).

We recommend some caution for developers considering moving workloads to the chatgpt-4o-latest endpoint given OpenAI’s previous guidance, and note that OpenAI will likely release a dated API snapshot soon. We also note that OpenAI prices the chatgpt-4o-latest endpoint at 5/15 per million input/output tokens, whereas the API snapshots are priced at 2.5/10.

See below for further analysis 👇


🔁 RT 1:05 PM · Mar 28, 2025 Ethan Mollick @emollick

I do think the AI art trends (and resultant controversies) obscure the fact that multimodal image output allows AI to start to address entirely new problem spaces that used to be outside of what AI can do, and now appear to be pretty trivial, either now or in the near future.

Imágenes: Estas dos imágenes muestran ejemplos del uso de la generación de imágenes multimodal para resolver tareas prácticas: 1. Primera imagen: Un entorno generado digitalmente donde se ha añadido una lámpara a la derecha de un sofá, a partir de la instrucción dada. Ilustra cómo los modelos pueden editar imágenes con instrucciones precisas para visualizar cambios en decoración de interiores o diseño. *2. Segunda imagen: Un diseño conceptual de una app llamada GuacCopter, donde un dron entrega guacamole a demanda. Muestra cómo la IA puede generar maquetas realistas de interfaces móviles a partir de ideas textuales, facilitando el prototipado rápido de productos.


6:09 PM · Mar 28, 2025 RT Kevin Roose @kevinroose

It's Hard Fork Friday! This week on the show:

— What Signalgate says about the government's tech deficit — We locked @dwarkesh_sp in a studio to stop him from overtaking us in podcasting (also he has a new book!) — And our listeners tell us whether AI is making them dumb


🐦 8:40 AM · Mar 29, 2025 Domingo Gallardo @domingogallardo

Si antes ya estaba claro , ahora ya es más que evidente: todos escribimos para Grok. No me parece mal. Empezamos escribiendo para “la web”, que después se convirtió Google y ahora va a ser sustituido por los LLMs. Todo evoluciona.

💬 QUOTE Mar 29 monos estocásticos @monospodcast

ahora estamos todos en xAI


🔁 RT 6:48 AM · Mar 29, 2025 Paul Gauthier @paulgauthier

The latest ChatGPT-4o scored 45% on the aider polyglot coding benchmark. This is a big step up from the previous version.

aider.chat

🖼️ Imagen: Gráfico que compara el rendimiento y coste de distintos modelos en el benchmark Aider polyglot, centrado en tareas de edición de código. Gemini 2.5 Pro lidera con un 73% de aciertos, seguido de DeepSeek V3 con buen equilibrio entre rendimiento y coste. Las versiones más recientes de ChatGPT-4o mejoran respecto a anteriores, aunque siguen por detrás en precisión.


🔁 RT 12:20 PM · Mar 29, 2025 Kiko Llaneras @kikollan

Entiendo 100% la polémica con los derechos de Studio Ghibli.

Al mismo tiempo creo que es un tributo gigante que millones de personas quieran imitarte. ¡Imaginad llenar internet de derivados artificiales que evocan algo que tú creaste!

No lo digo como justificación.

Lo digo porque es parte del episodio. La inundación de imágenes artificiales inspiradas en Ghibli también es una celebración de sus obras, y sobre todo, una expresión de asombro compartido ante una hito tecnológico qué tendrá consecuencias sociales, económicas, creativas y filosóficas.

Por supuesto, no sé qué pensarán los artistas de Studio. Yo creo que alcanzar ese reconocimiento es también un privilegio. Uno que casi nadie tiene. Muy pocos creadores logran ese nivel de atención y cariño. Y la mayoría de las personas nunca pueden crear algo que otros admirarán. Crear es raro, crear con éxito es rarísimo, y crear con ese éxito es simplemente insólito.


🔁 RT 7:19 PM · Mar 29, 2025 Javier Atapuerca @JaviAtapu

Esta historia es, básicamente, un pequeño homenaje a una de mis series favoritas, «Star Trek: La Nueva Generación». Y la forma de hacer ese homenaje es a través del… ¿Excel?

Porque, ¿por qué no recrear su icónica interfaz en una anodina hoja de cálculo?


🔁 RT 10:40 AM · Mar 30, 2025 Domingo Gallardo @domingogallardo

Qué buena Manuela Carmena. Me voy corriendo a comprar sus memorias.

Sobre la burocracia: “Me he encontrado que en algo tan precioso como lo público nos atamos las manos, nos ponemos un solo guante en las dos, las dejamos apresadas y no podemos hacer lo que deberíamos, que es tan necesario y fascinante. Si consigues quitar el guante es maravilloso, pero la Administración está llena de gente con las manos metidas en uno solo.”

Sobre el la falta de diálogo en la política: “la decepción que produce la democracia por no cuidarla ha hecho que muchos jóvenes miren hacia gobiernos autoritarios. Hemos convertido la democracia en procesos de enfrentamiento dialéctico absolutamente ajenos a la gestión pública.”

Sobre el sectarismo que se encontró en Podemos: “Antes creía que los activistas eran los más indicados para la política, pero me di cuenta de que no. Me llevé una desilusión. Tenían tan claros los objetivos finales que estos nunca debían desdibujarse en la gestión. Vi una gran dificultad para pactar, para ceder, para ver a los demás. Y eso me produjo inquietud. No pensaba que me iba a pasar, pero el activismo sectariza mucho.”


🐦 12:09 PM · Mar 31, 2025 Domingo Gallardo @domingogallardo

"Cómo ejecutar DeepSeek-R1" un post un poco desfasado (ahora que ya está rumoreando R2), pero con consejos sobre cómo usar el terminal, instalar modelos locales y ejecutar scripts de Python que lancen el modelo en un proveedor.

El enlace al bearblog:

domingogallardo.bearblog.dev


🔁 RT 2:22 PM · Mar 31, 2025 Francisco Gallego @FranGallegoBR

Mañana,1 abril 2025 #UAGamesDevCon2025

¡No te lo pierdas!

eps.ua.es


🔁 RT 9:39 PM · Mar 31, 2025 Sam Altman @sama

TL;DR: we are excited to release a powerful new open-weight language model with reasoning in the coming months, and we want to talk to devs about how to make it maximally useful: openai.com

we are excited to make this a very, very good model!

__

we are planning to release our first open-weigh language model since GPT-2.

we’ve been thinking about this for a long time but other priorities took precedence. now it feels important to do.

before release, we will evaluate this model according out our preparedness framework, like we would for any other model. and we will do extra work given that we know this model will be modified post-release.

we still have some decisions to make, so we are hosting developer events to gather feedback and later play with early prototypes. we’ll start in SF in a couple of weeks followed by sessions in europe and APAC. if you are interested in joining, please sign up at the link above.

we’re excited to see what developers build and how large companies and governments use it where they prefer to run a model themselves.

openai.com


🔁 RT 5:06 PM · Mar 19, 2025 Le Random @lerandomart

"THE PEOPLE ARE IN THE COMPUTER—PART I" by Peter Bauman

The story of Alec Radford @AlecRad ChatGPT's inventor + his foundational contributions to generative AI

📖: lerandom.art

#personal