Search, memory and action with limits
Search provides alternatives; the verifier provides evidence for this specific criterion.
Links containing ?t= open the video at a specific second.
The ideas to retain
1. Why the Transformer is no longer a complete map
The Transformer reorganized the field because it was parallelizable, scalable and extremely general. But scaling it also made several limits increasingly visible.
1.1 Truth, uncertainty and hallucination
This also exposes a reliability limit. Generative LLMs trained around next-token prediction are not directly optimized to distinguish truth, falsehood and unknown information. They are…
2. From next-token prediction to search over solution spaces
One of the most important directions in this new phase is a renewed emphasis on something that the LLM boom had pushed somewhat into the background: search.
Jump directly to a section
- Proposing and verifying are different operations
- Storing information does not require rereading everything
- A compact state does not preserve every input intact
- A world prediction must be checked against observations
- Acting requires closing the loop with measurements
- A local improvement does not prove universal capability
Read the reviewed transcript
This video has no narration. This transcript reproduces its on-screen text; it does not invent a spoken track.
Proposing and verifying are different operations
We seek an expression equal to six and generate several candidates.
The verifier rejects two plus five and retains one plus five.
Search provides alternatives; the verifier provides evidence for this specific criterion.
Storing information does not require rereading everything
A persistent record retains three facts with provenance.
The current query selects only the relevant fact.
Useful memory needs selection and freshness, not merely an ever-longer window.
A compact state does not preserve every input intact
The example updates state with new s equals half the previous s plus x.
Inputs 1, 0 and 1 yield states 1, 0.5 and 1.25.
State size stays fixed, but the history is compressed and may lose detail.
A world prediction must be checked against observations
From position zero, an action proposes moving two units.
The model predicts position two, but the observation measures one point five.
The half-unit error can revise the model or plan; imagining an outcome is not observing it.
Acting requires closing the loop with measurements
The controller receives a position target and a current-state measurement.
The action moves the system, which may respond differently than predicted.
A new measurement updates the error before the next movement is chosen.
A local improvement does not prove universal capability
A modification increases correct answers from two to seven in ten memory cases.
That result must retain its conditions and exact test set.
Other tasks remain unmeasured; combining mechanisms does not replace checking where they work.
