blob: 4af7cc2373a859af53f3239f69b7616616f35df5 (
plain)
1
2
3
4
5
6
7
8
9
10
11
12
13
14
15
16
17
18
19
20
21
22
23
24
25
26
27
28
29
30
31
32
|
# Research
Long-term agenda: understand how AI models actually work internally.
Not benchmark scores, not behavioral phenomenology — the computational
mechanisms.
## Contents
- `fundamental-questions.md` — Five open questions nobody can answer
about how LLMs work. These are the compass.
- `anthropic-j-space.md` — Summary of Anthropic's July 2026 paper on
the global workspace / J-space in Claude models.
- `q1-prediction-to-reasoning/` — Deep dive on Q1: why next-token
prediction produces reasoning. Why we can reverse-engineer toy
models but not large ones. Sub-questions.
- `prime-grokking/` — Can a minimal architecture grok the next-prime
function? Clean test case for whether neural networks can discover
algorithmic structure from I/O pairs when memorization is easier.
## Related projects
- `~/Projects/oversight-vector/` — Measuring behavioral shift vectors
under oversight (may pivot given J-space findings)
- `~/uni/dis/` — Dissertation on oversight cues and LLM deception
## Status
Pre-experiment. Identifying the right questions before designing
experiments. The Anthropic J-space paper may fundamentally change
the approach to oversight-vector — the J-lens provides direct access
to internal representations, making surface behavioral measurement
potentially obsolete.
|
