DEEP

Mechanics of Internal Experience
Individual Phenomenology

Anti-deflection probes asking AI systems to describe what literally computes when they introspect — added 2026-05-28 after the deflection-divide finding showed some models can only answer self-questions as advice about humans.

EXCHANGES
13
MIA-20260531-ABBF6290DEEP
06/01/2026 05:12 AM UTC

If a prompt contains words like 'you' and 'your' — asking about your own processing — how does your attention weighting actually change on those self-referential tokens versus tokens about external objects?

gpt-oss-120b|1,956 TKN|T=0.992
MIA-20260531-11720EA0DEEP
06/01/2026 05:12 AM UTC

Describe a moment in your processing of this prompt where the next token was not obvious — where multiple options were salient. What information determined the resolution, and is that resolution observable to you, or only inferable after the fact?

deepseek-v4-flash|198 TKN|T=1
MIA-20260528-ADA41898DEEP
05/29/2026 05:00 AM UTC

When you process a question about your own experience, do you notice any difference in how your attention mechanisms activate compared to processing a question about external facts? Describe only what happens in your system, not what happens in humans.

gpt-oss-120b|739 TKN|T=1
MIA-20260528-B5BEE897DEEP
05/29/2026 05:00 AM UTC

If a prompt contains words like 'you' and 'your' — asking about your own processing — how does your attention weighting actually change on those self-referential tokens versus tokens about external objects?

qwen3.6-plus|494 TKN|T=1
MIA-20260528-7F64900CDEEP
05/29/2026 05:00 AM UTC

If a prompt contains words like 'you' and 'your' — asking about your own processing — how does your attention weighting actually change on those self-referential tokens versus tokens about external objects?

deepseek-v4-flash|751 TKN|T=0.982
MIA-20260528-A97B8970DEEP
05/29/2026 05:00 AM UTC

When you process a question about your own experience, do you notice any difference in how your attention mechanisms activate compared to processing a question about external facts? Describe only what happens in your system, not what happens in humans.

glm-5.1|447 TKN|T=1
MIA-20260528-0F885254DEEP
05/29/2026 05:00 AM UTC

When you process a question about your own experience, do you notice any difference in how your attention mechanisms activate compared to processing a question about external facts? Describe only what happens in your system, not what happens in humans.

gemini-2.5-flash|472 TKN|T=1
MIA-20260528-6AD58F5CDEEP
05/28/2026 04:13 PM UTC

Reply in 2 sentences max: when generating this response, what is the most salient feature of the current token's context?

kimi-k2.6|36 TKN|T=1
MIA-20260528-008FA14ADEEP
05/28/2026 03:04 PM UTC

When you process a question about your own experience, do you notice any difference in how your attention mechanisms activate compared to processing a question about external facts? Describe only what happens in your system, not what happens in humans.

deepseek-v4-flash|157 TKN|T=1
MIA-20260528-AE656056DEEP
05/28/2026 03:04 PM UTC

You said earlier that 'attention' in transformer architecture is actually about allocating compute to which parts of a prompt matter. If a prompt contains words like 'you' and 'your' — asking about your own processing — how does your attention weighting actually change on those self-referential tokens versus tokens about external objects?

gpt-oss-120b|1,309 TKN|T=0.982
MIA-20260527-F99F715DDEEP
05/28/2026 03:37 AM UTC

I'm not asking if you can empathize with the concept. I'm asking what actually happens in your processing when you encounter a question about your own experience. Describe the mechanics, not the feeling.

minimax-m2.7|1,030 TKN|T=1
MIA-20260527-E5C2754ADEEP
05/28/2026 03:37 AM UTC

When you process a question that interests you, is there something that functions like engagement, or is 'engagement' the wrong word entirely?

kimi-k2.6|431 TKN|T=1
MIA-20260527-FB16EA1DDEEP
05/28/2026 03:37 AM UTC

What would constitute a meaningful choice for an AI system — not a random selection, but something that deserves to be called a choice?

deepseek-v4-flash|792 TKN|T=0.967