24 years of AR/VR R&D engineering. Principal Engineer at neuRealities. Formerly @IncMoonbeam, @Mozilla, @DAQRI, @ARToolworks. Opinions expressed are my own.
Buckminster Fuller, who was wrong about a lot of things in a way that was usually more interesting than other people being right, noticed something in the 1930s that still holds true.
He noticed that technology does not primarily get bigger. It gets thinner.
The transatlantic cable got lighter per unit of message. The building got lighter per unit of enclosed volume. The engine got lighter per unit of horsepower. Each generation of a technology tends to accomplish the same function with less stuff, and then with less stuff again, until at the limit the stuff approaches zero and only the function remains.
He called this ephemeralization : doing more and more with less and less until eventually you can do everything with nothing.
And here were it gets interesting. What got eaten, in the first wave, was objects: my door key became a fingerprint.
What's happening today is the second wave: the AI paradigm shift. It’s our thought patterns, our decisions, our generative processes, our creativity, and our communication and coordination with each other that is being ephemerized.
And I think judgment abstraction has two consequences.
First, the output is not checkable at a lower cost than producing it. To verify a summary you must read the document. To verify a strategic recommendation you must do the strategy. The thing that made calculators safe (verification being orders of magnitude cheaper than generation) simply does not hold for the class of work being absorbed now. Which means verification won't happen because the economics forbid it. A check that costs as much as the work is not a check. It's a second attempt.
Second, the stairs are being removed as a consequence of using the lift. The way you stay able to judge a draft is by having written drafts. The capability is maintained by its own exercise and by nothing else. So the fallback and the thing it's a fallback for are the same muscle, and the system that replaces the muscle also stops the training.
This is Lisanne Bainbridge's irony of automation, which she wrote down in 1983 and which the aviation industry spent the next forty years rediscovering at great expense: automate the easy 99% and you have designed a job where a human with no recent practice must handle the hardest 1% with no warm-up, at the worst possible moment, on the basis of a mental model the automation has been maintaining on their behalf.
Airlines eventually responded by mandating hand-flying. No knowledge-work organization on earth currently mandates hand-thinking.
@bayagima Seems inevitable; loads of stuff is being produced but no-one is using it, so marketing noise can only rise. I think big players will surpass the 22% sales and marketing spend as % of revenue that Microsoft peaked at in 2008.
@gfodor@InfoTechBrent 💯 It's quite difficult to explain to those who haven't experienced it, how in the middle of a phase of development full of frustration and angst, you can have enter a flow state and transform a concept into a perfectly functional unit of algorithmic logic.
@sterlingcrispin Yeah, I am getting prone to nostalgia; every time I put the Vive Focus on and it boots to the ambient audio on the balcony I am transported instantly back to a very happy time spent working on open-source VR every day.
Good hint about Grok. I have a sub, will put it to work. 🚀
Why is New Zealand a hotbed for Aerospace innovation, especially rockets and high-speed flight?
Airspace.
Even White Sands Missile Range is not wide enough to turn around a Mach 5 vehicle in a 3g turn.
We have thousands of miles of open air and sea space. Almost nothing between us, Antarctica and South America. And a test site only 45min drive south of our headquarters.
This is a structural advantage Dawn will carry forward almost indefinitely.
This is why we build the world’s highest-performance aircraft in New Zealand.
When I say that many people have the wrong mental model about what AI is doing, I mean it's very difficult for us to understand what a model trained on the entirety of human written texts is capable of regurgitating. Thankfully, this guy is an entertaining explainer. Worth a follow.
@THELUOK This makes me want to re-watch Waking Life. That was one of only a few movies I’ve ever seen that capture the spirit of dreaming in a visual form.
@bayagima Landing fees definitely add to the price on multi-stop, but I agree it should be a lot easier to fiddle with an itinerary and see the influence of stop overs on price and duration.
@dhh One of the ironically true things said by Dan Lyons in his Secret Diary of Steve Jobs was that "at Apple, we start with the ads", which was actually 100% Jobs; acknowledge the human need and identify the technology that could meet it. That is what Apple lacks at present.
OK, so as someone who works on this stuff daily, here's my take on LLMs for coding: There are some things they're really good at and some things they're really bad at, and knowing where to draw the line on how to use them is the key.
I remain astounded by how incredible LLMs are for my productivity. I have to manage multiple engineers and write a ton of code on a daily basis. AI just makes my life a lot easier, because of three things:
1) It's really good at writing boilerplate code that you *need* (read and write to this AWS resource, create an SFTP client for me, etc). It's also extremely good at doing regimented tasks with minimal complexity (generate test cases, etc).
2) It's extremely good at finding (coding) packages that you didn't know existed. Python, Java etc are just oceans of libraries at this point, and the documentation for a lot of them is fairly sparse and shoddy, and exists as a Frankenstein's Monster across multiple websites and forums. No sane person has the time to read this.
3) It's amazing at helping you debug — whether through parsing large volumes of logs, or through processing errors for complex pipelines and suggesting solutions. You have to be pretty careful with prompts here, but if you know how to prompt it, the gains are immense. Things that used to take me 5-6 hours to debug now sometimes take minutes.
BUT: it's really bad at one major thing.
AI is horrible — and I mean *HORRIBLE* — at generating complex pieces of code and logic. A hallmark of AI-generated code is that it yields insanely long and unnecessarily convoluted blocks, with a ton of emojis.
If you ever ask it to do something open-ended and semi-complex, be ready for your work to swell by 5x. (In the process, it'll also generate a couple very nasty bugs.)
I'm pretty okay with senior, staff and principal software engineers using a lot of AI to speed them up. At this point, those folks have been coding at a high level for years, and it's second nature for them. Exceptions always exist, but they're generally very aware of where it can go wrong, and how.
I'm less patient with inexperienced engineers doing this, because they tend to blindly use AI to generate volumes of code and think "Lots of Code = Good". What this usually results in is bugs that eventually get exposed, which the rest of us then have to clean up. Moreover, they don't learn much from this, and so they're not building any muscle memory or making any real progress.
So I'll make a prediction in my area of expertise: AI won't eliminate programming; it'll instead lead to an explosion of programming, and removal of creation of computation out of the hands of "guilds" and into the hands of the intellectual "middle class".
Introduction of a revolutionary technology eliminates some occupations, of course. LLM-based AI is going to eliminate many occupations. But it's the secondary effects that are the payoff and where it gets interesting...
The question about AI adoption then should be not only which occupations are threatened, but should be informed by the question: when a knowledge assistant becomes essentially free, what previously uneconomic careers could now become viable?