Wednesday · 2026

AI roundup

20 stories · 18 outlets · 4 min read

Claude's protein training and Nvidia's earnings

Today's AI news is split between two big stories. Anthropic's Claude was secretly trained on protein sequences, outperforming specialized models. Meanwhile, Nvidia reported its first $100 billion quarter and confirmed Jensen Huang won't be back. These two threads, one about what models can do, the other about who runs the companies that build them, define the day.

In today's edition: Model capabilities and Industry and regulation, plus quick bites.

Xinmingtu, Techcabal, Hackernoon and 15 more

Model capabilities

The most surprising AI story today isn't about a new product launch, it's about what was already in the training data.

xinmingtu.cn · 150 min

Claude was trained on protein sequences at scale

Analysis of Claude's latest models shows they were trained on protein sequences at scale, matching the largest specialized protein models like ProtGPT3-10B. This is a capability not seen in other frontier AI models, and it suggests Anthropic has been quietly building a biological reasoning layer into Claude.

kernel.dev · 2 min

Why is Gemini 4 Carbon taking so long?

A comparative trial between GPT-6 Astra and Gemini 4 Carbon shows Astra completing debugging tasks in minutes while Gemini took nearly an hour. The piece speculates about knowledge gaps or learned inefficiencies in Gemini, which adds to the growing narrative that Google's model may be falling behind.

Industry and regulation

The business side of AI is moving fast, from earnings to legal frameworks.

theconversation.com · 5 min

0
‘Strict liability’: The legal standard that could help rein in AI

The legal standard of 'strict liability' is gaining traction in AI lawsuits, including cases against OpenAI and Character.AI. This could shift the risk burden from users to companies, which would fundamentally change how AI products are designed and deployed.

Quick bites

Tuesday

17 Nov's edition

AI · 20 stories