I generally agree with you and find your perspective valuable. One thing that might be useful is to point out where you disagree with a few play money prediction markets on AI trends. Eg I "bet" against this awhile ago: https://manifold.markets/ScottAlexander/in-2028-will-an-ai-be-able-to-gener. I find it extremely unlikely that Hollywood is going to be automated within the next five years.
Dan, great article! I really value the retrospective.
This truism pains me: ".Every Ph.D. I talk to who went into data science says their days are filled with boring menial work. Still, they often like their job due to the nice offices, good pay, friendly co-workers, and amazing work-life balance"
This is a well written story -- I always find your articles interesting, cogent and supported by reality.
However..... maybe with regard to AI it is different this time around.... i.e., we will really hit AGI soon. By brute force, really, really interesting behaviors are occurring, enough soon to replace factory workers, to drive cars, and yes.... to practice good-enough medicine.
I presented with Peter Boltuc this summer (very well received at the main AGI conference) the concept of "Alien AGI", i.e, we are going to create alien-like AGIs that are not like us, will never relate to us, and by brute force do interesting AGI-like things. This is in comparison to "Natural-like AGI", i.e., from a mammalian-like architecture.
With regard to that, I toil in obscurity (I don't even have a cold office for that purpose, but go to a coffee shop each morning before the day job or else in my basement) but it *will* happen. Getting closer and closer.
Attached picture of robot embodiment (wifi control via Python code) for my latest model which should start showing "sparks" of superintelligence although admittedly would need $1B engineering to fill out all the basic routines I've short-cutted. If picture does attach here it is:
Don't give up.
The field really is still wide open, not for data science, etc (although that may pay the bills) but for real AI and real AGI.
Howard :)
Nope.... sorry.... can't attach the photos of my superintelligence embodiment I hope to show off the real thing (given airline kindness) in 2024 conferences.
Interesting!! I think a lot of people agree we are close to AGI, but an alien AGI. GPT-4 is already an "AGI" by some loose definitions, or at least shows "sparks" of general reasoning ability. The term "AGI" isn't very well defined. I like Chollet's definition, and I think by his definition we are still pretty far away and transformers and scaling haven't brought us much closer... but I have a lot of uncertainty about the matter.
I'm pretty optimistic about generative AI in synthetic biology. DNA synthesis is so cheap now that if you have a cheap and reliable assay (for example, on some biocatalytic enzyme) you'll be able to iterate on models very quickly.
I generally agree with you and find your perspective valuable. One thing that might be useful is to point out where you disagree with a few play money prediction markets on AI trends. Eg I "bet" against this awhile ago: https://manifold.markets/ScottAlexander/in-2028-will-an-ai-be-able-to-gener. I find it extremely unlikely that Hollywood is going to be automated within the next five years.
Yeah... Manifold Markets is suuuuper biased, especially on AI issues, since its mostly EAs and rationalists using it!! #EasyMana
Dan, great article! I really value the retrospective.
This truism pains me: ".Every Ph.D. I talk to who went into data science says their days are filled with boring menial work. Still, they often like their job due to the nice offices, good pay, friendly co-workers, and amazing work-life balance"
Hi Dan
This is a well written story -- I always find your articles interesting, cogent and supported by reality.
However..... maybe with regard to AI it is different this time around.... i.e., we will really hit AGI soon. By brute force, really, really interesting behaviors are occurring, enough soon to replace factory workers, to drive cars, and yes.... to practice good-enough medicine.
I presented with Peter Boltuc this summer (very well received at the main AGI conference) the concept of "Alien AGI", i.e, we are going to create alien-like AGIs that are not like us, will never relate to us, and by brute force do interesting AGI-like things. This is in comparison to "Natural-like AGI", i.e., from a mammalian-like architecture.
With regard to that, I toil in obscurity (I don't even have a cold office for that purpose, but go to a coffee shop each morning before the day job or else in my basement) but it *will* happen. Getting closer and closer.
Attached picture of robot embodiment (wifi control via Python code) for my latest model which should start showing "sparks" of superintelligence although admittedly would need $1B engineering to fill out all the basic routines I've short-cutted. If picture does attach here it is:
Don't give up.
The field really is still wide open, not for data science, etc (although that may pay the bills) but for real AI and real AGI.
Howard :)
Nope.... sorry.... can't attach the photos of my superintelligence embodiment I hope to show off the real thing (given airline kindness) in 2024 conferences.
Interesting!! I think a lot of people agree we are close to AGI, but an alien AGI. GPT-4 is already an "AGI" by some loose definitions, or at least shows "sparks" of general reasoning ability. The term "AGI" isn't very well defined. I like Chollet's definition, and I think by his definition we are still pretty far away and transformers and scaling haven't brought us much closer... but I have a lot of uncertainty about the matter.
You can email me the photo(s).
I'm pretty optimistic about generative AI in synthetic biology. DNA synthesis is so cheap now that if you have a cheap and reliable assay (for example, on some biocatalytic enzyme) you'll be able to iterate on models very quickly.