AI is Now Building AI

Dario Amodei, CEO of Anthropic, admitted that AI now writes "much of the code" at his own company. He's predicting AI systems that can do expert-level intellectual work across domains could be real by 2027.

Factory AI released a coding agent that analyzes its own performance and updates its own codebase daily. Meaning every day it looks at what it did, figures out how to do it better, and rewrites itself. No human involved.

Anthropic launched MCP Apps, which lets AI tools build and display working applications directly inside a chat conversation. So instead of just getting a text response, you could ask for something and the AI builds you a working tool right there in the chat.

Andrej Karpathy (one of the original minds behind Tesla's Autopilot and OpenAI) pointed out that AI "stamina" is a "feel the AGI" moment. These systems can grind on hard problems for hours without losing focus or getting frustrated. They just keep working until it's solved.

Sam Altman is promising a model 100x more capable, faster, and cheaper than today's. OpenAI's goal is to compress 25 years of scientific progress into 5, and they're already processing 8.4 million weekly messages on advanced math and physics.

The competition isn't slowing down. Alibaba's Qwen3-Max-Thinking now rivals GPT-5.2 and Opus 4.5 across 19 benchmarks. Moonshot AI's Kimi K2.5 claimed global state-of-the-art on agentic benchmarks, which measure how well AI can operate on its own without hand-holding. And Grok 4.20 is the only model actually making money on PredictionArena, a platform where AI systems place bets on real-world outcomes and get scored on accuracy.

Hobbyists are using GPT-5.2 to attempt solving all 675 open Erdos problems. These are famous unsolved math problems named after mathematician Paul Erdos that have stumped the field for decades. People are now treating these like overnight computing tasks - queue it up, let it run, check in the morning.

NVIDIA launched Earth-2, a fully open suite of AI weather models. Climate modeling used to require supercomputers running for weeks. Now researchers can test different scenarios and iterate quickly, almost like running simulations in a video game.

All this AI needs power and cooling. Karman Industries adapted SpaceX rocket engine tech to cool data centers with liquid CO2, cutting space requirements by 80%.

Saudi Arabia is pivoting its $500B Neom megaproject to become a data center hub. They're betting that computing power is going to be as valuable as oil was for the last century.

Microsoft won approval for 15 more data centers in Wisconsin and unveiled Maia 200, an inference accelerator they claim is 3x faster than Amazon's Trainium. Inference is the process of actually running AI models to get outputs - so faster inference means faster responses and lower costs.

NVIDIA and CoreWeave are investing $2 billion to deploy 5 gigawatts of AI compute capacity by 2030. To put that in perspective, 5 gigawatts is roughly enough electricity to power 3-4 million homes. That's how much energy is going into AI infrastructure alone.

Bill Gates-backed Neurophos claims their optical chip can deliver 470 petaFLOPS. A petaFLOP is a measure of computing speed - one quadrillion calculations per second. For context, NVIDIA's upcoming Rubin chip is expected to hit around 47 petaFLOPS. So Neurophos is claiming 10x that, and they're doing it with light instead of traditional electronics.

Demis Hassabis (CEO of Google DeepMind) predicts they're 18 months from "solving" humanoid robotics. That means robots that can navigate the real world and handle physical tasks the way humans do - not just in controlled environments, but in messy, unpredictable spaces.

On the longevity front, British researchers found that immune T-cells release something called "telomere rivers," and transplanting them extended mouse lifespans by 17 months. Telomeres are the protective caps on your chromosomes that shorten as you age. If this translates to humans, we're looking at aging as something you can potentially slow down or reverse.

The U.S. Department of Transportation plans to use Google's Gemini to draft new regulations in 30 days. These are processes that normally involve years of back-and-forth between agencies, public comment periods, and legal review. Now they're looking to collapse that into a month with AI assistance.

Washington State is moving to require 3D printers to detect and block gun manufacturing attempts. The surveillance and control infrastructure is scaling alongside everything else.

Clawdbot has been everywhere the last few days. It's an autonomous AI agent that people have been experimenting with - basically letting it run continuously, make decisions, and operate across different apps and services. It had to rebrand due to trademark issues and landed on a lobster mascot, which is funny because the sci-fi novel Accelerando (written back in 2005) opens with an AI lobster character. Reality is catching up to fiction.

What's driving the hype is that people are treating Clawdbot like it's a real entity. They're setting up dedicated Apple IDs, phone numbers, and Mac Minis just for the bot - essentially giving it its own digital identity and resources. Some are calling it the first digital employee. It's a glimpse at what happens when AI agents start operating persistently rather than just responding to one-off prompts.

Also worth noting: court records revealed Anthropic spent millions destructively scanning books - physically destroying them in the process - to create training data. The race for data is getting aggressive, and companies are doing whatever it takes to feed these models.