Open source.

If the open web doesn't have an answer to AI, AI will swallow the open web.

We released OLMo with everything: weights, training data, training code, evaluation suites. Open-source AI cannot be a marketing word, it has to be all the way down.

The future of enterprise AI is custom models trained on the company's own data. That requires open foundation models you can adapt, not closed APIs you have to depend on.

Bad actors can use open source models; good actors can use them too. The real risk is if regulation effectively bans open source.

The open-source AI ecosystem is the only credible counterweight to closed-frontier capture. Backing it is an act of structural diversification of the AI economy.

Open models are the only way to ensure that AI development remains a shared enterprise, not an oligopoly.

Open and decentralized infrastructure for foundation models is the only way to make sure academic research and small companies can keep contributing to the frontier.

“Open science and open source create a safer path for development of the technology by giving civil society, nonprofits, academia and policymakers the capabilities they need to counterbalance the power of big private companies.”

Context: Testimony to the US House Science, Space, and Technology Committee.

“We propose a unified framework that converts every text-based language problem into a text-to-text format. The simple objective lets us study transfer learning systematically.”

Open models are a public good.

Centralized AI poses risks no single regulator can solve. We need open infrastructure that lets users own their data and models.

We need open research and open models. The closed-lab approach is a danger to civil society.

“Science advances more quickly when we share our knowledge. When we put all of our knowledge in common, we push that boundary faster.”

If RLHF is opaque and proprietary, the open-source ecosystem is locked out of the most important phase of model development. Building open post-training tooling is the answer.

Open-weight models are the path to China's leadership in AI.

Open compilers like TVM are the only path to a diverse AI hardware ecosystem.

I think most developers want to use open-source AI. I think most of the AI ecosystem is going to be built on open-source models.

Regulating AI right now means regulating math. Open source is the antidote to the regulatory capture being attempted by the largest labs.

The biggest value of generative AI for most businesses comes from training and customizing models on their own data, in their own infrastructure. That requires open foundation models.

“BART pretrains by corrupting text and learning a model to reconstruct the original text. The denoising objective generalises across many language tasks.”

We can't talk seriously about AI safety while the post-training of frontier models is opaque to outsiders. Open-source post-training is the precondition for everything else.

Open-source AI is the path to keeping innovation accessible and trustworthy. The alternative, a small number of opaque models controlling the future of the technology, is the worse outcome for both users and societies.

Open data is the raw material of an AI future that is more equitable. Without it, AI development concentrates in whoever already controls the largest private datasets.

DSPy lets you treat LLM calls as differentiable building blocks, optimized end-to-end against a metric. The shift is from prompt-tuning by hand to compiling LLM programs.

Outdated laws should not be used to force platforms to police speech automatically. The principle that providers of communication tools are not responsible for individual user actions has to survive the AI era.

The AI revolution runs on open-source that nobody is funding sustainably.

Latent diffusion models compress the image generation problem into a learned latent space, making high-resolution generation accessible to consumer hardware.

Open replication is the lifeblood of an honest research field.

Open-source tooling and open educational materials are how the ML field reproduces itself. We can't outsource that to a few frontier labs without losing something important.

Open infrastructure is what allowed AI research to scale. It is what allows new entrants to compete.

Science doesn't work in secret. Open research is how safety work gets done.

Open infrastructure is what allowed many actors to build AI. Closing it now would be a fundamental mistake.

“The web has evolved into an engine of inequity and division; swayed by powerful forces who use it for their own agendas.”

If only frontier labs can run frontier models, then only frontier labs will shape what they do.

Frontier-lab tokenizers and pretraining mixtures barely cover African languages. The community has had to build its own datasets, evaluations, and benchmarks, often from scratch.

Open data infrastructure is the precondition for democratised AI work.

The smartest among us do not want to dominate the others. The idea that because a system is intelligent, it wants to take control is completely false.