AI Has a Discovery Problem: Why Finding the Right Model Is Getting Harder

The AI industry has spent years racing to build bigger and better models. According to developer Manuel Hacevedo, it has largely neglected the infrastructure for finding them. In an essay that circulated widely this week, Hacevedo argues that the number of available AI assets has grown from hundreds to tens of thousands in a few years, while the tools for searching, comparing and verifying them have barely improved.

The practical cost lands on working developers. Choosing between a fine-tuned Llama variant and a task-specific embedding model often means sifting through model hubs, GitHub repositories and forum threads with little to go on beyond a README file and a handful of benchmarks. That friction is not just annoying; it steers people toward whichever names they already recognize, which tends to entrench a small group of dominant models at the expense of smaller, more specialized alternatives.

Hacevedo's diagnosis is that the field has invested heavily in model capability and comparatively little in the plumbing that helps practitioners locate the right tool. The release cadence makes it worse: new models arrive weekly, documentation is uneven, and capabilities overlap in ways that are rarely spelled out. The gap is not confined to open-source ecosystems either. Commercial marketplaces and enterprise AI platforms curate their catalogs around partnerships and visibility, which does not necessarily map to technical fit.

Researchers face a sharper version of the same problem. Reproducing a published result requires knowing not only which model was used but its exact version, training data and configuration, details that are frequently scattered across papers, code repositories and release notes, or missing outright. Without that information, a result can be difficult to verify or build on, which slows the cumulative progress that research depends on.

Some remedies are emerging. Hugging Face remains the largest repository of models and datasets, but its search and filtering have not scaled with its catalog. Newer directories aggregate benchmark scores and offer side-by-side comparisons, yet they are fragmented and lean heavily on self-reported numbers. There is no standard metadata schema for AI models, no universal tagging convention and no accepted quality signal beyond human judgment, Hacevedo notes.

The stakes go beyond convenience. If the best tool for a job is hard to find, adoption consolidates around familiar names, and work from smaller labs or individual researchers can go unnoticed for lack of visibility rather than lack of merit. Hacevedo's proposed fix is a search engine for models, with strong filtering, comparable benchmarks and documentation standards, an ambitious ask that no single platform currently delivers.

What to watch is whether the major hubs treat discovery as a product problem rather than a side effect of scale. Standardized model cards, versioned metadata and third-party evaluation that does not rely on self-reporting would each chip away at the gap. Until then, the search cost stays with developers, and the next widely adopted model may be the one that is easiest to find rather than the one that works best.