Anthropic told the Pentagon to go pound sand. OpenAI picked up the phone before it stopped ringing. Alibaba dropped a model that makes last year’s frontier look quaint. Apple decided your laptop should run 70 billion parameters. And Andrej Karpathy says programming is “becoming unrecognizable,” which is a polite way of saying most of us are project managers for robots now.
March 2026 is wild. Here’s what actually matters.
1. Anthropic said no. OpenAI said how much.
This is the story. Everything else is noise by comparison.
The Pentagon offered Anthropic $200 million to let Claude AI analyze bulk commercial data on American citizens. Search histories. GPS movements. Credit card transactions. Anthropic’s response, paraphrased: “Absolutely not.”
The deal collapsed on a Friday afternoon. By Monday morning, OpenAI had its own Pentagon contract for classified AI systems. President Trump ordered all federal agencies to stop using Anthropic within six months and slapped the company with a “supply chain risk” designation, something that has literally never been done to an American tech company before.
Then the internet did what the internet does:
- #CancelChatGPT went viral. ChatGPT uninstalls surged 295%.
- Claude hit #1 on the App Store in the US, Canada, and Germany.
- Anthropic logged 503,424 downloads in a single day. Their all-time record. By a lot.
- Katy Perry (85 million followers) publicly switched to Claude. Yes, that Katy Perry.
- Nearly 100 OpenAI employees signed a protest letter against their own company’s deal.
The backlash blew up everywhere. Dario Amodei called OpenAI’s public messaging “straight up lies,” and the AI community largely sided with him.
Meanwhile, Donald Knuth (yes, the Donald Knuth) published Claude’s Cycles, a paper analyzing Claude’s mathematical reasoning. It became the most discussed AI paper of the week. The man is 88 years old and still putting out bangers.
Here’s why this is bigger than a PR crisis: it’s the first time AI ethics and safety have triggered a mass consumer migration. Not a Twitter thread. Not a petition. People actually uninstalling one product and paying for another because of values. That has never happened in this industry before.
2. Programming is dead. Long live… whatever this is.
Remember when Andrej Karpathy coined “vibe coding” a year ago? The idea that you just prompt an LLM, accept all suggestions, don’t read the diffs, and let the AI do the actual work? It was a joke. A fun observation about weekend projects.
It’s not a joke anymore. It’s the dominant paradigm.
Google Trends tells the story: search volume for Lovable is up 173,000%. Claude Code up 138,000%. MCP (the protocol that connects AI to tools) up 74,000%. Google AI Studio up 80,000%. These aren’t niche curiosities. This is a tectonic shift in how software gets made.
Karpathy himself has already moved on. His new term is “agentic engineering”: AI agents writing code independently, with humans providing direction but not keystrokes. His quote: programming is “becoming unrecognizable.” He’s not wrong.
Simon Willison’s Agentic Engineering Patterns is quickly becoming the textbook for this new discipline. The key insight: “writing code is cheap now.” The expensive part is knowing what to build and how to verify it works.
The numbers back this up. OpenAI says Codex hit 1 million weekly active users, with token usage up 5x since GPT-5.3 Codex launched. Their plan? Make Codex the foundation for enterprise AI agents. Not just coding. Everything.
The AI coding tools war is heating up: Cursor, Windsurf, Claude Code, Lovable, Replit, and a newcomer called Kiro (up 31,850% on Google Trends, seemingly from nowhere) are all fighting for developer mindshare. It’s a gold rush, and nobody knows which pickaxe will win.
3. Open-source is five points away from catching up. Five.
A benchmark study tested 94 LLM endpoints and concluded: “Open-source AI models are now within 5 quality points of proprietary.” Let that sink in.
Qwen 3.5 from Alibaba just dropped and it’s absurd. A full model lineup from 0.6B to 480B parameters. Mixture of Experts architecture. 1 million token context window. Native MCP and tool use support. The small models are generating enormous excitement in the local LLM community, with people calling the generational improvement “incredible.”
Oh, and PewDiePie fine-tuned Qwen2.5-Coder-32B to beat GPT-4o on coding benchmarks. We live in the strangest timeline.
DeepSeek V4 should drop any day now, reportedly with image and video generation. In the meantime, DeepSeek partnered with Tsinghua and Peking Universities to release DualPath, an inference system built specifically for complex agent scenarios. It hits KV-Cache rates above 95%, which is the kind of infrastructure work that doesn’t make headlines but makes everything else possible.
The geopolitical undercurrent is impossible to ignore. The debate over American closed models vs. Chinese open models is tearing the community apart. The departure of senior Qwen developer Junyang Lin has people wondering what it means for the project’s future. The AI race is not just company vs. company anymore. It’s ideology vs. ideology: open vs. closed, and it maps uncomfortably onto US vs. China.
4. Your next laptop can run a 70B model. Casually.
Apple’s M5 Pro and M5 Max landed this week. The headline for the AI crowd: 4x faster prompt processing than M4, and the M5 Max packs 128GB of unified memory. That’s enough to run a quantized Llama 3.3 70B locally. On a laptop. Without your fans sounding like a jet engine.
Some real numbers:
- 40 neural accelerators across every GPU core on the M5 Max
- ~3,000 to 3,500 tokens/second prefill on 7B Q4 models
- Memory bandwidth of 614 GB/s on the Max (double the Pro)
Someone on the LocalLLaMA community reverse-engineered Apple’s Neural Engine to train MicroGPT. Apple’s own MLX research team published a paper on leveraging the M5 neural accelerators for LLMs the same week. When a hobbyist and the company’s research lab are publishing about the same chip simultaneously, something interesting is happening.
For anyone who cares about privacy, latency, or not sending every keystroke to someone else’s server: local AI inference just went from “technically possible if you squint” to “genuinely practical.” The hobbyist era is over. This is professional tooling now.
5. We built 104,000 agents and forgot to lock the door.
The Universal Agent Registry now indexes over 104,500 AI agents across 15 registries. That’s not a typo. Over a hundred thousand AI agents are registered and (theoretically) usable.
GitHub Trending this week is basically an agent showcase: RuView builds knowledge graphs from repos with a Graph RAG agent. Airi is an open-source SuperAgent that can research, code, and create. Agent-Skills-for-Context-Engineering is a nano Claude Code clone built entirely in Bash, because of course it is.
Here’s the problem: nobody seems nearly as excited about AI agent security as building them.
- The OpenClaw Meltdown uncovered 9 CVEs and 2,200 malicious skills in the most comprehensive real-world test of the OWASP Agentic Top 10. That’s not a drill. That’s a fire.
- Researchers showed that invisible Unicode characters hidden in text can trick AI agents into following secret instructions. Tested across 5 models, 8,000+ cases. It works.
- Leo de Moura, the creator of Lean, published “When AI writes the software, who verifies it?” The question is catching fire, and nobody has a good answer yet.
The gap between what agents can do and what we can verify they are doing is widening every week. This will be the defining tension of AI in 2026, and we are nowhere close to solving it.
6. OpenAI fixed the one thing everyone complained about.
GPT-5.2 Instant had a reputation for being, to put it diplomatically, insufferably preachy. Ask it a factual question and it would add three paragraphs of caveats, assume you were in distress, and gently remind you to consult a professional. People hated it.
GPT-5.3 Instant shipped this week as the fix. The improvements: 26.8% fewer hallucinations with web search, direct answers without the hand-holding, and faster responses overall.
The tradeoff? It scored worse on OpenAI’s own safety benchmarks, particularly around content filtering. Which tells you something about where the industry is right now: the “safety vs. usability” pendulum is swinging hard toward usability, and the companies are following the users, not the other way around.
7. MCP: the most important thing nobody talks about.
The Model Context Protocol is not sexy. It does not have a viral moment. Nobody is making TikToks about it. But it is quietly becoming the plumbing that makes the entire AI agent ecosystem work.
Think of it as USB-C for AI. One protocol that lets any model talk to any tool, any database, any API. Before MCP, every agent integration was custom plumbing. Now there’s a standard.
The signals are stacking up:
- Qwen 3.5 shipped with native MCP support out of the box
- MCP is a rising search term in Google Trends, showing up alongside vibe coding queries
- Every serious AI coding tool (Cursor, Claude Code, Windsurf) relies on MCP for tool connectivity
If you’re building anything that touches AI, understanding MCP is not optional anymore. It’s not glamorous, but neither is TCP/IP, and that one turned out to be kind of important.
What’s next
DeepSeek V4 should land any day. If it delivers on multimodal generation, it could be the most capable open-source model ever released. Anthropic’s legal challenge against the “supply chain risk” label will set precedent for every AI company’s relationship with government. The EU AI Act is getting teeth, and the M5 benchmarks from the community will tell us whether Apple’s claims hold up outside a keynote slide.
AI in 2026 is no longer a technology story. It’s a power story. Who builds the models, who gets to use them, who profits, and who gets surveilled. March is proving that these aren’t abstract questions for panel discussions. They’re the kind of questions that make people uninstall apps and switch sides overnight.