AI Insider Weekly

This Week in AI: The Stories You Need to Know

Weekly Roundup  |  March 30 – April 5, 2026

Another week, another avalanche. From Google making AI video dirt cheap to OpenAI buying a podcast, from an AI agent hacking an operating system in four hours to Mark Zuckerberg quietly writing code again — the pace shows zero sign of slowing. Here’s everything that matters this week, including some tools going viral that your peers are already using.

1. Google Makes AI Video Ridiculously Cheap with Veo 3.1 Lite

Google has been on a quiet but relentless video AI roll. First came Veo 3.1 for quality, then Veo 3.1 Flash for speed — and now Veo 3.1 Lite for price. At just $0.05 per second at 720p and $0.08 per second at 1080p via the Gemini API, you can generate a 6–8 second clip for as little as 30–50 cents. That’s less than half the cost of Veo 3.1, and early users are saying the quality drop is barely noticeable.

For creators, marketers, and indie developers, this changes the economics of AI video overnight. Google also released its official prompt structure for best results:

[Cinematography] + [Subject] + [Action] + [Context] + [Style & Ambiance]

Example: Medium shot, a tired corporate worker, rubbing his temples in exhaustion, in front of a bulky 1980s computer in a cluttered office late at night. Lit by harsh fluorescent lights and the green glow of a monochrome monitor. Retro aesthetic, slightly grainy.

2. Google Gemma 4 Is Here — Free, Open-Source, Runs on Your Laptop

Google dropped Gemma 4 under the Apache 2.0 licence this week — free, open-source, and designed to run entirely on your own device. The 31B version is currently ranked #3 among open models globally, reportedly outperforming models 20 times its size on several benchmarks. Combined with PrismML’s 1-bit Bonsai, H Company’s Holo3, and Arcee’s Trinity — all of which shipped this week under open licences — the open-source AI ecosystem is rapidly making cloud subscriptions optional for a growing range of tasks.

If you have Ollama installed, you can pull it right now: ollama pull gemma4. No API key. No monthly fee. Just your hardware.

3. Claude Goes Agentic on Your Desktop — While You’re Away

Anthropic’s Claude desktop app now supports Cowork and Dispatch — a setup that’s generating a lot of buzz. The workflow: open Cowork in the sidebar, scan a QR code with your phone, and you can text Claude tasks from anywhere. One user shared that they sent this prompt from their phone: “Go to X, find 5–10 high performing posts from the last 24 hours, write a full script for each, put everything in a PDF in my downloads.” They came back to 7 posts, 7 scripts, and 7 virality scorecards — laptop untouched.

Pair it with /schedule and it runs every morning before you sit down. This is the agentic future that NVIDIA’s Jensen Huang was talking about at GTC — now available in a practical, everyday workflow.

Note: Anthropic also confirmed this week that Claude subscriptions will no longer cover third-party tool usage (including OpenClaw) starting April 4, to manage capacity. You’ll need a separate API key for those integrations going forward.

4. OpenAI Hits $852 Billion Valuation and Buys a Podcast

OpenAI confirmed a massive new funding round this week that values the company at $852 billion — up from $157 billion just 18 months ago. Alongside the capital raise, the company unveiled its ChatGPT Super App strategy: a unified platform combining chat, coding, search, and agent capabilities into a single interface for 900 million weekly users.

In a move that surprised many, OpenAI also acquired TBPN — a popular daily tech news podcast hosted by John Coogan and Jordi Hays that has featured guests including Mark Zuckerberg and Satya Nadella. The acquisition signals OpenAI’s growing ambitions in the creator economy and media distribution.

5. An AI Agent Hacked FreeBSD in Four Hours

In what researchers are calling a landmark moment for AI security research, an autonomous AI agent successfully exploited a vulnerability in FreeBSD — in just four hours. The agent navigated the entire attack chain independently: reconnaissance, vulnerability identification, exploit development, and execution. No human guidance mid-task.

This follows a warning from Ledger’s CTO that AI is fundamentally breaking the economics of cybersecurity — tasks that once took skilled researchers months can now be done in minutes with the right prompts. The implications for enterprise security teams are significant and immediate.

6. Meta’s Brain-Prediction AI: TRIBE v2

Meta released something genuinely unusual this week: TRIBE v2, an AI trained on 500+ hours of fMRI scans from over 700 people that can predict how a human brain responds to audio and visual stimuli. Feed it a video clip, and it tells you which regions of the brain would activate. It’s not generating content — it’s modelling human experience.

Meta has released a public demo at go.meta.me/tribe2. The research has obvious applications in neuromarketing, media, accessibility, and clinical settings — and raises equally obvious questions about privacy and consent.

7. AI for Video Storytelling: pai by Utopi Studios

One tool making rounds in creator communities this week is pai by Utopi Studios (pi.uti.film), which takes a different approach to AI video generation. Rather than asking you to generate individual clips, pai asks for a story. It then extracts characters, generates key frames, asks for your approval, and assembles a full multi-scene sequence — consistently. There’s even an auto-approval mode for the entire pipeline. Access is limited and waitlisted right now, but worth keeping an eye on if you work with video content.

8. Gartner: Enterprises Will Ditch AI Copilots by 2028

Gartner published a bold prediction this week: by 2028, over half of all enterprises will stop paying for assistive AI tools — copilots, smart advisors, and the like — and instead favour platforms that deliver actual workflow outcomes autonomously. In their framing, humans shift from completing work to supervising intelligent systems that execute on their behalf.

The firm also predicted that by 2030, legacy SaaS vendors that layer AI onto old software without redesigning for agentic execution could face margin compression of up to 80%. If accurate, this is the most consequential enterprise software prediction in a decade.

9. Yahoo Is Back — Built on Anthropic

Yahoo launched Scout this week, an AI-powered answer engine built on Anthropic’s technology. Rather than a conversational chatbot, Scout delivers direct answers with supporting links — a design choice aimed at activating Yahoo’s existing audience across news, finance, and email. It’s a long shot at relevance in a crowded market, but the Anthropic partnership gives it credible underlying technology.

10. Zuckerberg Is Writing Code Again

A small but telling detail emerged this week: Mark Zuckerberg has reportedly returned to writing code for the first time in roughly two decades, submitting three diffs to Meta’s internal monorepo. He is said to be a heavy user of Claude Code CLI. Whether it’s a PR moment or a genuine signal about how AI is changing what’s possible for non-practising engineers, it says something about where we are.

Editor’s Take

The theme this week isn’t any single story — it’s the velocity. Open-source models closing the gap. Agents executing autonomously. AI hacking systems in hours. Enterprises being told to rethink their entire software stack. Every one of these stories would have been headline news in 2024. In 2026, they all happened in the same seven days.

Weekly Roundup compiled as of 5 April 2026  •  AI Insider | aiinsider.in