Good morning. AI companies are worried their models are getting dumber from training on AI generated content. So they’re paying people to create high quality training data. But the work is low paid, and to keep up, people are using chatbots to do it. Which means companies are paying workers to feed AI more AI. The industry built the shortcut. Now it’s coming back through the front door.
Don’t give AI agents the company card. Princeton researchers put 14 AI agents in charge of a simulated software company with $1 million in seed capital and zero customers. The agents had 500 days to run the business, set prices, buy ads, fund R&D, and handle support. Almost all of them failed because they couldn’t make business decisions that compounded over time. Only Claude Fable 5, Claude Opus 4.8, and GPT 5.5 finished with more money than they started with. Then came the embarrassing part. A fixed rule script with zero intelligence earned $15.76 million and outperformed most of the LLMs. The script made the same boring decisions while the AI found new ways to lose money. So before spending six figures on loopmaxxing, maybe see if a simple cron can do the job better. (TechTimes, arXiv)
Brain implants are starting to look like overkill. The best brain-to-text systems still require surgeons to put hardware inside the brain. Meta’s Brain2Qwerty is trying to learn what’s on your mind without the skull opening part. The noninvasive system trained on brain recordings of people typing sentences while wearing an MEG scanner. In the peer reviewed paper, Brain2Qwerty cut the character error rate to 29%, beating EEG at 65%, which is good enough to make the implant crowd sweat. Meta says its newer version now translates brain activity into text with 61% word accuracy. Its best participant hit 78%. The catch is this is still a lab setup with healthy people typing, so nobody is walking out with a mind reading helmet. But brain-to-text is improving outside the brain, and every extra point of accuracy makes skull drilling a harder sales pitch. (Meta, Nature)
Fast APIs can still waste your users’ time. A quick wrong answer or retry loop won’t show up in a pretty benchmark table. You.com explains why raw latency misleads teams and how to measure what matters in production.
Stop optimizing for the demo. Download the guide.
👀 closer look
Physical AI will never have its ChatGPT moment. That kind of breakout only happens when the training data already exists. The best open source robot data has less than 5,000 hours of real world interaction, while LLMs were trained on trillions of data points scraped from the web. Getting more robot training data is expensive. Every new robot lesson has to be captured in the real world before a model can use it. So the race is on to make robot learning cheaper. Scale AI is building a library of people performing tasks. Nvidia is building world models. Ground Truth Machine adds biosignals, tracking brain activity, heart rhythm, sweat response, eye movement, breathing, and muscle tension while someone performs a task. Different tools, same problem. Physical AI has to manufacture experience before it can scale. Robot intelligence is about to make preschool look underpriced. (Forbes)
Ford tried to automate quality control, and it didn’t go well. To cut costs, the company tried replacing engineers with AI powered cameras that could spot defects during production. Ford bet automated checks could do the work of people who had spent decades learning where vehicles fail. Then the misses got expensive. Warranty repairs hit $4.8 billion in 2023 and recalls piled up, so Ford brought back more than 300 veteran engineers to catch failure points before parts reached the plant floor. They are now training the systems Ford thought could replace them. Recall and warranty costs are falling, and Ford just took the top mainstream spot in JD Power’s quality study for the first time since 2010. The cheap version of expertise got expensive fast. (BBC)
fun stats
☀️ 50% off. Discount Anthropic gave California to roll out Claude across state government, promising to make it more efficient.
🏗️ 70%. How much faster AI infrastructure spending is growing than cash earnings at the 5 biggest cloud builders: Microsoft, Amazon, Alphabet, Meta, and Oracle.
👤 1.2 billion. Live web profiles YC backed AI search layer Clado can query with natural language. Describe who you’re looking for, and it finds them, enriches the results, and hands over verified emails. Scary useful.