Governments, companies, nonprofits should invest in free, open source AI [pdf]
Calls for governments and institutions to fund free, open-source AI highlight a tension between public access to powerful models and the enormous capital and compute costs required to build them. Commenters compare AI to earlier open-source successes like Linux, but note that frontier LLMs resemble large-scale scientific projects (Manhattan or Apollo) more than typical software, raising doubts about whether volunteer or modestly funded efforts can keep up. Alongside proposals such as prize competitions and licensing mandates, others question whether AI currently delivers enough social value to justify massive public investment, especially given energy use, concentration of power, and unresolved issues around training data and copyright.
Open-Source vs Closed AI Economics
- Some argue governments and institutions should invest heavily in open-source / open-weights AI to avoid a future dominated by a few private labs and to capture the collaborative advantages seen in past FOSS success.
- Others counter that frontier models are extremely capital- and compute-intensive, so volunteer or lightly funded OSS can’t compete with firms investing billions and expecting returns.
- A middle view notes that much major OSS is already funded by corporations; similar models could work for AI, but only where the economics pencil out.
Commoditization and Where OSS Wins
- Comparisons to operating systems, databases, and compilers: in many of these domains OSS eventually dominated.
- Skeptics respond that OSS does best once a domain is commoditized and research is near the public frontier; where proprietary research and large capital outlays matter (e.g., frontier LLMs), OSS struggles.
- Some see token generation and mid-tier LLMs already trending toward commoditization, with multiple interchangeable models and downward price pressure.
Practical Barriers to Open Contributions
- Traditional OSS allows contributions from anyone with a laptop; LLM training requires expensive compute.
- This shifts “competition” from code quality to access to GPUs, raising the bar for meaningful contributions and making the classic OSS development model less applicable.
Alternative Funding and Prize Models
- Proposals include recurring inducement prizes for open models that meet strict benchmarks under fixed VRAM limits, possibly with corporate co-funding.
- Supporters think this could spur optimization and recognition; critics note evaluation creation, secrecy, and reproducibility would be costly and complex, possibly more than the prize pools.
Government Role: Invest, Regulate, or Restrain
- Some want large public investment or even state-led “Manhattan/Apollo-style” AI programs; others argue current political and economic conditions make such spending unrealistic or lower priority than climate, energy, or social needs.
- A different camp wants heavy taxation or tariffs on AI compute, seeing frontier AI as socially harmful, energy-intensive, and benefiting mainly “shovel sellers” (GPU vendors, clouds).
- Suggestions include mandated open weights (possibly noncommercial), publication of training data, or even state takeover in case of high risk—balanced by concerns about national-security competition and misuse risks from fully open frontier models.
Ethics, Data, and the Commons
- Ongoing disputes about whether training on public data without consent is “theft” or consistent with long-standing internet norms.
- Some argue commercialization of models trained on public data should require sharing outputs back to the commons; others highlight hypocrisy given widespread prior piracy and reuse of freely available information.