Artificial General Intelligence (AGI) is having one of its periodic Twitter outings, and so here is a thread about AGI - where we are, and in particular, why I think AGI is so far away. [1/17]

In 1838, British scientist John Herschel carried out a simple
experiment to figure out how much energy our Sun emits. Basically: see how long it takes to raise temperature of known quantity of water by 1 celcius. High school arithmetic then gives you your answer [2/17]

The results were mind boggling -- literally beyond comprehension. Every known source of energy would result in the Sun quickly burning out. [3/17]

Scientists of the time flailed around, constructing ever more eccentric theories to explain the unimaginable power of our star. [4/17]

My favourite theory of the time was that the Sun was being heated by a continuous influx of meteors crashing onto its surface... It sounds comical now but TBH at the time nobody had any better ideas. [5/17]

In the event, it was almost exactly a century before Hans Bether put forward the now-accepted theory of solar energy production by means of nuclear fusion. [6/17]

When we try to talk about AGI, I think we are in much the same position as scientists of 1838 were, trying to explain the power of the Sun. Those scientists were trying to explain something that they didn't yet have the scientific concepts to understand. [7/17]

We all witness human level general intelligence every day, just as the scientists of 1838 could spread their arms and witness the incomprehensible power of our star, 93 million miles away. [8/17]

We can articulate the manifold dimensions of human-level general intelligence, and wonder at its complexity and subtlety and nuance. (And for this purpose I recommend "Rebooting AI", the recent book by @GaryMarcus and Ernie Davis.) [9/17]

But we don't have the scientific concepts in place to truly understand it - how it came about or how it operates - still less how to construct it. [10/17]

Most of our attempts to explain and construct AGI are like the proverbial drunken man searching for his lost keys under a lamppost, not because he believes that is where they are, but because that is where the only light is. [11/17]

The truth is human intelligence emerged from an unimaginably long and complex evolutionary process. Mother nature has been doing a kind of evolutionary reinforcement learning in the wild with our ancestors, ever since life emerged 4.5 billion years ago. [12/17]

AGI will NOT magically arise by throwing a big pile of data at a big neural net. [13/17]

*IF* we manage to create AGI (and it really is an IF) then deep learning will be one ingredient, but the recipe will have many other ingredients, some of which we know but some that we simply can't imagine yet (cf. nuclear fusion in 1838). [14/17]

For these reasons, I'm deeply sceptical about head-on approaches to try to achieve AGI. Like trying to build a star with 1838 science+technology. [15/17]

For these reasons, I'm deeply sceptical about head-on approaches to try to achieve AGI. Like trying to build a star with 1838 science+technology. [15/17]

The achievements of AI (and deep learning in particular) this century (Hinton et al, DeepMind, Goodfellow, etc etc) are real and exciting and will be transformative. We should celebrate them. But we shouldn't confuse them with AGI. [16/17]

If you enjoyed this thread, then you might like my book: "The Road to Conscious Machines" (Pelican, 2020), which discusses these issues at length, with jokes and occasional profanity. US edition under the title "A Brief History of AI" in press with @flatiron, out soon! [17/17]