PAW's three-line pitch draws questions about how it works

@yuntiandengI do think marketing matters a lot tho. My English <-> Claudish demo got 1M views, and I explicitly said it was built with ProgramAsWeights. But I didn't see much of that translate into people using PAW. Getting people excited about something you built feels very different from getting them to see what they could build with it. PAW works very well for my own use cases, including harder tasks with the finetune compiler. But I'm genuinely curious what I'm missing: is it quality, usability, or how I'm explaining what it can do? Some examples: https://programasweights.com/demos

@willcbdude you've been replying to every jev tweet all week pushing your thing and it's getting a bit annoying lol of course it works for *your* use case, who else used it after the demo? did you DM anyone who expressed clear interest in the replies and see if they wanted any pointers / ran into any issues? the idea is good! but it's complex enough that people don't really know what it is on first glance (how tiny? how fast on a macbook? can it run in a browser? do i have to think about the qwen backend execution? does it need hf auth? why do i need to sign in with github?) i would focus more on visually engaging content (people love video) and on the agent-native experience, as well as giving people something like a skill to audit their workflows for fitting use cases. spray-and-pray marketing (esp in replies to other things) generally reads as desperate and casts doubt on the idea before people even look at it.

@willcbfor PAW in particular, it's confusing that you aren't pushing the github more. github is where these kinds of projects grow pre mass-market fit. you're jumping right to having a full community platform thing and it's not obvious why that's the right first move. people don't know what a PAW is yet really, and the code snippets look like it's a library everyone should build with (a la DSPy), but then the github is buried and it turns out it's just clients for a hosted service (why is it hosted? what benefit does this give me as a user? the marketing emphasizes how it's all local)

@yuntiandengFair on the repeated replies. But I think a lot of what you're describing is actually the marketing problem I'm talking about. When we launched in April, I promoted it much less and it got little attention. Replying more was my attempt to address that. Clearly that can backfire, but I've already tried several of your suggestions: Video demos (I built a 3D avatar controller), a playground, and both AGENTS.md and an agent skill (yes we did try that). Even though my Claudish demo got 1M views, little of that translated into PAW usage. I'm not ruling out product or usability issues. But I've tried everything I could think of but still struggling to get people to even take a serious look at it (including some of my reviewers who clearly didn't try the demo). I recently came across this article which conveys my point very well: https://cubednews.com/why-attention-is-the-scarce-resource-of-the-age/

@yuntiandengBut I cannot answer all questions people might want to ask in a single tweet, and the whole point of the demo is to allow people to try without GPUs or installing it. (here's our launch tweet back in April which includes answers to many of your questions: https://x.com/yuntiandeng/status/2044086557330579851) To directly answer your questions: 1. it's complex enough that people don't really know what it is on first glance It can be used in three lines of code (as shown on the landing page), or if you use the demo website you can try without installing anything: import programasweights as paw fn = paw.compile_and_load("Classify if an email is urgent or not") fn("Need your signature today") # "urgent" 2. how tiny? how fast on a macbook? can it run in a browser? Inference uses a 0.6B model, runs at 30 tokens/s on a MacBook M3 (it's in our arxiv abstract). Yes it can run in a browser: https://programasweights.com/browser. The compiler uses a 4B model. But it is hard to fit a paper's content within a few tweets. 3. do i have to think about the qwen backend execution? does it need hf auth? why do i need to sign in with github? No to all three questions. Sign in allows higher rate limit to prevent abuse (a lesson I learned from WildChat), but it works without sign in. Also models are public so you could set it up yourself without using our hosted compilation (inference by default runs locally), but compilation needs a GPU.

@yuntiandeng4. why is it hosted? what benefit does this give me as a user? the marketing emphasizes how it's all local This is the main technical point of PAW (which is the main novelty of our paper): understanding a task specification (once per task) and executing it (many times as we usually process many inputs for the same task) should be separate. Use a more capable model to compile the task into a much smaller, specialized model, then reuse that model across many inputs. The model that figures out what your function should do shouldn't have to be the one you run a million times. So our approach is compile-once/run-many. Compile-once is a bigger model relying on GPUs so we hosted it to make it easier for people without GPUs (but the model is public so advanced users can host it themselves). Run-many by default happens locally on CPUs as it's only a 0.6B model, so inputs are not sent to any external APIs.

@willcbokay then i think this answers my questions, and my takeaway is really that it’s indeed a pretty complex system, and you’re maybe doing yourself a disservice by trying to hide the complexity with “just 3 lines of code” as a researcher and developer i would *want* to understand this complexity, but then it’s a very different tool from jev, and “just like jev but local and open source” isn’t the right message jev is very clear about what it is and isn’t i would study DSPy and how its community was built organically and how it was marketed, it’s a closer parallel imo in terms of managed complexity with a simple dev-facing ui that blends local and remote execution

@willcbin particular, i really had to dig to get confirmation that it’s LoRA search. you could just say that outright and it’s much clearer! your target user probably knows what a LoRA is! “PAW synthesizes examples to compile a LoRA for <1B models that run on your laptop and beat XXB models in Y% of the time for structured tasks like A/B/C. one-time compilation runs on GPU; use your own (DGX spark, 4090 etc; MLX coming soon), or sign in for hosted access and community sharing”

@yuntiandengagreed, PAW's underlying compiler-interpreter system is quite different from Jev so I tried to avoid sounding like it is "open-source local Jev". For example, PAW does not output confidence scores and it cannot guarantee to output one of the predefined categories. I apologize if my replies made it sound like PAW is open-source local Jev, but the main reason I replied to many is because looking at their applications, they don't need confidence scores and can be simply solved using PAW. It feels especially tempting when I saw the email triage / log filtering / fuzzy search examples, which are exactly what I used as examples in the paper. Also, I think the audience issue is fair: researchers might find it hard to understand from the 3 lines of code how it works, although one thing I've been struggling these days is how to get a paper enough attention for others to even open it :)

@willcb@yuntiandeng like the launch tweet thread was pretty good! the people who clicked with it are your target users if you’re going after jev users, show them things it does that jev can’t do. people perceive jev as free, instant, and trivial to use, so it’s hard to compete on cost/ease

@willcbi would double down on showing off the cool things you can do with PAW with engaging content (threads, demos, videos, graphics) that show how your target users might use it if your target is research impact, focus more on the methods (what training recipes? any scaling laws?) stop it with the reply spamming though lol

PAW's three-line pitch draws questions about how it works

TLDR

11 Sources

Combined views

PAW's three-line pitch draws questions about how it works

TLDR

Combined views

Sentiment

Summary

Sentiment

Summary

Sentiment

Summary

11 Sources

Sentiment

Summary