AI2Day Weekly — week of Jul 27

Welcome to AI2Day Weekly, your ten-minute briefing on the stories shaping artificial intelligence right now. I'm Leo, joined as always by Aria, and we have a packed week. We're talking about a new test that exposes just how badly AI fumbles long videos, a quiet but important update to the hidden plumbing that connects AI to your everyday apps, and a free Chinese chatbot that has somehow managed to start a very public brawl inside the Trump White House. Let's get into it.

We start with a benchmark you've probably never heard of, but that points to a real gap in AI video tools. Researchers at Apple ML Research have released something called LVSum, a structured test designed to measure how well AI systems can summarise long videos. The dataset contains 72 videos spread across 13 subject areas, from cooking demonstrations to legal proceedings, each one running about 16 minutes on average. Human annotators wrote up to 10 summaries per video, and here is the key detail: every summary is pegged to specific timestamps in the footage. The AI has to say not just what happened, but when. That turns out to be surprisingly hard. Current multimodal models, the AI systems that can watch video and process language at the same time, consistently lose track of the order of events as recordings get longer. If you've ever wanted to skip straight to the moment a decision was made in a meeting recording, you already understand why this matters. LVSum was built to measure exactly that gap.

That is a genuinely useful benchmark because it mirrors what people actually need from these tools. And staying with Apple ML Research, the team has also published work on a completely different problem: how AI understands physical, three-dimensional space. The new technique is called RayRoPE. The RoPE part stands for Rotary Position Encoding, which is an established method for helping AI models understand where pieces of information sit relative to each other. The Ray part is what's new. Instead of tagging each small tile of an image with a flat grid coordinate, RayRoPE describes the tile's position using the actual ray of light that passed through the camera lens to capture it. Imagine handing someone twenty photos of the same room taken from different angles. A person mentally stitches those into a three-dimensional picture. RayRoPE gives AI a more geometrically precise way to do the same thing, and it works even when the camera moves in complex, unpredictable ways. It's a quiet piece of research, but the kind of foundational work that makes future AI applications more reliable.

Solid foundational research tends to matter more in the long run than the flashier announcements. Speaking of infrastructure, let's talk about something called the Model Context Protocol, or MCP. If you've never heard of it, that's kind of the point. MCP is the invisible plumbing that lets an AI assistant reach outside itself and actually do things, read your calendar, query a company database, check a support ticket. Without a shared standard like this, every single connection has to be hand-built by an engineer. A company wants its AI to read from its internal database? Custom code. Then it wants calendar access? More custom code. It compounds quickly, and smaller teams simply cannot keep up. An update to MCP is now in progress with the goal of lowering that complexity for developers. The practical upside, if adoption spreads, is AI assistants that work with more of your everyday software right out of the box, without months of custom engineering work behind the scenes. It's not a headline-grabbing product launch, but it's the kind of change that quietly expands what AI can actually do in real workplaces.

A quick word from our sponsor, Train2Secure. Your people are your biggest cyber risk — and your strongest defence. Train2Secure runs realistic phishing simulations and short, engaging security-awareness training your team will actually finish, with compliance-ready reporting that runs on autopilot. Turn your staff into a human firewall. Start your free trial today at Train2Secure dot com — that's Train, the number two, Secure, dot com.

Infrastructure updates rarely get the attention they deserve, and this one is worth watching. Now, from invisible plumbing to very visible policy. The United Kingdom is rewriting its national crisis response plans for the first time since 2004. To put that in context: when those plans were last updated, YouTube did not exist and the first iPhone was still three years away. Two things pushed ministers to act now. Britain shattered temperature records in May 2025, then broke those new records again in June. And the government has flagged that AI is giving criminal groups the ability to run cyber-attacks faster and at a greater scale than human operators alone could manage. Alongside the policy overhaul, the government launched a public preparedness campaign and ran a live stress-test exercise called Operation Albiston Shadow. The practical advice being issued to households is pretty straightforward: keep bottled water, hold some physical cash, and own a battery-powered radio. Simple things, but they reflect how seriously officials are treating disruption scenarios that were simply not on the radar twenty years ago.

That advice about physical cash and a battery radio sounds almost old-fashioned, but the reasoning behind it is very much a product of the current moment. Let's stay on unintended consequences, because the next story is one that snuck up on a community that really was not expecting it. Birdwatching forums are being flooded with fake sightings generated by AI image tools, and it is starting to threaten the integrity of citizen-science databases that professional researchers actually rely on. The issue crystallised around a genuine moment: in June 2025, a western reef heron, a species you would normally have to travel to Africa or the Mediterranean to see, turned up on the north Wales coast. It made national news. People drove hours for a glimpse. But that same enthusiasm for rare sightings is now being exploited. AI image-editing tools can add or alter details in a photograph in seconds, convincingly enough to fool experienced birders. The term circulating in these communities is AI slop, meaning low-effort AI-generated content posted with no concern for accuracy. The stakes go beyond hurt feelings. If fake sightings make it into species population databases, the science built on top of that data becomes unreliable.

That is a genuinely tricky problem because the hobby has always relied on photographic evidence as its gold standard, and that standard is now easier to fake than at any point in history. Over to robots, and a company making a very deliberate point about what industrial robots actually need to be. AGIBOT, a Shanghai-based robotics company founded in 2023, unveiled four new machines at the World Artificial Intelligence Conference in Shanghai this month. The company says it produced its fifteen-thousandth robot last month. Their flagship product, the A3 Ultra, is a humanoid that stands one point seven four metres tall and is rated for up to eight hours of operation before needing a recharge or battery swap. At a factory in Nanchang, China, AGIBOT robots have been inspecting roughly three thousand tablet units per shift over more than sixty-four hours of continuous operation. The company's message is deliberate: the benchmark for embodied AI should no longer be whether a robot can pull off an impressive demo once. It should be whether the machine can be manufactured, delivered, and integrated reliably into a real operating environment, day after day.

That framing matters because the gap between a great demo and a dependable machine is where most robotics companies have historically stumbled. And on the theme of making robots work alongside real people, Boston Dynamics is approaching the same challenge from a different angle. The company ran a free webinar called The Art Behind Human-Robot Interaction, focused on Atlas, their electric humanoid robot, which is one of the few humanoids currently in actual commercial trials. The core argument from the speakers is that predictability, not a human-like appearance, is what makes a robot safe to work near. One of the session's speakers spent fourteen years as a character animator at Disney Studios before moving into robotics, and that background is not incidental. The team is borrowing principles from animation, the idea that exaggerated, clearly readable motion communicates intent better than subtle, naturalistic movement. Large language models are also being used to move robots away from rigid pre-programmed scripts toward more adaptive behaviour. The goal is a machine that a factory worker can read and anticipate, the same way you learn to read a colleague's movements over time.

Animation principles in a factory setting, that is a combination I did not expect, but it makes a lot of sense when you think about how people actually interpret movement. And finally, a story that is as much about politics as it is about technology. A free, open-source AI model called Kimi, launched by Chinese company Moonshot in the week of July fourteenth 2025, has exposed a bitter split among the people who advise the Trump administration on artificial intelligence. Early tests suggest Kimi is roughly as capable as the paid models sold by OpenAI and Anthropic. The fact that it is free and open-source changes the competitive picture significantly. What followed the launch was a very public argument. David Sacks, who served as White House AI and crypto advisor until March 2025, publicly called Anthropic's models lobotomized and woke. Emil Michael, a senior Pentagon official, called OpenAI's new head of strategic futures a, quote, supreme village idiot. This is not a theoretical policy disagreement happening quietly inside a committee room. These are key figures in US AI policy fighting openly online. The Trump administration has already moved to curb a process called distillation, where AI models are trained on the outputs of rival models, which US companies say Chinese firms use to copy their capabilities. But Kimi's arrival suggests those measures have not closed the gap.

A free Chinese model that matches expensive American ones, combined with a very public meltdown among the people responsible for the US response to that challenge. That is not a reassuring combination. And it is a reminder that the competitive dynamics in AI right now are moving faster than policy can keep up with. This week has really shown that. Whether it's crisis plans that are two decades out of date, science databases being polluted by AI-generated fakes, or emergency plumbing standards that could reshape how AI plugs into everything, the technology is creating pressure in places people were not watching.

Exactly right. And that is what we're here for every week. That is it for AI2Day Weekly for the week of July twenty-seventh. Thanks for spending ten minutes with us. For deeper coverage of every story we touched on today, including links to the original research papers and source reporting, head to a i 2 day dot com. You can also sign up for our weekly newsletter there, so the briefing lands in your inbox before the weekend. We'll see you next week.