Daily Briefing — April 3, 2026

Google just dropped Gemma 4, a new family of open weight AI models that you can actually run on your own hardware. There are four sizes in the lineup: two beefier ones aimed at developers with serious GPU setups, and two lighter versions built specifically for mobile devices. The bigger variants, 26B Mixture of Experts and 31B Dense, are designed to run on an 80GB H100 at full precision or on consumer GPUs when quantized down. The 2B and 4B models are going after on-device use cases on phones and tablets.

The more interesting move here might actually be the licensing switch. Google has ditched its custom Gemma license and moved to Apache 2.0, which is about as developer friendly as it gets in the open source world. This was a real complaint in the developer community, and Google listened. Apache 2.0 means fewer legal headaches for teams trying to build products around these models.

The technical detail worth paying attention to is how the 26B MoE model works. It only activates 3.8 billion parameters during inference, which means you get much faster response times without sacrificing the depth that a larger model brings. That is a genuinely useful design choice for anyone building latency sensitive applications. The 31B Dense is slower but more capable, and Google is pushing it as a base for fine tuning on specific tasks.

A new academic paper out of arXiv is doing what a lot of AI research quietly avoids: actually naming which jobs are at risk and where. The study looks at agentic AI, meaning systems that don't just respond to prompts but take sequences of actions autonomously, and maps its exposure across occupations and regions. That's a meaningful distinction from the usual "AI will change work" hand-waving, because agentic systems can complete multi-step workflows without a human in the loop at each stage.

The findings land differently depending on where you sit in the labour market. Some roles have high task exposure not because AI will replace the person entirely, but because the repetitive, rule-bound parts of the job can be handed off to an agent running 24 hours a day. That quietly changes what your value proposition needs to be.

What makes this research worth paying attention to is the multi-regional lens. The disruption is not uniform. A task that gets automated away in a high-wage knowledge economy might look very different in a market where that same task is still done by hand at scale. For anyone building a career right now, the geography of where agentic AI hits hardest is as important as the job category itself.

Cursor just dropped version 3, and the headline is not about a new feature. It is about a company publicly admitting that the product which made them famous is already becoming less relevant. Cursor built its reputation as the go-to IDE for AI assisted coding, and now they are pivoting hard toward agentic workflows where you hand off entire tasks to an AI agent, not just ask it to autocomplete a line of code. They are calling this shift "agent first" and the internal codename was Glass, which honestly sounds like something you name a project when you know it has to be transparent about a big strategic bet.

The competitive pressure here is real. OpenAI and Anthropic, who are literally Cursor's suppliers, turned around and built competing products. Claude Code and Codex have been pulling developers away with heavily subsidized pricing, which is a brutal position to be in. Imagine your landlord opening a restaurant next door and undercutting your prices with their own ingredients.

What Cursor 3 signals is that the baseline expectation for coding tools is shifting fast. "Help me write this function" is becoming table stakes. The new benchmark is "go build this whole thing and come back when it is done." That changes what it means to be productive as a developer, and it changes what skills actually matter on a team.

OpenAI just closed what might be the largest private funding round in history, pulling in $122 billion at a valuation of $852 billion. To put that in perspective, that makes it more valuable than most companies that are actually listed on a stock exchange. The investor list reads like someone invited every major power player in tech and finance to a very expensive dinner: Amazon dropped roughly $50 billion, Nvidia put in $30 billion, SoftBank another $30 billion, and that is before you count Andreessen Horowitz, T. Rowe Price, Microsoft, and Abu Dhabi's sovereign wealth fund.

Here is the part that should make you pause though. The headline number is staggering, but the reason this keeps making news is not the money coming in. It is the money going out. OpenAI burns through cash at a scale that makes most companies look fiscally conservative. The company reportedly lost around $5 billion last year even as revenue grew fast.

So what you are watching is a company that needs institutional capital the way a city needs infrastructure. The funding is not a sign of dominance. It is a sign of how expensive it is to stay at the frontier of AI development. That dynamic shapes everything downstream, from how aggressively OpenAI prices its products to how much pressure it puts on competitors to do the same.