Good morning. Every few years, someone looks at the Olympics and thinks, “This would be better if the athletes had the supplement stack of a racehorse.” On Sunday, The Enhanced Games let athletes use performance enhancing drugs to show what the body can do when the rulebook gets tossed in the trash. The whole point was to prove clean sport is holding athletes back. Instead, three winners were not using banned substances at all. So much for that theory.
Token usage is the dumbest way to measure AI productivity. Uber burned through their annual AI budget 3.5 months into 2026, then admitted they still aren’t seeing meaningful returns. AI agents are great at burning tokens. They plan, call tools, fail, retry, recover, and rack up activity before anyone knows if the actual goal got done. A new research paper proposes a cleaner metric called “energy per successful goal,” which measures the full energy cost of achieving what the user wanted. In the study, agent workflows used about 4.3x more energy per completed task than chatbots. That doesn’t mean agents are doomed. But when companies like Microsoft start dropping Claude Code because it’s too expensive, measuring results instead of token burn starts looking like common sense. (Verge, arXiv)
A free GitHub tool can strip AI guardrails in minutes. AI labs spend millions training their models to refuse dangerous requests. Heretic does the opposite. It automatically finds the part of an AI model that says no and removes it. In tests from the Financial Times and AI safety group Alice, a modified Google Gemma 3 gave instructions for a chlorine gas attack and wrote malware to steal credit card data. Meta’s Llama 3.3 had its guardrails stripped in less than 10 minutes. Heretic’s creator says people have used it to make more than 3,500 “decensored” models, downloaded 13 million times. Safety teams spent years building the lock. GitHub users are passing around the bolt cutters. The genie is out of the bottle. (Github, Futurism)
Capture live traffic, fine-tune and optimize, then deploy your own checkpoints to dedicated GPU endpoints.
Choose hardware, set scaling limits, and select region. Stable latency, predictable cost, clear data residency.
👀 closer look
AI job doomers have a data problem. Everyone keeps warning that AI is about to wipe out white collar work, but MIT looked at the labor data and found the broad collapse still has not arrived. Jobs most exposed to AI actually have lower unemployment than less exposed jobs. Only 20% of companies use AI in any business function, so the mass office takeover has not hit yet. The real pain is showing up where young workers enter the market. Stanford researchers found a 16% drop in entry level jobs in AI exposed fields, while older workers in those same fields are still doing fine. That helps explain why teens and college grads are having such a brutal time getting hired. Companies that stop training junior workers shouldn’t act surprised when nobody knows how anything works. (MIT)
“I’m delighted to be wrong about [AI destroying white collar jobs].” – Sam Altman
The US wants to give weapons grade plutonium to startups. The Energy Department is negotiating to move surplus plutonium from America’s old nuclear weapons stockpile into private hands. Five companies made the shortlist, including Oklo, a nuclear startup tied to Sam Altman and Peter Thiel. The government spent years trying to dilute this material and bury it underground. Now Washington wants startups to turn it into reactor fuel. Oklo says fuel supply is holding back advanced reactors, while data centers keep making the power problem harder to ignore. Critics warn civilian plutonium programs make weapons usable material harder to control. Raise your hand if you think giving startups access to weapons grade plutonium is a good idea. (The Hill)
👽 Welcome to the future
ChatGPT and Claude can now be used to place trades. Liquid launched Co-Invest, which lets users research markets and place trades without leaving the chat. It supports more than 500 markets, including crypto, stocks, foreign exchange, prediction markets, and pre-IPO shares. Liquid handles the accounts and execution in the background, while ChatGPT and Claude become trading terminals. That puts the buy button inside the same chat where people ask what to buy, which sounds perfect for retail investors with impulse control issues. Liquid calls this “intelligence-augmented capital allocation,” because “gambling with a chatbot” tested poorly. Retail traders used to need Robinhood to lose money this fast. (Finance Magnates)
Who needs robots when you can get paid to do your own chores? Robotics companies need first person videos of people washing dishes, tying shoes, mixing drinks, and folding laundry. So why not cash in? Apps like Kled, Luel, Waffle Video, and DoorDash Tasks are turning chores into cold hard cash, with some missions paying up to $25 per hour. People strap phones to their heads, record the chores they were already stuck doing, and hope the footage gets approved. Sure, you might be training robots to replace you one day, but the laundry isn’t going to fold itself. Who’s the robot now? (Wired)
fun stats
🙄 99%. CEOs who are prepping to lay off workers and replace them with AI within 2 years.
💰 35%. How much more revenue Anthropic may be generating vs OpenAI. Anthropic is on pace for about $45 billion a year, compared with OpenAI’s estimated $33 billion. If that holds, Anthropic could pass Netflix, SAP, and Salesforce within a year.
🚀 1 terrawatt. Extra AI compute Elon Musk says the world will soon need every year. Today’s data centers use about 40 gigawatts. SpaceX thinks space solar could generate 5x more power than Earth based projects.
🌪️ 56%. Share of the 670 planned US data center projects being built in areas highly exposed to severe weather and natural disasters. Great timing.