AI2Day Weekly — week of Sep 14

Welcome to AI Today Weekly for the week of September fourteenth. I'm Leo, and coming up this week: a humanoid robot company wants a two-point-five billion dollar valuation on one-point-eight million dollars in sales — we'll explain why that math makes sense to some investors and not to others. DeepMind just mapped every possible typo in human DNA, all nine billion of them. And five major AI model releases landed inside a single week, so we'll cut through the noise and tell you what actually changed. That's all ahead. Let's get into it.

We start with humanoid robots and a filing that will raise some eyebrows. Agility Robotics, the company behind a two-legged warehouse robot called Digit, just submitted its financial disclosures to the Securities and Exchange Commission ahead of a stock-market listing. The headline numbers are striking. Net sales last year: one-point-eight million dollars. Operating loss: one hundred and forty million dollars. Planned valuation: two-point-five billion dollars. That gap — roughly fourteen hundred times annual revenue — tells you exactly where humanoid robotics sits right now. The company is going public through a SPAC, which stands for Special Purpose Acquisition Company. In plain English, a SPAC is a shell company that already trades on the stock exchange and merges with the private firm to get it listed without a traditional IPO. The deal is expected to raise more than six hundred and twenty million dollars, including two hundred million from Foxconn, the electronics giant that makes much of the world's consumer hardware. Agility's bet is this: Digit robots have now logged more than sixty-five thousand operating hours across nine customer sites, and the company projects twenty-five thousand robots deployed by twenty-thirty-five, generating roughly two-point-five-five billion dollars a year in subscription revenue. That is a very large projection sitting on a very small foundation. Whether you find that exciting or alarming probably says a lot about your appetite for early-stage technology investing.

Thanks, Elena. From robots to the human genome. Google DeepMind this month launched something called AlphaGenome Atlas, and the scale of it is genuinely hard to absorb. Human DNA can be altered at roughly nine billion individual points — each a single-letter change that might do nothing, or might cause a serious disease. Testing each one in a laboratory would take longer than human civilization has existed. DeepMind ran them all through its AI model instead and stored the results in a free, searchable database for academic researchers. The dataset is one petabyte in size — that's more than thirty times larger than the AlphaFold protein database DeepMind released a few years ago, which itself rewrote structural biology. Each variant gets a single score called an AVI — an AlphaGenome Variant Impact score — that summarises the predicted effect in one number, covering both the protein-coding parts of DNA and the stretches that regulate it. Early results are encouraging. Researchers at the Broad Institute used the score to flag a variant in a gene called DNM1 that was later confirmed in the lab to be linked to a severe form of epilepsy. A researcher at the University of Exeter applied the tool to whole-genome data from more than fifty-four thousand participants in the UK Biobank and found rare variants tied to common traits. Tools like this tend not to make headlines for a week and disappear. They quietly reshape what is possible in medicine over years. That is the practical takeaway here.

Our sponsor Zero2Hero is a poker site with the coach built into the table. Every mistake you make gets flagged and graded automatically as you play, with the cost in big blinds, and you can ask the coach about the hand while it is still live. The free tier is real: 50 hands a day against AI opponents, no card needed. Try it at Zero2Hero dot A I. Zero, the number two, hero, dot A I.

And speaking of reshaping entire industries — let's talk about Nvidia. Most people know Nvidia as the company that makes the chips powering almost every serious AI system on the planet. It is also now one of the largest strategic investors in technology. The company's equity portfolio — meaning the value of the shares it holds in other businesses — reached ninety-nine billion dollars as of late July this year, up from roughly seven billion dollars just a year before that. Two years ago the figure was two-point-two billion. In 2026 alone, Nvidia committed more than forty billion dollars to financing rounds across the AI industry. The biggest slice went to what the industry calls frontier AI labs — the research companies building the most capable models. Nvidia says it has invested nearly fifty billion dollars into that group. Its stake in Intel has grown to a market value of thirty billion dollars. Its SpaceX holding was worth twenty-one billion as of June. And as we covered on September third, Nvidia has agreed to acquire Hugging Face, the open platform where researchers share AI models, for twelve-point-nine billion dollars. The strategy is straightforward: Nvidia sells the infrastructure everyone uses to build AI. Now it also owns stakes in the companies buying that infrastructure. That creates an enormous feedback loop, and it is worth watching closely.

And those Hugging Face and model release stories connect directly to our next segment — because last week was one of the most cluttered weeks in recent AI history. Five major releases landed inside five working days. Anthropic put out two new models on Tuesday. Meta and Google followed on Wednesday. OpenAI released on Thursday. A research lab in Abu Dhabi also published work during the same period. And Nvidia's Hugging Face acquisition sat over all of it. For businesses trying to figure out which AI tools to use, the pace is genuinely disorienting. The CEO of AI startup Runpod, Zhen Lu, put it plainly to CNBC: quote — I feel like model fatigue is a real thing. There is just so much frothiness that you have to make noise — end quote. That framing matters. Not every release represents a meaningful leap. According to one technology executive we spoke with, the two releases that actually moved the needle in recent months were Anthropic's Claude Fable 5 back in June and a model called Kimi K3 from China's Moonshot AI in July. The practical takeaway for anyone buying AI tools: ignore the week of the announcement. Wait two to three weeks and see which models are still being talked about. The signal-to-noise ratio will be much better. And for context on the broader money flowing behind all this: Gartner projects total AI spending will reach two-point-five-nine trillion dollars this year — a forty-seven percent increase over 2025.

Now let's shift gears to something that does not involve a press release or a valuation. This is about what AI is actually doing inside scientific laboratories. A growing body of peer-reviewed research — and I want to stress that, peer-reviewed, not company announcements — shows that AI is now proposing experimental designs that outperform anything researchers came up with by hand. In quantum optics, which is the study of individual particles of light and how they interact, a program called Melvin was built back in 2016. It searched automatically for quantum light experiments. Researchers built some of what it suggested in real labs and confirmed the results. A more recent tool called PyTheus, published in 2023, generated one hundred distinct quantum optics experiments on its own, many involving photon arrangements that no researcher had previously thought to try. Moving into bigger physics: a 2025 paper in Physical Review X described AI-designed layouts for gravitational wave detectors — the instruments that sense ripples in spacetime — that outperformed the best human-designed next-generation versions. AI-designed instruments now span particle physics detectors, fusion reactor coils, and neutrino-detecting antennas. The practical implication is not that scientists are being replaced. It is that the space of experimental ideas a small research team can explore is expanding dramatically. That matters for drug discovery, materials science, and anywhere else the bottleneck is not funding but human imagination.

Over to an unusual segment now. We keep an eye on our own coverage statistics here at AI Today, and last week the numbers told a clear story about where the industry's attention went. In the seven days from August thirty-first to September sixth, we published thirteen stories mentioning Google's Gemini — the family of AI models that powers Google's chatbots, its search answers, and now parts of Gmail. The week before, we ran three Gemini stories. That is a three-hundred-and-thirty-three percent increase, week on week, and Gemini accounted for six-point-six percent of everything we published across one-hundred-and-ninety-eight pieces. More importantly, eight separate outlets fed into that coverage. This is not one reporter with a preferred beat or one press release doing all the work. When a story shows up across eight different newsrooms in the same week, something real is happening at the company. The short explanation is that Google shipped product. The biggest single item was the release of Gemini 3-point-8 Flash — its third new AI model in six weeks — a smaller, faster version of its main model aimed at high-volume, cost-sensitive use cases. The trend line is worth watching. Sustained product velocity from Google, combined with Nvidia's growing investment in the AI ecosystem, suggests the competitive pressure at the top of the market is not easing.

Next, a story about AI safety that has direct consequences for anyone using an AI assistant at work. A research paper published via Hugging Face — from a team at a company called Multiverse Computing — found something that should concern anyone who builds or deploys AI tools. Standard safety training is causing AI models to refuse perfectly reasonable requests at a surprisingly high rate. The study tested a mid-size language model called Qwen-3-8B, using political content as the test case. Before safety training, the model incorrectly refused safe prompts about two percent of the time. After standard safety training, that number jumped to seventy-four percent. Three out of four safe prompts were blocked. Why does this happen? Most safety tools work by blocking broad topic areas. If the word election or political manipulation appears near a prompt, a guard model — a second AI that checks responses before they go out — flags the whole thing, even if the question is completely benign. The paper also found that during training, nearly twenty percent of examples were silently dropped before anyone noticed. The fix the researchers propose involves something they call boundary pairs: matched prompts where one should be blocked and one directly beside it should not. Using this approach cut false refusals near problem areas from thirty-three percent down to four percent. The takeaway for anyone managing AI at an organisation: if you are only measuring how often your AI refuses bad content, you are missing half the picture. You also need to measure how often it refuses good content.

And we close with a story from the factory floor. The common argument in AI research circles is that robots will get smarter when they get more data. Rudy Cohen, co-founder and CEO of a robotics startup called Inbolt, says that argument is wrong for anyone actually running a factory. His company builds a real-time control layer for industrial robots. Most factory robots follow a fixed, pre-programmed path. If a part sits at a slightly different angle on a conveyor belt, the robot misses it. Inbolt's software adds computer vision and AI that watch the part's actual position and adjust the robot's movement mid-action to compensate. They call it in-hand localisation. The company has deployed this across more than one hundred factories and logged over forty million robot cycles in production, working with lines at Ford, Toyota, and Stellantis. Cohen's argument, which he will make at the RoboBusiness conference in Santa Clara on October twentieth, is that the real bottleneck in physical AI is not data collection. It is the time and cost of installing systems on real factory floors without bringing production to a halt. That is a different problem than the one most AI researchers are focused on, and it is the one that is actually slowing commercial deployment. If you are in manufacturing, that distinction matters a great deal.

That is our week. Thank you for listening to AI Today Weekly. If you want the full stories on anything we covered today, along with our daily briefing, head to the newsletter — link in the show notes. That is the week in AI. Full stories and the daily briefing at A-I-2-Day dot live. That is A, I, the number two, D-A-Y, dot live. See you next Monday. If you got something out of this, a thumbs up and a subscribe genuinely helps.