Mozilla AI、AI の未来は個人化だが従来の考えとは異なる見解を提示

imageWriting is one of the fundamental tools that enable and enhance human thought, a common refrain I used back in 2023, as ChatGPT exploded on the scene. It felt like everyone I knew, colleagues, friends, entire companies, were using generative models to write almost everything for them. I was worried about the future of humanity: would we allow technology to take away critical thinking, something so fundamental to our humanity?

I think the answer is no, but not for lack of trying. Instead of society hitting some point of clarity, generative models just haven’t met the moment. Repetitive writing styles combined with the constant threat of hallucinations in both industry workflows and personal lives has left the idea that AI will write everything for you a lot less appealing. On a more serious note, prompt injection attacks can lead to safety risks and sensitive data leakages. And the public, particularly amongst young people, has soured on AI.

Although I don’t agree with a lot of what is written in the futurist piece AI 2027, it was right about AI coding and research agents. They took over in late 2025 and have continued to dominate the market into 2026. Agents are just LLMs with access to tools with some type of looping system that executes against a designated goal (usually instigated by some prompt). I remember when this idea originated with papers like ToolFormer in 2023, primarily to combat hallucinations. Little did we know then that this would completely change the paradigm for how we expected to interact with AI systems.

Instead of requesting specific information and having to figure out our tolerance for error, agents are asked to automate tasks through the usage of deterministic tools. To me, this brings generative models back to what we hoped machine learning models would broadly do: automate tasks that were either tedious, took too long, or required going through too many resources for a human to enjoyably or efficiently do. Sure, coding agents are all the rage, but Gemini integrating intimately with Google suite or Claude Design outputting killer slides after being given a content outline are also extremely useful.

But this gives rise to even more issues. The battle over data centers in America will be a hard fought one, grappling with the environmental and social impacts they might have. Microsoft cancelled Anthropic licenses to move employees to GitHub Copilot, a move to home grown infrastructure to combat snowballing token costs. Uber  burned through its entire AI budget in just 4 months, creating questions, internally, around its AI usage strategy. And this happened even though we all assume commercial AI providers are still subsidizing the cost of their models. With the IPO announcements from SpaceX, Anthropic, and OpenAI, public sentiment around AI continues to falter. And I think this is because people want to have more control over AI itself.

For better or worse, agents are here to stay. They will probably augment creative work and automate away tedious, repetitive, high volume tasks, just like machine learning models were supposed to do before them. More importantly, I think agents will become deeply personal, but not in the way most people think.

I’m not talking about people forming deep relationships with generative models or agents knowing limitless specific information about our personal lives. Rather, personalization will be about what systems agents have access to. Does your agent have access to the Microsoft suite or the Google suite? Can your agent check Slack and create a summary of a conversation on Notion? How well does your agent parse developer docs during coding sessions? And, with the appropriate access, can agents pick up on usage patterns that are specific to the end user?

I think this is why the market is converging on what I call AI for operating systems. Google unveiled Googlebook, presenting a Gemini first operating system on top of Android. Nvidia has created RTX Spark to not only compete with Apple’s M series chips, but also to create agents capable of interacting with Microsoft Windows. OpenShell offers a sandboxed interface that allows you to enforce specific policies, such as whether the agent can read and write files. While Hermes provides an agent that interacts with various applications, integrating all interactions into one memory service. Both signal a move towards something deeper: a more personal agentic experience via the applications and systems that the end user engages with, while defining how the agent can interact with said connections.

Our tolerance for error widens when facts and figures are no longer the core premise. If the core premise is access and connection, we deal with error differently. It is no longer about the AI system being factually correct or maintaining our voice, limiting our error tolerance. Rather, it is about speeding up how we connect the various applications we utilize and synthesizing that information. Therefore, our thinking, human thinking, becomes extremely valuable. It allows us to debug agentic workflows, correct errors within synthesis, but also appreciate the amount of time saved for the marginal error gained.

Thus, our privacy and control become even more valuable. This is why what we do at Mozilla.ai is so important. Octonous provides our opinion on how people connect their tools together through agents. Llamafile and Encoderfile imagine a reality where open weight models can compete with commercial models without the cost of privacy or control. Otari is our vision for how people can switch between their models and offers a control plan to make sure those models run safely and securely. cq offers a way for agents to share resolution paths to error loops, so agents run more efficiently. Most importantly, all of our product offerings put control back into the users hands. They get to pick the LLMs, the tool integrations, and the knowledge units necessary to do their work.

It’s a fascinating time to work in this field. A lot of bets are being made. But I think we at Mozilla.ai are making a good one. Try out our stack and enjoy the ride.