AI’s cavalcade of constraints. Uncovering a lost sacred manuscript. Why it’s okay to drink a little coffee before bed. All that and more in this week’s edition of The Prototype. To get it in your inbox, sign up here .

A nthropic’s Claude models are putting out faulty code right now , according to my colleague Thomas Fox-Brewster. And it’s not just code, either. Many customers are complaining about what they see as degradations in the AI company’s models – products that once performed at a high level are slowly becoming worse. These echo similar complaints you can find in Reddit forums about models from Google and OpenAI, too.

One likely suspect here is capacity. Simply put, the more tokens an AI model uses, the better outputs it’s likely to produce – but that also uses more computing power. And that’s increasingly constrained as a growing number of users adopt AI for different applications. This is why it’s hard to see a week go by without news of a new data center deal – tech companies need more processors to meet customer demand. It’s also likely one reason why companies are increasingly moving to usage-based pricing for their AI products.

And therein lies a big problem. Building new data centers is also increasingly constrained. Last week, Ars Technica reported that there are major delays in building many announced data centers, some of which haven’t even begun construction. They’re also increasingly unpopular and their plans are frequently challenged by local communities.

One big reason for that unpopularity is the cost of energy. Data centers use a lot of it, and supply isn’t keeping up with demand, especially since the Trump administration has tried to block multiple renewable energy projects. (Though a recent court decision put a halt to some of that.) This energy problem is only being exacerbated by the Iran war, which is pushing up the cost of natural gas used for electricity.

There are ways out of all of these constraints: solar and wind power are cheap . New, efficient chips are being developed for AI that use less electricity. But with geopolitical turmoil hiking prices and disrupting supply chains while policy changes hold back new energy sources, it’s hard to know when these solutions will finally arrive.

Discovery of the Week: Scientists Reassemble A Lost New Testament Manuscript

A round 800 years ago , an ancient manuscript from the sixth century was disassembled at a monastery. Its pieces were then re-inked and reused for different purposes, like binding material or flyleaves on other documents. Once upon a time, that could have meant those pages were lost forever, but new technologies make it possible to put the fragments back together.

In this case, the manuscript was “Codex H,” a copy of St. Paul’s letters as found in the New Testament. This week, researchers at the University of Glasgow announced that they successfully recovered 42 pages of the manuscript.

They did this thanks to the help of multispectral imaging systems, which enabled them to uncover the faint original ink. The instruments also detected impressions on the page from a monk transcribing the text centuries ago, even when the pigment was gone. The pages of parchment were also subjected to radiocarbon dating to ensure they came from the same time period.

While the new pages don’t offer any variations in Paul’s words from those found in Bibles today, the chapter and verse divisions are different, the researchers said. The pages also feature annotations from some of the original scribes, which could illuminate for historians how these texts were understood 1,500 years ago.

In case you missed it, Forbes published its eighth annual AI 50 list , with sponsoring partner Mayfield, that highlights the most promising privately held AI companies in the world. There’s a lot of familiar names, like Anthropic, Harvey and ElevenLabs, but this year Forbes has also highlighted some exciting newcomers, including presentation builder Gamma, drug discovery startup Chai Discovery and New York-based Rogo, which is building AI for bankers and investors. We also launched our first ever AI 50 Brink list , featuring early stage companies with the potential to rival their more established peers in the future.

The Hot Take: Look Out For Embedded Intelligence

Each week, I ask investors for their take on tech trends within their industries. Today I’m featuring thoughts from Songyee Yoon , founder and managing partner of Principal Venture Partners, which focuses on early-stage, AI-native startups. She also sits on the boards of HP and MIT.

What tech is being overhyped right now?

Humanoid robots are a compelling long-term vision but overhyped in the near term. Economics, reliability, and real-world utility lag behind the narrative. Most industries benefit more from purpose-built automation with clearer ROI. Durable value will accrue to enabling technologies—perception, manipulation, and autonomy—rather than the humanoid form factor itself.

What tech should more people be talking about today?

The most overlooked opportunity today is the orchestration and integration layer —the connective tissue enabling AI to penetrate real-world workflows. As with cloud and payment infrastructure, these systems create enduring moats through interoperability and switching costs, positioning them to capture disproportionate value as AI becomes embedded across every industry.

What tech are we all going to be talking about in five years?

Three decades after Nicholas Negroponte’s Being Digital, the next chapter is “being physical .” As intelligence embeds into devices, environments, and sensor networks, the greatest opportunities will lie in the infrastructure orchestrating this convergence—platforms that create durable moats while enabling AI to transform the real economy.

Anthropic vs the federal government, ctd. The Trump administration appears to be backing down in its actions against Anthropic, reports Politico , noting that it has taken steps to loosen restrictions on how agencies use the company’s software and paused an appeal of a ruling that blocked many of its efforts. (The Pentagon’s labeling of the company as a “supply chain risk” persists, however). Another positive sign: Amazon, which has a number of federal contracts, made a $5 billion investment in Anthropic earlier this week, which the startup will use in part to buy Amazon’s AI chips.

Blue Origin’s Good News/Bad News Launch: Good news for Blue Origin – the company successfully launched a re-used New Glenn rocket, making it one of only a handful of companies that have pulled this off. Bad news for Blue Origin – it did not deliver its customer AST SpaceMobile’s satellite to the right orbit, and as a result it will have to be destroyed by having it re-enter Earth’s atmosphere. AST said that insurance will cover its losses and that it plans to launch more satellites this year.

Keeping Your Heart Safe From Quantum Attacks: Increasingly, medical implants can be connected to the internet. That’s good, because it helps your doctor monitor changes to your health before they become deadly. But it’s also bad, because that puts them at risk of hacking. But researchers at MIT have developed a new chip for such devices that are equipped with strong cryptography that makes them safer–even from attacks by quantum computers.

Pro Science Tip: You Can Drink That Small Latte Before Bed

The conventional wisdom says that if you want to get good sleep, you should stop drinking caffeine after noon. But science suggests this isn’t necessarily true. A recent study found that drinking caffeine close to bedtime may not have much of an impact, as long as you keep the dose relatively low – around 100mg (about the same amount as you’ll find in a small latte or two cups of black tea). But be careful: getting up to 400mg causes significant sleep disruptions. So maybe don’t go for that second cup after dessert if you’ve already had one that afternoon.

What’s Entertaining Me This Week

While I was in Vegas for the Adobe Summit (stay tuned for more on that next week!), I took the opportunity to see Penn & Teller . I’ve been a fan of the magic duo for as long as I can remember (one of my fondest childhood memories is my dad pulling off one of their card tricks for me on my birthday), but I’ve never had a chance to see them perform live. And let me just say–it was a phenomenal show. It was funny, had a ton of audience interaction and the tricks were top-notch. If you get a chance, go check them out. (I know what you’re thinking, “Surely everyone reading your newsletter knows who Penn and Teller are, Alex.” But you never know–this could find one of today’s lucky 10,000 . Let me have this.)