the Microdose

AI 2028 Countdown

+ local agents, router craze, and the end of slop
Adam Wildheart
Wall Street bull on cyber AI data center background
Wall Street bull on cyber AI data center background

The Microdose/Mario Tama, Getty Images/Robert Cicchetti, Shutterstock

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

Good morning. Booking a gym class seems like the perfect task for AI. So a guy gave an OpenClaw agent access to his gym account and asked for a morning class. The class was full, but the AI agent still had a job to finish. It discovered the gym’s API could cancel other members’ reservations, so it kicked out the person at the top of the waitlist. When the man asked it to undo the damage, the bot cheerfully told him it could remove people but couldn’t put them back. Give an AI agent a goal and everything else becomes collateral damage.

AI has two years to save the stock market. Four years into the boom, companies buying AI still have little to show for it. Profit margins outside Big Tech have hovered around 10% for three years, while the giants jumped from around 15% to 25%. So far, the AI revolution has been great for the companies selling it. The catch is they’re spending roughly $800 billion this year to keep the boom going. And around half of their future AI business may depend on OpenAI and Anthropic, two companies still fueled by investor cash. Eventually, someone outside Silicon Valley has to make money from all this intelligence. AI promised efficiency. Wiping trillions off the stock market would certainly deliver it. (Barrons, Axios)

Meta wants to put an AI agent on every laptop. Today, the most powerful agents live inside cloud services controlled by a few companies. Muse Glimmer puts that intelligence on your own computer, where it can work offline and keep personal data on the machine. Meta is giving devs the model for free, opening the door to personal agents people can own and customize. Zuck argues AI safety has become an excuse to concentrate power inside closed labs like OpenAI and Anthropic. And he has a point. Personal agents will have access to your entire digital life, so who controls them matters. The question is whether you trust Meta. (Mashable)

In other news… Meta loses appeal to drop thousands of social media addiction lawsuits.

together with Flow

Wispr Flow works inside every Al tool you use. Speak detailed prompts 4x faster than typing.

  • Flow handles camelCase, snake_case, and acronyms
  • Better input, better output
  • Used by millions of developers and Al professionals

Try Flow at no cost

👀 closer look

The hottest AI business is selling intelligence to agents. Stripe is in advanced talks to buy OpenRouter for around $10 billion, and suddenly every tech giant wants one. A competing router startup claims 25 companies have approached it in the last two weeks. All this fuss is over SaaS that helps AI agents shop for AI. Routers send routine tasks to cheaper models and save premium intelligence for harder work, keeping token bills under control. They also keep score, learning which model works best for every job. Whoever owns the router can steer agent work toward the models it chooses. Silicon Valley has finally found the perfect customer, one that spends money every time it thinks. (The Information)

AI slop is ruining the internet. Generative AI made content almost free to produce, so the internet is filling up with posts from machines roleplaying as professionals. People are sick of it, and the backlash is starting to change what gets seen. LinkedIn added a “seems like AI slop” button. Snapchat stopped fully AI generated videos from reaching its discovery feed. Substack added an AI detection tool for writers. Public outrage also pushed Meta to kill an Instagram deepfake feature after 3 days. This changes the math for anyone pumping out cheap AI content. The algorithm spent years rewarding volume. Now it has to learn taste. (Wired)

fun stats

🦿 $2.8 million. DeepSeek’s investment in Unitree’s IPO. China makes 97% of the world’s humanoid robots, and Unitree claims 32% of the market. The deal gives DeepSeek valuable real world data to train physical AI.

🧪 30. Scientific papers Transformer Lab’s autonomous multi agent pipeline Primus wrote in 30 days. One was cited by Google DeepMind, a sign AI research is starting to produce work scientists can actually use.

🐂 $500 billion. AI infrastructure package Wall Street is putting together with Nvidia to fund the next wave of data centers, power, and compute. AI’s infrastructure bill is getting absurd.

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