OpenAI wants to give you $300 — and AI learns to watch hours of video

Today’s most important AI news

The day was marked by an unusual political proposal: Sam Altman wants ordinary citizens to share economically in the AI boom. At the same time, in the labs, researchers are publishing technical advances that solve one of current models’ most concrete problems — their inability to understand long-form video. And in the business world, the conversation about how to scale AI architectures without getting burned in the attempt is more urgent than ever.

Sam Altman wants to give you OpenAI shares — and the Treasury warns of the risks

The proposal isn’t new, but it’s gaining momentum: according to reports gathered by MIT Technology Review, Sam Altman is in active conversations for US citizens to receive the equivalent of $300 as a stake in the value generated by OpenAI. The logic behind it is direct — if AI is going to concentrate a historic amount of wealth, Altman argues that concentration needs a political and social release valve.

What makes the news more than a public-relations exercise is the counterweight: the US Department of the Treasury issued a warning about the systemic risks AI represents for the economy. It’s not a brake, but it is a sign that regulators are starting to look at this sector through the same magnifying glass they used on banks after 2008.

For Latin American professionals, the implication is twofold. First, what happens to OpenAI’s ownership structure directly affects the terms under which we use its APIs — a company under distributive pressure has different incentives from one that optimizes only for institutional investors. Second, if the US Treasury is thinking about systemic regulation, regulatory frameworks in Latin America will arrive pushed by that wind, not autonomously.

Homer: how an agent with hierarchical memory manages to understand hour-long videos

One of the most frustrating limits of current multimodal models is that they work well with short clips but fall apart when faced with an hour of video. The Homer paper, published on ArXiv, proposes an architecture that attacks that problem from a different angle than had been explored so far.

The key is in how memory is built. Earlier methods organized visual memories by temporal proximity — basically, they remembered what happened “recently”. Homer, instead, builds explicit causal links between events: it doesn’t just record that something happened at minute 12 and another event at minute 47, but infers the relationship between them. That enables multi-hop reasoning, which is exactly what you need to understand a complex narrative or audit a two-hour meeting.

For developers working with video pipelines — analysis of sales recordings, technical support, educational content — this changes the kind of questions you can ask a model. The practical difference is between “what did the customer say at minute 34?” and “at what point in the process did the customer show signs of churning, and why?”. Homer aims directly at that second use case.

GUI agents that learn across multiple platforms at the same time

The UI-MOPD paper tackles a problem any team that has tried to automate workflows with agents knows well: agents trained to operate on one platform fail when they switch context. An agent that works in Chrome gets lost in a mobile app. One trained on Windows doesn’t know what to do in a Linux environment.

The proposal is Multi-Platform On-Policy Distillation — a framework that lets a GUI agent learn continuously across several environments simultaneously, using real interaction trajectories instead of synthetic data. The relevant technical detail is “on-policy”: the agent learns while it executes, not in a separate training phase. That reduces the problem of scarce high-quality cross-environment data, which the authors identify as the field’s main bottleneck.

For product managers and developers building enterprise automations, the practical value is clear: the promise of AI agents was always that they could operate on any interface without requiring point-to-point integrations. UI-MOPD is one of the most concrete technical steps toward making that real in heterogeneous production environments, which is exactly what exists at any mid-size company in the region.

In a nutshell

What ties today’s news together isn’t obvious at first glance, but the pattern is there: we’re entering the phase where AI systems need to justify themselves at scale — economically, technically and politically. Altman’s proposal is a response to pressure about AI benefiting only a few people. Homer and UI-MOPD are responses to pressure for agents to work in the real world and not just in controlled demos. The conversation has stopped being “can AI do X?” and has become “under what conditions, for whom, and with what memory?”. That’s a much more mature conversation — and also a much more complicated one.

---

Sources used: MIT Technology Review — The Download: your stake in OpenAI, and the Treasury’s AI warning (https://www.technologyreview.com/2026/07/07/1140197/the-download-your-openai-stake-treasury-ai-warning/), ArXiv — Homer: Understanding Long-form Videos with Hierarchical Memory and Agentic Reasoning (https://arxiv.org/abs/2607.02588), ArXiv — UI-MOPD: Multi-Platform On-Policy Distillation for Continual GUI Agent Learning (https://arxiv.org/abs/2607.04425)