@mnl@hachyderm.io
@mnl@hachyderm.io avatar

mnl

@mnl@hachyderm.io

🏴‍☠️ I like computers! Silicon is alive! All hail abstraction! 🏴‍☠️

🌺 💐 🌸 he/him 🌸 💐 🌺

This profile is from a federated server and may be incomplete. Browse more on the original instance.

mnl, to webdev
@mnl@hachyderm.io avatar

So I’m genuinely puzzled by . Do you folks who use it return html fragments? Or do you just render a whole page and extract single nodes? If you do the first, how is this “better” than just returning json and rendering it on the frontend?

I don’t really understand the architectural premise and the hypermedia systems book doesn’t really help me click.

mnl,
@mnl@hachyderm.io avatar

@jochen pretty cool frameworks and creativity. I wish I had time to delve into this stuff more.

I was doing something "related" in 2010: https://www.goldeneaglecoin.com/

Mustache is rendered with JSON, on the server side they get compiled to PHP and the HTTP requests are routed "internally" as function calls. In JS, the same templates are fed a proper XHR.

We're replacing it with react 🤣 but I didn't have to modify a single REST endpoint from 10 years ago.

The react feels just as snappy, btw.

matt, to random

I'm getting tired of simplistic, indignant characterizations of generative AI like this one: https://social.ericwbailey.website/@eric/111584809768617532 "a spicy autocomplete powered by theft that melts the environment to amplify racism and periodically, arbitrarily lie"

It's a tool like any other; it can be used for good as well as bad. Yes, the copyright issue is real, but we can presumably overcome it by using models whose developers are more scrupulous about their sources of training data, not throwing out the whole thing.

mnl,
@mnl@hachyderm.io avatar

@maria @simon I corroborate the information with other sources, or validate it experimentally. I find google results harder and much more exhausting to parse through. ChatGPT can also provide citations, although tools like perplexity.ai are better.

One way to bypass the whole "facts? true or not" problem is to tell ChatGPT to ask questions/be the student/propose exercises/suggest other topics of interest.

I find the "most people don't know how to do that" a difficult argument to carry.

mnl, to LLMs
@mnl@hachyderm.io avatar

My mental image of : they are an encoded version of the part of human thought that arises from language, and an encoded version of previously made thoughts and knowledge. It’s an imperfect encoding, but it is generative.

Using llms is discovering, examining, experimenting and ultimately applying language’s intrinsic carrying of meaning and reasoning.

1/

mnl,
@mnl@hachyderm.io avatar

We are all multilingual: be it actual language (French, German, Swahili, …), levels of speech (formal letter vs shooting the shit with homies, …), programming (lisp vs rust vs js …), stem (maths notations vs physics notation vs engineering jargon), etc…

Using is not asking the machine to think for us, it’s putting our own thoughts through a kaleidoscopic prism and harvesting its reflections.

2/

mnl,
@mnl@hachyderm.io avatar

The input is human, the output is human in as much as a human ascribes meaning to it. If you picture as a language exploration and transposition machine, a lot of the things commonly seen as problems disappear.

Think of a proper input language, think of which reflections make most sense to harvest, and leverage that, instead of thinking it “creates” language and meaning.

3/3

mnl, to random
@mnl@hachyderm.io avatar

this conference on genAI for coding is so weird.

mnl,
@mnl@hachyderm.io avatar

@kellogh @vonneudeck if feels like uh, no one really knows what they are talking about. Really really weird. Still able to flip “ai lead” people’s brains out with 2 prompts.

after a talk like: ai is good at this and not at this and so blablabla, you can go “have you tried asking it xyz?”

Also, there’s a lot of historical blindness. I came with a stack of books because whatever, I didn’t have to pack many clothes, and all the ideas from the 80ies seem like they’re getting reinvented.

mnl, to LLMs
@mnl@hachyderm.io avatar

On the off chance, anybody at the codeforward.ai conference in Arlington, VA?

mnl,
@mnl@hachyderm.io avatar

@kellogh it’s today and tomorrow. I fly in and then this https://x.com/connormaj/status/1731853442056929425?s=46&t=raYrrDMm3ELWVH16gKzj4Q 😭

mnl, to fountainpens
@mnl@hachyderm.io avatar

This pocket printer is so nice to put links into your sketchbook. They have paper that supposedly holds 20 years (not this one, that’s cheap). But I expect most links to not hold up that long anyway.

mnl, to LLMs
@mnl@hachyderm.io avatar

I just had chatgpt write a ghidra plug-in that renames all the variables to something sensible and I’ve never used ghidra before and uh I reverse engineered my printer driver (and by I I’m not sure who I’m referring to).

@byte_swap

mnl, to random
@mnl@hachyderm.io avatar
mnl, to LLMs
@mnl@hachyderm.io avatar

If anything, LLMs show that the duality language/thought is porous. That's already embodied in the chestnut that "writing is thinking."

Anybody using ChatGPT can't deny that there is meaning encoded in its output, even if it just a derivation of its training corpus.

Give language a bit more credit.

I have a hard time understanding the "it's just fancy autocomplete/stochastic parrots/a language model" as a criticism.

1/

mnl,
@mnl@hachyderm.io avatar

@kellogh my toot or the boundary between thinking and language?

The fact that you can eloquently say some dumb ass shit imo shows that language on its can encode meaning, be it dumb ass shit or clever thoughts. Thoughts are not "separate" from speech otherwise you would never think of saying dumb ass shit.

And yeah those 3 are the questions that LLM surface, and they're sooo interesting.

1/

mnl,
@mnl@hachyderm.io avatar

@kellogh
Just saying "LLM's language doesn't encode meaning because there is no thought" is missing a bit part of why they are so valuable.

For one, they make it possible to play with this "dumb ass shit" vs "interesting thought".

And second, they show how you can get useful things purely out of language that has been converted to numbers and then transformed to other numbers, and that say probabilistic representations of language can general useful formal code, for example.

2/

mnl,
@mnl@hachyderm.io avatar

@kellogh doesn't have to be an inner dialogue, once you communicate in symbols for example the pure fact that you only have a limited set of symbols already does some shaping of your thoughts for you. I'm a noob in this domain, but I find this "it's just autocomplete" framing extremely reductive.

mnl,
@mnl@hachyderm.io avatar

@kellogh so for example i think a lot of my "engineering" is done by visualizing shapes in my head. They appear and then I draw them and then I turn them into words. Here I have a clear distinction and words often fail me. But also I can wax eloquent about abstraction without really thinking about it.

For example: this random garble of a thread :)

mnl,
@mnl@hachyderm.io avatar

@kellogh I like the separation of "meaning" from thinking, i don't think i can get anywhere thinking (!) about what thinking is. But the words that an LLM produces convey very useful meaning (at times).

Saying it's just random autocomplete is a bit like saying "search results are just hashtable values".

which like, true, also that's fucking cool

mnl, to random
@mnl@hachyderm.io avatar

THINGU

https://youtu.be/wrZ7PnolbQ4?si=7blxyyVG1hBTXtkI

johncarlosbaez, (edited ) to random
@johncarlosbaez@mathstodon.xyz avatar

Recently my work has switched to developing agent-based models for epidemiology and climate change, and helping the Fields Institute set up a "Mathematics of Climate Change" program, which may become part of a consortium between different math institutes.

These tasks are really absorbing me now. They're very different than anything I've done! I used to be focused on developing and explaining cool math and physics ideas. I still do that, but now:

  1. I'm spending a lot of time thinking about grants and trying to get institutions to collaborate. I'd like to blog about this - but I feel I shouldn't say much about any given step until it's succeeded. I'm used to being an outsider, free to blab about anything. But now I'm not. The people I'm working with might not like seeing blog articles about the details of what's going on, especially if things don't work.

  2. A lot of the math required for better agent-based models needs has already been created by applied category theorists. A lot of it just needs to be brought down to earth and applied! So part of my job is to learn this math, talk to people about it, synthesize it, and try to apply it.

  3. I'm more focused on concrete applications. For example, while I'm glad to have helped create modeling software based on category theory, I really want to see it get used. We're getting close, but we're not there yet. (Our software may get used for a project on the health of truckers, which would allow me to write an article called "Category theory for the hard-working trucker".)

Here are some more thoughts about the math of agent-based models:

https://johncarlosbaez.wordpress.com/2023/11/23/agent-based-models-part-4/

mnl,
@mnl@hachyderm.io avatar

@johncarlosbaez I'm not sure who to turn to, and I don't have the background nor the time nor the will (to be quite honest) to learn the necessary maths to be able to follow these articles, but I know I am fairly "mathematical" in my approach to writing software (because composing components only works well if the underlying mathematics are sound).

I've been sketching and refining what the underlying mathematical abstraction would be for (LLM) agents. 1/

mnl,
@mnl@hachyderm.io avatar

@johncarlosbaez There is a big runtime component that I can fairly easily abstract away with monads (including all the real-world nitty gritty of logging, usage metrics, at-scale deployment, persistence).

Critical to proper LLM applications however is the management of context. Not just that we need to pass the proper context down the monad (easy), but really that context is a multi-faceted composable object itself.

Agents could be part of context management.

2/

mnl,
@mnl@hachyderm.io avatar

@johncarlosbaez (By context, I mean what the additional information should be that the model is prompted with, which can be in a trivial case the result of the previous step, but a summary of the previous conversation,technical documentation, pieces of source code (that get updated by the agent itself).

In fact extrapolating further, since conversations can be branched, context could be "from the future" of another evaluation of the agent.

3/

mnl,
@mnl@hachyderm.io avatar

@johncarlosbaez
I feel there is much more to this that I could model with a monad to make composition elegant and practical (what would a UI for this look like? what would logs look like? What would at scale deployment look like if I have 100M agents?)

I would love to discuss this with someone coming from the more theoretical side. Do you have an idea where I could look?

4/4

mnl, to LLMs
@mnl@hachyderm.io avatar

openai's nonsense shows how much we need to focus on local models.

I played a bit with mistral and codellama on anyscale, and these things look very useful already (for the work I do), even though they require more prompting effort. These things are fast and run on my local computer as well.

I'm fine using some commercial service here and there for my exploratory work, because GPT4 can't be beat, but for building tools I rely on local models are necessary.

Very encouraging.

  • All
  • Subscribed
  • Moderated
  • Favorites
  • •
  • megavids
  • rosin
  • thenastyranch
  • GTA5RPClips
  • tester
  • InstantRegret
  • DreamBathrooms
  • ngwrru68w68
  • magazineikmin
  • everett
  • Youngstown
  • mdbf
  • slotface
  • kavyap
  • JUstTest
  • cisconetworking
  • khanakhh
  • normalnudes
  • osvaldo12
  • cubers
  • tacticalgear
  • Durango
  • ethstaker
  • modclub
  • anitta
  • provamag3
  • Leos
  • lostlight
  • All magazines
    •