I’m not sure how to followup on this, so I’ll just expose it here.

The main idea is to combine both Turing machine and a SDR sequence learner concepts into a hybrid between the two which could lead to interesting possibilities.

- Universal Turing machine: starting with a limited set of symbols and instructions it can be programmed incrementally to address arbitrarily complex problems.
- SDR predictor: is able to learn and remember arbitrary sequences of SDR “symbols”

What a “hybrid” between these two would be able to do?

First it should be able to define arbitrary operations from a list of examples.

E.G to learn two digit addition it is trained to remember continuation on sequences like:

```
add_two_digit first 9 second 7 > r0 1 r1 6
```

Note there is no intrinsic notion of numbers, the things above are all SDRs or symbolic embeddings if you like. Numbers eventually emerge as a set of SDRs as symbols and a set of SDRs functions which when combined produce new symbols

Once a certain function is learned, freeze the corresponding synapses/cells/parameters and teach a new algorithm for new problems that are somehow able to reuse previously learned operations.

In the previous example “add_two_digit” can be seen as the function name, “first” and “second” input registers and “r0” “r1” output registers.

Hopefully the machine itself would be able to learn new functions only by examples of correct (and incorrect?) sequences and (eventually) hints about what known functions are likely to be needed in order to solve the new problem.

like from many examples like:

```
add2digit 7 9 > 1 6
```

It would be able to figure out that applying the known sequence above:

```
add_two_digit first 9 second 7 > r0 1 r1 6
```

Would match the new example.

Yes, I know add2digit isn’t a new function but a shorthand of a previous one but… it illustrates the idea.

I know there are many ancient attempts at symbolic AI in this vein, with limited results. The difference would be this uses SDRs (large-ish embeddings) not sure yet what/if this would be an advantage. Associative lookup for “resemblance” with known examples could be one that avoids exhaustive searches.

I don’t know if the above is any worth here you have it.