In partnership with
This week we explore:
Hark: the Figure AI founder's next swing at replacing your phone with something smarter
A soldier raids World of Warcraft using only his brain. No keyboard. No mouse. Just thoughts.
Andrej Karpathy's AI org chart went viral. Then he deleted it.
Anthropic is shipping faster than anyone expected. Claude now uses your Mac while you're away.
Cursor launched a "proprietary" model. It was secretly Chinese.
Elon announced a $25 billion chip factory to build chips for robots and space. Wild.
You can now rent human brain cells in the cloud. This is not satire.
Artificial Intelligence
Brett Adcock Thinks Your Phone Is Pre-AI
Brett Adcock, founder of Figure AI (humanoid robots) and Archer Aviation (eVTOL), just announced Hark. The pitch: today's AI models are too dumb, and the devices we use to access them predate AI entirely. So he's building both at once.
Hark will co-design its foundation models, hardware, and interface as a single vertically integrated product: persistent memory, real-time speech and vision, proactive intelligence that thinks ahead of you. Adcock put in $100M of his own money. Key hire: Abidur Chowdhury, lead designer of the iPhone Air, now running hardware design. Team is 45 engineers and designers from Meta AI, Apple, and Tesla. First model releases expected this summer. No device timeline yet.
The race: OpenAI is building AI hardware with Jony Ive. Meta ships AI glasses. Hark is the latest entrant. Whoever cracks the post-smartphone AI interface wins the next platform.
TechCrunch / Bloomberg
Anthropic's Best Month Ever
Three agentic features, four weeks. Here's the stack:
Claude Code Channels (March 20): Developers can now hook Claude Code directly into Telegram and Discord, message their agent from anywhere, and get pinged when a task finishes. This is Anthropic's direct answer to OpenClaw, the viral open-source agent that blew up after letting users text an AI worker through WhatsApp and iMessage. OpenClaw's creator, Peter Steinberger, was poached by OpenAI. Channels is polished, permission-gated, and built on MCP. Currently Telegram and Discord only (OpenClaw covers nearly every platform).
Dispatch (March 14): Assign Claude tasks from your iPhone. Come back to finished work on your desktop.
Computer Use (March 23, research preview): Claude Code and Claude Cowork can now click, scroll, open apps, fill spreadsheets, and run dev tools on your Mac autonomously. Connectors for Slack, Google Calendar, and other supported apps come first. If no connector exists, Claude falls back to controlling the screen directly. Mac-only for now. Pro and Max subscribers.
The stat that puts it in context: Claude Code creator Boris Cherny told Lenny Rachitsky's podcast that Claude Code writes 80-90% of all code at Anthropic, productivity per engineer is up 200%, and Claude Code now accounts for 4% of all public GitHub commits, projected to hit 20% by end of 2026.
He Ranked Every Job in America by AI Risk. Then Deleted It.
Andrej Karpathy, who co-founded OpenAI and ran AI at Tesla, built an interactive chart scoring 342 American jobs on how likely they are to be replaced or reshaped by AI. Scale of 0 to 10.
Medical transcriptionists: 10. Lawyers, accountants, financial analysts: 9. Software developers: 8-9. Plumbers and roofers: 0-1. The surprising finding: the higher the salary, the higher the risk. Jobs that happen on screens and in documents are the most exposed. Jobs that require being physically present in the real world are the safest. About 60 million American workers sit in high-risk roles.
He took it down within a day, calling it a quick experiment being misread. It was instantly archived and forked everywhere. Elon Musk commented that AI might eventually make all jobs optional.
The other Karpathy story this month: he built a script that ran 50 scientific experiments overnight on its own, found a better result, and filed it automatically. No human involved. AI is now doing science, not just assisting with it.
Cursor Said It Built Its Own AI. It Hadn't.
Cursor is a $29 billion coding tool used by over a million developers daily. Last week it launched a new AI model called Composer 2, describing it as built in-house with "frontier-level" intelligence. The announcement didn't mention where the underlying model came from.
Within 24 hours, a developer dug into the code and found it: the model was Kimi K2.5, an open-source AI built by a Chinese company called Moonshot AI, backed by Alibaba. The license on Kimi K2.5 requires any company making over $20 million a month to display clear attribution. Cursor makes around $166 million a month. There was no attribution anywhere.
Cursor's co-founder acknowledged the omission and said the company contributed significant additional training on top of the base model. Moonshot publicly congratulated Cursor and confirmed a licensing partnership. The controversy quieted but didn't disappear. The lesson for everyone watching the AI industry: "built in-house" doesn't always mean what it sounds like.
Quick AI Hits
As AI tools get more powerful, they also become bigger targets for security attacks. (Hacker News)
Boris Cherny, the engineer who built Claude Code, said on Lenny Rachitsky's podcast this month: "Coding is largely solved." He predicts the job title "software engineer" starts fading by end of 2026, replaced by something closer to "builder." Worth hearing directly.
Here’s how I use Attio to run my day.
Attio is the AI CRM with conversational AI built directly into your workspace. Every morning, Ask Attio handles my prep:
Surfaces insights from calls and conversations across my entire CRM
Update records and create tasks without manual entry
Answers questions about deals, accounts, and customer signals that used to take hours to find
All in seconds. No searching, no switching tabs, no manual updates.
Ready to scale faster?
Biotech
He Has a Brain Chip. He's Playing WoW With His Mind.
Jon Noble is a British Army veteran with severe paralysis. This week marked 100 days since he received Neuralink's N1 brain implant. His update: he's now playing World of Warcraft, exploring full raids, at full speed, with no mouse and no keyboard. Just thought.
The first session was clunky. Once his brain and the device found their rhythm, he called it "pure magic" and "addictive." For someone who needed physical assistance for most daily tasks, this is the difference between dependence and independence.
Neuralink has 12 patients globally. The company also received approval from the FDA to fast-track development of a system that would translate thoughts directly into written words for people with ALS, stroke, or spinal cord injuries. That may be the bigger story long-term.
Biocomputing
A Melbourne Startup Is Renting Out Living Human Brain Cells. Online.
Cortical Labs, based in Australia, just opened a cloud service where you can run code on computers made of actual human neurons. Each unit, called a CL1, contains about 800,000 lab-grown brain cells sitting on a silicon chip in a data center in Melbourne. You log in, upload your code, and it runs on living biology.
The cells learn. They adapt through electrical signals, the same way a brain does. Earlier this year, Cortical Labs got them to play DOOM. A second facility is being built in Singapore. Weekly access starts at $300. The cells consume less power than a pocket calculator.
Why it matters: biological computing could eventually learn from tiny amounts of data while using almost no energy, two things traditional chips can't do. This is early and experimental. But the fact that it exists and is available for rent online is remarkable on its own.
Infrastructure
Elon Musk's Plan to Build the World's Largest Chip Factory
On March 21, Elon Musk announced Terafab: a joint venture between Tesla, SpaceX, and xAI to build the largest semiconductor factory ever conceived. It would be located at the Giga Texas campus in Austin. Cost: $20-25 billion. Goal: produce enough chips every year to power millions of autonomous vehicles, humanoid robots, and eventually, AI satellites in orbit.
Musk's argument: every chip manufacturer on Earth combined is producing only about 2% of what his companies will need. So he's building his own. The factory would make two types of chips: one for cars and robots, one hardened for the radiation of space.
The honest caveat: Musk gave no start date, no completion timeline, and the cost isn't yet in Tesla's budget. He has a history of big announcements on long timelines. But the underlying demand problem he's describing, not enough chips to run the AI future, is real regardless of whether Terafab delivers on schedule.
Our Vision
This week's stories share one thesis: the interface between humans and machines is being rebuilt from scratch.
Hark thinks the phone is pre-AI. Anthropic thinks the desktop is pre-agent. Cortical Labs thinks silicon is pre-biology. Musk thinks the supply chain is pre-Musk. Neuralink thinks the mouse is a workaround for people who haven't yet learned to use their brain directly.
The practical read for you: the skills that made white-collar work valuable are eroding faster than most institutions admit. Karpathy's deleted map made that concrete. The leverage now is not what you know. It is how fluently you direct systems that know more than you do.
The next computing platform is still open. AI glasses, AI pins, AI hardware from Hark or OpenAI's Ive project, something not yet announced. No one knows. Stay oriented, stay curious.
What do you think is coming next? Reply and let me know.
How did you like this week's edition?
Are you tracking agent views on your docs?
AI agents already outnumber human visitors to your docs — now you can track them.
Facts. Without Hyperbole. In One Daily Tech Briefing
Get the AI & tech news that actually matters and stay ahead of updates with one clear, five-minute newsletter.
Forward Future is read by builders, operators, and leaders from NVIDIA, Microsoft, and Salesforce who want signal over noise and context over headlines.
And you get it all for free, every day.