the Microdose

A New Type of AI

+ 6x premium, agent swarms, and vibe science
Adam Wildheart
OpenAI Jerry Tworek terraform planets
OpenAI Jerry Tworek terraform planets

Jerry Tworek/The Microdose AI

Subscribe to The Microdose
Subscribe -- post
Cheri Wildheart
Adam Wildheart

Good morning. You might’ve seen Cursor bragging that their AI built a web browser from scratch in under a week. They even dropped this flashy video. But before you start handing pink slips to your dev team, there’s a little something you should know. It’s mostly hype. Despite burning through millions of dollars worth of AI tokens, the code is a “tangle of spaghetti” that won’t ever resemble a real world web engine. There’s no question agents can print 3m+ lines of code. Making it work is another story.

A former OpenAI VP wants $1 billion to reinvent AI. Jerry Tworek says today’s models are fundamentally broken. They learn by brute forcing mountains of data, then promptly forget it all when something new pops up. He’s pitching a new type of AI that will continuously learn, self upgrade, and use 100x less compute. And since he’s already dreaming big, Tworek’s plan includes self-replicating factories and terraforming planets. Core Automation, his startup, is the latest AI neolab raising piles of cash without a working product. Who needs customers when VCs will fund your sci-fi dreams? (The Information)

Brand name AI models cost 6x more than open models. 80% of all AI tasks run on expensive, closed models from companies like OpenAI, Google, and Anthropic. But open models from Moonshot, DeepSeek, and Mistral do pretty much the same job for a fraction of the cost. A new MIT study says switching to open models could save around $25 billion a year. So why aren’t companies jumping on this? They’re stuck thinking open models are less reliable. Sometimes it’s easier to burn piles of cash than train a new AI. (MIT)

Wispr Flow helps teams work 4x faster in every app. Speak prompts, dictate PRDs, generate code, and reply to messages using your voice. Flow precisely captures your ideas at the speed you think.

Used by teams at Clay, OpenAI, and Amazon to execute 20% faster. Because it works, magically.

Stop typing. Start talking to your apps. Try Wispr Flow for free.

Google Chrome’s latest feature: impulsive AI shopping. Powered by Gemini AI, Auto Browse takes over tedious tasks like clicking and booking flights for you. The bot asks permission before swiping your credit card, which seems reassuring except AI agents are consistently unreliable. If Google’s bot screws up, surprise! You’re responsible. It’s like handing your credit card to a distracted teen and hoping for the best. No worries about privacy, Google already knows everything about you. (Wired)

China’s new Kimi K2.5 outperforms ChatGPT 5.2 and Claude Opus 4.5. Moonshot AI’s latest open source model sets the standard for performance. Trained on 15 trillion tokens, it handles all the same tasks as your favorite proprietary AI, plus it can orchestrate swarms of up to 100 sub agents at once. Benchmarks show it crushes frontier models in agent tasks, multimodal, and deep search accuracy. It’s available on Hugging Face, but you’ll need serious hardware to run it locally. Time to start stacking Mac Minis. (Agent Native)

OpenAI dropped Prism, an AI workspace for scientists. Before GPT-5.2 researchers had to manually test hypotheses, grind through mathematical proofs, and cite real sources. Now anyone with a ChatGPT account can produce credible looking science. OpenAI says Prism will do for science what AI did for coding. Vibe science sounds like an incredible win for data manipulators everywhere. (OpenAI)

Prediction markets are cashing in every time Musk is wrong. Degens on Polymarket and Kalshi discovered an easy money making hack: betting against Elon’s predictions. AI smarter than people by now? Nope. Robotaxis driving themselves everywhere? Still waiting. Starship rockets gently touching down on Mars? They’re still exploding. One user made $36k betting Musk would keep missing his targets. Even Musk himself admitted betting against Tesla’s robotaxi timelines was an easy profit. At least he’s consistent. (Futurism)

Anthropic’s CEO says AI could win Nobel prizes and take over the world. Elon predicts AI will make money obsolete and create infinite prosperity. Google is busy using AI to crack quantum computing and revolutionize drug discovery. And OpenAI is… launching ads. Strange days indeed.

fun stats

🏃‍♂️ $20 billion. Anthropic has doubled its funding target as Claude wins enterprise buyers.

☝️ $30 billion. SoftBank is in talks to invest even more money into OpenAI, as the startup seeks a $100 billion war chest.

🚘 0. Tesla will stop making the Model S and Model X to pivot to Optimus robots. 73% of Tesla’s revenue last year was from cars.

subscribe today

Get an edge with The Microdose

Skip the prompts, get the signal. We’ll send you short daily updates about the real AI + future tech you need to know. Fast, smarter, mildly addictive – and free. 

Subscribe -- page