USAF, IBM, Chase Manhattan, NASA; old stuff @ https://nitter.gq/t.co/cxBmPFAL4V. Maker, governance, evolutionary algorithms, and aerospace geek. https://nitter.gq/t.co/OnX3Q8e4HZ
California
Joined February 2009
- Tweets1.1K
- Following1.1K
- Followers332
- Likes3.9K
Likely unpopular opinion: All git diagrams might be drawn wrong -- commit as node and repo content as edge is lossy, maybe misleading.
A more useful representation appears to be commit as a hyperedge -- the parent commits and the previous repo content are all hypertails. 1/3
The commit message itself, then, is the event that triggers the graph transition. Tree content is a location in an n-dimensional space, and filenames are axis names in that space, with file content being the n-coordinate's location on that axis. 2/3
If the OpenAI board had used GPT-4 to game things out, they would have seen that growing market footprint is how they set safety standards. They aren't the "IETF of AI" -- that would have needed OpenAI to be a decentralized, open system from the start.
After years of tilting at the Haskell windmill and a long conversation with GPT-4, I'm starting to get a horrifying realization that Haskell's idea of a monad can be understood and implemented using basic UNIX concepts: pipes and filters, 'make', even 'bash -e'.
Sigh.
Grokker still has warts but with VIM it turns out to be really, really useful; finishing projects in a weekend instead of weeks, brainstorming and coding for complex problems while keeping history in local files, versionable in git. Added GPT-4 support: github.com/stevegt/grokker
Lightbulb moment: Go interfaces are simply callback signatures.
(In Go, best practice is to define interfaces in the caller. Interfaces allow complex callback objects with multiple methods as opposed to simple callback functions. That's it, no magic.)
'There is a type of person [...] who “hates the feeling of dependence, especially when it is a direct result of his not understanding something."'
Yep. Homer links to a very, very good article. Reminds me of my '68 Mustang days, and part of why I liked working on F-15s.
I've never been mechanically oriented. Other than rockets when I was a kid, I was never interested in machines. Books & writing were my interest, not counting scuba diving & exploring the world. But when my house was destroyed in 2017, I got busy, learned how to rebuild it & did, righted & fixed my upside down washing machine, and patched up my old Suzuki tossed a hundred yards by the storm. You can do what you need to do & there's much satisfaction to it. Love this article & wish more people did their own stuff. They'd be happier. wired.com/story/vintage-van-…
Grokker lets you have a conversation with a set of documents or code. Works great for quickly absorbing concepts/papers, as an authoring tool for turning complex ideas into text, and as a chat client in your editor.
It helped me write its own README.
github.com/stevegt/grokker
Steve Traugott retweeted
Basing mask recommendations on prevalence of community spread has always been a terrible idea. Even with sufficient testing, it’s dumb. A great way to ensure that low levels of transmission don’t stay low.
This timeline just gets weirder and weirder. ChatGPT only knows older versions of itself so we're having an interesting dev conversation about its new API: "Yes, you support 4096 tokens per query now... No, that doesn't slow you down; with the 3.5 model you're actually faster..."
Steve Traugott retweeted
I said it last week but state machines are a lost art.
We make them all the time but in the worst ways & often not explicitly. Just by modelling a bunch of objects and events or vaguely connected micro services and lambdas.
Simplicity comes if you think in state machines IMO.
TIL Github Copilot + #OpenSCAD is DWIM CAD. You can write a comment that says "make a 3mm od tube with .1 mm wall thickness" and it will do the thing.