With more time, could it explore more?
Reading The Bitter Lesson made me ask what a method could actually do with a larger computation budget.
Another chance to think
What changes when a problem-solving system receives more computation? Does it return the same answer later, or examine candidates it has not yet seen?
That question stopped me while reading Richard Sutton’s 2019 essay The Bitter Lesson. Looking back at AI research, Sutton argues that general methods able to use growing computational resources have proved stronger over the long run than approaches built around explicit human domain knowledge.
He emphasizes search and learning.
The essay makes my impulse to supply more good answers feel a little uncomfortable. While working to put known answers into a system, how much effort would I put into enabling it to find alternatives?
This 2019 argument is not a guarantee that longer computation wins on every present-day task. I want to bring one of its questions into how I think about a design.
Supplying an answer and finding one
Take a simple hypothetical choice among plans. One method follows priorities written in advance by a person and selects immediately. Another generates possible plans and compares them using the same criterion.
The first may do better right away. Priorities written by someone who understands the problem can be useful. The second could waste time generating poor candidates or comparing them by the wrong measure.
Now suppose we give both more computation in the same execution environment. What can the fixed-priority method do with it? Does the candidate-comparison method have room to examine plans it has not yet considered?
The ability to search is not sufficient. We would need to see whether it repeatedly generates the same candidates, explores meaningfully different ones, and makes better choices with the added computation. I have not implemented or measured this example.
That is the question I want to keep. Is the answer we supplied correct, and does the structure allow further discovery when resources become available?
Chess offers an image for it: writing down another piece of advice about a good move, or enabling more continuations to be examined. Those are different places to spend effort.
Does that leave nothing for people?
It might sound like we should specify nothing. But the rules of chess and a preference for a particular move have different jobs. Legal movement and a judgment about promising moves cannot simply be removed together.
The hypothetical planning system still needs mandatory conditions and a criterion for a good result. Producing many candidates does not excuse violating those conditions. People have decisions to make when opening a space to explore.
The work seems to move: from supplying each result in advance toward defining a problem that permits comparison and examining how results are distinguished.
This is a design question I am extending from Sutton’s essay, not a successful implementation I can claim. A system working harder against the wrong criterion has no reason to head where I hoped.
Beyond the moves I know
Real tasks also have limits on waiting time and resources. A method’s long-term scaling potential does not grant an unlimited response time today.
I would therefore compare within the computation budget we could actually allow: what result it produces there, and what changes with a little more. Without evidence of improvement, that part should remain unknown.
I picture a sheet beside a chessboard listing moves I favor. It makes me pause between adding another line and enabling a continuation I have not written down.
If there is a little more time, I would like it to explore beyond my list.