Getting AI Back on Track

 

“Altman admits scaling isn’t the answer” - how long do we have to wait until Altman admits LLMs aren’t the answer? What it shows is that a large percentage of people working in Computer Science have no understanding of what they are doing when reading or writing. How can that be – they can write programs. The symbols used in programs have a single meaning, so no confusion. But a specification is different – it uses words, and many words have multiple meanings. How do they handle this? Humans have a very limited bandwidth, so most of the parsing is done unconsciously, leaving enough bandwidth to understand the meaning of the message. This is how they can be fooled into thinking an LLM is “reading” the text, when it is doing no such thing.

An example:

A high jump frame, over which competitors have to jump. “Raing the bar” means increasing its height, increasing the difficulty of jumping over it. “Raising the bar” can be used as a figurative allusion, meaning something is harder to do.

Taiwan Semiconductor raised the bar on track widths by using far ultraviolet as the light source, making 2 nm widths possible.

The meaning is clear.

Another example:

Fred raised (the subject of) the bar on forever chemicals in drinking water at the meeting, saying the benefits did not justify the costs.

Raising a subject is a perfectly valid use of words, even when the subject happens to be a bar on something. The parson’s unconscious parsing will apply the correct meaning – the LLM has no such luck. To make matters worse, the LLM only uses words up to this point to guide it on the next word, whereas a human will use the words after the next word to provide context.  The Simple Simon approach to extracting meaning from text shows how limited LLMs are – in complex text it may take many pages to establish the meaning being created by a few words (they may not be next to each other).

What this demonstrates is the lack of ability of many people in Computer Science to understand reading and writing. Computer Science will carry this stain of foolishness or poverty of training for a long time – probably better to find a new standard – SAI Science (Semantic AI) perhaps.

At the sight of so many suckers eager to be fleeced, the large consultancies threw away their ethics and “went for it”. Getting AI back on track will be long and arduous.