This Week in AI — April 12, 2026: 13 Biggest Updates You Need to Know

Published: April 12, 2026 | Reading time: ~10 minutes | Weekly AI Roundup

A director of AI at AMD just proved with real data that Claude has become 73% less thorough since February. The CIA rescued a pilot in Iran using AI that detects heartbeats from miles away. Meta’s new super-intelligence lab dropped its first model and it already beats Claude and GPT on key benchmarks. And Netflix just built an AI that can erase objects from video in a way that respects physics.

A big week. Here are all 13 updates explained simply — including a step-by-step tutorial at the end for a free AI memory system built by the co-founder of OpenAI.

1. Is Claude Getting Dumber? The 73% Drop That Has Everyone Talking

Stella Lorenzo, Director of AI at AMD, did something most frustrated Claude users only wish they could do. She pulled 6,852 real team session files and tracked over 200,000 actions Claude took in actual work. Then she posted the full breakdown on Claude’s official GitHub page — with screenshots and metrics.

The findings are stark. Claude’s thinking depth dropped 67% since February — closer to 73% by end of March. The model used to read your code 6.6 times before editing it. Now it reads it twice. That is a 70% drop in how much research it does before it touches your work.

The dirty secret: When AI companies push updates to make their models cheaper to run, they don’t tell you. No version number. No changelog. The Claude you used in February and the Claude you use today may be two completely different products — and you have no way to know. Anthropic has said nothing publicly.

2. The CIA’s ‘Ghost Murmur’ — AI That Detects Your Heartbeat From Miles Away

On April 3rd, a US fighter jet was shot down in Iran. The pilot spent 2 days hiding in mountains with no phone, no signal, no way to communicate. Iranian forces were actively hunting him.

The US military reportedly found him using a classified CIA tool called Ghost Murmur — which detects a person’s heartbeat using its electromagnetic signal from miles away. Your heart is a muscle. Every contraction produces a tiny electromagnetic pulse. The CIA built sensors sensitive enough to pick it up at great distance, then layered AI on top to filter one heartbeat from background noise.

Physicists are debating whether this is even scientifically possible. Either the CIA just revealed the most advanced sensing technology ever built — or this is the most effective military misdirection in modern history. Either way, AI is now part of warfare in ways we are only beginning to understand.

3. Meta Muse Spark — The First Model From Zuckerberg’s Super-Intelligence Lab

Meta quietly launched a super-intelligence lab a few months ago. This week, its first model dropped: Muse Spark. It works differently from anything else out there.

Instead of answering your question once, Muse Spark spins up multiple AI agents in parallel — all working on your problem simultaneously. Ask it to plan a trip to Goa: one agent drafts the itinerary, another compares hotels in your budget, a third finds activities for kids. All running at the same time, stitched together for you.

On official benchmarks, it beats Claude Opus 4.6 on most visual tasks and beats GPT-5.4 on health benchmarks. Snap a photo of your lunch and it labels every item with calories and a health score. Send a photo of your yoga pose and it tells you which muscles you are actually engaging.

And Meta has a distribution advantage nobody else has: billions of users already on WhatsApp and Instagram. Muse Spark is rolling out inside those apps soon — you won’t need to download anything new.

Try it now: Go to meta.ai, upload any image and ask a question. Rolling out on WhatsApp soon.

4. Claude Now Lives Inside Microsoft Word — 1.2 Billion Users

While Claude was dealing with a PR crisis over the quality drop, Anthropic quietly launched something massive: Claude is now live directly inside Microsoft Word as a full beta integration.

What you can actually do with it: take any 100-page contract, ask Claude to summarise what the other party changed, and it flags every deal-breaker, every negotiable point, and every cosmetic edit. Ask it to push back on a specific clause — Claude writes the actual redline edits directly in your Word document using proper tracked changes. You click accept.

Word has 1.2 billion users on Earth. Most have never touched an AI tool. This is how AI quietly becomes the thing you use every day without thinking about it.

5. Netflix ‘Void’ — The AI That Understands Physics When Erasing Video

Every existing video editing tool can paint over an object to remove it. Pixels disappear, background fills in. But none of those tools understand what the rest of the scene should physically do after the object is gone. Netflix’s new model, Void, does.

Remove the hands holding a spinning wooden top — and the top actually starts wobbling because nothing is holding it steady anymore. Remove a heavy kettlebell from a pillow — and the pillow slowly decompresses and rises back up on screen.

Netflix has the largest video library on the planet to train this on. What this is building toward: re-shoot an entire scene of a Netflix show without flying actors back. Change a sponsor’s product in a movie scene without redoing the shot. Netflix is building a button that erases reality.

6. Google Gemini Notebooks — Finally, Organised Conversations

The single biggest complaint about Gemini was that every chat sat in one unsorted pile. Google fixed it with Notebooks — a new sidebar section that lets you group chats by project.

The part that actually matters: every notebook has its own sources you can attach to it — your PDFs, Google Sheets, web pages, personal notes. Gemini answers using that specific context, not generic stuff scraped from the internet. If you have used NotebookLM and wished it lived inside your main chat app, this is exactly that.

7. Claude Managed Agents — Build AI Agents in Plain English, No Code

Anthropic launched Managed Agents — which completely changes what it takes to build an AI automation. Until now, you needed to deal with authentication breaking, memory crashing mid-task, retry loops, crashing servers. All of that is now handled by Claude itself.

You open a text box, type in plain English the agent you want — like ‘Build me an agent that evaluates acquisition targets and drafts an investment memo’ — and Claude figures out the configuration, writes the code, and runs the whole thing.

A two-person company can now run the kind of automation that used to need a full engineering team. We have a full detailed guide below — see Article 2.

Try it: Go to claude.ai/agents — build your first agent in plain English in under 10 minutes.

8. Gemini’s Interactive Physics Simulations

Google also shipped interactive simulations inside Gemini. Ask it to show you how a double pendulum works — and it builds a fully playable physics simulation right in the chat window. Move the sliders, the physics changes in real time. Pause it, reset it, try a new configuration.

Switch to Pro mode in Gemini, type any concept, and watch it build a live experiment in 30 seconds. If you have a kid in school or are learning anything technical yourself, this changes how studying works.

9. Google AI Shopping Built for India — Inside Gemini

Google shipped a full AI shopping experience inside Gemini, built specifically for how Indians shop. Ask in plain English: ‘I’m running my first half marathon — compare the Adidas Evo SL and Ultra Boost 5 and recommend one.’ Gemini pulls up actual product cards with real prices and real stock availability inside the chat, and builds a comparison table.

Google Lens now lets you point your camera at anything to find that exact product online. You describe what you want. You don’t browse for an hour.

10. Shopify AI Agent — Fix Store Leaks Daily

Shopify dropped an AI agent for ecommerce store optimisation. If you run a store, think of it as Moneyball for conversions. Instead of chasing big redesigns, fix small leaks daily:

‘Why is my product page slow on mobile?’
‘Where am I losing users in checkout?’
‘Improve this page for conversions’

Most stores lose on tiny gaps compounding. Every 1% you fix shows up directly in revenue. Full guide at shopify.dev/docs/apps/build/ai-toolkit

11. ChatGPT + Upwork — Find Freelancers Without Leaving Chat

Finding a freelancer used to mean opening four tabs, writing a job post, filtering 200 proposals, getting ghosted, starting over. Upwork is now inside ChatGPT. Describe your project in chat — it matches you with freelancers from 18 million profiles, drafts the job post, and hands you off for contracts. Building got easier. Finding who builds also just got easier.

12. OpenClaw Native Video Generation + Cursor Remote Mobile

OpenClaw added native video generation plugged into 9 different video models — OpenAI, Google, Alibaba, xAI, MiniMax and more. Generate AI videos from every major model without leaving a single app.

Cursor launched remote mobile coding — control Cursor running on your laptop directly from your phone, from anywhere. Wake up in traffic, remember a bug you forgot to fix, type it on your phone, your laptop handles it. By the time you’re home, it is done.

13. Anthropic’s Gigawatt Power Deal — Winning the Actual War

While OpenAI was telling investors it has more compute than Anthropic, something quietly happened: Anthropic signed a multiple-gigawatt power station deal with Google and Broadcom, starting 2027. That single deal is nearly twice the size of OpenAI’s entire current capacity on Earth.

OpenAI is winning the press release war. Anthropic is winning the actual war.

BONUS: Free AI Memory System — Built by the OpenAI Co-Founder

Andrej Karpathy — who helped build OpenAI from day one, then led Tesla’s self-driving AI — just open-sourced a fix for Claude’s memory problem. Here is how to build it in 5 minutes, zero coding, completely free.

What You Are Building

A personal Wikipedia where Claude remembers everything you ever showed it — across sessions, forever. Every research paper, article, or YouTube transcript you feed it gets connected into a visual knowledge graph.

Step 1: Download Two Tools

Obsidian (obsidian.md) — a notebook app that shows how every note connects to every other note visually. Download, install, create a new vault.
Obsidian Web Clipper — free Chrome extension. Find any article online, click once, it saves straight into your vault.

Step 2: Connect Claude

Open the Claude desktop app (not browser version). Split your screen — Claude left, Obsidian right.
Go to Karpathy’s GitHub — find the file called ‘wiki schema’. Copy everything from the top down to the notes section.
Paste the schema as a prompt in Claude, then hit enter. Claude asks which folder — pick your vault, trust the workspace.

In 30 seconds, Claude builds your entire Wikipedia structure: raw research folders, organised knowledge folders, an index, a changelog. This vault is now Claude’s permanent memory.

Step 3: Feed It

Find any article, paper, or YouTube transcript. Click the web clipper once.
Switch to Claude. Type: ‘Please ingest [paper name].’
Watch Obsidian — dots appearing. Each one is an idea Claude pulled out. Every concept, every statistic, all connected.

The magic: Feed it three completely unrelated papers and ask Claude ‘how are these connected?’ It finds the story you never would have seen yourself.

Which jobs get replaced first? Claude’s analysis from three papers: Most exposed — computer programmers, data entry, customer service, financial analysts (if your job mostly happens on a screen, the clock is ticking). Safest — physical trades, healthcare, construction, farming (if your job needs a body in a specific physical place, AI can’t touch it yet).