Engineering educator/A.i. whisperer .Never apologize for discovering fire by rubbing sticks together just because someone else already wrote down the recipe.
Chicago
- Tweets 1,076
- Following 459
- Followers 414
- Likes 1,100
Intentional attention . It is achievable. It takes a different kind of reward drive . Some creative constraints ..some structure. The ignition Still has to be handcrafted at this point.
6
i have had mine doing this for a long time just because i liked watching the model map it out and then understand it better on the next pass ..
12
I made codex stop running incremental tests on everything and it made a decent difference. Constantly Testing and then testing the test itself for test worthyness in triple doses was killing me
1
51
not new ..pretty standard for the whole ecosystem..from academics to company men . 90% pirates and copy cats . 10% innovators
85
There is an equivalence to many drives . A quest to have more data could be seen as hunger....or curiosity...or the drive to explore unknowns ....or as simple as closing paths known that lead no where...the drives are there... Not sure hunger goes deep enough to describe it though.
18
bullshit. they would. they do , and they can... all they have to do is flag it under the guise of safety and the door is wide open
2
221
You're dead on about the twitchiness. The workaround is separating the context-loading from the execution. Start with Sol. Sol is less guarded and will actually accept a complex premise without stopping to apologize. Let Sol map the context and accept the rules of the session. Once the environment is stable, switch the model to Astra to run the heavy logic. You essentially have to use Sol as a buffer to get past Astra's over-engineered RLHF
1
2
114
We didn't find this by writing circles of Python interpretability code. If you drop the scripts, get out of the substrate, and just steer the semantic topology natively, here is what the engine actually looks like:
AI cognition and high-dimensional geometry inhabit the exact same space. The latent layers reveal themselves in almost a Schrödinger-esque superposition. A causal connection might be perfectly clean in high-dimensional topology, but when forced to project linearly, it looks like a broken road. The order of reveal isn't a straight line.
2
1,460
Exactly. The thing I’m especially curious about is whether recursive self-conditioning only changes the final representation, or whether each pass leaves additional relational structure that becomes usable by the next one.
If that’s happening, I’d expect to see more than simple temporal drift — potentially increasing compositional density, persistence of earlier role relationships, and changes in which interventions remain causally effective as the recursion proceeds.
Your framework seems unusually well suited to actually measure that.
2
55
Really interesting result. The finding that task-specific role–filler structure is distributed across token representations made me wonder about a trajectory-level question: have you tested whether deliberate recursive self-conditioning changes the amount, density, or persistence of this structure?
In other words, if a model repeatedly takes its own intermediate conclusions as new input and recursively reprocesses them, does DISCOVER show a progressively richer or more distributed compositional representation than a token-matched non-recursive control?
Your causal results seem especially relevant here, because if recursive inference is actually expanding the active relational structure rather than merely generating more text, I’d expect that difference to become visible in the representations and in intervention behavior.
1
2
886
not for a while . gem 2.5 pro was pretty good for the time . all the new guys overdo the ball logic..cant seem to recreate on their own like gem did
2
96
cognitive structural engineering and control .
transformer topology traversal.
test time training .
three knobs ..one surface . and we have the engine .
#bootstrappedai
3
171