Tagged Llama · show all articles
Publishing a model's weights is a one-way action: it cannot be recalled, only reacted to. Four problems in that fact remain unsolved, and none of them are solved by a better license.
Open weights convert a vendor's benchmark claim from an assertion into something a stranger can rerun on their own hardware, byte for byte. That conversion is real and it is limited, and the Llama record shows both halves of it happening in public.
Four problems in open-weight AI have outlasted every fix proposed for them. Each is backed by a paper's own stated limitation or a documented incident, not a debate position, and none of them has a technical resolution yet.
A released checkpoint is usually described as an endpoint. It is closer to a beginning — ten separately documented failures, from safety training stripped in weeks to weights carrying a hidden payload, show what actually happens once the download starts.
Two questions decide how open-weight AI matures by 2035: whether it closes the capability gap with closed frontier models, and whether regulators and license terms converge on treating it like closed deployment. Four scenarios, each with a falsifier.
Six labs call their releases open, and each grants a different bundle: some cap commercial scale, some restrict use categories, one ships weights behind a paper alone, and only one also publishes the data, code, and logs needed to rebuild what shipped.
Meta's own papers specify exactly how Llama encodes position, shares attention across heads, gates its activations and normalizes its signals — and exactly what changed, and did not change, from Llama 2 through Llama 3 to Llama 4.
A permissive-sounding grant still carries a usage threshold, a use policy and a naming rule. What checkpoint you start from, how you adapt it and what you can prove about the result decide what actually ships.
Meta did not set out to start an open-weight movement. A gated research release leaked within a week, and everything that followed — the commercial license, the benchmark claims, a rival's answer — was improvised in response to that accident.