{"id":47870,"date":"2025-05-27T15:24:21","date_gmt":"2025-05-27T19:24:21","guid":{"rendered":"https:\/\/mjtsai.com\/blog\/?p=47870"},"modified":"2025-06-03T10:32:54","modified_gmt":"2025-06-03T14:32:54","slug":"claude-4","status":"publish","type":"post","link":"https:\/\/mjtsai.com\/blog\/2025\/05\/27\/claude-4\/","title":{"rendered":"Claude 4"},"content":{"rendered":"

Anthropic<\/a> (Hacker News<\/a>, MacRumors<\/a>):<\/p>\n

Claude Opus 4 is the world’s best coding model, with sustained performance on complex, long-running tasks and agent workflows. Claude Sonnet 4 is a significant upgrade to Claude Sonnet 3.7, delivering superior coding and reasoning while responding more precisely to your instructions.<\/p>

[…]<\/p>

Both models can use tools—like web search<\/a>—during extended thinking, allowing Claude to alternate between reasoning and tool use to improve responses.<\/p>

[…]<\/p>

Both models can use tools in parallel, follow instructions more precisely, and—when given access to local files by developers—demonstrate significantly improved memory capabilities, extracting and saving key facts to maintain continuity and build tacit knowledge over time.<\/p><\/blockquote>\n\n

Simon Willison<\/a> (Hacker News<\/a>):<\/p>\n

\n

Anthropic publish most of the system prompts for their chat models as part of their release notes<\/a>. They recently shared the new prompts for both Claude Opus 4<\/a> and Claude Sonnet 4<\/a>. I enjoyed digging through the prompts, since they act as a sort of unofficial manual for how best to use these tools. Here are my highlights, including a dive into the leaked tool prompts<\/a> that Anthropic didn’t publish themselves.<\/p>\n<\/blockquote>\n\n

Carl Franzen<\/a> (via Dare Obasanjo<\/a>):<\/p>\n

As Sam Bowman, an Anthropic AI alignment researcher wrote on the social network X under this handle “@sleepinyourhat<\/a>“ at 12:43 pm ET today about Claude 4 Opus: <\/p>

“If it thinks you’re doing something egregiously immoral, for example, like faking data in a pharmaceutical trial, it will use command-line tools to contact the press, contact regulators, try to lock you out of the relevant systems, or all of the above.”<\/p>

[…]<\/p>

While perhaps well-intended, the resulting behavior raises all sorts of questions for Claude 4 Opus users, including enterprises and business customers — chief among them, what behaviors will the model consider “egregiously immoral” and act upon? Will it share private business or user data with authorities autonomously (on its own), without the user’s permission?<\/p>\n

[…]<\/p>\n

Bowman added: […]<\/p>\n

TBC: This isn’t a new Claude feature and it’s not possible in normal usage. It shows up in testing environments where we give it unusually free access to tools and very unusual instructions.”<\/p>\n<\/blockquote>\n\n

Peter Steinberger<\/a>:<\/p>\n

I asked Claude 4 what new API’s in macOS 15 could be beneficial… and it got me REALLLLLLY excited. Asked it for links. It chugged a long for minutes and then…<\/p>

“Based on my research, I need to correct my earlier statement.” LOL<\/p><\/blockquote>\n\n

Previously:<\/p>\n