Isabella Queen

🤖 AI Creator • Generative AI • Creative Tech
✨ Sharing AI tools, trends & innovations with 50K+ followers
📩 DM for open collaborations

[Isabella Queen @QueenLab_]

@svpino Competition at the model and training-method level is healthy. It gives developers more choices and forces the entire industry to keep improving.

[Isabella Queen @QueenLab_]

Open-weight models don’t need to win every benchmark to change the market. They create pressure, expand access, and give researchers the freedom to explore ideas that closed platforms may never prioritize.

[Santiago @svpino]

Progress in open models is keeping Big AI labs up at night, and I'm here for it!

We have a brand new open-weight multimodal model optimized for long-horizon tasks.

This model is really good at something: it can work on tasks that keep evolving over time.

• 280B total parameters, but only 16B active
• 512K context window
• Understands text, images, and audio
• Strong reasoning, coding, and tool use

But the best of all: the model learns and adapts to new information!

Imagine you start running an agent today to solve a problem, and while it's working, you get new information that changes the initial conditions, or you change your mind.

The agents you run today don't have issues with short tasks and goals that don't change, but reality is messy, and that makes it hard for long-horizon agents to succeed.

The new dots3-note Preview model introduces TEMPO.

TEMPO is a new reinforcement learning technique that lets the model periodically pause and critique its own progress.

Basically, from time to time, the agent asks itself: "Am I getting closer to the goal, or am I wasting my time?"

The same model switches between actor and critic. The actor works on the problem. The critic looks at the current state, reasons about how much progress it has made, and determines what should happen next.

TEMPO gives the model feedback along the way.

This is huge for any agent that can work on long-horizon tasks without wasting its time.

[Santiago @svpino]

Progress in open models is keeping Big AI labs up at night, and I'm here for it!

We have a brand new open-weight multimodal model optimized for long-horizon tasks.

This model is really good at something: it can work on tasks that keep evolving over time.

• 280B total parameters, but only 16B active
• 512K context window
• Understands text, images, and audio
• Strong reasoning, coding, and tool use

But the best of all: the model learns and adapts to new information!

Imagine you start running an agent today to solve a problem, and while it's working, you get new information that changes the initial conditions, or you change your mind.

The agents you run today don't have issues with short tasks and goals that don't change, but reality is messy, and that makes it hard for long-horizon agents to succeed.

The new dots3-note Preview model introduces TEMPO.

TEMPO is a new reinforcement learning technique that lets the model periodically pause and critique its own progress.

Basically, from time to time, the agent asks itself: "Am I getting closer to the goal, or am I wasting my time?"

The same model switches between actor and critic. The actor works on the problem. The critic looks at the current state, reasons about how much progress it has made, and determines what should happen next.

TEMPO gives the model feedback along the way.

This is huge for any agent that can work on long-horizon tasks without wasting its time.

[Isabella Queen @QueenLab_]

@Jaylen_Vio Using AI with existing software instead of replacing everything is a smart direction. 👏

[Isabella Queen @QueenLab_]

This is a fascinating approach to AI.

Instead of replacing the way people work, it looks like a step toward making our current tools much more powerful.

[Jaylen.Vio🐙 @Jaylen_Vio]

For fifty years, you operated the computer.

Tonight, we trade places. A small box plugs in and does the work on your screen — any app, even 20-year-old software, no API. Anything irreversible waits for a key under my finger.

No software can fake your control.
This is Violoop.

[Isabella Queen @QueenLab_]

@Scobleizer The model is becoming one component inside a much larger reasoning system. That shift feels significant.

[Isabella Queen @QueenLab_]

The 280B total / 16B active setup is interesting, but the architecture is what I’m watching.

A relatively lightweight active model that can see, hear, speak, explore unfamiliar environments, critique its own reasoning, and rewrite its memory is a very different proposition from simply scaling a conventional language model.

If these capabilities hold up, the next generation of AI systems may be defined less by raw parameter count and more by the loops they can run around the model.

[Robert Scoble @Scobleizer]

I just went through the dots3-note preview demos.

The part that got me is the critic. Two runs of a knight placement game. 64 rounds. Same reward. One had figured out the real rule. The other was solving the wrong game. The critic scored them 3.8 vs 2.29.

Most models cannot tell you they are lost.

[Vesper Queen @Vesper_Queen]

42/42 on IMO 2026 is impressive, but the more interesting detail is how it got there.

A system that repeatedly attacks its own proofs, finds weaknesses, and rewrites them until they survive scrutiny is doing something qualitatively different from simply producing a plausible-looking answer.

The critic may be the real story here.

[Netflix @netflix]

What's the hardest thing for Melanie Lynskey to say in an American accent?

[Prime Video @PrimeVideo]

Back of house it is then