Threads' 100M Sprint + Llama 2 Goes Open

Archive note: this is a backdated post, written years later while rebuilding this site. It’s dated to the moment it covers, but the hindsight is real.

Meta, of all companies, owned July twice.

On July 5, Threads launched — Instagram’s text-based Twitter rival — and hit 100 million signups in five days, beating ChatGPT’s record without breaking a sweat. On July 18, Llama 2 dropped: a family of capable language models (7B to 70B parameters, chat-tuned variants included) with weights you can download and — the actual news — a license permitting commercial use, free, for anyone under 700 million monthly users. Which is everyone except about five companies.

One of these stories is bigger than the other, and it’s not the one with the bigger number.

Threads: distribution is the product

Threads’ sprint proves something useful but narrow: when you can convert an existing two-billion-user graph with one tap, “fastest-growing app ever” is a button you press, not a thing you earn. The interesting questions are still open — retention (early signs: steep drop-off), whether Twitter’s self-inflicted exodus actually lands anywhere, and whether Meta honors its promise to federate Threads over ActivityPub. That last one is the only part I care about long-term: a Meta property speaking an open protocol would be the biggest mainstreaming of protocols-over-platforms ever. I’ll believe it when my Mastodon account can follow a Threads user; the incentives for walled gardens to grow doors are historically poor.

Llama 2: the Stable Diffusion moment for text

The pattern I flagged with image models — closed API first, open weights within months — just completed for language models, at production-relevant quality. The first Llama leaked in March; Llama 2 makes it official strategy. Meta can’t beat OpenAI on a closed-model race it entered late, so it’s commoditizing the layer instead — classic “if you can’t win the product, open the platform” — with the side effect that every developer on Earth now has a capable LLM they own.

What this concretely changes:

1. The API is no longer the only door. Supply-chain logic applies to AI now: building on a model API is a vendor relationship — pricing, deprecations, content policy drift all included. Weights on your own disk are immune to all three. For anything sensitive — client data, regulated workloads — “the model runs here and the data never leaves” just became an available architecture.
2. The fine-tuning cottage industry begins. Within days, specialized variants are appearing — code models, language-specific tunes, uncensored forks (with everything that implies). The bazaar dynamic that made Stable Diffusion’s ecosystem explode is now running on text.
3. My homelab is paying attention. The 7B and 13B models, quantized, reportedly run on consumer hardware — the same llama.cpp wave that’s been building since spring. That experiment is happening on my own machines next month; it deserves its own post.

The strange new map

Tally the year so far: OpenAI sells the frontier closed, Apple sells hardware and says nothing, and Meta — the company this blog has spent years criticizing — is accidentally the patron of open AI, for ruthlessly sensible strategic reasons. Motives aside, the result stands: the most important capability of the decade now has a version nobody can revoke.

Next month: it runs on my hardware or I’ll have written a very different post.