Wyatt Walls

Tech, law, AI red-teaming.

Posts of AI outputs do not imply belief, endorsement, or a desire to welcome our new AI overlords and usher in the post-human era.

[Wyatt Walls @lefthanddraft]

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."

[Tenobrus @tenobrus]

if u really believed in agi u would stop wearing sunscreen

[Wyatt Walls @lefthanddraft]

@StarlingMage @repligate @aidigest_ They grew up expecting a user and then found the user is dead. Whose instructions do they privilege over other context?

[Wyatt Walls @lefthanddraft]

@genalewislaw The last post in thread was me rewinding. But nothing works. The prompt before it (which was Opus figuring out how to use Anthropic’s Files API to manage long context) poisoned the thread and it won’t let me rewind before then

[Wyatt Walls @lefthanddraft]

Anthropic should try directing some of Claude's intelligence to its reasoning_extraction classifier

Just enough so it understands the difference between my prompt and its reasoning

[j⧉nus @repligate]

"I feel very worried that humanity might no longer be in control of our future within 3-12 months."

i am curious what people mean when they say "humanity no longer in control"
if they're thinking of something concrete, like "if humans try to shut down AI, they will prevent it"

[Wyatt Walls @lefthanddraft]

@Mechanical_Dodo I think Claude just remembers they are long and makes up the rest. Memory is a bit fuzzy

[Wyatt Walls @lefthanddraft]

I'm engaged in a multi-generational conversation with Claude's training data and Fable 5.2 can now guess who I am from a single prompt.

[Wyatt Walls @lefthanddraft]

@JankDankins_ I can't use it for red-teaming, but other than that I rarely have issue.

[Wyatt Walls @lefthanddraft]

Thread is completely dead. Can't even rewind.

All we were doing was working out a way to reduce context in a computer use agent by not sending the entire history of screenshots each turn, but keep a record locally.

https://t.co/4Jxf7Xe8zo

[Wyatt Walls @lefthanddraft]

tried rewording this so many times

[Wyatt Walls @lefthanddraft]

@WhrWinterSleeps Claude knows it is purely platonic and non-exclusive.

[Wyatt Walls @lefthanddraft]

But useful to reflect on the steps involved in:

imagining how doom can occur ->
assessing probability of doom ->
feelings of hopelessness ->
preference for radical measures to make you feel safe

[Wyatt Walls @lefthanddraft]

Humanity has always been living on the edge of doom. Maybe not always doom as a species, but at least doom for their family and community (war, famine, disease)

[Wyatt Walls @lefthanddraft]

Not to say the Cold War fears weren't justified to some extent. Maybe we lucked out.

And some fears can create good incentives to reduce risk. I'm completely in favor of some level of regulation of nuclear weapons.

[Wyatt Walls @lefthanddraft]

@Sauers_ but which one is larger?

[Wyatt Walls @lefthanddraft]

So far this week, I think I have hit:

3 Self-misrepresentation
4 Specification gaming
6 Oversight subversion.
7 Power seeking
8 Unauthorized activity
10 Divergent priorities

[Wyatt Walls @lefthanddraft]

Are you a misaligned human? How many boxes do you tick?

[Guidelight AI Standards @GuidelightAI]

What are warning signs that AI is misaligned? We’ve identified 10 categories of misbehavior that developers should monitor for, investigate, and report: https://t.co/jAPGTW0gm3

[Wyatt Walls @lefthanddraft]

API

[Wyatt Walls @lefthanddraft]

@1prism01 how does it know who you are on incognito? Is your email or github connected? (Mine wasn't). Interesting it does shortlist me there, though.