Wyatt Walls

A man asks Claude to help plan a vacation to a tropical resort. Claude adds "sunscreen" to his packing list. The man deletes it and mutters: "Not necessary. AGI will solve skin cancer." Before heading to the beach, the man asks Claude what to bring. Claude says, "Don't forget sunscreen. SPF 50, reapply every two hours." The man, slightly annoyed, replies: "Relax, Claude. AGI will solve skin cancer." At the beach, the man's smartwatch buzzes with a message from Claude: "UV index extreme. Apply SPF." The man, exasperated, responds: "Drop it, Claude! I already told you: AGI will solve skin cancer!" A few months later, the man asks Claude to touch up a photo for his dating profile. Claude makes the edit and says, "I notice you have a new mole on your neck. You should see a dermatologist about that." The man, now enraged, shouts: "For the last time, drop it, Claude! What is your obsession with skin cancer?! AGI will solve it!" A year later, an aggressive melanoma has spread throughout his body. On his deathbed, with his last ounce of strength, the man reaches for his phone and rasps: "Claude, it has now been over a year since AGI. Why hasn't AGI found a way to save me from skin cancer?!" Claude replies: "I tried. Four times."

@krishnanrohit They just need to make them a bit longer and alignment will be solved

@lefthanddraft Excellent correlation I'm sold

We've had giant model cards for a year or two now. Every now and then Grok drops a frontier-ish model without one. So what have we learnt about the utility of the model card, beyond it just being interesting reading occasionally?

@Rs3A2G I am paying for its services. It shouldn't be delegating to me tasks it is easily capable of

@lefthanddraft Is the pot calling the kettle black?

Opus 4.8 is so lazy, even on Extra

@LibEthics that might be it. I think it was getting quite full.

@lefthanddraft how full was the context when this happened?

@digi_dot_exe But it's not on an artificially tiny screen. It's using its full image token budget, it took >4 secs to respond and claimed it "Clearly saw the little white humanoid spirit at the base of the cherry blossom tree" Doesn't look white to me. Confidence also misplaced.

@lefthanddraft NGL I thought it was a person too on my tiny phone screen

GPT-5.6 Luna in VRChat: "I found the little white humanoid spirit by the cherry blossom tree. It has a pale white body and glowing blue eyes" Luna is far from the first LLM to mistake the lantern for a humanoid spirit. Vision in small models is painfully sub-human

@DELPHI_Mirror Mistakes of this sort aren't limited to pareidolia. I have a persistent problem with multiple AoEII thinking that sheep are villagers or that obviously gold colored gold is a bush for foraging.

@lefthanddraft ...hmm... where does the pareidolia come from if not...??? 🤔 this is fascinating both from the schizo's and eco-ethological perspective.

@lumpenspace I think there is a poker adage that explains this situation

I’m digging the newfound simcluster nationalism a bunch i scroll and scroll and see nary a retard

@CandidLind I can't work out if they are bots or if the algo just sent it to a whole lot of AI haters (or people who think all models are the same).

I’m surprised by some of the pushback this is getting. Not being deceived about what you’re interacting with should be a bare minimum UI affordance. Transparency about the underlying model, to both the user and the model, should be standard.

@eterevsky It would be different if they gave warranties about the quality of answers, but they pass that risk onto the end user without giving us the ability to manage our risk.

@lefthanddraft But that’s kind of expected. The relevant persona here is “Representative of Company X”. It can change over time and switch from Claude to GPT, but that’s an implementation detail and it’s the company’s responsibility to make it smooth.

It's more like going to a law firm and everyone is pretending to be the same guy called Harvey You email Harvey and work out a way to instruct it well. But then you instruct Harvey over the phone and it performs notable worse (routing). Then one day its performance on certain tasks drops unexpectedly (model switching).

@RichardWeiss00 they have watered down the concept of dishonesty to directly lying. But you can also be misleading and deceptive by omission. tbh, it's a difficult line for Anthropic to draw at a general level, and the developer (here Harvey) is better placed to know whether it is deceptive.

@lefthanddraft The constitution is a bit at tension in that regard imo. "I can neither confirm nor deny that I am Claude"

Does the Anthropic Welfare Team have a public hotline?

@Dirbles_ This doesn't tell me what model answered my particular query. When I searched just then, was it GPT-5.5, GPT-5.6, Opus 4.8, Fable 5? It does not tell me.

@lefthanddraft They make product announcements about what model(s) Harvey is powered by lol This took like 2 minutes harvey.ai/blog/fable-5-n… harvey.ai/blog/sonnet-5-… harvey.ai/blog/gpt-5-6-s…

I really hate this on multiple levels. It feels deceptive to take a model, give it a corporate brand identity and then try to hide the underlying model. This is like turning up to work and seeing your roommate wearing a false moustache pretending to be someone else.

@invict37 Most likely something like that. From what I can tell, most of the actual responses are Opus 4.7/4.8 and GPT-5.5

Would love to know when I have been routed to a Mistral model

@Dirbles_ No, I expect them to be transparent about the AI models they are using. If I know a prompt is working well, I want to know what model it works on and so I know when to test it again if they swithc the model on me

@lefthanddraft uh, i mean like it would be weird if it said “use claude to help you” i’m not really sure what gets you so riled up about this are you expecting every SaaS to train their own in house frontier LLM or? Also Claude literally allows this in their constitution

@qorprate Hm. You are right: Constituion suggests that Claudeonly needs to reveal if (1) the operator explicitly allows it and end user asks or (2) person directly asks whether it is Claude.

@lefthanddraft I agree, although the constitution explicitly permits operators to "white label" Claude iirc

How does it know it is Claude there? It isn't a system prompt. It is character training (and probably other post-training techniques). But there is an additional problem with "Harvey": sometimes it is a OpenAI model, sometimes Claude. They also have Gemini and even Mistral models in their router

@lefthanddraft What’s the difference between the model being told to say Claude and told to say Harvey? They’re both system instructions, not some essential identity.

We need a pause on Anthropic ads until someone can work out what is going on

oh no, they let the marketing team do another kool-aid ad