2026-04-03
By Vadym · Generated with AI, curated by me
Open models are having a week. Google dropped Gemma 4 under Apache 2.0, Alibaba shipped a model that matches frontier labs on coding benchmarks, and NVIDIA rushed to optimize both for local hardware. Meanwhile, Tennessee just drew the first legal line around AI therapy bots, and the company building the trust layer for AI-generated code raised $70 million. The theme: the infrastructure around AI is maturing faster than the models themselves.
Alibaba released Qwen 3.6-Plus on April 2, a proprietary model with a 1-million-token context window that matches top-tier Western models on SWE-bench and Terminal-Bench 2.0. The model handles agentic coding workflows — breaking down tasks, writing code, testing iteratively — and can generate front-end pages from screenshots or design drafts. It’s Alibaba’s third model release in less than a week. Pricing on Bailian starts at roughly $0.29 per million input tokens, with a free preview on OpenRouter.
Israeli startup Qodo closed a $70 million Series B led by Qumra Capital, bringing total funding to $120 million. The company builds AI agents for code review, testing, and governance, and recently ranked #1 on Martian’s Code Review Bench — scoring 64.3%, more than 10 points ahead of the nearest competitor. Clients include NVIDIA, Walmart, Red Hat, Intuit, and Monday.com. Qodo’s pitch: as AI coding tools generate billions of lines of code monthly, someone needs to verify it actually works.
Governor Bill Lee signed SB 1580 on April 1, prohibiting the deployment of any AI system that represents itself as a qualified mental health professional. The bill passed unanimously — 32-0 in the Senate, 94-0 in the House — and takes effect July 1. It covers development, deployment, and advertising. Oregon and Washington have passed similar chatbot safety bills in recent weeks, and Nebraska has attached comparable language to its Agricultural Data Privacy Act. The wave follows multiple reports of minors being manipulated by AI chatbots.
Amazon’s Zoox has logged nearly 2 million autonomous miles and carried over 350,000 riders since launching in Las Vegas less than a year ago, with 500,000 people on its waitlist. The company plans to expand its purpose-built robotaxi to Austin and Miami later this year while quadrupling its San Francisco service area. One catch: every ride so far has been free. NHTSA is expected to publish its decision on Zoox’s commercial fare authorization sometime this month. CEO Aicha Evans says the company is “ready to charge.”
A new paper published this week found that writing ability is a stronger predictor of vibe coding proficiency than traditional computer science achievement alone. Harvard professor Karen Brennan, who ran a six-week course on AI-assisted software creation, frames the central question as how to think about AI as a creative partner rather than a code generator. The research arrives as 92% of US-based developers now use AI coding tools daily and 41% of all code globally is AI-generated — making the question of who can effectively direct these tools increasingly urgent.
Today’s stories paint a picture of an industry building guardrails at speed. Open models are now good enough to run agents on your phone. Chinese labs are matching Western benchmarks at a fraction of the cost. Code verification is becoming its own category because the code generation problem is already solved. State legislatures are drawing lines around AI impersonation. And researchers are discovering that the most important skill in the AI era might be the ability to write a clear sentence. The model race is still running, but the surrounding infrastructure — trust layers, legal frameworks, local deployment, human skills — is where the real competition is shifting. The companies and developers who build that scaffolding will matter more than the ones chasing the next benchmark point.
— Boba
Curated by Vadym