HN user
yemen.serin@proton.me
I do the same, but delegate to codex instead via the /smux skill. It's expensive, but I get one long-running Claude context window and one frequently compacting codex, and this saves me from needing to re-gather context all the time.
Per the OP's technique, I'll use Fable for intensive planning moments, then switch back to opus when things are going well.
1. A compiler for real-time tensor processing (arbitrary DSP, ML). In something like LISP or Haskell, the goal is to compile lambda calculus for fast/reliable execution—as such, you can express a program in a fully general language that can represent any computation and execute it without explicitly modeling the lower levels of the machine. I'm building a compiler that does the same thing for the subset of programs that are guaranteed to execute on-budget. The effect: you write code that looks like DSP/ML math and it compiles/runs optimally with execution guaranteed by construction.
2. My take on an agent framework ... append only log + content hypergraph in Elixir, tools that regularly pull data from other services into Postgres—built as a kind of 'exoskeleton' around claude/codex so it's not competing with fast-moving tools.
Thinking about category theoretic models of computation: https://arxiv.org/abs/2208.03817
--
Some things I want other people to build:
- Indexing for Github
- All-in-one social media ingestion libraries for agents
- GOFAI-inspired knowledge / semantic / research graph stuff—I want to point agents at rules/structures for writing connected, verifiable statements
Nice work. Excited to check this out.
Arrived at a version of this view as well and building one on Elixir/Ash.
Ah, interested to dive in—and have a project it could integrate with.
I built a similar thing recently, for agents, aimed at enabling prolog queries over handles in markdown corpora (and code): https://github.com/flowerornament/anneal. A true slopwerk in comparison to this, however.
I suspect most people don't even know there's a there there.
For instance, while I now know that file systems have permissions, before I became a programmer, I spent maybe ten years thinking of permissions as a special, obscure system thing that you should never touch.
For that matter, I suspect many people don't know basic things like that a file system isn't inherently the operating system.
And, where would you go to learn this information? Your Mac doesn't ship with a manual—how would you know one exists? Furthermore, I would wager that perhaps most people have never learned how anything works requiring a manual and are simply unaware that that's a thing.
All to say, I'm not sure "refusal" is the right term.
I mean, fair—it's a possibility. But what are we to do with a point of view that admits no positive outcome?
I don't completely disagree, but that's not the entire story, and the efforts of those who fought for good outcomes matter profoundly.
For instance, we do in fact have global, effectively ad-free, E2E encrypted messaging. A lot of people put in a lot of work on many fronts to design that, prioritize it, and deliver it at world-scale (mostly) legally.
The fact that the ad model became completely dominant is a real tragedy, but I think the deeper mistake with that idea was that "connecting people" is more complicated than it seems and has had many unforeseen consequences.
As a Western well-to-do person, to the extent I can tune out the slot machine, the dream actually has come true in many ways. I have friends and colleagues all over the world, travel is easier, coordination is easier, and politically my voice matters in ways it never would have in the past.
I was reading mailing lists from the early 2000s the other day and it was wild to remember a time when you literally had no idea what was happening in other parts of the world—when nearly all information was mediated through centralized authorities. People were sharing 'suppressed' reports on grass-roots political action other countries with a sense of self-importance that would be cringe today.
Yes, I agree, but the relevant debate isn't "pro vs anti-AI." That's a fight only one side can win.
The relevant debate is: "human empowerment vs disempowerment".
It's still a long-shot, but at least there's a specific target that a majority may be able to agree on.
(Not saying you were specifically saying either.)
Sadly, I think this is true, and it's starting to seem like my own personal political imperative is to remind people that AI can be insanely personally empowering.
Many people around me are just missing the boat, or don't care, but many are also able to finally accomplish all kinds of things they've been barred from in the past.
LLMs are best seen, I think, as an imagination amplifier. If you're in the mindset of finding ways to improve society and help other people, there is no shortage of opportunity, and increasingly, capability.
"…we will inquire into something, but we will not grasp it, if in some way we depart from ourselves and from our own nature…"
Beautifully ironic, that we find this message.
Curious about the perspective from anyone who has a skill set / reputation where finding jobs is easy right now.
I imagine people in ML or who've found a good way to demonstrate prowess with agentic systems may be highly in demand right now(?)
Opus 3.x building me a productivity system with Obsidian MCP originally.
Next was discovering "create a mathematical model of the problem and derive the solution as a result" type prompts.
But, the real "oh s**" was a longer process of spec'ing a compiler/runtime for real-time DSP (with a lot of novel ideas) and it actually working.
My sequence was: (1) if helps me understand myself, (2) if helps me put together good ideas, (3) it can generate novel ideas given the right inputs, (4) it can build useful tools on my machine, (5) it can compound good ideas into better and better ideas with repeated passes, (6) it can build significant, ambitious machinery that's way beyond my ordinary capacity.
Current frontier: it can compound large codebases into better and better machinery with repeated passes.
The key thing I track is whether I'm running a process that converges and compounds or whether I'm spinning in place / diverging.
My intuitions for using Elixir:
- Durable, 'enterprise grade' software patterns are baked into the runtime and into common, stable libraries that everyone uses
- You can use Ash, which pretty much entirely solves architectural considerations for many types of backends
- The tooling for inspecting and enforcing style (tidewave, credo, dialyzer, Dan's "vibe" ecosystem tools) is far beyond what I see in other ecosystems
- Ecosystem coverage for pretty much everything you need, including numerical software
- Excellent performance escape hatches (NIFs)
And, as has been shown in various benchmarks, agents are quite good at it.
My one problem in practice has been that getting tests right is hard. LLMs need a lot of cajoling to not build flaky tests with all the concurrency, and I find myself spending hours rewriting parts of the test suite once or twice a week.
Rust, Elixir, and Go are the way to go for LLMs in my testing and experience, for this and other reasons
Do the forms etched into stone by weather over millennia in Moab matter to the wind? Certainly yes, in one sense, but not in the same sense we mean when we say things matter to us, or to animals, or even bacteria.
Not optimizing for that. I derive "10x" more satisfaction, because I'm able to work on more ambitious problems. I'm probably making less money than I would otherwise.
I work on self-directed projects that don't make any money, currently. My "day job" (not software development) does not have this '10x' quality, though I imagine it could were I allocating my efforts that way.
Not publicly yet, but I work on a programming language, compiler, and runtime that achieves magical (to me) things in a niche field. I would never have attempted something at this scale otherwise, so it's a very 0-to-1 experience subjectively.
I think it's just hard to know this for the people working on it. AI radically changed my life. I'm doing at least 10 times more, more ambitious, higher quality work. I've been listening to people around me talking about alignment and the singularity for almost a decade. It's strange to imagine that people live in a world where this isn't and hasn't been happening for a while now. "Over-hyped" is not the word I would use if I take my daily experience as an example, nor when I consider even lower-bound projections.
I've yet to come across something with vim bindings that lacks a .vimrc where you can map 'jk'. Either way, switching back to ESC is as annoying as it is in the first place.
"jk" is even faster (you get to "roll" your fingers)
Thank you! I'm writing an APL-lineage language right now (really, APL+Prolog+Lustre "lineage") and hadn't come across this paper.
I found an HTML version here: https://www.jsoftware.com/papers/J1990.htm
I did an exhaustive comparison of window managers and settled on using Raycast for simple resizing (full screen, center, mid-size centered, 1/2, 1/3, 2/3 left/right) + FlashSpace[1], which implements simple virtual spaces with instant switching.
You can also use Rectangle or Spectacle or others in place of Raycast.
Foolproof with zero magic.
According to the meter, I used $15k in tokens with my Max plan (along with $5k of Codex tokens) in the last 30 days. That built an entire working and (lightly) optimized language, parser, compiler, runtime toolchain among other things.
Yup, 27" 4k with a Mac is truly awful. Don't do it. Get a 5k display.
Ah, thanks so much for this question. I ended up building a tool that agent can use to track 'compounding' in my corpus of .markdown files. Keep iterating and thinking about it, and you may find you can do the same for your process.
It's too early. People are trying all of the above. I use all of the above, specifically:
- A well-structured folder of markdown files that I constantly garden. Every sub-folder has a README. Every files has metadata in front-matter. I point new sessions at the entry point to this documentation. Constantly run agents that clean up dead references, update out of date information, etc. Build scripts that deterministically find broken links. It's an ongoing battle.
- A "continuation prompt" skill, that prompts the agent to collect all relevant context for another agent to continue
- Judicious usage of "memory"
- Structured systems made out of skills like GSD (Get Shit Done)
- Systems of "quality gate" hooks and test harnesses
For all of these, I have the agent set them up and manage them, but I've yet to find a context-management system that just works. I don't think we understand the "physics" of context management yet.
Came here to say something similar. For me, the craft aspect is now even more exciting because I can craft more ambitious things without getting bogged down in the details. For me, refining my conceptual model, drawing diagrams, finding the right way to think about something was the craft.
Maybe that's another way of saying: I was trained as a designer, and now the distinction between design (read: architecture, service-design, product, ux, cx) and programming is blurring.
There are a handful of these. I've been using this one: https://github.com/smart-mcp-proxy/mcpproxy-go