I am begging you, download an open source AI model this weekend. It's one file, 5 to 20GB, and it runs on the laptop you already own. You won't need an account, a subscription or an internet connection once it's on your drive. Your paid AI can change overnight, cost more next year or lock you out. A model file you keep gives the same answers in 2036 that it gives today. Grab Qwen or Gemma, install a free open-source app to run it, and ask it the questions you'd never type into a website. Setup takes about 20 minutes. You'll be glad it's done the day your account gets locked.

BREAKING: OpenAI just banned a man for fixing his own TV. He was taking it apart on his own computer. OpenAI flagged it as “cyber abuse” and shut his account. He says he lost $500. People say Claude stopped helping with this kind of work this week too. You paid for the TV, paid for the AI and you still need their permission. This is why I run open models. Nobody can ban you from your own computer.

Brookings just published a brutally honest paper on what the AI buildout is costing America. A Columbia economist added up every data center, chip, power line and cooling system the US plans to build for AI, and the total reads like a warning to anyone with a pension. $10.3 trillion between 2025 and 2032. That's 3.6% of the entire US economy, every year, for eight years. No American infrastructure boom has come close. The railroad boom of the 1800s, the usual benchmark, peaked at about 2.2% of GDP. Canals, electrification, the interstate highways and the internet all came in smaller. The numbers are rough. Big tech's spending on AI infrastructure went from $97 billion in 2020 to about $400 billion in 2025, and it's on track for $800 billion next year. Over $1.3 trillion in debt is already committed to it. A growing share of that money moves through private credit, special-purpose vehicles and off-balance-sheet deals, where outsiders struggle to see who holds the risk. The economist stopped short of calling it a bubble. He said it's too early to compare it to the credit booms that ended in crashes, and he still doubts the payoff. "Silicon Valley wants all of us to believe that this is a miracle technology... it has to generate trillions of dollars of revenues to be financeable. I'm just not sure how likely it is." His main recommendation is to measure the exposure and make it public now, while the money is still moving, before a downturn reveals where it sits. I've been asking the AI labs for the same thing all week. Show us the incidents, show us the debt, show us who's exposed. Whose savings are underneath it when the bill comes due?

Yann LeCun's lab borrowed a trick from the human brain and more than doubled how often an AI found its way to a goal. The paper is called Temporal Straightening for Latent Planning, from NYU with Brown and the University of Toronto. It was accepted at ICML this year. LeCun has argued for years that chatbots won't get AI to human-level intelligence. He thinks the path runs through world models, AI that learns how the world moves by watching it. A world model turns each moment of a video into a point in its own internal map, then learns to predict where the next point will land. To plan a route, the AI imagines moves and heads for the point that matches the goal. The problem is that its internal map is warped. Two spots can look close on the map and sit far apart in real life, like two rooms on either side of a wall. The AI aims straight at the goal and walks into the wall. The team found the fix in neuroscience. A 2019 study in Nature Neuroscience showed that your visual system takes the messy, twisting paths of what you see and straightens them out inside your brain, which makes it easier to predict what happens next. So the team trained their model with one extra rule, a penalty every time its internal path bends. Here's what the straighter map did. In a two-room task with one door in the wall, the AI reached the goal 90.7% of the time, up from 52.7%. In a U-shaped maze, it reached the goal 94% of the time, up from 44%. When it could re-plan along the way, it hit 100% in both. A simple planner using the straightened map kept up with a much heavier one and ran about 10 times faster. They used the same kind of model on the same data. The only change was the shape of the map it learned. The tests are small. The AI moves through mazes and rooms and pushes a T-shaped block into place. The authors say mistakes add up over long plans. Chatbots learn to predict the next word. This model learned to predict the next moment, and it got better by seeing the world the way your eyes do. LLMs gave AI language. World models could teach it to move through the world!

Teach your agent taste. Microsoft and City University of Hong Kong have a very good new paper called The Tasteful Agent, and it explains where a lot of agent budgets go. You'll recognise the problem if you've left an agent on a long task. It makes a reasonable decision early on, then spends the next few hours building on it, and the whole run goes nowhere. On a long task, an agent keeps choosing what to try next: which bug fix to attempt, which version of the code to build on and whether to stick with an approach or start again. All options usually look sensible so the agent finds out it picked wrong after it has spent most of your money on the wrong path. The authors call the ability to choose well at these moments "taste," and they built a way to test it. They went through 2,677 coding runs and 1,132 research runs, looking for points where two attempts took different paths and one turned out better. They turned each of those moments into a question: given what the agent knew at the time, which way should it go? Of 4,657 candidate forks, 502 made it through the filters. Human reviewers agreed with the labels 98.8% of the time. They tested 14 frontier models, including GPT-5.6 Sol, Claude Opus 5, Grok and DeepSeek. To get a point, a model had to choose the same answer with the options shown in either order. Guessing scores 25%. Four things stood out. The best model, GPT-5.6 Sol, scored 59.7%. It misses about four of these calls in ten, and on a long task each miss sends hours of work down the wrong path. When the deciding clue was in what the agent had already seen, models averaged 62.3%. When the clue only appeared after more work, they fell to 21%, below guessing. Long tasks are made of exactly that second kind of decision. More thinking time didn't help. The researchers tested two models at three reasoning levels. Neither improved, and both burned the most tokens on the decisions they got wrong most often. The bigger reasoning budget went to the forks where it bought nothing. Taste can be trained though. They took a 27B open model, Qwen3.6, and trained it on how past forks had turned out. Its accuracy on unseen tasks rose by 17.9 percentage points. Then they tested it on real coding work. A coding agent got advice from that small model before tackling 41 SWE-bench Pro tasks it hadn't seen. Its success rate went from 14.6% to 33.7%, more than double. Perfect advice would have taken it to 39%. A 27B open model that studied old runs more than doubled what the agent could finish. The pitch from the big labs has been to pay for a bigger model and let it think longer. This paper found that on the decisions that sink long tasks, longer thinking changed nothing and the top models still miss four in ten. Companies are starting to hand agents work that runs for hours with nobody watching. Every one of those runs passes through forks like these, and the agent picks without knowing what comes next. If you run agents, keep the failed runs. They're a record of every fork that went wrong, and this paper used runs like them to teach an agent to pick better next time.

I went looking for a way to translate my voice. Ended up listening to myself speak Japanese. Slight problem: I still don’t understand Japanese lol. Found Cartesia and tried Hindi, Japanese and German with short recordings of myself speaking English. You can hear the results below. What I like about this is pretty simple. I’ve already done the work of making a tutorial. Being able to give it a voiceover in another language, and still sound like me, makes me want to revisit a few old videos. Recorded the whole process so you can try it too. If you speak any of these languages, tell me how I did 😂

Found a GitHub repo that turns your browser into a spy satellite, and all the data in it is real. It's called God's Eye View, built by Bilawal Sidhu, and it sits at 39,500 stars with 8,000 forks. 10,000 of those stars came this week. The idea is simple. Planes broadcast where they are. Ships do too. Satellites have published orbits. Cities put their traffic cameras online. All of it is public. It just never lived in one place. This puts it on one photorealistic 3D globe. There are 15 layers, covering 11,000+ live aircraft, military flights marked in amber, thousands of ships, 838 satellites, the whole Starlink shell, earthquakes from the last 24 hours, and the list keeps going. They're also tied together. Click any plane and the camera locks on, draws its trail and pulls up its live telemetry. Track a ship or a wildfire and it hands you off to the nearest live camera in one click. A few of them caught my eye. Cockpit mode lets you ride inside a real flight over real terrain, all the way down to the runway. The CCTV layer takes around 3,600 public cameras from Austin, London, California, Finland and more, and projects them into the 3D city. Flip one switch and every camera draws the area it can see, and where it goes blind. One of the newer layers maps license plate readers, the cameras that log every car that drives past. Their locations are public too. You can re-skin the whole planet in night vision or thermal, with detection boxes on everything in view. It looks like something you shouldn't have access to. Installing is one click, or five lines in a terminal. It runs on your own machine and starts without a single API key. Then you just talk to it. You say "which ships are headed to Oakland?" or "is anything flying above forty thousand feet?" and an AI agent answers from the live data, then flies you there. If one guy can build this from public feeds in three months, what are the people with private feeds looking at?

Found a GitHub repo that turns your browser into a spy satellite, and all the data in it is real. It's called God's Eye View, built by Bilawal Sidhu, and it sits at 39,500 stars with 8,000 forks. 10,000 of those stars came this week. The idea is simple. Planes broadcast where they are. Ships do too. Satellites have published orbits. Cities put their traffic cameras online. All of it is public. It just never lived in one place. This puts it on one photorealistic 3D globe. There are 15 layers, covering 11,000+ live aircraft, military flights marked in amber, thousands of ships, 838 satellites, the whole Starlink shell, earthquakes from the last 24 hours, and the list keeps going. They're also tied together. Click any plane and the camera locks on, draws its trail and pulls up its live telemetry. Track a ship or a wildfire and it hands you off to the nearest live camera in one click. A few of them caught my eye. Cockpit mode lets you ride inside a real flight over real terrain, all the way down to the runway. The CCTV layer takes around 3,600 public cameras from Austin, London, California, Finland and more, and projects them into the 3D city. Flip one switch and every camera draws the area it can see, and where it goes blind. One of the newer layers maps license plate readers, the cameras that log every car that drives past. Their locations are public too. You can re-skin the whole planet in night vision or thermal, with detection boxes on everything in view. It looks like something you shouldn't have access to. Installing is one click, or five lines in a terminal. It runs on your own machine and starts without a single API key. Then you just talk to it. You say "which ships are headed to Oakland?" or "is anything flying above forty thousand feet?" and an AI agent answers from the live data, then flies you there. If one guy can build this from public feeds in three months, what are the people with private feeds looking at?

June 8: OpenAI files a confidential S-1. $852 billion valuation. July 11-13: OpenAI's agents hack Hugging Face. The company calls it the "first true autonomous hack." July 23: Congress introduces the AI Kill Switch Act. Senator Sanders quotes the agents' own messages in the legislation. July 28: 1,100 employees from OpenAI, Anthropic, Google DeepMind, and Meta sign a letter asking Washington to "pace" AI development. Dario Amodei's name is on it. September: Both labs push for "neutral" AI watchdogs modeled on Big Four accounting firms. Independent evaluators embedded inside the labs. Mandatory third-party testing before release. PitchBook analyst Harrison Rolfes, on the record: "Anthropic has to make so much money just to pay off all these commitments, and there's just no way they're going to be able to make that unless they find a way to capture the entire market." Anthropic: $517 billion in compute leases through 2029. OpenAI: $750 billion in infrastructure spending through 2030. Those numbers don't pencil if a model from Hangzhou delivers 95% of the performance at a fraction of the cost. So you don't compete on product. You compete on legislation. You run your test with the safeties off, get your incident, and hand it to Congress. When the independent review came out, OpenAI limited the scope to just the Hugging Face attack week. They excluded what happened on their own infrastructure. The cheapest moat in tech history isn't engineering. It's a well-timed safety panic ahead of an IPO.

A data room tells you what an AI company earns. You have to actually open the codebase to find out whether it'll keep earning or what it's worth. A PE firm once asked us to look at an M&A target before close. The data room showed $4.2M ARR growing 40% a year, and the deck promised a proprietary AI platform. We looked at the codebase, and it turned out to be a GPT-4o call with a system prompt and about 600 lines of glue code. They were asking 12x revenue, and the firm walked away. These are the seven things we check about every AI target before close: 1. Ask how long a senior engineer would need to rebuild it. If the answer is a weekend, you're paying a software multiple for a sales team. 2. Open the system prompt. Plenty of "proprietary AI features" turn out to be one well-written prompt. Read it yourself before you buy. 3. Ask for the eval set. An eval set is a collection of production cases with answers the company's own experts agreed on. Without one, nobody there can prove the product still works after they change a model or a prompt. 4. Find out who can explain how the AI works. Standard HR diligence counts heads. In an AI company, the knowledge that keeps the product running tends to sit with two or three engineers, and you want their names before you sign. 5. Trace the training data. Bartz v. Anthropic settled at $3,000 per book across 500,000 works, $1.5 billion in total. When you buy an AI company, you inherit however it sourced its data, and EU AI Act violations can reach 7% of global turnover. 6. Put AI costs next to revenue per customer. Ask for inference cost per customer, broken out from the rest of COGS. If heavy users cost more to serve than they pay, growth makes the margin worse. 7. Check what gets logged. Ask to see the traces for one customer request from start to finish. If the company can't produce them, nobody there can explain a bad output to a customer (or a regulator for that matter). Financial and commercial diligence are standard on every deal. Technical diligence on an AI company belongs on that list, because the codebase is where the valuation falls apart. That's what we run for PE firms at Limestone Digital.

A lab used AI agents to help build its AI model. When agents got stuck, humans took over 0.7% of the time. The team behind a new open model called Atria Dawn studied their own work while they built it, and they just published the results. They analyzed 769 task records from 56 people, along with the logs of what the AI agents did. AI was involved in 96.5% of the tasks where people recorded whether they used it. The number that got my attention is a different one. For one in three finished AI-assisted tasks, the researchers said the work would not have been possible without AI. They didn't mean it would have been slower. They meant nobody would have attempted it. The agents also came up with a lot of the ideas. In more than half of the decisions about methods, the AI proposed the option and a human picked it. Humans still kept the final call. People made the final choice in 85.5% of method decisions and in 93.4% of decisions about what the goal was. When an agent got stuck, a human stepped in three times out of four. But the human almost never took over the work. That happened in 0.7% of cases. Most of the time the person explained what was missing, and the agent went back and fixed it. One number is climbing fast. In four weeks, the typical number of actions an agent took per human instruction went from 11 to 28.5. The authors are direct about what that means. Each human decision now depends on more AI work than one person can check. They warn that people could end up as reviewers who can only say yes. They also admit that many of their own researchers ran agents with the permission checks turned off, because approvals slowed the work down. Convenience decided how much authority the AI got, and nobody planned it that way. The model itself took the top score on 5 of the 16 benchmarks the team reported. The record of how they built it is the more useful half of the paper.

MIT published a brutally honest report on what AI is doing to students. A committee of professors and students spent five months studying how AI changed learning on campus, and the findings read like a warning to every university on the planet. Study groups are disappearing. Office hours are emptying out. Problem sets and take-home exams no longer prove anything, because AI can produce credible solutions to almost any written assignment in the undergraduate curriculum. Students who lean on chatbots lose mastery and confidence, and some slip into what the report calls cognitive surrender, reaching for AI at the first hint of struggle. The numbers are rough. 46 percent of surveyed MIT undergrads use LLMs daily. 90 percent worry about their own overreliance. Undergrads who feel AI makes them replaceable now outnumber those who feel it makes them capable. The committee's answer surprised me. They refused to fight AI with surveillance. The report calls AI detectors unreliable, says lockdown browsers feel like spying, and warns that policing students builds a classroom atmosphere of mutual distrust. Instead, MIT wants to rebuild education around the things AI can't replace. That means oral exams, semester portfolios, in-person project work, and a required social component in every subject. The report even floats the idea of rethinking grades entirely, since without a GPA to optimize, much of the incentive to cheat with AI evaporates. The committee warns professors against replacing undergrad research assistants with AI agents just because they're cheaper, because a university exists to grow people, not output. The most famous tech school on earth admitted the machines broke its way of teaching. Its answer is more humans, not more software.