When @sama told me at the AI summit in Paris that they were serious about releasing open-source models & asked what would be useful, I couldn’t believe it. But six months of collaboration later, here it is: Welcome to OSS-GPT on @huggingface! It comes in two sizes, for both maximum reasoning capabilities & on-device, cheaper, faster option, all apache 2.0. It’s integrated with our inference partners that power the official demo. This open-source release is critically important & timely, because as @WhiteHouse emphasized in the US Action plan, we need stronger American open-source AI foundations. And who could do that better than the very startup that has been pioneering and leading the field in so many ways. Feels like a plot twist. Feels like a comeback. Feels like the beginning of something big, let’s go open-source AI 🔥🔥🔥

Context is king for AI Agents. It’s clear that the more relevant context the agent has to complete the task at hand, the more differentiated and successful the agent will be. Thus, when building AI Agents, you need to work backwards from that ideal context, and figure out what kind of product it would take to deliver it. This is a core part of the moat you’re building in AI. Here are a few of the major categories starting to emerge: * Corporate knowledge: Agents are going to require access to a wide degree of corporate data to make decisions properly, in which case the AI systems with the best access to this data will often win. This can either be because they’re directly have the data in their system already, have a “right” to index or get embeddings from another system (especially for code bases and other knowledge), or have nailed integrations across platforms via MCP, A2A, or otherwise. * User memory: As we’re seeing in the earliest stages of memory inside of products like ChatGPT, having a complete history of what that user has done with the AI can be incredibly powerful. The more the AI Agent can build on from past interactions, and understand all of the nuances about a user, the better results it can deliver. We can also imagine new sources of context beyond chat history, like what is that user’s role in the company, who do they tend to work with, what type of data do they tend to work on, and more. * Domain expertise: Ultimately, AI Agents are going to have the most impact when they’re not just generalists, but when they augment work in particular lines of business or verticals (like financial advisors in banking or researchers in life sciences). This means that the more the agent understands all of the specifics about that industry’s workflows or proprietary data at a granular level will be critical. One can imagine a future where 10,000s or 100,000s of tokens are used up in the context window just explaining the job function and specific instructions for that agent. * Tool use: The best AI agents are going to know their way around a variety of external systems to be able to get work done successfully. Knowing when to search the web vs. answer questions from internal knowledge vs. talking to a specific software platform will make all the difference in successful AI Agents. This problem is *not* easy as we can already see in consumer AI apps. * User experience: While this is hard to sustain as a differentiator, nailing the right Agentic UX is critical for adoption, and successfully completing workflows. Knowing if the agent should be interacting with a user via chat, or in-line with an existing UX, or needs a new interface altogether can make major differences in the customer success and adoption. There are going to be multiple ways to build moats when building AI Agents, but context is going to be king. Some of mix of the above -and probably plenty more- will be necessary to win.