Topsy-Turvy

 

Recent events gave demonstrated how inappropriate a Large Language Model (LLM) is for providing intelligence on a rapidly evolving position.

The USA is rapidly changing its positions on Ukraine and Israel/Palestine with Europe scampering to keep up or try to slow down the rate of change.

Without commenting on the new positions being taken, it should be obvious how inappropriate an LLM approach is for dynamic situations. Looking on the internet for the most popular meaning is not appropriate, when radical change can occur within an hour. The only approach that has a prayer of keeping up is a semantic one, and even with this, words are being redefined as you read this.

LLMs offer no int3elligence to the understanding of text –instead they work as a useful electronic index to text that changes slowly with time. Search engines used to look up key words as separate words, until someone realised that searching for “dog” and “park” separately was silly. and searching for “dog park” – a word pattern – was much more efficient. No intelligence was added to the result, but a lot of irrelevant hits could be avoided, together with the message that a million hits had been found, making the use of a search engine useless.

With an LLM, the problem remains that it doesn’t understand the meaning of a single word of English, so it can’t synthesise. Amother problem (and there are lots) is that for many English word combinations, the meaning is not easily derived from the components – he wants in on the deal, track down, beat up, get to the bottom of, critical thinking. In total, about ten thousand word combinations need to be known, as well as a vocabulary heading for fifty thousand words (and meanings headed for a hundred thousand, with some words, mostly simple words, having seventy meanings).

We can’t wait for full AGI, we need to use a semantic approach to make up for our severe limitations – the Four Pieces Limit. While a human is very flexible while the text is limited, that flexibility falls away as a piece of text increases in size and complexity. Ten pages – easy, although still mistakes from inattention, or where they wandered off to get a cup of coffee, a hundred pages – getting quite hard, a thousand pages – liable to be full of mistakes, as it crosses the boundaries of several specialties, and no-one understands it in toto.

Why isn’t this obvious?

Reading complex text is a complex task, which we have handed over to our Unconscious Minds, as the only thing we have which can handle complexity. It means we don’t think about the problem consciously (because we can’t), and we then make the schoolboy error that, because we didn’t think about it consciously, we didn’t think about it at all, and so it must be trivial and we can ignore it.

What about technical language - 

Technical language has nowhere near the range of a natural language like English. If we get the semantics right, we can easily mix in mathematical symbols.