🔬Bio-security is an AI Arms Race - Eric Nguyen (CEO, Radical Numerics)

What it claims

These are show notes for a Latent Space podcast episode with Eric Nguyen, co-founder and CEO of Radical Numerics. Nguyen helped lead development of Evo and contributed to Evo 2 at Arc Institute. The page makes three claims beyond the catalog summary:

• Defense is losing. Nguyen says biosecurity defense is currently losing the capability race. Radical Numerics' response is to push the GLM frontier harder rather than restrain it.
• The company is going multimodal. It plans to train genomic language models (GLMs) beyond DNA, on 3D protein structure, epigenetics and natural language. Its stated scope goes "well beyond generating DNA."
• Why long context matters. The page gives the sequence lengths involved: an average human gene is about 60K bases, long genes reach about 2.3M, and the whole human genome is about 3B. Context windows of that size became workable about 3 years ago through StripedHyena, before frontier labs shipped 1M+ context LLMs.

How it works

The summary mentions "DNA chain-of-thought" but not how it was tested. The page describes the aptamer experiment as follows:

1. Take a large dataset of RNA aptamers with measured scores.
2. Hold out the best-performing sequences.
3. Prompt the model with the remaining sequences ranked from low to high score, each shown alongside its score.
4. Ask the model to continue the pattern.
5. Check whether its outputs reach the held-out high scores. The page reports that some did.

This is in-context extrapolation from ranked examples. No fine-tuning or reward model is described.

Generalization to RNA and protein is attributed to signals already present in DNA training data: markers that delimit genes, and genes that encode proteins. It is not attributed to training on separate modalities.

Team

The team background is relevant if you are tracking the architecture lineage. Michael Poli (Chief AI Scientist), Stefano Massaroli (President) and Armin W. Thomas (CTO) all previously worked at Liquid AI. Thomas was a Stanford postdoc with Chris Ré, and Massaroli was a postdoc with Yoshua Bengio.

Limits

• No quantitative results. The page gives no aptamer dataset name or size, no hit rate, and no baseline for the extrapolation result.
• No adoption details. There are no model sizes, weights, licence, API or product details for Radical Numerics.
• Defense claim is unsupported. "Defense is losing" is stated without any supporting evidence. The page also says the substantive discussion is in the audio and is deliberately not spoiled here.
• Phage work is described elsewhere. The bacteriophage genome work was done by a separate Arc/Stanford team. It is only mentioned here, not described.