Late September long reads

Debunking claims about job displacement, towards Act Two for GenAI, and my fabulous adventures with Midjourney.

Hello readers! I’ve been busy in my consulting practice, meeting deadlines for two big projects which, thankfully, are now complete. Next, I must prepare for a month of public speaking on topics as diverse as AI and the creative process and the concentration of power in technology.

Update on the newsletter. I have prepared a fresh series of articles on copyright in the age of artificial intelligence. That project got delayed because I ran into an unexpected roadblock: I was working with a hypothesis that turned out to be incorrect. I wanted to fix that before I publish. It required some additional research. The next batch of copyright articles, on the evolution of copyright in the United States, will be posted soon.

In the meantime, I continue to be fascinated by the very near future. Brett King and I have enjoyed conversations with excellent guests on The Futurists podcast recently. If you haven’t checked out our awesome new web site, inspired by Jack Kirby comic books from the 1970s, please take a look. The new site went live yesterday and our dev team continues to make updates. We are still squashing bugs!

We have the good fortune to live in a time of rapid change. I find it very exciting despite the social chaos and political turbulence. One of my motivations for publishing this newsletter and the podcast is to encourage people to think differently about the future. The more you practice foresight, the better equipped you will be to surf the ceaseless tide of technological change.

In that spirit, here are some of the longreads that captured my attention in the past week.

1. Carl Benedikt Frey and Michael Osborne on how AI benefits lower skilled workers.

These guys again. You may recall that ten years ago, this duo of Oxford University researchers made a big splash when they published a paper that concluded that 47% of US employment was at risk of “computerisation”.

This finding was widely hyped in the press during 2013 and 2014 as a claim that 47% of American jobs would be automated or that 47% of American workers would be displaced by AI and robots. That’s not entirely accurate reporting. Welcome to journalism in the era of clickbait headlines.

It did not help that their paper was riddled with bombast: “In the present study, we will argue that legal writing and truck driving will soon be automated.” So far, neither has occurred, though AI maximalists will argue as they always do that these changes are imminent, just around the corner, just three to five years away.

Here we are, ten years later, and quite obviously Frey and Osbourne’s forecast was inaccurate.

Being first is risky because you might be wrong.

The paper set the admirable goal of being the first to examine the impact of technology on future employment, making this claim: “To our knowledge, no study has yet quantified what recent technological progress is likely to mean for the future of employment. The present study intends to bridge this gap in the literature.”

And in this respect the paper was a success, because it opened up a vast new topic of doom mongering about the inevitable demise of human labor in the face of artificial intelligence, robotics and other technological disruptions in the workplace.

Alas, hard data defeats this thrilling narrative. When Frey and Osbourne’s paper was published in 2013, the US unemployment rate was well above 6% according the US Bureau of Labor Statistics. Today unemployment is hovering near 3.5%, trending steadily downward with a slight uptick last month.

True believers in the AI apocalypse may insist that the deluge is just around the corner, but a reasonable reader might conclude that the bursting of the giant bubble of media hype would diminish the credibility of this dynamic duo.

Have they been chastened? Perhaps. In their new article in the Economist (free, registration required) the Oxford duo seems to be sailing pretty close to the shoreline in their forecasts, limiting their claims to some very obvious, non-controversial conclusions: “Machines make producing average content easier.” No kidding!

Anyone who has used ChatGPT has already observed that the output is stunningly pedestrian. If you are looking for elevated prose, keep searching because so far AI is not capable of it. Because large language models (LLMs) are trained on a vast amount of human-generated content, the output seems to be incapable of surpassing the average. AI-generated content is rarely fresh with novel insight. Frey and Osbourne call this the “average in, average out” dilemma.

Which is why professionals at the peak of their prowess tend to sneer at AI chatbots. I’ve spoken to screenwriters, novelists, poets, songwriters, legal authors and academics who have each independently arrived at the conclusion that AI will never challenge professionals who have mastered their craft.

Maybe. I remind these folks that never is a very long time. Watch this space. Generative AI continues to make progress at a rapid clip.

In the meantime, Frey and Osbourne note that GenAI confers an enormous advantage today on low-skill workers by increasing their productivity.

I call this finding uncontroversial because it is so obviously true. A good writer has no need for a chatbot to help her compose an email to a customer. Her work is superior to the output of an AI. But a below-average writer may find that task difficult or unpleasant; for this writer, the chatbot is an easy way to dispatch a tedious chore. It is a productivity gain. Moreover, everyone wins because chatbots will elevate the niveau of business writing.

Cheers to fewer badly-written emails!

Still, Osbourne and Frey cannot resist the temptation to speculate again about widespread technological unemployment. To support their doomsday narrative, they force-march the Luddites and the Writers Guild as Exhibits A and B, making trite arguments that made me wonder if these two researchers relied a little too heavily on ChatGPT while composing this article.

(Edit: in the emailed version, I mixed up the name of one of the authors of the paper. Odd because I was just re-reading it today. I need an editor!)

2. Generative AI’s Act Two

A couple of readers sent me links to this useful analysis by Sequoia’s Sonya Huang and Pat Grady.

(The authors credit ChatGPT as a co-author. Is this now a thing?)

Huang and Grady explain that the first round of GenAI progress was “technology out” which is a fancy way of saying that scientists came up with a new tech solutions and then set out on a quest to find a problem to solve. That happens pretty routinely in the technology field.

They predict that the next wave of progress in GenAI will be “customer-back” instead of “technology out.”

This forecast will make sense to anyone who has been exploring new AI tools during the past year. It can be hard to figure out how the current tools fit with your workflow; often you need to adjust your workflow to the AI rather than the other way ‘round. Moreover, you may may find that you need to invest a certain amount of time and effort to understand how to use a generative tool, only to conclude that it doesn’t really solve a significant problem.

Your mileage may vary. What works great for one person is a damp squib for another.

Now a new generation of applied AI toolbuilders are focusing their attention on customer workflows and working back to AI to solve real problems. I’d bet that this approach will be far more useful for far more people.

As the Sequoia authors write:

We now believe the market is entering “Act 2”—which will be from the customer-back. Act 2 will solve human problems end-to-end. These applications are different in nature than the first apps out of the gate. They tend to use foundation models as a piece of a more comprehensive solution rather than the entire solution. They introduce new editing interfaces, making the workflows stickier and the outputs better. They are often multi-modal.

This article is useful in a number of ways. First it includes two excellent diagrams that map out the landscape of companies by sector. Second, the authors share what they got wrong in their initial report on this trend. This kind of candor is refreshing.

Third, they document the “value gap” in Generative AI, which is illustrated by charts that reveal diminishing levels of user engagement and growth. These are not indicators of robust health.

The upshot is that the current crop of GenAI apps initially succeeded in garnering massive attention and awareness, but they failed to hook most of those users into a routine habit. Which is precisely why the next generation of apps must begin by analyzing user workflows and finding routine tasks to streamline, boosting productivity by making the interface more intuitive and natural.

If you believe, as I do, that generative AI and other artificial intelligence tools can provide a major productivity boost (and even a creativity boost, see the next item!) to human workers, then you’ll find much to appreciate in this article.

Midjourney for cinematic storytelling

For about 16 months, I have been paying attention to the rapid progress in generative AI for images. I started with DallE 2 when it was first released in early 2022, but found it cumbersome and the results uneven and disappointing.

A few months later, Midjourney was released, along with a bunch of GenAI apps built on the open source Stable Diffusion model. So I tried them. To my surprise, I got hooked.

This is pure fun for me because I am an oil painter, so I have a certain level of skill to benchmark against the AI.

The most recent three versions of Midjourney have blown my mind. The quality of the imagery is very good. You still get bizarre artifacts like extra fingers or random illogical details, but that’s part of the wonder and oddball joy of collaborating with a robot brain.

I am aware that other folks are able to obtain similar results or better results with other tools, and that there is a new version of DallE available, and also that multi-modal generative AI will soon mean that you can generate images from a chatbot and so on. If I tried to convey all of that here, I would need a different newsletter. This space evolves in weeks, not months.

Back to Midjourney. The newest version convincingly renders an stunning range of outputs. On Discord, you can see Midjourney users cranking out logo designs, product concepts, IKEA catalog mockups, plausible knockoffs of Otomo and Studio Ghibli anime, origami figures, and even adorable sockpuppets.

For me, the most interesting developments are found in generative photorealistic imagery. As I wrote recently, generative AI seems to be on track to merge with the built-in AI features in smartphone cameras. This will confer a new kind of expressive literacy on anyone with an iPhone.

Today the interfaces are clunky, but that will eventually get fixed. After that, these tools will be integrated into our workflow (think: PowerPoint is about to get a massive upgrade).

You can experience a preview of this today by using Midjourney to conjure up images from fictitious motion pictures. To me this kind of prompt writing is definitely a creative act, because the human user needs to have sufficient imagination to conceive of a non-existent film and describe it vividly enough for the bot to crank out visuals. You do the typing, then the AI magically brings to life the images in your head.

I found the tutorial in this YouTube video Advanced Midjourney V5.1 Guide (Ultra Realistic Cinema AI Photography) particularly inspiring. In just eleven minutes, the author packs in a lot of useful tips. Intrigued, I clicked through on the link in the notes and ended up purchasing his richly detailed book of prompt recipes and tips which is continuously updated for the newest versions of Midjourney. The book is a whopping 65 pages long, packed with info and plenty of pictures. I recommend it if you have the slightest interest in this subject.

At this point, some readers will be wondering if I am contradicting myself, because I started this article with the observation that generative AI creates “average” results. To me, there is no inconsistency: for images, I am perfectly content with the average output that is Midjourney’s default mode. I am delighted that it seems to be improving monthly. I cannot take photos like these, so for me this is an upgrade.

I use Midjourney to generate images for this newsletter. I cannot say that this tool “increases my productivity” because what tends to happen is the exact opposite. It reduces my productivity by absorbing hours. I go down the rabbit hole and end up spending hours tweaking text prompts in the hope of obtaining imagery that may be relevant to my needs. Sometimes it works. If my goal were efficiency, it would be much faster to download an image from Unsplash.com.

On the one hand, the resulting images are pleasingly weird because they come from an alien mind. Plus the process is mesmerizing and fun: there is a lottery-like fascination that comes from seeing the result generated before your eyes. The reveal is genuinely entertaining in a way that Netflix no longer seems to be.

On the other hand, these systems remain riddled with the fundamental flaws that I mentioned above, like: cliched imagery; some obvious racial and gender bias in the default ethnicity or identity for certain roles: a persistent inability to draw fingers, hands and other human anatomy correctly; and some absurd logical gaps that reveal that the AI has no idea what it is actually drawing.

To paraphrase Dr Johnson, the AI doesn’t necessarily draw perfectly well, but what’s remarkable is that it can draw at all.

Below are a couple of Halloween-themed images that I cranked out yesterday. Effortlessly. As you can see, Midjourney seems to have conflated my initial prompt about Day of the Dead costumes with some visual concepts from Venetian Carnival (not that there’s anything wrong with that!). Together we steered towards a kind of foggy blue neo-Victorian skyline as the background. What surprises me is the cinematic flavor of these images. They look like stills from a movie that does not exist.

Enjoy your weekend. The next installment of articles on copyright and generative AI will be published in October.