←
Daanish Khazi1,132 posts
Daanish Khazi avatar

Daanish Khazi

@bertgodel

ceo @paperinstr

◷ Joined February 2018▧ 51 Photos and videos
Daanish Khazi avatar
Daanish Khazi@bertgodel·Oct 14

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.

1:17
Vincent Weisser avatar
⟳ Daanish Khazi retweeted
Vincent Weisser@vincentweisser·24m

Introducing Model Factory. Train, deploy, and continuously improve your own models on one unified stack. Run on-demand, on dedicated clusters, or bring your own. Own your intelligence. Own your model factory. Own the means of production.

2:12
Daanish Khazi avatar
Daanish Khazi@bertgodel·6h

If gta 6 was a data company it would leak a $270B ARR to The Information

JUST IN: GTA 6 is now expected to make ~$5.2 billion in its first week

Daanish Khazi avatar
Daanish Khazi@bertgodel·Nov 17

huge +1. doesn't make sense to have multi-agent to handle different problem types we went with 130+ versions of a single agent (from accounting to supply chain) instead of many personified bots for our cowork product: paperinstruments.com/work

This by Anthropic seems both right & a condemnation of many “agentic solutions” & harnesses being built at companies. Rather than a sprawl of specialized subagents built by different teams, you often get better & cheaper results using one agent to do it all claude.com/resources/article…

Post attachmentPost attachment
LIORDANOV avatar
⟳ Daanish Khazi retweeted
LIORDANOV@IAMLIORDANOV·Nov 9

"... 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

Daanish Khazi avatar
Daanish Khazi@bertgodel·Nov 9

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.

Alex Shaw avatar
⟳ Daanish Khazi retweeted
Alex Shaw@alexgshaw·Nov 8

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

Daanish Khazi avatar
Daanish Khazi@bertgodel·Nov 5

New research from Paper Instruments (@qifang804) and Berkeley Sky Lab (@shulynnliu, @shagarw21) exploring the limits of score centering for async RL

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:

Post attachment
Shu Lynn Liu avatar
⟳ Daanish Khazi retweeted
Shu Lynn Liu@shulynnliu·Nov 5

🤔 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. 🧵

Post attachment
Qi Fang avatar
⟳ Daanish Khazi retweeted
Qi Fang@qifang804·Nov 5

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:

Post attachment