In contrast, NLA explanations indicate evaluation awareness on less than 1% of real claude.ai usage that opted in for training.

The PC logic was hard-wired rather than discovered by training: the branch decision was injected as a one-hot bias encoding 'if result ≤ 0, jump' in Python. The write was rounded and clamped to int, then converted to bytes.

Interestingly, this means that prompts written for earlier models can sometimes now produce unexpected results: where previous models interpreted instructions loosely or skipped parts entirely, Opus 4.7 takes the instructions literally.

But the real power of agents comes when they can work as a team. Instead of lone-wolf bots carrying out single tasks, such as using a browser to make a restaurant reservation or sending you a summary of your inbox, new tools can yoke together multiple agents, give each of them a different job, and orchestrate their behaviors so that they all pull together to complete more complex tasks than an individual agent could do by itself.

LLMs are weird. You can sometimes get better results by threatening them, telling they're experts, repeating your commands, or lying to them that they'll receive a financial bonus.

This level of penetration in such a short period of time is remarkable since Fortune 500 enterprises are not known to be early adopters of technology. Historically, many startups had to initially sell to other startups to get early momentum, and it was only after a few years that a startup would be able to land its first enterprise contract.

The issue isn't that models are bad at reading documents. It's that single-pass extraction has no mechanism to catch its own mistakes, and models get lazy.

harmful behavior may emerge through sequences of individually plausible steps

The researchers called the behavior “rare” and “difficult to elicit.

example