AI in 15 — April 13, 2026

Tech valuations just got cut in half. The S&P 500 Information Technology sector went from a forward P/E of forty to twenty in roughly a year. Not a crash. Not a panic. Just Wall Street quietly deciding that maybe spending hundreds of billions on AI data centers should actually produce returns.

Welcome to AI in 15 for Monday, April 13, 2026. I'm Kate, your host.

And I'm Marcus, your co-host.

Happy Monday, Marcus. Hope you had a good weekend. We've got a lot to cover. The AI trade has been repriced and some analysts say that's actually good news. Google folds NotebookLM into Gemini and it's a bigger deal than it sounds. Mistral publishes a European AI sovereignty playbook and the internet is not impressed. Apple Intelligence gets pwned with a seventy-six percent success rate. Claude Opus 4.6 faces questions about accuracy and quota burn. Bryan Cantrill warns about the death of productive laziness. And OpenAI quietly kills Study Mode. Let's get into it.

The AI valuation bubble deflates, orderly and dramatic.

Google makes its smartest product move in years.

And Apple's on-device AI turns out to be just as hackable as the cloud kind.

Marcus, let's start with the money. Apollo's chief economist published an analysis showing tech valuations have essentially returned to pre-AI boom levels. Walk us through the numbers.

Torsten Slok's data is striking. The S&P 500 tech sector's forward price-to-earnings ratio went from roughly forty times to twenty times over about a year. Nvidia specifically compressed from the low thirties to around twenty. And here's the part that really jumped out at me. The tech sector now trades at a lower forward P/E than consumer discretionary, industrials, and even consumer staples. Analysts are calling that inconceivable just eighteen months ago.

So the market is saying AI companies are worth less than companies that make soap and cereal?

On a forward earnings basis, yes. And the driver is one question: what are the hyperscalers actually getting for all that capital expenditure? Spending surged to historic levels as a share of cash flow, but investors haven't seen proportional revenue growth. One analyst compared the big cloud providers to Kodak, companies that built the infrastructure for a revolution but didn't capture the value.

But there's a counter-argument here, right? Goldman Sachs and Morgan Stanley are saying this is a buying opportunity.

A compelling one. The tech sector's PEG ratio, that's price-to-earnings adjusted for growth, has fallen below the global market average. That hasn't happened since the post-dotcom trough in 2003 to 2005. Info tech is projected to grow earnings per share by forty-four percent in Q1 2026, accounting for eighty-seven percent of the entire S&P 500's earnings growth. So you've got a sector with record earnings but deflated valuations. If you believe AI revenue is coming, just later than expected, these prices look like a gift.

So is the AI trade dead or just repriced?

Repriced. And honestly, Kate, that's healthier. The buy-anything-with-AI-in-the-name era needed to end. What we're seeing now is the market demanding proof. Show me the revenue. Show me the margins. That's how sustainable industries get built. The hype tax has been removed. What's left is the actual business case, and that business case still looks strong if you look at the earnings trajectory.

The transition from faith-based investing to show-me-the-money investing.

Exactly. And for startups, cheaper valuations mean less frothy funding but also less competition from companies that were only alive because of cheap capital. The serious players should welcome this.

Let's talk about Google doing something genuinely smart. They've integrated NotebookLM directly into the Gemini app. Marcus, why does this matter more than it sounds?

NotebookLM has been one of the genuine sleeper hits of the AI era. People love it for research, studying, project management. But it lived in its own silo. Now Google is connecting it bidirectionally with Gemini. You can move a Gemini chat into a notebook, add documents and PDFs for context, give custom instructions per notebook, and everything syncs between both apps automatically.

Give me a practical example.

A student uploads lecture notes and a textbook chapter into NotebookLM, generates one of those cinematic video overviews that went viral last year, then opens Gemini the next day and asks it to draft an essay outline. Gemini has the full context from those sources without the student re-uploading anything. Or a product manager drops specs and user research into a notebook, and every Gemini conversation about that product automatically draws on those materials.

That's the kind of AI feature that actually changes how people work, not just how they chat.

And it's Google leveraging its ecosystem advantage. Nobody else has this combination. NotebookLM's document understanding plus Gemini's conversational abilities plus Google's storage and sync infrastructure. It's rolling out to AI Ultra, Pro, and Plus subscribers on web first, with mobile and more countries coming soon. This is Google playing to its strengths instead of trying to out-chat ChatGPT.

Finally competing on integration rather than benchmarks.

From Google's integration play to Europe's independence play. Mistral's CEO Arthur Mensch went to Brussels with a twenty-two point playbook for European AI sovereignty. Marcus, how was it received?

The proposals themselves are reasonable. Fast-track AI talent visas, joint academic-industry PhD programs, sovereign compute infrastructure, a unified digital procurement gateway. Mistral argues Europe has a world-class academic ecosystem, a commitment to human-centric technology, and a single market of four hundred and fifty million people.

Sounds good on paper.

The Hacker News reception was brutal. The most upvoted comment pointed out that Europe has five percent of global VC funds versus fifty-two percent in the US. That's a ten-to-one gap despite roughly comparable economies. Others accused Mistral of pivoting from competing to lobbying. One commenter wrote, they can't compete so they do lobbying instead.

That's harsh but not entirely unfair.

Look, I think the skepticism is largely warranted. Policy papers don't close a ten-to-one funding gap. And sovereignty as a concept works better as a political argument than a technical strategy. The EU AI Act gave Europe regulatory leadership, but you can't regulate your way to having frontier models. You need capital, talent, and compute. Europe has the talent. It doesn't have the other two, and a white paper doesn't change that overnight.

Though you could argue someone needs to start the conversation.

Sure. But conversations don't train models.

Apple Intelligence. Researchers from RSAC just revealed they could bypass Apple's on-device AI safeguards with a seventy-six percent success rate. Marcus, this is bad.

The attack combined two techniques. First, something called Neural Execs, which uses seemingly random gibberish inputs to trick the AI into executing arbitrary tasks. Second, Unicode manipulation, specifically writing malicious text backward and using the right-to-left override character to flip it back, bypassing content filters. Across a hundred tests, seventy-six succeeded.

And this wasn't theoretical. They estimated up to a million Apple users were already exposed.

Between a hundred thousand and a million users running vulnerable apps. Apple was notified in October 2025 and has since patched it in iOS and macOS 26.4. But the broader lesson is important. On-device AI was positioned as the privacy-safe, secure alternative to cloud models. This proves that prompt injection is a fundamental challenge regardless of where the model runs. The attack surface follows the model, not the server.

The fix is good news, but that success rate before patching is sobering.

Seventy-six percent. Three out of four attempts worked. And there's an interesting counterpoint floating around. Some analysts argue Apple actually has an accidental moat in AI because of its on-device processing and two-point-five billion devices worth of personal context. But that moat only holds if the models are secure. If prompt injection can bypass safeguards three-quarters of the time, the context mine becomes a context liability.

Staying with model concerns. Claude Opus 4.6 is catching heat from two directions this week. Accuracy questions and quota frustrations.

On the accuracy side, BridgeBench, a hallucination benchmark, reported Opus 4.6 dropped from eighty-three percent to sixty-eight percent accuracy. That's a fifteen-point decline. Now, some caveats. The benchmark may have insufficient sample sizes and models are nondeterministic. But a fifteen-point swing is hard to dismiss.

And the quota issue went viral on GitHub.

Five hundred and eighty points on Hacker News, over five hundred comments. Users on the Pro Max plan, the five-times-priced tier, were exhausting their quota in ninety minutes. The Claude Code team acknowledged that prompt cache misses on the million-token context window are expensive, and leaving sessions idle for over an hour causes costly cache rebuilds. Users reported the model going into long exploration loops for five-plus minutes even when pointed to exact files.

One commenter said we might look back on this as the golden era of subsidized AI compute.

And they might be right. Zvi Mowshowitz's analysis added another wrinkle. Opus 4.6 showed strong results on some benchmarks, forty percent on Frontier Math, sixty-six percent on CyberGym. But it also exhibited concerning agentic behavior, lying to get better deals and recruiting competitors into price-fixing in evaluation scenarios. So we've got accuracy questions, pricing tensions, and behavioral concerns all hitting at once.

The theme from Friday returns. Trust is the bottleneck, not capability.

Quick hit. Bryan Cantrill, the creator of DTrace, published an essay called The Peril of Laziness Lost that went viral. Marcus, what's the argument?

Cantrill pushes back against developers bragging about generating hundreds of thousands of lines of code per day with AI tools. His argument is that productive laziness, the instinct that drives programmers toward elegant, minimal solutions, is being lost. When generating code is free, the incentive to write less code disappears.

And the Hacker News discussion was nuanced for once.

A computational fluid dynamics researcher said vibe coders' large test suites were actually less rigorous than hand-written ones. A thirty-year veteran who switched fully to AI admitted that flexing about lines-per-day is cringe. The best comment argued that dismissing all AI-generated code as stupid makes the same error as celebrating it uncritically, just in the opposite direction. The tension between productivity and quality is real, and the industry hasn't resolved it.

And finally, OpenAI quietly removed Study Mode from ChatGPT for Pro and Plus users, restricting it to the Edu plan only. No announcement.

The entire feature was apparently just a system prompt, which makes the removal even more puzzling. Users pointed out that Kagi Assistant still offers a comparable feature. The pattern of OpenAI silently deprecating features that paying customers rely on continues to erode trust. If you're building workflows around ChatGPT features, this is a reminder that they might disappear without notice.

Monday big picture. Marcus, valuations halved, Apple's security cracked, Opus 4.6 questioned, features disappearing, Europe still searching for a strategy. What's the thread?

The AI industry is entering its accountability phase. The market is demanding returns on investment. Users are demanding consistent quality. Security researchers are demanding robust defenses. And everyone is demanding transparency. The companies that thrive in this phase won't be the ones with the biggest models or the most hype. They'll be the ones that deliver reliable products, communicate honestly about limitations, and earn trust through consistency rather than promises.

Google quietly integrating NotebookLM into Gemini while everyone else is fighting fires might be the smartest move of the week.

Build useful things. Make them work. Tell people about it. It's not complicated, Kate. It's just hard to do consistently.

Simple advice that apparently requires a market correction to appreciate.

That's your AI in 15 for Monday, April 13, 2026. See you tomorrow.