John Ball @jbthinking

@TrueAIHound Yes. Catching up fast? Like how radiologists are no longer needed (obviously!) and how nobody will own a car by 2000 because driverless cars are ubiquitous?

@GaryMarcus Never ask a mathematician to explain biology, I guess.

Science used to use evidence to make outlandish claims. Here a mathematician is ignoring most of the animal kingdom and leaving only humans and statistical artificial neural networks as ‘intelligent’. Science would also consider the options and full scope of the claim. To refute the claim of consciousness for LLMs, start with its ‘anatomy’. You can see how the program generates tokens using statistics that are dumped in by a human-training step designed by people. Where is sensory input? Where is the storage of context that overrides the statistics? Why are other apes and animals excluded since they are far more aligned with humans than computer hardware and software!

I blame the tech CEOs for their language claiming that ‘the AI’ did this or that and will soon take our jobs and be worth trillions. Better to ground these programs properly with ‘the statistics made this mistake and that one, and in time its failures will be fixed with good technology’

@HedgieMarkets Well, maybe those robots will keep learning and become better soon. I mean, that’s how magic works and the robots could get bored.

"our sensors do most of the heavy lifting for the neocortex" - this would mean that bat's sonar does the heavy lifting for them, as vision does for us. And also human language that can come from almost any sense (vision, touch, hearing, ...). That needs each sense to have the same capabilities, rather than other brain regions like the human neocortex. My model, Patom theory, uses the strengths of senses, but puts the recognition into the general purpose regions of the cortex as the final arbiters to deal with commonality between senses (so we recognize our spouse by sound if blinded). Yes, I appreciate this is not aligned with your model. I like "The world computes itself" other than this removes all computation, since the world does what physics and biology has it do without additional effort! What is cheaper computation than that? I see brain damage giving a lot of support for the "book keeping system" aligning all our senses with what we experience, and using stored patterns of (simplified) experience to make decisions. In NLU (Natural Language Understanding) our system aligns the linguistic syntax to semantics and back in context. Both do heavy lifting that I think shows the need for more than just the sensory recognition in our brains. Anyway, I look forward to your next posts!

‘This AI race’ isn’t a good description. Whoever wins the race for statistical AI isn’t a solution for human-like capabilities. Real AI is the next generation that aligns with the human brain. Today’s LLMs are expensive in power and cost. The brain is orders of magnitude cheaper to run and accurate at the same time. Lossy AI isn’t the goal for humanity.

The Turing machine is often cited as the reason computers can do anything, but my brain research shows the fundamental deviation from the computational model to what our brains do. We live in the computer paradigm age that is slowing progress in AI. Too many false assumptions to deal with causes lazy designs. Now LLMs demand exponential increases in power to do tasks that only require a laptop. It is bad science and worse engineering!

@TrueAIHound If you’re not designing code that replicates human capabilities I’d imagine not much use for Turing’s work. My point was that people thinking the imitation game is just conversational chat haven’t understood what he designed. Again, I’m not a fan of those behaviourist tests.

Well, in 2023 AGI was really close too. Better - save predictions until something like it already can be done. Otherwise it can look like hype to drive stock price instead of being a scientific claim. Geoffrey Hinton predicted the demise of radiologists in 2016 and that only caused an under supply of them. And driverless cars due by 2020 didn’t happen either.

@realBigBrainAI Or at least stop those calling statistical programs AI and only call AI things that are ‘intelligent’ or that at some level duplicate cognition.

So my question to you on patterns is this: say your claim is right that there are no patterns stored but learned on the fly. When you see the same thing again it resolves to the same novel pattern. Now without a previous change recognised somewhere, how would a brain recognise, say, your mother? Wouldn’t that be the symptom of Alzheimer’s? How would your model explain the visual recognition loss in facial agnosia where their visual face cannot be recognised but their voice and all other memories remain? I agree at the sensory level that there is no need to learn sensory patterns (that you study?) but at the global level, I cannot see how that would lead to human levels of memory and recognition.

The weakness is most obvious when trying to build robots with machine learning. Its model is alien to how brains work, while cybernetics and feedback loops vastly simplify the scope of robotic tasks like sensory motor control. Today’s well funded companies continue to try to get the wrong tool to fit everything in AI. That’s not how engineering works as it is best built on scientific principles.

The reason I don’t pretend that ai relates to ‘intelligence’ is exactly this. First, brains simpler to humans without language still perform amazing feats of sensory and motor control. Intelligence’ is the wrong word for this essential skill set. Our capabilities build on this, so the symbolic computer model has nothing to be grounded to except symbols that don’t symbolize anything!

I prefer a scientific lens to look at technology problems. We can imagine lots of possibilities that have no evidence, which is why ockham’s razor is so important to remove theory that is beyond observation. Lots of things may be possible, but science should be built on what we can see and measure, not ideas that may be made up by the brain from impressions being composed - per Hume.

The problem with that paradigm is the associations our brain brings with that. If the brain works on computable problems, it still doesn’t mean the brain is like today’s computers. I mean, does the brain encode data and transfer it around like computers do? Information theory doesn’t apply to brains. My research shows why that model isn’t how brains work. Holding onto the computer model starts AI on the wrong path that differs from brains.

@TrueAIHound I prefer to think of it as science versus engineering. F = ma is science while the use of it is engineering. Without good science there is thrashing as seen in LLM limitations.

Sure, humans are flawed in many respects. Our brains choose the wrong statistical choices based on psychology system 1 and 2. Refer kahneman thinking fast and slow. That tends to be the excuse for failures with LLMs. It isn’t valid in that brains have a far more accurate representation of the world at the lowest levels, while word vectors are lossy statistical sequencing models. When today’s lossy models are replaced with those holding accurate representation, the current hallucinations will be replaced with human-like errors. But for now, statistics just isn’t the right tool to select from the myriad possibilities to deal with in the real world.