April 2022 BenGoldhaber(.com) Newsletter
Last month OpenAI released a new version of their text-to-image model DALL-E. I haven’t seen any quantitative measures of its performance, but my reaction is ‘this is incredible’. Some of my favorites:
A baroque painting of a man who is angry because Starbucks made his drink incorrectly
A kid and a dog staring at the stars
Rabbits attending a college seminar on human anatomy.
Detailed plans for reversing climate change.
This came on the heels of the release of a number of benchmark breaking models. PaLM, etc, along with even more funding for research and startups. One that caught my attention was a new startup from ex-googlebrain employees attempting to train transformer models using ‘learning-from-demonstration’ techniques to use software to automate business ops tasks. I think this is going to be extremely successful, and is also a bad precedent for anybody concerned about AI Safety (“lets give large un-interpretable models unrestricted access to software, what could go wrong”)1.
I’ve seen some attempts to predict the short term impact of this tech (ex. tweet thread). I want to be a bit cautious, as five years ago I thought we were on the verge of mass self-driving car adoption and the extinction of radiologists and ophthalmologists at the hands of automated computer vision systems. Adoption cycles for new technology are much, much longer than you’d intuitively expect. It took about a hundred years from Faraday’s discovery of electricity generation to mass public electrification. Similarly we’ve been in an information technology ‘revolution’ for several decades, and it’s still not clear if IT has actually had a positive impact on productivity
But I’ll venture out on a limb and make several semi-vague observations/predictions for the next four years of AI tech adoption:
The marginal price of generating words at the level of your average newspaper article or internet columnist is approaching zero, the marginal price of generating new images is now on the same track. Anything where the price point or transaction costs of hiring a mediocre ghostwriter or illustrator were too high will be possible.
Indie graphic novels are going to become much more common; every self-published kindle novel will have custom book covers.
‘Good Design’ had been an honest signal of a companies quality. If they had the budget and bandwidth to hire a good illustrator, they probably invested well in other parts of the business like the quality of the product. This will be less true moving forward, so it should be easier for new entrants to move into semi-established markets (ex. like instagram dropshipping style businesses), but longer term might push people back to well known brands (if you can’t use other signals of quality, you go with what you know has traditionally had quality products).
Robotic process automation will finally stop being a buzzword, and a product will be released where you train an AI to fully imitate and execute a business process.
Basically every business on Fiverr will actually be a front end for a large language model.
Autocomplete and text prediction will be a default expected feature for new products. (note: I really enjoy Github Copilot for coding.)
It will be impossible to the casual internet reader to determine what comments are bot-generated vs. human written. Advertising and PR will employ large networks of bots to create simulated grass roots movements around products and people (this is happening now, but I think this will rapidly get adopted).
The… seedier parts of the internet are going to get much more creative.
[Least confident] I think new RL models have cracked the hard part of robotics, and we’ll see consumer robotics take off, with new consumer level robots approaching initial roomba level sales but for more complex tasks (e.g. laundry, dishes).
Following up from an article last month, and having investigated it more, I feel more certain that semaglutide is going to become very widely used for weight loss. Telehealth startups are already starting to sell it, there’s a new competitor close to FDA approval that will push down prices, all of which will push it into mainstream consciousness in next twelve months. I’ll go so far as to say I think we’ve hit ‘peak obesity’ in the US, and that we’ll see a general decline from here on out.
Inside the New Right: A profile of the emerging ‘new right’, a collection of thinkers and political entrepreneurs drawing from neo-reactionary and other, primarily internet based intellectual communities. Based on some anecdotes from NY and LA friends, I’m inclined to believe that it’s true, that among ‘hip trendsetters’ these ideas are getting a lot of traction.
Is China in trouble? Over exuberant enforcement of Zero-Covid policies across many different cities and region seem to be having second order effects that are causing huge economic disruptions.
Never put off till tomorrow what may be done the day after tomorrow just as well. - Mark Twain
Lose an hour in the morning, chase it all day. - Yiddish Proverb
We are what we pretend to be. - Kurt Vonnegut
Sorry to Bother You: This was a fantastic film. A dark comedy with ‘five minutes into the future’ sci-fi elements, the pacing was near perfect. I feel like most comedies now feel rushed, like they’re constantly trying to keep your attention with big obvious jokes. I loved how Sorry to Bother You would invest in minute long jokes about entering a long keypad number, or little visual gags of a photo changing in the background. It also had a really good cast.
The Batman: A little underwhelmed - perfectly fine entry into the Batman canon, and I appreciated its attempts to bring in more noir elements, but I feel like I’ve already forgotten most of the plot. Maybe because in my heart there’s only room for the Jim Carrey Riddler.
Discworld: I’ve read a few books set in Terry Pratchet’s famous fantasy universe, but I recently downloaded the full series and am enjoying them. He had a mastery of a specific type of dialogue that is fun and easy to read.
Until next month!
No, I don’t think this specific startup is going to create AGI, but my read from their initial announcement was they want to train goal directed agents to execute tasks in the real world, which I expect will be less safe than non goal directed tool AI (more on the distinction). Mostly it’s funny that after all the time rationalists spent arguing about whether advanced AI could ‘escape from the box’ the idea that it would ever be boxed was wildly optimistic!