It's not simple weights and numbers all the way down. The available output is pre-set by the tokens we allow it to predict.
There was a whole bit in there about not having a language module or using words. But it does. We tell it.
Humans do not come pre programmed with a set of possible "tokens". We just figure it out and I believe that fact captures something very essential. Maybe the missing piece of AGI. The fact that humans can just be awash in pure sense data, and somehow just figure out what is important and what to do. Never ceases to amaze me.
Sure there is some minimal marginal cost, but it's so close to zero that it's usually negligible, and the incentive is to basically give it away and "monetize" something else. Your point about games actually just makes my original point. Software is already usually free or dirt cheap, which is why reducing the cost to make the software can't create some "low cost / low value" quadrant. Unless your talking about bespoke software that has such a small market size it isn't worth making today. I could maybe see that area opening up, but even that software would not fit the OP's description of software that "has no owner and is not meant to be maintained"
But focusing on production cost is silly. The cost to consumers is what matters. Software is already free or dirt cheap because it can be served at zero marginal cost. There was only a market for cheap industrial clothes because tailor made clothes were expensive. This is not the case in software and that's why this whole industrialization analogy falls apart upon inspection
One thing that has become clearer to me over the years is that reasoning by analogy (like this article does) sounds a lot smarter than it is. If you look from first principles, it's clear that physical goods and software don't share the same properties and thus the analogy falls apart.
Physical goods like clothes or cars have variable costs. The marginal unit always costs > 0, and thus the price to the consumer is always greater than zero. Industrialization lowered this variable cost, while simultaneously increasing production capacity, and thus enabled a new segment of "low cost, high volume" products, but it does not eliminate the variable cost. This variable cost (eg. the cost of a hand made suit) is the "umbrella" under which a low cost variant (factory made clothes) has space to enter the market.
Digital goods have zero marginal cost. Many digital goods do not cost anything at all to the consumer! Or they are as cheap as possible to actively maximize users because their costs are effectively fixed. What is the "low value / low cost" version of Google? or Netflix for that matter? This is non-sensical because there's no space for a low cost entrant to play in when the price is already free.
In digital goods, consumers tend to choose on quality because price is just not that relevant of a dimension. You see this in the market structure of digital goods. They tend to be winner (or few) take all because the best good can serve everyone. That is a direct result of zero marginal cost.
Even if you accept the premise that AI will make software "industrialized" and thus cheaper to produce, it doesn't change the fact that most software is already free or dirt cheap.
The version of this that might make sense is software that is too expensive to make at all because the market size (eg. number of consumers * price they would pay) is less than the cost of the software developer / entrpreneurs time. But by definition those are small markets, and not anything like the huge markets that were enabled by physical good industrialization.
This video was fascinating. I didn't know about "open endedness" as a concept but now that I see it, of course it's an approach.
One thought... in the video, Ken makes the observation that it takes way more complexity and steps to find a given shape with SGD vs. open-endedness. Which is certainly fascinating. However...
Intuitively, this feels like a similar dynamic is at play with the "birthday paradox". That's where if you take a room of just 23 people, there is a greater than 50% chance that two of them have the same birthday. This is very surprising to most people. It seems like you should need way more people (365 in fact!). The paradox is resolved when you realize that your intuition is asking how many people it takes to have your birthday. But the situation with a room of 23 people is implicitly asking for just one connection among any two people. Thus you don't have 23 chances, you have 23 ^ 2 = 529 chances.
I think the same thing is at work here. With the open-ended approach, humans can find any pattern at any generation. With the SGD approach, you can only look for one pattern. So it's just not an apples to apples comparison and sort of misleading / unfair to say that open-endedness is way more "efficient", because you aren't asking it to do the same task.
Said another way, I think with the open-endedness, it seems like you are looking for thousands (or even millions) of shapes simultaneously. With SGD, you're kinda flipping that around, and looking for exactly 1 shape, but giving it thousands of generations to achieve it.
I did a chat with Gemini about the paper, and tldr is...
* They introduce a loop at the beginning between Q, K, and V vectors (theoretically representing "question", "clues" and "hypothesis" of thinking)
* This loop contains a non linearity (ReLU)
* The loop is used to "pre select" relevant info
* They then feed that into a light weight attention mechanism.
They claim OOM faster learning, and robustness acro domains. There's enough detail to probably do your own PuTorch implementation, though they haven't released code. The paper has been accepted into AMLDS2025. So peer reviewed.
At first blush, this sounds really exciting and if results hold up and are replicated, it could be huge.
It's not simple weights and numbers all the way down. The available output is pre-set by the tokens we allow it to predict.
There was a whole bit in there about not having a language module or using words. But it does. We tell it.
Humans do not come pre programmed with a set of possible "tokens". We just figure it out and I believe that fact captures something very essential. Maybe the missing piece of AGI. The fact that humans can just be awash in pure sense data, and somehow just figure out what is important and what to do. Never ceases to amaze me.