A look at BigScience, a global effort of 900+ researchers backed by NLP startup Hugging Face, that's working to make large language models more accessible
Kyle Wiggers / VentureBeat : Tweets: @thom_wolf Tweets: Thomas Wolf / @thom_wolf : Great overview of what's been going on in BigScience! @BigscienceW is taking an under-explored path I find very exciting: what if building large language model was more like #NASAWebb & @CERN LHC and less like the not-peer-reviewed drones in #DontLookUp? https://venturebeat.com/...
Context & Ripple Effects
BigScience began in mid-2021 as a 500-researcher plan to build an open-source large language model independent of any single company, and by this profile the effort has grown past 900 researchers under Hugging Face's stewardship. Thomas Wolf frames the ambition explicitly: build frontier models the way CERN and NASA run big science — peer-reviewed, shared infrastructure — rather than the closed, unaudited releases coming out of corporate labs.
First-order effects
- Academic and independent researchers get a path to train and study GPT-class models without negotiating access through a company, directly countering the closed-lab gatekeeping that defined 2020–2021 NLP.
- Hugging Face converts its startup brand into scientific legitimacy, positioning itself as the neutral host for a research consortium rather than just a tools vendor.
Second-order effects
- Corporate-led alternatives look different by contrast: Facebook's earlier AI Language Research Consortium shows the incumbent pattern of company-convened partnerships, which BigScience's independence-first structure now competes against for talent and credibility.
- If the consortium ships a competitive open multilingual model — as the later BLOOM release did — closed labs face pressure to justify restricted access on safety grounds alone rather than capability scarcity.
Third-order effects
- A durable split emerges in how frontier models get built: concentrated capital inside a few labs versus distributed, CERN-style consortia, with access policy becoming the battleground between them.
- Wolf's own later argument that current paradigms won't produce true scientific breakthroughs suggests the open-science bet is also epistemic — that outside-the-box progress requires exactly the independent, non-corporate research lane BigScience is building.
The trend: Frontier language-model development is bifurcating between closed corporate labs and Hugging Face-style open research consortia, making who controls model access a structural question rather than a product feature.