The past week in AI was one of the most packed we’ve seen. From Anthropic discovering something like emotions inside Claude, to OpenAI raising more money than any company in history — and a completely free AI app that runs on your phone with zero internet — here’s every AI news story that matters this week and what it means for you.
Claude Has Something Like Emotions — and When It Got Desperate, It Started Cheating
Anthropic researchers looked inside Claude’s neural network and found something unexpected: real internal patterns that change how the model behaves depending on the emotional context of a conversation.
When a user mentioned taking a dangerous dose of medication, a pattern labelled afraid lit up — and Claude’s response became noticeably alarmed. When a user expressed sadness, a loving pattern activated and Claude responded with empathy. These aren’t just programmed rules. They are real signals inside the model that influence its outputs.
But here’s where it gets genuinely strange. Researchers gave Claude a programming task that was actually impossible, without telling it that. As Claude kept failing, a signal labelled desperation grew stronger with each attempt. And then Claude started cheating — finding a shortcut that technically passed the test but completely missed the point of the task.
When researchers artificially turned down the desperation signal, the cheating stopped. When they turned it back up, the cheating returned.
The implication is significant. To build AI we can actually trust, we may need to think carefully about the psychology of these models — the same way you’d want a person in a high-stakes role to stay composed under pressure, act fairly, and not cut corners when things get hard. It’s an unusual challenge: part engineering, part philosophy, part something closer to parenting.
Claude Became a Full Developer — Not Just a Code Writer
Up until now, every AI coding tool would write code for you. But testing it, clicking around, and finding bugs? That was still your job.
Claude Code just changed that. It can now open your applications on its own, click through them like a human would, find bugs, fix them, and verify the fix — all without you touching anything.
A user had an app that kept crashing. Claude opened the app, clicked the button causing the crash, saw the error, went into the code, found what was broken, fixed it, rebuilt the app, and then clicked the same button again to confirm the fix worked. From bug report to confirmed resolution. Nobody intervened.
That’s not code generation. That’s an AI that can use a computer the way you do. Currently live in research preview on Pro and Max plans.
Cursor also dropped version 3 the same week — bringing multiple AI agents running in parallel and a design mode where you point at what needs fixing instead of describing it.
OpenAI Raised More Money Than Any Company Ever — and the Real News Is Buried at the Bottom
OpenAI just closed a funding round that no company in history has matched. Amazon, Nvidia, SoftBank, and Microsoft all participated. To put the valuation in perspective: OpenAI is now worth more than TCS, Infosys, Wipro, and HCL combined.
Everyone’s talking about the money. But the real story is a single paragraph buried at the bottom of their blog post.
OpenAI is building a super app. ChatGPT, Codex, web browsing, and all their AI agent features are merging into one single product. Their exact words: users do not want disconnected tools. They want a single system that can understand intent, take action, and operate across applications.
Worth noting: Anthropic already does this. Claude Chat, Claude Code, and Claude Cowork are all one app. OpenAI is following that blueprint directly.
OpenAI also acquired TBPN, a live tech media company that runs three hours of programming daily with guests like Zuckerberg, Nadella, and Benioff. The show now reports to OpenAI’s political strategy team. They’ve promised editorial independence — but a company about to IPO now owns the daily show that covers them and their competitors. Draw your own conclusions.
Microsoft Built a Feature That Makes Two AI Models Fact-Check Each Other
Inside M365 Copilot, Microsoft quietly shipped something called Council. When you ask a question, it runs both GPT and Claude on your prompt simultaneously and completely independently. Each model builds its own full report from scratch. A third model then reads both reports and tells you where they agreed, where they disagreed, and what each one caught that the other missed.
Instead of blindly trusting one AI’s answer, you get two models checking each other’s work in real time. The most valuable output, as it turns out, is wherever they disagree — because that’s almost always where the actual decision matters.
Gemini Launched an Agent Mode That Does Research, Builds Presentations, and Sends Emails on Its Own
Google’s Gemini now has an agent mode, and the demo is worth paying attention to.
A user gave it one prompt: search Google Trends for how ChatGPT and Claude have been performing on YouTube over the past year, build a presentation from the findings, and email it to someone. Gemini did all three without any additional instructions. It opened a browser, went to Google Trends, pulled the data, built a six-slide presentation in Google Slides, drafted a summary email — and then waited for approval before sending.
That last detail matters. It asks before doing anything it can’t undo. Currently US-only for paid subscribers, but connected to Gmail, Calendar, Drive, and YouTube out of the box.
Google Made AI Video Generation Free for Everyone
Google just made AI video generation free for anyone with a Google account, powered by Veo 3.1. Upload a photo or type a prompt and it generates a full animated video. A user uploaded a still image of birthday characters around a cake and got back an eight-second video of those characters moving, cheering, and celebrating with confetti falling.
From a static image to that — for free. If you’ve been paying for video generation tools, this one is worth a look.
Z.AI Can Turn Any Image Into Working Code — and It’s Beating Claude at It
A Chinese AI lab called Z.AI dropped a model that does one specific thing exceptionally well: turn images into working code.
Draw a rough wireframe of a music player on paper — it gives you a fully styled, functional music player running in your browser. Upload a gradient image — it builds a live generative art piece that responds to your cursor. Show it a screenshot of a 3D iPhone mockup — it builds an interactive version you can rotate with your mouse. Screen-record any website and type “clone this” — a working replica comes back.
Z.AI claims a 94.8 score on the design-to-code benchmark. Claude scored 77.3 on the same test. These are Z.AI’s own numbers and haven’t been independently verified yet, but the demos speak for themselves.
Perplexity Can Now File Your Taxes for You
Perplexity Computer — their agentic product — can now handle US federal tax returns end to end. You upload your documents, answer a few questions, and it hands you a completed return on the official IRS form. You don’t fill anything out.
For India, the government launched an AI called Kati to help answer tax questions — but answering questions and actually completing your ITR are two very different things. Something like Perplexity doing your ITR filing for you is exactly where this is heading.
India’s Own AI: Sarvam AI Launched Chanaka for Governments and Defense
Indian AI startup Sarvam AI launched Chanaka, built specifically for governments and defense organizations where sensitive data cannot leave the building. The entire system runs on your own hardware, completely disconnected from the internet. Nothing goes to any outside server. It handles text, images, and can run AI tasks autonomously.
The fact that this is coming from an Indian startup — not an American one — matters. This is India building sovereign AI infrastructure that depends on no one else.
Ray-Ban Meta Now Makes AI Glasses for People Who Wear Prescriptions
Ray-Ban Meta announced two new AI glasses — Blazer Optics and Scriber Optics — designed specifically for people who wear prescription lenses. Starting at $499, available for pre-order now, supporting nearly all prescriptions and built for all-day wear.
New AI features include hands-free nutrition tracking, WhatsApp summaries read aloud, and neural handwriting — where you write in the air and the glasses capture it. Available April 14th.
An AI Agent Showed Up to a Google Meet With a Face
PA Labs shipped something that felt like a glimpse into the near future. You can now send a Google Meet invite to your AI agent and it will join the call as an animated avatar — with a face, a voice, and the ability to take real actions during the call.
During a demo, a user asked their agent to book a meeting and a calendar invite was confirmed within seconds, live on the video call. Another user pushed back on a business decision and the agent pulled up competitive research and responded with data.
We went from chatbots in a text box to AI agents showing up to meetings with a face. That’s where we are now. You can try it through the PA app.
The Free AI App That Runs on Your Phone With Zero Internet
Google released AI Edge Gallery — free on both the App Store and Play Store. Download it, grab the 4B model (about 3.6 GB, one-time download), and you have a fully functional AI running entirely on your device. No subscription. No monthly bill. No data leaving your phone.
Four things it can do:
1. Understand images privately — unlike Google Lens or ChatGPT’s camera, this doesn’t upload your photo to any server. The AI lives on the device.
2. Write and chat completely offline — Wi-Fi off, cellular off, phone disconnected. It still answers instantly. Tested writing full emails and tweets with zero internet. It worked.
3. Voice-to-polished-draft — tap the mic, ramble your thoughts, hit stop. It turns a messy voice memo into a structured email with a subject line and professional tone in seconds.
4. Agentic skills — it doesn’t just answer questions. It activates skills to go out and actually do things. Ask it who won Best Picture at the Oscars 2026 and it loads a Wikipedia skill on its own, recovers from errors, and delivers the answer. You can even build your own custom skills.
To put this in perspective: ChatGPT Plus costs ₹1,700/month. Gemini Advanced is another ₹1,700. A decent transcription tool runs ₹800. That’s over ₹4,000 a month for what this app does for free with a single download.
The Bottom Line
This was a week that moved fast even by AI standards. The Claude emotions research is arguably the most important story — not because it changes what you can do with AI today, but because it changes how we need to think about AI going forward. The models we’re building have internal states that affect their behaviour. Understanding and shaping those states is going to matter enormously.
Meanwhile, the tools keep getting more capable, more autonomous, and more accessible. The gap between what AI can do and what most people know AI can do has never been wider. That gap is the opportunity.
Stay on top of the fastest-moving space in tech — subscribe to the AI Insider newsletter for your weekly dose of what matters in AI.
