PAW's Viral Demo Fails to Convert, Sparking Developer Marketing Debate

The outreach struggles of ProgramAsWeights (PAW) author yuntiandeng triggered a public discussion about how to market developer tools: his English<->Claudish demo hit 1 million views and was explicitly labeled as built with PAW, yet converted almost no users, and the author himself publicly reflected on "where the marketing went wrong."

Confirmed

• Developer willcb publicly criticized yuntiandeng for spamming PAW promotion under related tweets all week, calling it off-putting and pointing out that a million views still brought no users.
• willcb argued the project's concept is so complex no one can grasp it at a glance: questions like how small the models are, how fast they run on a Mac, whether you need to manage the Qwen backend, and why GPT is needed are all left unanswered; he had to work hard just to figure out that PAW is essentially LoRA search/compilation, and suggested saying so outright, since target users already understand LoRA.
• willcb offered a ready-to-use formulation: PAW searches for LoRAs for models under 1B using synthesized examples. He opposed slogans like "just 3 lines of code" or "like jev but local and open-source" that mask complexity, arguing that researchers and developers precisely want to understand the underlying complexity, and suggested PAW adopt a DSPy-style narrative.
• willcb also questioned PAW's open-source strategy: such projects should grow on GitHub until they find mass-market fit, yet the author jumped to a self-built community platform; the code looks like a general-purpose library like DSPy but is actually just a client for a hosted service.
• yuntiandeng responded to the architecture criticism: the core innovation of the PAW paper is separating "understanding the task specification" from "executing the task"—a task is compiled once, then repeatedly executed locally on large volumes of inputs by a 0.6B small model (compile-once/run-many).
• yuntiandeng clarified that PAW's architecture differs considerably from Jev's: PAW doesn't output confidence scores, nor does it guarantee outputs belong to one of a set of predefined classes, so it shouldn't be treated as an "open-source local Jev"; though he observed many applications simply don't need confidence scores.

Unconfirmed

• The performance details willcb mentioned (e.g., running speed on a Mac, model size limits) have yet to receive specific data-backed responses from the author.

Why it matters

The dispute reflects a common dilemma for research-driven open-source tools: real technical innovation exists (task compilation, local execution with small models), but unclear positioning and mismatched channels can severely blunt outreach—a demo with a million views and near-zero conversion is a cautionary case for independent developers.

2026-09-21 ~ 2026-09-21 · 11 related posts

• [source] 1M-view demo, no users: PAW creator opens up on the build-vs-adopt marketing gap — yuntiandeng · 2026-09-21
• 'Spray-and-pray marketing reads as desperate': dev critique of 1M-view demo that converted no users — willcb · 2026-09-21
• PAW critique continued: burying the GitHub repo undermines pre-product-market-fit growth — willcb · 2026-09-21
• [source] PAW Author Explains Compile-Once/Run-Many Design: 0.6B Model Runs Locally on CPUs — yuntiandeng · 2026-09-21
• Dev Argues PAW Shouldn't Hide Its Complexity, Points to DSPy as the Better Playbook — willcb · 2026-09-21
• 'Just 3 lines of code' backfires: dev argues tools should expose complexity, not hide it — willcb · 2026-09-21
• [source] Dev urges PAW to just say it's a LoRA compiler for sub-1B models instead of hiding the mechanism — willcb · 2026-09-21
• PAW author clarifies: not open-source Jev, just 3 lines of code for triage tasks — yuntiandeng · 2026-09-21
• Dev advice for PAW launch: say 'LoRA search' outright, differentiate from Jev on capability — willcb · 2026-09-21
• Tweets are the new abstract: how to actually get people to read your paper — willcb · 2026-09-21
• How to promote a research tool: demos for builders, methods for researchers — willcb · 2026-09-21