Every week brings ten new AI announcements. Most don't matter. Here are the four from this week that actually do, and why.
1. OpenAI launches its agent platform
OpenAI rolled out a public agent platform that lets developers chain GPT calls with browsing, code execution, and third-party tools without writing orchestration code.
Why it matters: this is OpenAI's clearest move into the territory currently owned by LangChain and CrewAI. If it works as advertised, expect a wave of consumer apps that "do things for you" by summer.
2. Gemini 3 details leak ahead of Google I/O
A Google blog post went up briefly and confirmed Gemini 3 will ship with a 2M-token context window and native video understanding at 60fps. The post was pulled within 20 minutes.
Why it matters: real-time video reasoning is the missing piece for genuinely useful AR glasses and accessibility tools. If Google ships this in May, it leapfrogs current multimodal leaders.
"The model that understands video in real time wins the next platform shift."
3. EU AI Act enforcement begins
The first round of AI Act fines started this week. Two mid-sized European startups were warned over insufficient transparency in high-risk hiring tools.
Why it matters: this is the first real enforcement signal. Expect every B2B AI vendor to scramble for compliance documentation in the next 90 days.
4. The open-source release nobody noticed
A small Chinese lab released "Yi-Lightning 2," a 70B model that matches GPT-4-class performance and runs on a single consumer GPU after quantisation.
Why it matters: open models are closing the gap faster than ever. By year-end, "good enough" AI will run locally on your laptop — with massive implications for privacy, cost, and what businesses can build.
What to actually do this week
- If you build software: try OpenAI's agent platform on a real workflow.
- If you run a business in Europe: ask your AI vendors for an AI Act compliance statement.
- If you care about privacy: download Yi-Lightning 2 or wait for Ollama to add it.
The pace is exhausting, but the pattern is clear: AI is moving from "interesting" to "infrastructure." Treat it accordingly.