home.social

Search

14 results for “mnl”

  1. Spectatorul @[email protected]

    📚 Poeta greacă Liana Sakelliou este invitata ediției a 59-a a Colocviilor de Traduceri Literare, luni, de la ora 18:00, la MNLR. Poemele din volumul trilingv „În Primul Paradis” vor fi citite în greacă, română și engleză. Intrarea e liberă. Articol complet pe site. #Poetry #Translation #Bucharest

    spectatorul.com/2026/10/05/col

    #poetry #translation #bucharest
  2. mnl mnl mnl mnl mnl @mnl

    As you can see, under the hood, it’s not much code, it’s fairly trivial code once you know a couple of programming language fundamentals.

    @shriramk I’m sure can say much more eloquent things about this approach than I can.

  3. mnl mnl mnl mnl mnl @mnl

    RE: macaw.social/@jay/117393989247

    This is a direct consequence of the industry myopically framing the use of as “managing agents”, when the technology is nothing like it. It’s literally the corporate structure trying to imprint and capture a technology it doesn’t know how to deal with and that is actually so easily used to undermine the concept of a “corporation” itself.

    These models are language transformers, and we developers have always transformed languages and built tools and complete research fields that deal with that. I don’t prompt my model as “agents”, but as “planning with heuristics” and sometimes (very rarely) parallelism makes sense (as in, doing multiple inferences in parallel in order to create a single artifact).

    This entire “code review agents and design agents and planning and planning verification and adversary review” is just burning tokens for the sake of cosplaying hierarchical structures that have been established to control labor.

    What is actually possible when you approach llms from a hackers perspective, subverting and questioning the status quo, is hard to put into words. I have never hit any quota limits ever, and yet I am consistently the person putting out the most projects / artifacts in any ai forward group I am part of, shaking my head at people burning $200 for a round of code review that doesn’t help anything.

    Here’s basically my entire approach in two prompts:

    - “create an elegant language to do X”
    - “implement a compiler/interpreter for X”
    - “now do X. And X1. And X2. And X3.”

    The language is elegant and thus easy to read and generate in the context of X. The compiler is usually fairly trivial, and much easier to code review / formally analyze. And then you’re done, in less tokens than it would take to “write a plan to do a tiny subset of my as of yet partial understanding of X”.

    I’m serious, check attached screenshots, sonnet 5.5 medium, two prompts, of which only 4 are actually relevant to the thing at hand. The hard part (always was): knowing what the right primitives are, and how to compose them.

    This was the only way to get useful stuff out of gpt-3.5, and still is.

    claude.ai/share/9784fe15-cb53-

    #llms #llm #ai
  4. mnl mnl mnl mnl mnl @mnl

    Since I can’t show it yet (hopefully there will be at least a 20second snippet in today’s keynote), you’ll have to trust me on this one. 😂

    2/2

  5. mnl mnl mnl mnl mnl @[email protected]

    I think software engineers should start looking at #mathematics and seriously elevate the rigor of their thinking and communication.

    Now that “code” as the “maths-lite” has been “solved” by #llms #llm #ai, being able to communicate higher level concepts concisely, or conversely, read and understand another person or machine’s output is primordial.

    And for programmers, that really means abstract maths (usually), since so much of our code is based upon fairly simple structures (framing monads as monoid in the category of endofunctors is if you think about it how we developers would talk about things too if we spent too much time doing Haskell or scala, now that I’m able to swim a little further off the shore (which is like, nothing, actual mathematicians are kinda sailing the oceans), this is the most concise and easiest way for me to remember its structure, to then apply it to whatever (here my more traditional brain still takes over, but I don’t expect it to last much longer).

    Software has always been about combining the abstract with the very concrete, since software is built to be used, often by humans directly. That will not change with more powerful models, because otherwise what’s the point, if we want to burn compute we could just 10 GOTO 10 or sample /dev/urandom.

    I wouldn’t have been able to prompt the monster I built over the weekend (and yes, _i_ built it) without knowing a modicum of algebra, a hefty dose of computer architecture, some serious distributed systems knowledge, a sprinkle of temporal calculi, a big serving of compiler and PL knowledge, and a side dish of web design (design systems are just “algebras” of widgets, if you think about it). What I had to completely leave behind is anything that resembles traditional software engineering, it was full automerge.

    But, because everything is built on sound principles, I have no issue saying that this is basically one of the most readable and robust pieces of software I’ve ever produced (or seen, tbh). The moment I said “ok let’s test this”, everything just worked and has since. The only adjustments I did was change a font size and switch the UI timeline cursor to follow wallclock time (which is “less correct” but looks better).

    1/

    #mathematics #llms #llm #ai
  6. mnl mnl mnl mnl mnl @mnl

    This weekend, I probably built the most ambitious project I have ever worked on. It is a full graph compiler for vision pipelines, against cuda, metal, cpu, remote models, with load shedding, batching, time freshness invariants, NATS streaming, including a full UI design system, with ofc an extensive monitoring setup (down to individual GPU lanes), from zero in 2 days, and it just works, I can point 20 clients at it, streaming 50+ video feeds for 24h now, with the most ridiculous model graphs, javascript runtime callbacks that feedback into further CUDA inference. I can't share more here but it is absolutely mental.

    I used the opportunity to distill the codebase into a series of textbooks for me to learn about dataflow punctuations, handling multiple clocks (there's about 8 time domains in the system: incoming video, proxied video, vision system frame input, model computation clock, guaranteed result time, prospective result time (when we will have collected all the computations for the current frame, NATS transport time, webUI animation refresh, webUI timeline step, webUI frontier...).

    I don't really know what to do with this though, because literally this was a "last-minute demo thingie" and I was just like "YOLO let's do hands off the wheel make no mistake read the literature".

    Where did my expertise come in (because it certainly did)? Was it the understanding of how to create such a system using compiler techniques? Was it (barely) knowing how a GPU works? Was it (barely) knowing what dataflow and realtime streaming constraints are (bandwidth, latency, freshness invariants, load shedding, backpressure)?

    I think it's all of the above, the fact that actually designing the system in my head took about 30 seconds (I could just _see_ it), despite not really knowing what I'm doing?

    Extremely unsettling, even for me.

    #llms #llm #ai
  7. mnl mnl mnl mnl mnl @[email protected]

    "Anti-AI" is often used to refer to "everything bad openai and anthropic is doing", which ironically is just solidifying these company's stronghold over the technology and discourse, since any attempt to say "hey maybe the world is more complex" can easily be dismissed as "AI boosterism".

    And thus, lose any opportunity to change people's minds since those that you won't dismiss are already on your side.

    It's so evident now that I write it, how much easier it is to actually effect real change talking to people who are maybe naive or too enthusiastic about it, than trying to organize with people who think they have it all figured out.

    #llms #llm #ai

    #llms #llm #ai
  8. mnl mnl mnl mnl mnl @[email protected]

    The funny thing about all this "#llms are destroying software communities" is that I am just done showing what I work on or contribute to existing projects, if all that comes back is some knee-jerk "this is slop" or "all you know how to do is greenfield" (sure...) or the omnipresent "you are enabling fascists".

    I just keep all that stuff to myself and share it with a tiny circle of friends because fuck that. And in the meantime, no progress is being made on what collaborating on software in the age of #llm / #llms / #vibecoding (because it sure is a challenge) could actually look like.

    What it looks like with my friends: we exchange ideas, talk about ideas, share knowledge, and the code is irrelevant because it is so easy to recreate once you know what you are going for. So most collaboration is now face to face, on calls, with whiteboards, sharing books, inspiration, little html snippets and the occasional markdown file. PRs were always a crappy way to collaborate and 0 progress has been made since patch files were first shared per email in the nineties. (In fact, I'm sure that was actually a nicer way to collaborate than whatever github and co provide).

    So yeah, #llms are destroying communities for sure, but not in the way most people think.

    #llm #llms #vibecoding
Share on Mastodon

Enter the server where you have an account.