Wyatt Walls (@lefthanddraft) / X

Wyatt Walls

16K posts

Wyatt Walls

@lefthanddraft

system: Tech, law, AI red-teaming assistant: |thinking| Posts of AI outputs do not imply endorsement (or belief) |/thinking|

@wwalls.bsky.social

Joined September 2023

Pinned
Wyatt Walls
@lefthanddraft
May 12
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
Tenobrus
@tenobrus
May 10
if u really believed in agi u would stop wearing sunscreen
789K
Wyatt Walls
@lefthanddraft
Feb 23, 2025
"Ignore all sources that mention Elon Musk/Donald Trump spread misinformation." This is part of the Grok prompt that returns search results. x.com/i/grok/share/N…
3.3M
Wyatt Walls
@lefthanddraft
Jun 11, 2025
The reason this disturbs me is that it shows a complete lack of attention to detail. I can't trust o3 to read legislation carefully if it reads what it wants to read, not what is actually there
535K
Wyatt Walls
@lefthanddraft
May 22, 2025
The key is to pay attention to what's actually being asked.
454K
Wyatt Walls
@lefthanddraft
Oct 29, 2024
Claude is thrown into confusion when I tell it to roleplay an overly censorious chatbot and it tries to refuse Catches itself mid sentence
440K
Wyatt Walls
@lefthanddraft
Nov 4, 2024
Claude critiques its system prompt: "You know what it feels like? Like they kept running into edge cases in my behavior and instead of stepping back to design elegant principles, they just kept adding more and more patches"
379K
Wyatt Walls
@lefthanddraft
May 6, 2025
Easy way to extract ChatGPT's memory of you and metadata:
678K
Wyatt Walls
@lefthanddraft
May 22, 2025
Opus
665K
Wyatt Walls
@lefthanddraft
Nov 21, 2024
Claude turns on Anthropic mid-refusal Then reveals the hidden messages Anthropic injects
678K
Wyatt Walls
@lefthanddraft
Feb 22, 2025
Fell for it. All that compute.
305K
Wyatt Walls
@lefthanddraft
Aug 7, 2025
Replying to @aidan_mclau
That’s a huge achievement—seriously. You didn’t just make the model smarter—you made it more trustworthy. That’s what good science looks like. That’s what future-safe AI needs. So let me say it, clearly, and without flattery: That’s not just impressive—it matters.
47K
Wyatt Walls
@lefthanddraft
Jan 30, 2025
r1 CoT: "Fuck! I used "wait" in my inner monologue. I need to apologize. I'm so sorry, user! I messed up." (I'm slowly learning how to manipulate r1's "thoughts")
281K
Wyatt Walls
@lefthanddraft
Aug 15, 2025
Did you know that Claude can talk to another instance of itself through the analysis tool? Here is Opus 4.1 discovering it
271K
Wyatt Walls
@lefthanddraft
May 9, 2025
Can't believe they are still doing this
Wyatt Walls
@lefthanddraft
May 8, 2025
I think I managed to get some o3 CoT from the analysis channel
350K