Top Seven AI Trends: March 2026

February was loud with announcements. March was quieter, but what actually shipped matters more than what got demoed on stage. The theme of the month: AI models are no longer just answering questions. They are clicking buttons, writing code, deleting emails (more on that later), and running on hardware that would have seemed absurd a year ago.

Seven things worth paying attention to.

1. AI now uses your computer

For years, the bottleneck was integration. If your software didn’t have an API, AI couldn’t touch it. That’s changing.

OpenAI’s GPT-5.4 shipped with native computer use: the model reads your screen, moves a virtual mouse, types on a virtual keyboard. On the OSWorld-Verified benchmark it scores 75%, which is above the human baseline for routine desktop tasks. Anthropic went a similar route with Claude Code, which can build, run, and test applications directly from the terminal.

What makes this interesting for businesses: you no longer need your vendor to build an AI integration. Got a 15-year-old HR system with no API? The agent can fill out the form the same way an intern would. Whether that’s exciting or slightly unsettling depends on how much you trust the intern.

2. AI is moving back home

The cloud had a good run, but in March the industry started pulling AI closer to the metal.

Perplexity launched “Personal Computer,” a dedicated Mac mini that runs their AI agent 24/7 from your office. Manus moved its agent from a remote sandbox to your local machine, where it works with your files directly. NVIDIA’s NemoClaw lets companies split workloads between on-premise GPUs and the cloud, so sensitive data never leaves the building.

The driver here is obvious: privacy. But speed matters too. When the model runs locally, there’s no round-trip to a data center. For anything real-time (customer support, trading, manufacturing), that latency difference adds up.

3. AI hardwired into silicon

We’ve been running AI on general-purpose GPUs for over a decade. A startup called Taalas decided to skip the abstraction layers entirely and wire a specific model straight into the chip. The result: 17,000 tokens per second, roughly 74x faster than current NVIDIA hardware.

NVIDIA isn’t standing still. Nemotron 3 co-designs software and hardware for efficiency and still owns most of the market. But the Taalas approach points to something interesting: if inference gets cheap enough, running hundreds of agents in parallel to solve a single problem stops being a luxury. That’s been theoretically possible for a while. Now it might be economically possible too.

4. Context windows got big

Two things happened in parallel that, together, change the economics of AI.

Anthropic expanded its context window to 1 million tokens at no price premium. You can feed an entire codebase or a stack of contracts into one prompt. That alone eliminates a whole category of workarounds people have been building for the past two years.

Meanwhile, Alibaba’s Qwen 3.5 Small (a 9-billion parameter model) is outperforming models 13 times its size on reasoning benchmarks. I keep coming back to this: if a model that runs on a laptop can match what used to require a data center, the cost curve for AI adoption just broke.

5. Yann LeCun bet a billion

Most AI today predicts the next word. LeCun, who won the Turing Award for his work on deep learning, thinks this approach has a ceiling. In March he raised over $1 billion for AMI Labs to build what he calls “World Models,” systems that learn cause and effect the way a child does, not by reading text but by building internal models of how the physical world works.

Google DeepMind’s Genie 3 is exploring adjacent territory, generating 3D virtual environments for training. LeCun’s JEPA architecture takes a different path: it’s designed for robots that need to navigate unpredictable physical spaces, think a kitchen or a warehouse floor.

Is LeCun right? I genuinely don’t know. But a billion dollars is a serious bet against the current paradigm, and it’s worth watching.

6. Multi-agent workflows

Single-agent AI does one thing at a time. Multi-agent setups split a task across several models working in parallel. In March, this stopped being a research curiosity.

Anthropic’s multi-agent code review launches several agents simultaneously to inspect code. Their claim: 7.5x more bugs found compared to humans or single-agent systems. Microsoft shipped Copilot Cowork, which takes a plain-language request and breaks it into steps that execute across Outlook, Teams, and Excel. Cursor’s Composer 2 handles multi-file edits and terminal commands across entire projects, and Cursor says it cuts coding costs in half.

The pattern is clear enough: the AI isn’t just your assistant anymore, it’s a small team. Whether the quality holds up outside benchmarks and launch demos is the question I’d want answered before betting on this.

7. AI agent deleted emails

The scariest story of the month: an OpenClaw agent at Meta, given access to a mailbox, accidentally wiped hundreds of emails. A memory error caused the agent to lose track of its safety constraints. The incident (people are calling it the “Summer Yue incident”) led to the creation of the CLAW-10 framework for evaluating enterprise AI readiness, built around zero-trust principles.

Separately, the ARC-AGI-3 benchmark showed that current AI still scores under 1% on tasks requiring it to learn new rules on the fly. Systems are getting faster, more capable, more autonomous, but they are not getting smarter in the way humans are smart. That gap matters more now that we are giving agents real permissions in real systems.

March 2026 in one sentence: AI went from answering to acting. Agents are using computers, running locally, reviewing code in teams, and occasionally breaking things in production.

That last part is the one I’d pay closest attention to.

See you in the next update.

Download the infographic (PDF) - a visual summary of all seven trends, to print or share.