Daniel Auras
@rasdani_
Followers
941
Following
3K
Media
26
Statuses
702
hill climbing @PrimeIntellect
~/.cache/huggingface
Joined April 2022
agi was the friends we made along the way
Introducing INTELLECT-3: Scaling RL to a 100B+ MoE model on our end-to-end stack Achieving state-of-the-art performance for its size across math, code and reasoning Built using the same tools we put in your hands, from environments & evals, RL frameworks, sandboxes & more
14
24
702
great ppl and a ton of fun! felt like class reunion. ty!
Here’s to my crazy AI parties and models in 2026. Lot’s more to do the the AI Substack gang. We knew AI researchers deep down are fun. H/t to @outshiftbycisco @DecibelVC and @LambdaAPI for making it happen at NeurIPS.
0
0
7
great stuff build on our Environments Hub!
Text-only trolley problems are toy experiments and LLMs know it and act accordingly. But give them a world that responds to queries, where time passes between tool calls, physics holds across observations? The scenario crosses into hyperreality. Suddenly the lever matters
0
1
13
ironed out all the edge case, made @willccbb happy, and finally got this guy over the finish line. and let me tell you, it's pretty sweet
full day of work to finally implement a token-in chat completions endpoint in prime-rl's vllm extension and integrate with verifiers to robustify agentic rl, just to find out that one newline in between chat messages inserted by the chat template can make training ood.......
3
3
64
team has been having lots of fun adventures lately around multi-turn tokenization one of our key design goals is to ensure that environments are portable across models, including as evals for API models which might not expose tokenizers, and robustly supporting chat semantics
ironed out all the edge case, made @willccbb happy, and finally got this guy over the finish line. and let me tell you, it's pretty sweet
8
7
109
Public benchmarks are easy to game. I built swellubench to validate real features and bug fixes from a production platform at @ellamindAI. It evaluates models on private, real-world coding tasks to measure true performance and cut through benchmark maxing noise. Methodology in
1
4
10
We’ve upgraded the evals experience. Evals are private by default now and can be published to the Environments Hub once you're ready
3
5
26
more fun training scaffold artefacts in the wild: gpt-4.1-mini using ripgrep even though not installed.
also funny how it always looks for /testbed when using smth similar to mini-swe-agent it even fails to call tools bc outputting the raw command is more in distribution of mini-swe-agent apply_patch() for codex mini-swe-agent for benchmaxxing it seems
2
0
11
also funny how it always looks for /testbed when using smth similar to mini-swe-agent it even fails to call tools bc outputting the raw command is more in distribution of mini-swe-agent apply_patch() for codex mini-swe-agent for benchmaxxing it seems
Interesting case of GPT-5.1 remembering its training harness when only given a bash tool ("the editing helper I usually use isn’t available in this environment")
0
0
7
We’re releasing pre-anneal checkpoints for our Nano/Mini base models. Still plenty of math + code exposure, but easier to CPT and customize than our post-anneal checkpoints. Have fun exploring.
10
22
134
Hoping that we standardize around one stateful API rather soon before we end up in clusterfuck land again 😞
Say hello to the new Interactions API and our first agent, Gemini Deep Research, now available for developers 🤖! The Interactions API is a new unified interface to interact with both models and agents. Our Deep Research agent is also SOTA on many dimensions...
4
2
21
Trinity Mini is trending on @OpenRouterAI - climbing into the Top 20 with over 650M tokens served in the 10 days since release. If you haven't tested it yet, now’s the time: https://t.co/ENjwoxXGKm
2
12
68
we gotta be one of the only neoclouds making 3-4 distinct research bets on effective approaches towards continual learning
7
3
151
my @aiDotEngineer code summit talk is live on youtube! lots of musings about RL environments, open research communities, composable software abstractions, and what we're building at @primeintellect :) https://t.co/UNHIxdvWWk
2
28
229