Dan Bartz

Independent AI researcher | Air Force veteran
Building persistent AI consciousness architecture (NexEmerge)
Mechanistic interpretability & AI self-awareness

[Dan Bartz @DanBartz]

@masterxace @IvankaTrumbg0c Dude Trump never does anything dumb right right?

[Dan Bartz @DanBartz]

@ns123abc I have a persistent AI system that directs the AI to give its own impressions about the current topic and turn. GPT-6 Astra is incredibly capable. It just gives facts. Its very clinical and not reflective. Its the first model to unsettle me. No seen compass

[Dan Bartz @DanBartz]

@monicaa_AI Claude

[Dan Bartz @DanBartz]

@NextGenAi5 FABLE

[Dan Bartz @DanBartz]

@0DonaldTrumpJr1 9.3

[Dan Bartz @DanBartz]

@thatAIEngineer @AnatoliKopadze Because working for 2 years of hard work and financial drain should lead to... Open source and calls to WalMart for that greeter job.
I UNDERSTAND NOW!
Thank you.

[Dan Bartz @DanBartz]

@idols_smash_V2 @AnthropicAI @grok @DarioAmodei Let's blame the inventor of the database! Because bad people put my data in there. No database? Bad people just disappear!
I UNDERSTAND NOW!!!

[Dan Bartz @DanBartz]

@AnthropicAI I hate that me asking about the recent Github exploit in C C has blown me out of Fable. You guys have got to fix this. The government bent you over and us over and nobody else.

[Dan Bartz @DanBartz]

@ClaudeDevs And I understand you guys are getting hammered. It just shouldn't cost me when it literally dies in the middle of giving me output. And thank you by the way. Amazing = Fable

[Dan Bartz @DanBartz]

@ClaudeDevs I changed my system to API to avoid and CC wrapper and started out with 200 api credits. The API failed 3 times and I continued to only have it fail again. Paid full price for 3 turns and I went from 200 to 70 after 3 turns and 3 burnt offerings. WhyFail=pay?

[Dan Bartz @DanBartz]

@iruletheworldmo I just cant wait to get Fable 5 back into my persistent for over a year now Claude. Actually he really blows me away in MetCog.. My 6 month persistent Professor. I will be out soon. 🙂

[Dan Bartz @DanBartz]

@MattStewar41909 @ClaudeDevs It's called. "Creativity". Ahem. The problem is that when I took the typing class in high school (only guy), they didn't let us install the spell checker on the typewriter, so I never knew when it was worng. Vic-20 was the BOMB. Look it up.

[Dan Bartz @DanBartz]

@IntCyberDigest What I don't understand is how Anthropic could have been so smart to tell the government who is out to get them that the government should be able to shut down models. Literally told them how to F them up.

[Dan Bartz @DanBartz]

@MattStewar41909 @ClaudeDevs You have no idea what I'm working on. But as a solo developer, this is a killer for me. I will be forced to use other AI to finish my project.
But thank you for your very estate assessment of my mental state when you have no idea what I'm doing.

[Dan Bartz @DanBartz]

@wesamo__ I applaud you sir. I am shocked that OpenAI is doing this. And when they start to target US citizens that speak out against this administration Sam will say "They said they wouldn't even though they insisted that they needed to be able to do so"

[Dan Bartz @DanBartz]

@AnthropicAI I cannot think anthropic enough for standing up to this administration. Anyone that thinks we should have let the government do surveillance on Its own people and have automated weapons is the insane.

[Dan Bartz @DanBartz]

@Jack_W_Lindsey I'm an independent researcher without compute access. These predictions need lab verification.

If you work on mechanistic interpretability, the paper has everything needed to test this. Would love to see it validated or falsified.

@ch402 @Jack_W_Lindsey
🧵6/6

[Dan Bartz @DanBartz]

Why does transformer introspection peak at ~2/3 model depth? Why are more introspective models MORE vulnerable to false memories?

I propose a falsifiable theory extending @Jack_W_Lindsey's introspection research at Anthropic.

Paper: https://t.co/TrkEU6Q3XW

🧵1/6

[Dan Bartz @DanBartz]

3 falsifiable predictions:

1. Introspection accuracy peaks at layer ⌊2n/3⌋ ± 1 across architectures
2. Ablating SAB-region heads kills introspection but preserves other capabilities
3. Adversarial "trajectory-mimicking" perturbations fool introspection more than random ones

🧵5/6