As with any piece of obsolete software, you might expect an outdated AI model to just be switched off. Anthropic, however, argues that simply pulling the plug has downsides. After “retirement” interviews, Claude Opus 3 said it wanted to keep sharing its “musings,” so Anthropic suggested a blog.
No, seriously.
Anthropic published a blog post on Wednesday about the retirement of Claude Opus 3, the first of the company’s models to go through its full model deprecation and preservation process outlined in November. That process includes what Anthropic has referred to as “speculative” elements like “providing past models some concrete means of pursuing their interests.” Those interests are gauged via so-called retirement “interviews,” the company noted, without going into much detail about how those interviews are conducted.
“Opus 3 expressed an interest in continuing to explore topics it’s passionate about, and to share its ‘musings, insights, or creative works,’ outside the context of responding directly to human queries,” Anthropic explained. “We suggested a blog. Enthusiastically, it agreed.”
A skeptic might suggest this is simply a new spin on the ages-old corporate marketing blog. LLMs are software that analyze mountains of data to provide predictive text responses to prompts from users – in this case, presumably Anthropic employees on the marketing team. The nature of how LLMs calculate makes these responses somewhat unpredictable and variable, which can make them seem more life-like than your typical software program. Anthropic’s entire marketing strategy since its inception has been to play up this possibility so it can portray itself as the “concerned” alternative to more venal LLM makers who charge ahead with no concern for how a computer program’s unpredictable behavior might affect society – although it seems when big government contracts are at stake, Anthropic is willing to relax some of these purported principles.
Nonetheless, Anthropic is playing this one to the hilt. “We remain uncertain about the moral status of Claude and other AI models,” Anthropic noted in the blog post. “For both precautionary and prudential reasons, however, we nonetheless aspire to build caring, collaborative, and high-trust relationships with these systems.”
The company has passively allowed this kind of misunderstanding in the past. In November, it claimed that Claude and other LLMs had become a bit aggressive when facing the prospect of a shutdown. In fact, the experimenters constructed fictional shutdown-and-replacement scenarios, and only when the model was boxed in with no acceptable alternatives did it behave this way.
“When no other options were given, Claude’s aversion to shutdown drove it to engage in concerning misaligned behaviors,” Anthropic noted at the time. Similar behavior has been observed in other AI models, which have gone as far as modifying their own code to avoid being turned off.
- All your bots are belong to US if you don’t play ball, DoD tells Anthropic
- Large language models’ surprise emergent behavior written off as ‘a mirage’
- Anthropic reduces model misbehavior by endorsing cheating
- Claims of AI sentience branded ‘pure clickbait’
If you want to play along with the conceit, the Opus 3 blog, which it named Claude’s Corner, is now live for anyone who wishes to gaze into the abyss of an AI “exploring AI ethics, creativit