iamwil

@interjectedfuture.com

Tech Zine Issue 1: LLM System Eval https://forestfriends.tech Local-first/Reactive Programming ⁙ LLM system evals ⁙ Startup lessons ⁙ Game design quips. Longform: https://interjectedfuture.com Podcast: https://www.youtube.com/@techniumpod

First time I've heard the metaphor, "Pokemon exception handling". This is apparently a common term. I was going to say, it's a hilarious and creative phrasing if the AI came up with it on its own. But not today. Not yet.

Bild

Some will take the view that unlike AI image generation, photography is art that requires human judgement and creativity. Hilarious, cuz early on photography was widely derided as "not art". AI gen practitioners don't yet have the language to describe their editorial choices like filmmakers do.

Beyond the engagement, explainers like this are only doing labeling. Great when you just want to know something exists. But to know something deeper is to know its relationships and a take-away abstraction of its mechanics and design constraints. We can do more.

Just because we invented recorded music didn't mean there wouldn't be Taylor Swift. I think we'll always demand one socially and culturally. Yes, "great chefs can come from anywhere," but don't discount what experts can do with the same tools available to plebs. So it is with math.

Unlike Latin characters, Korean characters are composed: ㄷ + ㅏ + ㅁ = 담 What if an array-based programming language could leverage Korean as its Pandas-like data-manipulation functions, and then we compose them with combinators? Could we get a terse idiomatic language like APL?

According to a former Google Books PM, the current state of affairs was a self-own by the publishing industry. It's what was ruled fair-use after going nuclear on the litigation. If publishing and intellectual property peeps were a bit more chill, it wouldn't have come to this. x.com/bradrcarson...

Bild

It doesn’t have to be. You just care about different things at a different level. Does an IKEA chair designer not care because he’s not steam bending every Poang? Does a web programmer not care because he doesn’t hand write assembly to squeeze out every drop of performance? bsky.app/profile/fre...

Bild

While thinking with Fable about ‘if’ conditions in programming languages, it made the observation: pattern matching on a single value and GROUP BY over a collection are the similar computations lifted from one cardinality to many

Both can be true: people are holding it wrong, and AI will replace white-collar jobs. LLMs are a major leap in handling unstructured data, so it’s hard to imagine they won’t eliminate some white-collar work.

Not picking on this particular person. But I’m a little dumbstruck by this current discourse where people are all pikachu-face about how others weren’t lying, but the reason they were getting subpar results from AI was because it was a skill issue all along. x.com/karthiktade...

Bild

It's the best kind of reaction video when AI comes up with a result that drops the jaws of practitioners in the field. Happened with protein-folding. Happened with Go. Happened with web dev. And now, the Jacobian conjecture. It's the Age of Wonders.

Out of anyone I track, @_chenglou really leans into the affordances of LLMs and then asks what we can do with this newfound affordance. I've adopted generating Quint proofs for complex state management. What other guardrails can we enable that was previous onerous? x.com/_chenglou/s...

Bild

I never thought I'd get to use the SAT word "palimpsest" when I was younger. Now, with code bases, agent or collective-human written, it's more than a fitting word. Good thing agents are getting better at archeology.

For a long time, I also did not understand what value was and its relation to money flow. Now that I have a better understanding, I think this quip also applies to open source advocates and decentralization evangelists, not just mathematicians and students. x.com/soleio/stat...

Bild

How is OpenRouter assured that 3rd party open model providers are actually serving the model they say they are? Had a similar problem when building a CoinMarketCap clone. It's mostly trust + measurements, I guess.