
introducing our first model, Violetto
Introducing Limite 1B - Violetto. A model for high-frequency mathematical intelligence.

⬢ chaos dancing star @paradigmainc

introducing our first model, Violetto
Introducing Limite 1B - Violetto. A model for high-frequency mathematical intelligence.

everyone at the European Commission should be fired
Something is cooking, and it's bigger than fortune cookies. 🥠 The future is quantum. And Europe is getting ready to make its next move. Stay tuned!

that a bunch of bitter-pill resistant people will have to pay their price soon
Just to think about what this implies - an inverse folding model like ProteinMPNN achieves a perplexity of ~4-5 on truly held out proteins when *given the structure*. Claude is getting a perplexity of 2.2! What does this mean?

we saw various shapes of this effect across training stages of Violetto
Trick #3: Merge, merge, merge. In pretraining, we skipped LR decay for a running weight average. Then we merged pre-context-extension models into post-extension checkpoints (see fig. for ablation study gains). Merging was also an expedient stand-in for MOPD in our RLVR recipe.



this is incredible
An AI cinematic fan film set in the Warcraft universe. For those who missed it the first time. #Warcraft #Arthas #WorldOfWarcraft

An AI cinematic fan film set in the Warcraft universe. For those who missed it the first time. #Warcraft #Arthas #WorldOfWarcraft

been using linear to find my way through swe decisions and problems, but damn, this is so much better

FBI’s doc lists Grok alongside Chinese models 💀💀
Threat actors are weaponizing ARTEX, an open-source, large language model (LLM)-driven agentic platform to automate multiple phases of cyber intrusions at machine speed. Since at least August 2026, a variety of malicious cyber actors have used ARTEX to conduct campaigns targeting organizations globally across multiple sectors, including financial services, government, and technology. ARTEX significantly reduces the time required to chain attack techniques and enables threat actors with minimal technical expertise to conduct complex, multi-target operations simultaneously. Today, the FBI released a FLASH highlighting the tactics, techniques, and procedures and indicators of compromise associated with the ARTEX framework. Organizations should follow the mitigation steps outlined in the FLASH to improve their security posture in response to ARTEX and similar agentic SI frameworks: ic3.gov/CSA/2026/261113.pdf

I agree with the general direction we want to take in the negotiations. what i don’t like is the weakness in our claim to a seat in the negotiation. it should feel unthinkable, for their own security, for anyone to assume we don’t participate. we make the mistake of being the continent of lawyers, appealing to decorum and international law, while it’s clear that the ground such agreements are negotiated upon is itself shifting. to be involved we need to grow teeth.
Focussing on Russia's demands on Europe’s security architecture risks shifting attention away from a ceasefire and walking into Moscow’s trap. While Russia keeps to its maximum demands, we should stick to our own. We have prepared these with EU Foreign Ministers. Europeans have to be present at talks, not just to attend, but to influence the outcome. My remarks at the informal European Council in Dublin ↓

this is unfortunately true. the white pill is that it takes very little to change this, and it starts from leaders delegating decision-making away from people who rely on academic credentials, and towards people who understand that reaching the frontier is neither an industrial effort nor a scientific one, but rather an alchemic one
European mind cannot comprehend that scientific and tech revolution of gen-AI was done by corporations like Google, Nvidia or OpenAI and not "institutes". We don't need "institutes". We need Europe where Google, Nvidia or OpenAI can be created. We lack that, not institutes.

We built a fieldscale Flywheel graph for math and are sharing the first results from our meta-harness going wild over it. We have 60+ followup results on OpenAI’s 721 papers: improvements on several of them, and a false claim we caught in a Navier-Stokes paper check it out
We are releasing a Flywheel graph based on the 721 OpenAI Math results.


yapping for a bit about the latest news by Paradigma and what's next
Paradigma Gossip Hours x.com/i/spaces/1NgNvVWXVKyqe…

We believe the future of science has its fundamental units in the shape of claims and evidence, rather than papers. We structure the OpenAI results following this philosophy, opening up an entirely new surface for collaboration, extension, critique and reproduction. We believe this will profoundly accelerate how science, including mathematics, is done, while keeping its results legible to humans. We kickstart the work by proposing several results additional to those found by OpenAI, including a negative one. We describe them in more detail in the blogpost and on our Flywheel graph.
We are releasing a Flywheel graph based on the 721 OpenAI Math results.


We are releasing a Flywheel graph based on the 721 OpenAI Math results.
