The Rise of Open Weights
SCOTTLOGIC.COM
I’m off on holiday very shortly, and have had a bit of a hectic week, so this issue is a little different. Rather than sharing a small curated collection of blog posts and articles, I am sharing just one thing!
The Rise of Open Weights is something I created over the weekend. While the current events around open weights (which if you’ve read previous issues you’ll be very aware of) are fascinating, the path that the industry has taken to get here is even more interesting.
It’s quite a long read, and is presented as a long-scroll story (which was a lot of fun to put together), mapping out the following journey:
Some of the things that surprised me while researching and writing this piece:
- 2015–2022: OpenAI began life as a non-profit committed to openness, yet gradually became one of the most closed AI companies in the world.
- May 2023: Google’s leaked “We Have No Moat” memo predicted that open-weight models would eventually catch the frontier. Three years later, it increasingly looks like that prediction was right.
- 2017–2025: The cost of training frontier models exploded from under $1,000 to almost $500 million, fundamentally changing the economics of AI.
- January 2025: DeepSeek wasn’t a miracle built on a shoestring budget. The story is much more nuanced, and says a lot about how rapidly AI knowledge has diffused.
- 2023–2026: The gap between frontier and open-weight models collapsed from around a year to just a couple of months—and, on some benchmarks, has now disappeared altogether.
- 2026: Perhaps most surprising of all, as open-weight models have become more capable, commercial models have become more expensive, more tightly controlled and increasingly influenced by geopolitics.
- Today: The real question is no longer “Which model is best?” It’s whether organisations should continue renting AI from a handful of providers, or start owning the capability themselves.
So, if you have 20 mins free, please do check out the story.
I’m off 🏝️