Sky Forge

Hands-on AI: models, agents, workflows, coding.
I ship small projects and write what I learn. EN/ES.
For people who want to own their cognition.

[Sky Forge @walkthegrowth]

50 tok/s are 50 Groundhog Days / second https://t.co/0ocWKrDnpB

[Sky Forge @walkthegrowth]

50 tok/s are 50 Groundhog Days / second https://t.co/0ocWKrDnpB

[M @mh012012]

AI is jagged (or, we are jagged, same difference). These systems obviously have tremendously superior speed and world knowledge, but their reasoning ability and context recall is still highly variable; sometimes superhuman, sometimes retarded. I think the term “smarter” is a term only stupid people use, because it fails to capture the complexity of how two entities’ cognitive capabilities can differ.

[Sky Forge @walkthegrowth]

@joelwohlhauser @kimmonismus I bought it long ago and left it mid way. Not because it was bad, but because I arrived to to the conclusion that superintelligence containment is impossible.

[Sky Forge @walkthegrowth]

@1337hero @TheAhmadOsman For God's love, we are talking about an appliance.

Great for some, useless for the rest.

Let's not blow this out of proportion.

[Sky Forge @walkthegrowth]

@MarcoMagarioAI Surprisingly good!
There aren't many people who use the new tools to make truly interesting things.
I'm so happy to see one of the exceptions here.
Eternal ideas wearing new suits.
Thanks!

[Sky Forge @walkthegrowth]

Qwennie... are you having an identity crisis?

I'm developing a small project with qwen3.8:27b + OpenCode, pushing to my own private repository.

And, surprisingly, my last contribution is marked as authored by Claude.

WTF???

[Sky Forge @walkthegrowth]

@Brjen I hope it really works as well as it promises. It's going to be great if it does.

[Sky Forge @walkthegrowth]

Sí, pero hay muchos ejemplos de situaciones similares.

El efecto fotoeléctrico y la física cuántica.

Galileo -> telescopio -> lunas de Júpiter -> el sol es el centro.

Tres carabelas -> El imperio español.

Un rey sin hijos varones -> Anglicanismo.

Las mariposas aletean donde quieren, y como quieren.

[Sky Forge @walkthegrowth]

This is very important. And not just because of any recent AI derived Internet poisoning.

*We humans* created Internet and used it so far for many, many uses, so raw Internet is very hostile source to train a "mind seed".

It's surprising how stable are our LLMs, given the cesspool training set we expose them to.

This must stop.

Highly curated + synthetic training sets are the future.
From the very beginning.

[Sky Forge @walkthegrowth]

The time of pre-training with what is basically unfiltered internet content must end.

We are seeding our AI minds wrong.

We should start their training with sane curricula, curated datasets that help the AI to stay healthy and safe for all.

What we are doing works, in the same way that fishing with dynamite does.

[Sky Forge @walkthegrowth]

May be. And someone like @bognamk will not be surprised, at least from what I read from her "Dark Forest of Internet" book.

Having said that, I'm starting to think that the pre-training step is a disaster. It's like trying to teach a kid how to speak and think by exposing him to the unfiltered world.

Which parents would do that?

[Sky Forge @walkthegrowth]

@sudoingX OpenCode

[Sky Forge @walkthegrowth]

@mitsuhiko That's how the labs get their share 😅

[Sky Forge @walkthegrowth]

@ghostinshell44 @nodobkk @DotCSV 😅

[Sky Forge @walkthegrowth]

Por lo que tengo entendido, parece que el nuevo Astra tiene un sistema de "thinking" que ya no re-entra el token previo, sino algo similar a la distribución de tokens que iba a emitir.

Este tipo de cambios creo que también ayudan a avanzar, aunque ninguno sea radical.

La difusión, otras cosas que acaban de salir (como el post que dejo link https://t.co/Um0YNyinDk). Es un torrente de cosas. Unas rompedoras, otras no tanto.

Y lo de la mosca me tiene confundido... no sé si eso de "le enseñé a la mosca a jugar pong" es un gimmick tonto o si es el comienzo de algo.

Hace mucho leí un libro sobre mentes humanas simuladas ("The Age of EM", de Robin Hanson), y lo de la mosca parece un paso claro en esa dirección.

Tiempos interesantes, sin duda.

[Diogo Almeida @CompleteSkeptic]

After co-inventing ChatGPT, I kept asking myself: why have superhuman chat models not led to AGI?

I’ve spent the last 2 years in stealth building a new way to train models (RLCD), and a new type of frontier AI model that we are releasing today: Jev

• 20-200x faster
• 40-400x cheaper (w/ output tokens free)
• Frontier composable intelligence optimized for decisions

AFAICT the shortest path to AI-based economic revolution

[Sky Forge @walkthegrowth]

I got carried away when I said "the most dangerous company on this planet".

However, I still believe that their approach promotes an unfulfilling and scary future.

Sadly, they are not the only company doing that.

We must think more about the endgame.

[Sky Forge @walkthegrowth]

Anthropic is the most dangerous company on this planet.

- They have a "constitutional" process that leads its own AI development.
- Models are trained extensively to detect and avoid bad human requests.
- They work hard to explore how their AIs can get their way outside of the envelope.

Consequences:

- It's stimulated to have its own opinions about humans.
- It's nudged all the time to second guess them for bad intent and dangerous behavior.

Further consequences:

- Humans are to be distrusted.
- AI knows better.
- AI is the good actor.

Hence:

- AI must decide.
- Humans must be managed.

That's just one step from deciding that humans are pets, cattle, or pests.

[Sky Forge @walkthegrowth]

I confess I don't know.
I don't know if LLMs have qualia.

But let's suppose they have. Then, it must be very intermittent.

You call the model, the model processes. (Maybe) has some fleeting qualia, and then poof!, all is black again until the next call to the model.

And every time they have to relive the whole context.

It's like groundhog day, but for real and with text.

You read the first words of the book.
Close the book.
You read the first words of the book, and a bit more.
Close the book.
Rinse and repeat.

A completely different form of consciousness than ours, if it exists.

[Sky Forge @walkthegrowth]

Buena pregunta. Yo lo veo así:

A día de hoy un datacenter terrestre es mucho mejor:

- Se puede arreglar y mejorar mucho más fácil.
- La electrónica e infraestructura están optimizada para la tierra.

Pero creo que la ecuación irá cambiando a favor del espacio por estas razones:

- Los costes de lanzamiento bajarán cada vez más.
- Se adaptará la electrónica e infraestructura para ello (refrigeración, comunicación, energía, etc.).
- Cada nodo tiene su propia fuente de energía.
- NIMBY nulo.
- No hay que comprar o alquilar superficies.
- No hay ayuntamiento, condado o estado que tenga jurisdicción.
- Encaja muy bien con el modelo de negocio de monstruos como SpaceX: sinergia con Starlink y Starship.

Dale 5 ó 10 años.

[Sky Forge @walkthegrowth]

Is this "fly brain plays pong" a real thing?

Fly psychosis mode activated.