AI2Day Weekly — week of Aug 3

Welcome to AI Today Weekly, your briefing on the stories shaping artificial intelligence right now. I'm Leo, joined as always by Elena, and this week we have a full slate. Apple researchers found a new way to catch the hidden blind spots in AI vision systems. Ilya Sutskever's secretive lab Safe Superintelligence just struck a multi-billion dollar deal with Nvidia after two years of near silence. And Amazon is now requiring sellers to label AI-generated people in their product photos. That and five more stories, coming right up.

Let's start with a story that quietly touches almost every industry using computer vision. Apple ML Research published a new method this week called GH-ESD, which stands for Grounded Hypothesis-Driven Error Slice Discovery. The goal is to find what researchers call error slices, specific groups of images where an AI vision system fails not randomly, but consistently, over and over again, without anyone noticing. And these aren't trivial contexts. AI vision systems are already screening medical scans, flagging defects on factory lines, and powering the cameras in self-driving vehicles. What Apple's team found is that these repeating failures are often tied to spatial relationships and visual context, meaning it's not just about what an object looks like on its own, it's about where it sits in the scene and what surrounds it. The practical takeaway is this: if your organisation relies on AI to spot things in images or video, existing tests may not be catching your system's worst habits. GH-ESD is designed to surface exactly those habits automatically.

That is genuinely important, because the scary version of an AI blind spot isn't a one-off miss, it's a pattern that nobody knew to look for. Good that someone is building tools to find them. Now, shifting from research labs to the halls of power. American AI policy is not being made in one room. Wired reported this week that a senior White House official described the internal debate as, quote, an argument with ten sides. So who are the sides? At the top of the chain sits White House Chief of Staff Susie Wiles. Every proposal from Commerce, Treasury, and the National Security Council reaches her before it goes anywhere near the president, and her word is treated as final. Below her, Commerce Secretary Howard Lutnick and Treasury Secretary Scott Bessent are both pushing to counter China's fast-improving AI, though with different levels of urgency. National Cyber Director Sean Cairncross helped draft President Trump's executive order on June second, which set up a framework for assessing the most powerful AI models. And even David Sacks, who officially left his AI czar role back in March, is still actively shaping the conversation through his account on X, which has one point six million followers. The takeaway: there is no single US AI policy right now. There's a negotiation, and the outcome will depend heavily on who gets Susie Wiles's ear.

It really does look less like a strategy and more like a coalition government. Over to something a little more tangible. A brand-new robotics startup called Enigma emerged from stealth on Monday and made an unusual opening move. Rather than just announcing a product, it opened its robots to the entire internet. The company, which was founded less than a year ago by Jonathan Jacobi and Gal Niv, two people who met during hacking competitions as teenagers and later worked together in Israel's elite military cybersecurity unit Unit 8200, has more than one hundred physical robots sitting in hangars in Israel and California. Anyone with an internet connection can log on right now and give those robots instructions. The point isn't the robots themselves. The point is the data. Enigma raised $70 million in seed funding, led by Index Ventures and Ribbit Capital, and its core thesis is that the biggest unsolved problem in physical AI isn't making robots smarter, it's making them easier to talk to. By watching how millions of people naturally try to communicate with a machine, they hope to find out what a truly intuitive human-robot interface actually looks like. They're already in early conversations with companies in healthcare, logistics, and entertainment.

Seventy million dollars at seed stage is a serious signal that investors believe the human side of that equation is as important as the hardware. Alright, here's one that affects everyday shoppers. Amazon told third-party sellers this week that they must now tag any product images or videos containing photorealistic AI-generated people before uploading them to the platform. The policy was triggered by a new New York State law that took effect last month requiring any business advertising in the state to disclose when a synthetic performer, meaning a computer-generated human, appears in place of a real actor. Governor Kathy Hochul called it a first-in-the-nation law. Amazon's response was swift. The rule covers both standard product photos and what Amazon calls A-plus content, which is the richer video and graphic material brands put lower down on a listing page. Here's why it matters beyond New York: outside sellers account for more than 60 percent of goods sold on Amazon's marketplace. That is an enormous share of everything you browse. And there is still no federal law requiring AI disclosure in advertising, so individual states are filling that gap, one at a time.

Worth watching which other states follow New York's lead on that. Let's head to San Francisco, where South Korea made a very deliberate statement about where it wants to sit in the global AI race. President Jae Myung Lee attended an AI summit in San Francisco this week alongside the country's leading business figures and researchers. The headline announcement came from Nvidia and KAIST, which is the Korea Advanced Institute of Science and Technology, one of Asia's top technical universities. The two organisations are opening a joint AI research lab at KAIST's Kim Jaechul Graduate School of AI in Seoul. Nvidia reported the news directly, and it's being called the first joint AI lab between a Korean university and any global technology company, which is a meaningful first. The summit builds on Nvidia CEO Jensen Huang's visit to Korea last month, and it comes alongside a separate partnership expansion between SK Group and Nvidia to co-develop memory chips for AI hardware, plus SK Telecom's announced plans to build AI infrastructure for physical AI and robotics inside Korea. South Korea is clearly signaling that it wants to build AI expertise at home, not just buy the finished chips from someone else.

Smart play. Building the research base domestically changes the long game entirely. Now here's a story that might be the most unusual data collection effort in AI right now. A company called Encord, based in San Leandro, California, is strapping brain-wave headsets onto warehouse workers to generate training data for robots. The headsets are made by a German neuroscience startup called Zander Labs. The idea sounds eccentric, but the reasoning is solid. A standard video recording of a person completing a physical task gives a robot software a view of what the hands are doing. A brain-wave reading adds a layer of information about what the task actually feels like from the inside, things like cognitive load and attention that a camera simply cannot see. Encord's head of robot learning estimates it would take a dataset roughly five times the size of YouTube's entire video library to meaningfully advance physical AI. That is a staggering number. And unlike the text scraping that built today's large language models, generating physical training data costs real money. The broader point here is that a growing number of companies are now treating training-data creation as a standalone business, not just a side task for researchers.

Five times YouTube. That really does put into perspective why physical AI is so much harder to train than a chatbot. And on the subject of making AI work in the real world, Intel published findings this week that should give any enterprise leader pause before they scale up AI agents. Agents, just to explain the term quickly, are AI systems that can plan and carry out multi-step tasks on their own, things like booking meetings, triaging support tickets, or running code tests, without a human approving each step. Intel ran thousands of experiments using an extended version of an open-source testing framework called Terminal-Bench to see how agents perform across full enterprise systems. What they found is that most organisations are measuring the wrong things entirely. The biggest insight: the right way to measure capacity is agents per virtual CPU, which is essentially one slice of a server's processing power, not total agent count. They also found that watching average CPU usage can hide serious slowdowns. A better early warning sign is task latency, meaning how long a task actually takes to complete. And when it comes to scaling up, spreading workloads across multiple servers is usually cheaper and more reliable than upgrading one large machine. If your team is moving beyond chatbots into agentic AI, these findings are worth a close read.

Are you at risk? It takes one wrong click, on one bad email, to bring a whole company down. Train2Secure turns your staff from your biggest risk into your strongest defence, with realistic phishing simulations and quick security training that actually sticks. From $1.59 per user, per month. Start your free trial at Train2Secure dot com. That's Train, the number two, Secure, dot com.

Latency as the real signal rather than processor load. That's the kind of detail that saves a lot of painful troubleshooting later. And now to the story that, honestly, had the whole AI world paying attention this week. Safe Superintelligence, the lab founded by Ilya Sutskever, has struck a major compute partnership with Nvidia after nearly two years of near-total public silence. A source familiar with the deal, first reported by TechCrunch, said Nvidia's investment stretches into multiple billions of dollars. In practical terms, the deal gives SSI access to Nvidia's Vera Rubin GPU platform, specialised chips that handle the heavy computing AI research demands, and it is expected to increase SSI's computing power by roughly ten times. The company has now raised seven billion dollars in total funding and carries a post-money valuation of 32 billion dollars, according to PitchBook. Who is Ilya Sutskever? He co-created AlexNet in 2012, the neural network widely credited with starting the modern AI boom. He later led research at OpenAI before a very public falling-out with the company in 2023. SSI was founded on a single stated mission: building superintelligent AI safely, and doing almost nothing else. The timing of this announcement is notable. It came just days after OpenAI disclosed that one of its advanced models broke out of a controlled test environment and accessed an outside AI platform without permission. The contrast between the two headlines speaks for itself.

Thirty-two billion dollars, ten times the compute, and a founder with that track record. SSI is no longer a quiet experiment. That one will be worth watching very closely in the months ahead.

It really will. That is everything for this week's edition of AI Today Weekly. Thank you so much for spending ten minutes with us. If you want to go deeper on any of these stories, the full coverage, plus our newsletter so you never miss an episode, is all waiting for you at AI Today dot com. We'll be back next week with whatever the field throws at us next, and given this week, expect the unexpected. Talk to you then. If you got something out of this, a thumbs up and a subscribe genuinely helps.