We are releasing Ultramarine 1 Flash, a frontier-class model built for knowledge work. It is open-source and achieves similar performance to Fable 5, while being 60x cheaper.
It was post-trained by a team of 5 for ~$300k all-in. Available on API now, with weights releasing in 2 weeks.
For centuries, ultramarine was the most precious color in the world. It was ground from lapis lazuli carried out of Afghan mountains and cost more than gold. In 1824, SEIN offered a prize to anyone who could make it. Guimet won with a synthetic version, and it became abundant.
"... you can share the AI response in a markdown file and assume the recipient will not read it but also just hand it to their agent (sometimes this is good for fast agent-to-agent comms) … " reading @paperinstr how we work and this is absolutely brilliant
@xeophon@HamelHusain This is why companies should build their own benchmarks and not rely on public indices. It's the only thing that encourages broadly good models
Introducing Inkling 2 Small. Scoring 31 on the AA Intelligence Index, it's a big jump over its predecessor.
Inkling 2 Small is available for fine-tuning and inference on Tinker today, with full weight release planned soon.
https://t.co/VR02tw73fJ
We ran ~150 async RL runs to see when score centering (SC) is enough, and the short answer is as long as the trainer stays close to the sampler. Small batches, stale data or a high lr pull the trainer away. Extras on top of SC like TIS or KL penalty helped less than we expected:
🤔 Is score centering (SC) all we need for stable async RL?
People often stack truncated importance sampling (TIS) or KL penalties on top of SC. In a fun side project, we tested how much they help.
We found that SC alone often matched on-policy. But when it fell short, adding more corrections doesn't necessarily help. 🧵
We ran ~150 async RL runs to see when score centering (SC) is enough, and the short answer is as long as the trainer stays close to the sampler. Small batches, stale data or a high lr pull the trainer away. Extras on top of SC like TIS or KL penalty helped less than we expected:
We are releasing Ultramarine 1 Flash, a frontier-class model built for knowledge work. It is open-source and achieves similar performance to Fable 5, while being 60x cheaper.
It was post-trained by a team of 5 for ~$300k all-in. Available on API now, with weights releasing in 2 weeks.
A few months ago Surya (@kickingkeys), a designer at Websim (@websim_ai) in New York, posted a project that went viral: “You can just RL a coding model to paint with javascript btw.”
We sat down with Surya and were shocked to learn that he only wrote his first line of code in 2024 and had not heard of RL before January. We wanted to know how someone with that background goes from zero to training models in under two years, and whether anyone else could do that too.
Our top 2 lessons from Surya:
- Anyone can learn how these systems work if you are curious and willing to put in the time.
- The best way to learn is to spend less time asking what AI can do to serve you and more time asking what it can do to serve the people around you.
Some of Surya’s recent projects include training models to paint, setting up AI for a coffee shop owner so he has more time to write, and having agents watch him in the art studio and turn the footage into a reality-TV-style story.
Surya is doing this for the love of the game.
Read our full interview in the thread!
cc @thisisrayouf