Sometimes I think that LLM:s are like SQL in that they were thought of as end-user technologies, but they are not easy enough to use so end up being building blocks instead.
When Inventing on Principle came out, the examples felt impossible since they were extremely niche interfaces built for a very specific set of problems.
Now, if only we had some magic technology that could follow some principles to build niche interfaces for specific problems…
I don’t necessarily think physical paradigms is the point. Affordance and direct manipulation are what matters. LLMs make those so much harder.
This is fully consistent with his views wanting the ultimate human agency in computing. LLMs are not that. They are useful but they are not about giving humans agency and transparency of computing substrate. They are about abstraction and almost all of Brets work has been a rallying cry against it.
He literally hates how glass flattens all our interaction and wants more physical paradigms of computing. Ones that use all our senses. LLMs are the equivalent of glass screens in software in a way.
He has often said how Apple did not actually get why Alan wanted Dynabook and hence we ended up with iPad.
It’s kind of ironic to see people wanting to do that to his work now.
Have we actually talked enough about what ai will do to (engineering) organizations? Like, in detail, how will the structure, roles, power dynamics etc, change when AI does most of the "work". I think we haven't even scraped the surface tbh.
Andreas Eldh retweeted
My dear friends, I’d like to show you my latest toy project
Introducing Handwired: turn your regular JavaScript code into a neural net!
- no training
- exact outputs vs the JS program
- fully differentiable
All runnable in a browser, without GPU
Hypothesis: We'll soon live in a world where everything can be built by anyone, so the defining feature of any organization will be what they decide not to do.
Andreas Eldh retweeted
The only impressive thing left to do in software is building a product that feels great to use. That still requires real, commendable effort and grind - and is maybe harder to do than ever. Everything else is a prompt away, and even if valuable, it just does not impress.
Met up with the Linear team in Montreal last week, then went cycling in Mont-Tremblant area. It was very nice.
Andreas Eldh retweeted
we’re currently in the bundling era of agents since every cycle now is faster than ever before. general purpose agents are all the rage because they can theoretically do everything but that also creates an intense usability problem.
for normal ppl no amount of proactivity, suggested prompts, or feeds of ideas completely fixes that esp when the entire capability surface is hidden behind a chat box.
i think we’ll see the pendulum swing back very quickly. the product layer will rapidly unbundle into specialized agents built around specific jobs, contexts, triggers, & interfaces. there is so much empty room still.
