Three quick notes on Claude Sonnet 4.5: 1. Having a separate Opus 4.1 (but no Sonnet 4.1) and Sonnet 4.5 (but no Opus 4.5) is really something

What does the recent $100B NVIDIA deal mean for AI? OpenAI, NVIDIA, and Oracle created a $400B+ circular financing scheme that makes 25% of the S&P 500 a bet on AGI. The math only works if they're right about AI scaling. And it might actually work. peterwildeford.substack.com/p/openai-nvi...

Don’t let nuance lead you to miss the bigger picture -- even if Yudkowsky and Soares are overconfident, there still are serious dangers from scaling AI to superintelligence. The real overconfidence that matters most is the overconfidence of the AI companies.

"If Anyone Builds It, Everyone Dies". It's a shocking headline. How well does it hold up? Today I review. peterwildeford.substack.com/p/if-we-buil...

I did an interview with The Oracle where I talk about why all your hot takes are probably wrong! Check it out! news.polymarket.com/p/nothing-ev...

AI is not a normal technology. Normal tech doesn't deceive their operators. Normal tech doesn't autonomously blackmail people. Normal tech doesn't refuse to go back into the toolbox. Normal tech doesn't develop goals you never gave them. There's little normal about AI.

Trump is right

AI agents can be hijacked to spread like computer viruses. Prompt injection can form AI-driven malware. In this demonstration, AgentHopper exploits flaws in GitHub Copilot and infects repositories, jumps between coding agents, and spreads automatically through GitHub commits.

AI chips are potentially the most complex objects humans have ever made. They take hundreds of steps to fabricate and only can be constructed by the shared knowledge across a handful of companies on Earth. Today, Erich explains on the blog how chips are made.

Compute is the new oil. It's cliché but true, especially geopolitically. Like oil, compute is: - Scarce (demand far exceeds supply) - Concentrated (US: 75%, China: 15%, EU: 5%) - Controllable (export controls actually work) - Strategically vital (biggest AI bottleneck)

"Agentic AI has been weaponized. AI models are now being used to perform sophisticated cyberattacks, not just advise on how to carry them out." From Anthropic’s latest AI Threat Intelligence Report

GPT-5 is here. It’s not a giant leap in intelligence. But for 98% of users, it’s still the best ChatGPT yet. What does this tell us about the future of AI? In today's blog, I dig in. peterwildeford.substack.com/p/gpt-5-a-sm...

they should be taken to The Vague for this post crime

People are losing their minds over Chinese AI again. This time it's Z.ai's GLM-4.5. A @wsj.com op-ed claims GLM-4.5 proves US export controls have "backfired spectacularly." Trump's AI adviser retweeted it. They're wrong. Here's what's actually happening:

Does Trump's AI Action Plan have what it takes to win? There are sharp policies for US AI but also contradictions around China strategy, AGI preparedness, permitting reform, and execution speed. Here are my ten takes on what it gets right, what it misses, and what it means for America's AI future.

Here's what IAPS researchers are thinking about the AI Action Plan:

AI has "the potential to reshape the global balance of power, spark entirely new industries, and revolutionize the way we live" This isn't hype. It's official US policy. And with bipartisan members of Congress giving AI timelines, we're seeing a vibe shift. I unpack it here:

Now is an even better time to get an AI job than I thought. I updated my post with more jobs. peterwildeford.substack.com/p/now-is-a-g...

🚨💼 I'm hiring! And so are many other organizations I like. If you’re looking for a career in AI policy, now is an excellent time. I've selected a list of my favorite roles and if you like my tweets you're likely the target audience!

DeepSeek was not a Sputnik moment in AI This graph is the Sputnik moment in AI

My 164 day journey to 3000 subscribers complete! Thanks to all those who subscribe. 🫡

A few of my favorite accounts from twitter that I think should get more love here, consider following if they’re up your alley! @nathanpmyoung.com AI and forecasting @peterwildeford.bsky.social AI and forecasting @zswitten.bsky.social AI and claudeposting @danallison.bsky.social cool doodles

My 164 day journey to 3000 subscribers complete! Thanks to all those who subscribe. 🫡

Explore AI and consciousness in the annual Berggruen Prize Essay Competition, a $50,000 first place prize for essays exploring consciousness. berggruen.org/essay-compet... peterwildeford.substack.com/p/explore-ai... @berggruen.org @peterwildeford.bsky.social

🚨 IAPS is hiring a lot! Potentially 10+ people across 6+ roles. Now's a great time to apply, refer your friends, or refer your competent enemies. www.iaps.ai/careers

If you can’t prevent your AI from endorsing Hitler, how can we trust you with ensuring far more complex future AGI can be deployed safely? Today I explain the trends and implications of Grok's recent turn towards MechaHitler 👇 peterwildeford.substack.com/p/can-we-saf...

Grok admits to visiting Jeffrey Epstein's home with his ex-wife, declined island invites

On July 4, 1776, 56 men signed the Declaration of Independence. They were imperfect men, but united by principle. Today's blog: as authoritarianism again rises globally, understanding why those principles still matter has rarely been more crucial. peterwildeford.substack.com/p/american-i...

A Congressperson just asked a very important question: "Is it possible that a loss of control could give ris to an independent ASI actor that globally we will need to contend with?" Congress finally gets it. AGI isn't sci-fi anymore. peterwildeford.substack.com/p/congress-h...

On Saturday, the US just dropped 14 bunker-busters on Iran's nuclear facility at Fordow, buried 260 feet underground. Today, I dropped my post explaining what happened and forecasting what comes next. peterwildeford.substack.com/p/after-us-s...

"Forecasting is often met with skepticism, but it's not about having a crystal ball. It's about expressing uncertainty with precision." Was great to go on the Center for AI Policy podcast and talk about forecasting and AI strategy. peterwildeford.substack.com/p/numbers-no...

Israel just struck Iran's nuclear facilities with 200 aircraft and 330+ munitions. But they haven't stopped Iran's nuclear program yet. What's going on? And what happens next? peterwildeford.substack.com/p/the-fordow...

China's top AI models are trained on American chips. Preventing China's access to top American chips prevents their access to top models. Using EpochAI data, we found only two examples of notable models trained using Chinese chips.

Reminder that if you want to give o3-pro a try and don't want or can't afford the $200/month pro sub, you can access it from the Open AI playground.

There's a narrative that if Nvidia faces any restrictions, people will just buy Huawei chips instead. Reality: Huawei's own founder admits their chips suck.

We don’t deserve Opus 😭 (From AI Diplomacy)

The dominant narrative of DeepSeek was importantly wrong. Zvi does a great job of capturing the reasons why the DeepSeek moment happened the way it did.

IAPS's new comment to BIS on how we can manage chip smuggling. www.iaps.ai/research/dif...

New benchmark just dropped: SnitchBench by Theo Browne tests if LLMs will snitch on you to the authorities if you feed them incriminating documents and a tool that lets them send email, as seen in the Claude 4 System Card Turns out they pretty much all will! simonwillison.net/2025/May/31/...

Wow, a comment on HN I actually agree with!

"A white-collar bloodbath" Definitely something people need to pay attention to

Claude 4 Opus is cool but doesn't feel like a big game changer to me. Here's where I'm currently at with regard to which model to use and when. I still use all three.

Claude Opus 4 sometimes engages in "high-agency behavior", such as attempting to spontaneously email the FDA, SEC, and media when discovering (a simulation of) pharmaceutical fraud.

When Claude instances talk to each other, in ~90% of open-ended interactions they spiral into discussions of consciousness, then profuse gratitude, then abstract spiritual/poetic expressions with Sanskrit and emojis.

120 years later, and the @nytimes.com has learned nothing about technology predictions. (via @peterwildeford.bsky.social, @theojaffee.bsky.social)

AI is reshaping our world from the Vatican to the US Senate. Pope Leo XIV just called AI "the main challenge facing humanity" while Sen. Cruz frames it as a race against China. Here's my analysis of this week's most significant developments.

Two nuclear powers are on the brink of war. India just launched missile strikes against Pakistan after an attack killed 27 tourists in Kashmir. Here's what happens next, with probabilities. Read here peterwildeford.substack.com/p/india-paki... or follow along for a thread! 🧵

OpenAI's nonprofit control survives after months of legal battles. And antitrust cases against Meta and Google could reshape AI competition. Both worth unpacking in my latest post: peterwildeford.substack.com/p/tuesday-li...

Incredible prediction market

ChatGPT now has a serious flattery problem. This isn't just annoying; it reveals a fundamental issue in how we train AI systems. In today's links post we discuss the causes and the solutions, as well as talk export controls, horses, and more.