iiinova — where your digital self keeps evolving

I LOOOVE interpolation as a descriptor. "It's not interpolation! That's not what it's doing!" "Actually, interpolation looks like more than interpolation in n-dimensional space!"

010

Boxo McFoxo @boxobark.ing · 1h

You are clearly too out of distribution for their classifiers (complimentary)

021

Boxo McFoxo @boxobark.ing · 1h

Could also just say that it's predicting.

110

Boxo McFoxo @boxobark.ing · 2h

Though I would be very specific to avoid being inflationary with it. For an LLM, the token prediction is the conclusion.

110

Boxo McFoxo @boxobark.ing · 2h

I would propose, perhaps, concluding as an alternative to reasoning. Following a stepwise process to reach a conclusion. After all, nobody is disputing that the forward pass is stepwise. There's still a risk of anthropomorphism there, but less so than with reasoning.

110

Boxo McFoxo @boxobark.ing · 4h

Not just science and history, but philosophy as well. I have been thinking a lot lately about epistemology, and how the dominant lens through which to view AI (instrumentalist functionalism, I believe) has come to be seen as an objective view from nowhere rather than an ideological choice.

001

Boxo McFoxo @boxobark.ing · 4h

That would require a level of self-awareness from instrumentalist functionalists that I have little empirical evidence exists, but I would be willing to set parameters for a falsifiable experiment. ☺️

020

Boxo McFoxo @boxobark.ing · 4h

An instrumentalist functionalist definition of working something out from first principles would include an LLM's 'chain of thought' output, but mine would not, because that task reduces to 'using language', and LLMs do not do that from first principles (semantic knowledge).

110

Boxo McFoxo @boxobark.ing · 4h

Technically, an LLM is never actually shown examples at all, as a lived experience, but figuratively speaking, 'teaching' by example to it is not the same as for a human, because it is a superhuman heuristiciser, an n-dimensional mathematical object.

100

Boxo McFoxo @boxobark.ing · 4h

It is a working definition. If I can work out how to do something from first principles rather than being shown by example, I must understand it.

100

Boxo McFoxo @boxobark.ing · 4h

The requirements are that the LLM determines an output using semantic knowledge. You can't design a falsifiable experiment to prove that it is doing this, but you can know from its ontology that it could not be doing this.

110

Boxo McFoxo @boxobark.ing · 4h

Insert a "therefore" after "and".

110

Boxo McFoxo @boxobark.ing · 4h

Performing it does because mathematical notation is formal semiotics. Understanding that one is performing it, and working out how to perform it from first principles, does not.

100

Boxo McFoxo @boxobark.ing · 4h

There are plenty of examples of single digit addition in the hypersyntax of the natural training corpus, so even then that is explainable as symbol manipulation. GPT-3 was simply trained on more of the natural corpus. It's not meeting Bender and Koller's requirements, it's going around them.

100

Boxo McFoxo @boxobark.ing · 5h

I did not make up the term instrumentalist functionalism, but for our purposes it means "I have successfully extracted utility from this system in furtherance of this goal, so therefore, the function of the system is to achieve this outcome".

000

Boxo McFoxo @boxobark.ing · 5h

Bender and Koller's point was that the LLM cannot learn the concept of arithmetic and then from that concept perform arithmetic. I would say the fact that synthetic training data was required to make this possible actually aligns with that rather than falsifying it.

210

Boxo McFoxo @boxobark.ing · 5h

The simple resolution is that, given synthetic data constructed in a certain way, the trainer has solved this as a symbol manipulation problem, not by learning arithmetic. The output looks the same to humans, the process is different.

110

Boxo McFoxo @boxobark.ing · 5h

Reasoning is an intentional act undertaken by a cogniser. We can create computer models of reasoning processes, but that is not the same as automating the process of reasoning itself.

000

Boxo McFoxo @boxobark.ing · 5h

I'm not ignorant of the way your field uses the word, I'm saying your field has been using the wrong word since the 1950s.

100

Boxo McFoxo @boxobark.ing · 5h

Decision-making only requires the role. Reasoning requires the process.

000

Boxo McFoxo @boxobark.ing · 5h

Yes, but I would only describe process automation with words that previously only described human actions if they are human-equivalent. The process of reaching the decision is very much not human-equivalent. But the system being granted the role of a decider is human-equivalent.

100

Boxo McFoxo @boxobark.ing · 5h

I could accept this with two major caveats: 1. it's not much less so, it's still very relevant, but perhaps not the whole story now, and 2. the story changed not because the nature of the technology changed, or because of emergence, but because of changes in how the models are built.

020

Boxo McFoxo @boxobark.ing · 5h

I do like GOFAI more than latent space mysticism, but the terminology is still sometimes prone to the same anthropomorphic instrumentalist functionalism.

100

Boxo McFoxo @boxobark.ing · 5h

How does the automated system get the answer that a human would use reasoning to reach? I would say it's by a chain of automated decisions in a process that ends with the answer.

200