AIThe LLaMA Leak Was the Most Important Release of the Year
In late February, Meta announced a family of language models called LLaMA and made the weights available to researchers on request.
Within about a week they were on the public internet. Within a month someone had them running on a laptop, and a group at Stanford had fine-tuned one to follow instructions for a few hundred dollars of compute.
Whatever anyone intended, that sequence has answered a question the industry has been arguing about for three years.
What Actually Happened
Three things in quick succession, and each one mattered.
The weights escaped. However you feel about that, the practical consequence is that capable model weights are now permanently in circulation. You cannot recall a file.
Somebody made them run on ordinary hardware. A project appeared in March that reimplemented the inference code in C++ with aggressive quantisation, reducing the precision of the weights so the model fits in far less memory. The effect was that models people assumed needed server-grade accelerators started running on personal machines at usable speed.
Somebody made fine-tuning cheap. Stanford’s Alpaca work took the base model and tuned it to follow instructions using a modest set of examples, for a compute cost in the hundreds of dollars. The alignment step that made hosted assistants feel useful turned out not to require anything like the resources people assumed.
“In five weeks the field went from ‘you need a research lab’ to ‘you need a laptop and a weekend.’ Nobody planned that sequence and nobody could have stopped it.” — Sameer Gupta
What This Means for a Business
I’ve been writing for years that the interesting question is not what the frontier can do but what an ordinary team can afford to run. This is the largest single movement in that line I’ve seen.
On-premises deployment is now real. The blocker for a great many organisations was never enthusiasm, it was that using these systems meant sending data to a third party. Legal said no, and legal was right. Running a capable model inside your own network changes that conversation entirely.
Fine-tuning on your own material is affordable. Not a research project. A task. Combined with low-rank adaptation, which makes the training itself cheap and the resulting adapters small, specialising a model for your domain is now within an ordinary team’s reach.
Vendor dependency becomes a choice rather than a condition. I flagged this risk when GPT-3 launched as an API. Your roadmap ran through somebody else’s pricing page. It doesn’t have to any more.
Costs become predictable. Per-token pricing is fine in a pilot and unpleasant at scale. Hardware you own has a known cost and a marginal request price near zero.
The Part I’m Less Comfortable About
I want to be honest that I hold the positive view above with less certainty than the tone suggests.
The safety work is not in the weights. The refusals, the guardrails, the moderation layers that hosted services apply are largely wrapped around the model rather than baked into it. An open model does what it’s asked. If you deploy one, that entire problem is yours, and most companies adopting these have not budgeted a single person for it.
The licence position is murky. These weights were shared under research terms and then distributed in ways those terms did not contemplate. Building a commercial product on that foundation is a legal question I would not answer confidently, and I’d want it answered before it’s load-bearing.
Capability spread is genuinely dual-use. I made an argument in 2019 about what happens when plausible text becomes free, and concluded then that the capability was arriving regardless of any lab’s release decision. This month proved that in a way I’d rather have been wrong about.
“I’ve argued for years that the open version always arrives and that waiting is a viable strategy. I did not expect the gap to close in five weeks, and I’m not sure the field is ready for how fast that was.” — Sameer Gupta
What I’d Do Now
- Run an evaluation, internally, on your own data. Not to deploy. To calibrate. Most people’s mental model of what open models can do is six months out of date and it’s now out of date every six weeks.
- Architect for portability. Keep the model behind an interface. Keep your prompts, your retrieval layer, and your evaluation set independent of any provider. The right position is to be able to switch, cheaply, in either direction.
- Budget for safety if you self-host. Content filtering, prompt injection defence, output validation, and someone who owns it. This is the cost that makes self-hosting less obviously cheap than it appears.
- Get the licensing reviewed before you build. Not after.
Final Thoughts
The pattern I described last year around image generation has now repeated with language, and it repeated considerably faster. A capability appears as a rented service, and within months an open version arrives that runs on hardware people already own.
I expect that to keep happening, and I expect the interval to keep shrinking.
For anyone building on these systems, that’s the planning assumption I’d use. Not which vendor to commit to, but how to stay able to move. The frontier will always be rented. Everything eighteen months behind it is about to be yours, and for most business problems eighteen months behind is entirely good enough.