What is WOB?
WOB began life as a simple idea relating to peripheral vision. The idea spawned a simple rule, which was tested by modifying a known object detection algorithm, to embody the rule into the code.
It worked very well. See the video experiment elsewhere. The rule was then tested alongside another idea around object detection. Due to an experimental mishap, the rule looked like it was the only real factor in the experiment. It wasn't - it was only 1 factor, but further testing was (mistakenly) justified. The rule was tested against other code repos which had code which processed a lot of data, and the rule performed well when embedded into the code. The mistake ended up being the right one to make.
This was done using LLM-based code scanning techniques. Not for optimising in terms of speed, but in terms of focussing the workload on data where it is worthwhile.
The rule was shortened to 5 words - they are not disclosed, but a useful and useless analogy might be "LAZINESS MADE SMARTER".
A second fortuitous accident - when amending code, my LLM stupidly misinterpreted its prompt, which went unnoticed. The rule was tested and sparingly applied at every level in the test code base (when appropriate), and the results were even better. It took a while to analyse the "fault" - again, this was the right mistake to make.
Weirdly, the rule itself made its way into the LLM prompts (lack of experimental rigour) - this tainted some experiments, but when they were re-run, the rule had somehow rejected the experiment branches that were actually dead ends - without running them - somehow, it knew!
The WOB method was born - a generic technique in code factoring to improve data-heavy code. Not speed - efficacy - making laziness smarter.
WOB has been tested on a number of known algorithms, and has produced positive results. See my assorted repos for more information. Not all codebases are suitable. Monte-Carlo searches are already highly optimised, as are a lot of long-established codebases, such as Stockfish.
An additional, novel technique, codenamed Nitpicker, has sometimes found some small, but often significant additional improvements. The Nitpicker technique is not disclosed.
WOB BRAIN
The WOB project now has a long-term goal - WOB BRAIN - the rule, embedded in a neural network.
Early experiments were attempts to show learning and intelligent behaviours. They failed.
BRAIN 1 - An entirely new network substrate, borne of the rule itself. Sadly, that simple rule made BRAIN 1 determine that learning was pointless, so it refused to do so.
BRAIN 2 - Great at battleships, terrible at chess (better at endgames, but couldn't survive to get to that stage).
BRAIN 3 - Traditional graph substrate. It became clear the the original WOB rule was great for code analysis and focussing but for a learning mechanism, it was hopeless. It inevitably determined that things weren't worth the effort. Even with artificial nudging, the rule always won out in the end.
BRAIN 3a - new implementation method, overlay training - all changes to BRAIN3a were eventually rejected due to the rule being enforced. Tiny WOB3a brains were inserted into code, replacing the whole section. Due to the vicious nature of the rule, these worked just as well as standard WOB code factoring techniques, demonstrating that the rule is pervasive if applied in the right way. For WOB Brain to grow, new rules needed to be created. Again, they are not disclosed.
BRAIN 4 - Scrapped - some promising progress was made, in replacing the entire ASLib library (used in automatic algorithm selectors) with a single WOB Brain 4 - classic deferred credit problem came up again. A solution is proposed. Early spin tests indicated that the idea had legs, and this was found to work well with the original WOB Brain 1 - the brand new substrate, so Brain 4 was mothballed.
WOB Brain 5 - back to the original substrate - entirely new to the world, alongside a bold solution to the deferred credit problem and other machine-learning classic challenges. Due to lack of compute resources, the decision was made to focus WOB Brain 5 on real world scenarios, so ASLib was once again chosen since this is a defined problem, and suited to WOB Brain 5. Currently under development.