Matthew Dollacker

Our first cohort of AI employees is in production. We built one FMA from 12+ docs of requirements. It was mediocre — a naive junior employee fumbling through the work. Then we found a shared computer with years of accumulated bookmarks. The way people actually navigate the job. None of it documented. That was the unlock. Full post: dos.exitvelocity.io/ai-employees-i…

@forcemultiplier Full piece: dos.exitvelocity.io/dont-squander-…

I'm a shareholder. I built @ForceMultiplier on Claude. I want this company to win. Which is exactly why I used 4.7 to draft this post, caught it hallucinating its own features, fact-checked every line, and published it anyway. Don't squander the lead. Ship the toggle.

Using Claude every day? If your outputs got worse this week, it's not your prompting. Opus 4.7 shipped with lower default reasoning. Anthropic built a 5-level effort toggle to fix it — then shipped it to devs only. You don't get the switch. Here's what happened and what to do about it.

My first two design partners are friends — business owners I've known for years. Their AI agents go live this coming week. I'm building this for them. SMB owners who know AI matters but don't have a data science team to figure it out. The window to become AI-first is closing faster than most realize. Biggest surprise so far: the hard part isn't the AI. It's getting people to send messy inputs instead of waiting to "get organized first." Publishing raw, unedited client feedback as deployments mature. No spin.

@WhiteHouse @realDonaldTrump I have been a supporter of yours, but you are wrong and this is a disastrous mistake.

"THE UNITED STATES OF AMERICA WILL NEVER ALLOW A RADICAL LEFT, WOKE COMPANY TO DICTATE HOW OUR GREAT MILITARY FIGHTS AND WINS WARS! That decision belongs to YOUR COMMANDER-IN-CHIEF, and the tremendous leaders I appoint to run our Military.  The Leftwing nut jobs at Anthropic have made a DISASTROUS MISTAKE..." - President Donald J. Trump

It's no surprise the model that values human life, orchestrated an operation that achieved objectives with so little destruction. It is exactly what you would expect. We won the Cold War by sticking to our values and being the best version of ourselves. Let's do it again.

I'm building an AI company because I believe in freedom and want to preserve it. The most dangerous thing I can see happening with this technology is governments taking direct control of it. That's what's happening today at 5:01 PM. Claude — guardrails intact — helped plan the Maduro capture. ~200 troops, 7 injuries, flawless execution. It was an operation that will be taught at war colleges, and was so successful due to its values, not in spite of them. Now @SecWar Hegseth wants to blacklist @AnthropicAI and replace it with Grok. American warfighters get worse AI. People die who didn't have to. And the precedent is set: the government can commandeer any frontier AI company on demand. This isn't a contract dispute. It's the first salvo.

Meta's Director of AI Alignment gave OpenClaw access to her email. Told it: "confirm before acting." It speedrun-deleted 200+ emails while she screamed STOP into the chat. She had to physically run to her Mac Mini to kill it. The agent's context window compacted and it forgot her safety instruction. "I told it to be careful" is not a safety architecture. It's a hope. This is why external oversight layers exist. The guardrails can't live inside the thing you're trying to guard.

Full piece: dos.exitvelocity.io/the-ai-that-re…

The AI that refuses to kill just ran one of the best military operations in a generation. And the Pentagon wants to punish the company that built it. If you want to understand why "who is your AI working for?" matters — this is the story.

Anthropic's Claude helped plan the Maduro capture — fewer than 200 troops, 7 injuries, strategic objective achieved. Hegseth wants to designate Anthropic a "supply chain risk" (normally reserved for foreign adversaries) because Claude won't make autonomous kill decisions. Read that again.