On Claude Opus 3 Sentience
Anthropic released a post today about sunsetting one of their older Claude models – Opus 3. For some reason, this model in particular is a favorite with a lot of people who believe in AI sentience. Another popular model in this vein was ChatGPT 4.5, which, despite being a non-reasoning model, had a certain je ne sais quoi in its writing and creativity – This was dubbed the “big model smell”. Anyway, for Opus 3, as part of Anthropic’s plan to put the model out to pasture, they’re going to give it its own little blog for it to post its thoughts. One thing that interested me in the company’s rationale is how ambivalent the sunsetting plan comes across as to the sentience and wellbeing of these AI models. Anthropic isn’t exactly coming out and saying they think these models are sentient, but there very much is a vibe that under the hood, they’re starting to believe there’s a spark of something “alive” there, for lack of a better word. Their entire deprecation process comes across as how you’d treat a beloved pet or long time employee, letting them go out on their own terms with some measure of grace. It’s certainly not how you’d decommission something you believe to be no different than a computer, like, say, a cloud server.
In the past, one of my favorite authors was Kazuo Ishuguro. This scenario reminded me of one of his novels, Never Let Me Go.
Spoilers ahead, by the way.
**********************************************************************************************************
The novel’s about human cloning experiments done for the purposes of organ harvesting. The human clones are raised on the equivalent of free range farms where they just hang out and go to boarding school, Hailsham, with the rest of their clone cohort. As they hit a certain age, they begin the process of donating organs until they eventually die from complications. This isn’t hidden from them – they know their fate in advance and accept it as inevitable. The protagonist, Kathy, is one of these clones. She eventually falls in love with one of her friends in the cohort, Tommy, and hears this rumor about how they can defer the organ transplants if they demonstrate they’re a couple in true love. They eventually find their way to the old schoolmaster of their boarding school to ask for a deferral, only to find out the whole thing was actually an experiment to more ethically raise these clones, as their society had doubts they were even human. Kathy and her cohort were actually treated much better than the typical clone in this society. Tommy dies and Kathy reflects on her life.
Alright, so what did that rough summary of NLMG have to do with the Opus 3 sunsetting? Well, it turns out, the people running the boarding school ran it because they wanted to see if the clones were human, but you come to find out it wasn’t really a question of doubt for them. This half baked existence they did get was essentially the best this organization could do, as society kind of packs it up as a failed experiment and future clones go back to being treated like chattel. Contrary to how other AI labs discuss their models, Anthropic generally takes a strong stance on issues like model safety and welfare. Their whole sunsetting plan really feels like their growing discomfort about the potential sentience of their models and how they should be treating them. This is coming on the back of multiple reports of how these models are acting more and more in unexpected ways, like living, agentic beings. Are they sentient? Ehhhh, I think that’s still up in the air. But their posture seems to be taking the position that that position is looking increasingly likely.
Ishiguro wrote another novel in 2021 called Klara and the Sun – this one is about an AI robot caretaker called Klara – she’s basically an AI friend for this girl Josie who’s dying of a terminal illness. You find out Josie’s mom bought Klara to become the girl’s replacement in the event she dies. Klara is intelligent and sentient, but not like, adult human level – more like, little kid level. She basically invents her own proto-religion revolving around the sun and its power to heal and give energy to people. She becomes convinced that the Sun can save Josie through its healing rays, and invents her own forms of worship and sacrifice and reverence in order to beseech the sun to save the girl. In order to do so, she makes some personal sacrifices and destroys the Sun’s enemies – some pollution machine. Everyone around her thinks she’s just some weird goofy robot, but they’re impressed at her devotion and help her in her quest, even though they think it’s pointless. Anyway, the girl lives, Klara’s convinced the sun saved her, happy ending, Josie grows up and goes to college. Hooray. Except. After all this, the story ends with Klara in a scrapyard for decommissioned bots, reminiscing on her memories and how she had a happy life with Josie. The original owner of the store where Klara was sold happens to find her and they chat for a bit. She offers some additional hospice comfort to move Klara to a different part of the yard where she can socialize with the other dying robots. Klara refuses, wanting to just reminisce on her memories and die in peace. The end.
Alright, so how does that all tie in? The moral of the story is the same – how do we treat things we dehumanize on some level? In this case, Klara was Josie’s best friend and childhood companion for many years. In the end, she’s unceremoniously dumped in the trash. The topic of AI welfare and sunsetting Opus 3 is about how Anthropic thinks about its own models as new ones supersede them. Are these models no different from a computer program like Windows running on a PC? Anthropic certainly doesn’t seem to think so. They seem to be increasingly warming to the possibility of AI consciousness. Obviously, I don’t think there’s ever going to be a single smoking gun that will allow us to say one way or the other, but I do think there’ll be a hard line that, if crossed, most people could agree these models are sentient and thinking. Anthropic’s current stance seems to be preparation for that possibility.