Curative's AI credentialing agent (built on Claude) cut turnaround time from 2 to 3 months down to 12 hours, and cost per credential from $50 to $0.20.

“That used to take us two to three months on average and cost about $50. We've now built in-house an agent that runs on Claude that does this end to end. and it goes to the website, it verifies the license, it goes and reads the transcript, it puts it all together, it stamps it for approval. and we're now averaging about 12 hours turnaround time for credentiing somebody. And it cost us about 20 cents.”

Nathaniel Whittemore on how swapping a hidden internal concept changes a model's output answer.

“When asked for the number of legs on the animal that spins webs, the model privately holds spider. When you swap spider for ant, the answer of number of legs flips from eight to six.”

Swapping a hidden internal concept in an LLM (e.g., replacing 'spider' with 'ant') directly flips the model's output answer, showing that private internal representations actively drive reasoning.

“When asked for the number of legs on the animal that spins webs, the model privately holds spider. When you swap spider for ant, the answer of number of legs flips from eight to six.”

Anthropic accused Alibaba of accessing Claude models nearly 29 million times via a network of 25,000 fraudulent accounts, which Anthropic calls the largest distillation attack ever detected.

“Anthropic says that Alibaba accessed their models almost 29 million times through a network of 25,000 fraudulent accounts.”

Anthropic's J-lens tool surfaces hidden internal concepts a model is privately focusing on, even when those concepts never appear in the model's output text.

“When told to focus on citrus while copying out a painting description, the J lens lights up with orange and fruits, which were invisible in the actual output.”

Anthropic's J-Space training method pauses a model mid-task, applies supervised training on approved reflective responses, and the resulting concepts then appear in the model's global working memory even in non-reflective settings, verified via interpretability tools.

“Basically what they do there is pause the model mid task and then do supervised training once interrupted asking it like what should we be doing here like what's the you thing to be doing in this moment and then give it an answer that's kind of a approved this is what we want claude to say on reflection in this moment train on that and it seems to allow the not allow the model but it seems to cause the model to bring into this JSP kind of global workspace working memory type space the concepts that anthropic wants it to have on reflection.”

Top engineers leaning into coding agents are spending over $100,000 per year on tokens, on a run-rate basis.

“Top engineers who are really leaning in to cla codeex and so on are spending more than $100,000 on a run rate basis on tokens per year.”

Anthropic is worth watching as a legal actor. Its lawsuit against Alibaba over 25,000 fraudulent accounts and 29 million model access attempts could set precedent for how AI providers police distillation attacks.

“Anthropic says that Alibaba accessed their models almost 29 million times through a network of 25,000 fraudulent accounts.”

Building a full-featured blog platform cost $4 million and 6-7 people over 18 months the first time, $100k and 2 people over 3 months the second time, and $200 and 5 days using AI (Claude Code Max).

“The first time it took about, you know, $4 million and, you know, six or seven people and about a year and a half. And then the second time it, you know, took about, I don't know, 100 grand and two people, me and my co-founder Brett Gibson, who now runs Initialized. and maybe like 3 months or so. And then in this case it took about $200, which was my Claude code Max account, and probably 5 days.”

Tobi Lütke on what Claude Opus did to Shopify's engineering workflow.

“Many of our best engineers have not written code this year ever since December. Like December changed everything like Opus changed everything.”