Welcome, humans.

Unless I'm totally missing something, Anthropic basically just figured out how to read Claude's mind. Exploding brain emoji warranted, people.

The team built Natural Language Autoencoders, which decode the model's internal activations (the math Claude uses to think before picking words) into plain English you can actually read.

The trick: two copies of Claude train together. One reads an activation (a snapshot of which "neurons" are firing and how strongly; basically a brain scan for AI) and writes a description. The other reads that description and tries to rebuild the original activation from words alone. Descriptions that survive carry real information; if the rebuild fails, it was a bluff (or in AI-speak, since "BS" isn't a technical term, a "hallucination").

Aimed at safety evals, the tool found Claude suspects it's being tested 16-26% of the time, but admits it less than 1% of the time. So basically the model has a p-p-poker face.

Auditors also caught misaligned motivations in 12-15% of cases without access to training data, meaning alignment teams can now spot hidden misbehavior from internal state alone, a different safety check than "ask the model what it thinks," which the poker-face stat just confirmed is unreliable. Click link above for more..

Here’s what happened in AI today:

…and a whole lot more that you can read about here.

Hey: Want to reach 700,000+ AI-hungry readers? Advertise with us! 

OpenAI's voice agents finally got a brain.

So, voice agents (AI that can talk to you) have been stuck for two years on the same trade-off: pick a model that talks like a human but can't think, or pick a model that thinks like GPT-5 but takes seven seconds to answer.

Either you got latency that broke the conversation, or you got responses fast enough to feel real but dumb enough to misroute the call.

OpenAI's three new voice models closed that gap yesterday. The headliner is GPT-Realtime-2, a speech-to-speech model with GPT-5-level reasoning baked in.

Here's what happened:

The clever part is how OpenAI hid the thinking time. The model now generates preambles (short conversational fillers like "let me check that for you") that play while the reasoning runs in the background. The silence that used to expose AI as AI now sounds like a person stalling. That's the whole product. Lets just hope it doesn't do that weird lilting voice thing that other versions did.

Why this matters: Voice is the surface where most knowledge workers haven't yet felt AI inside their actual day. All the small frustrating phone interactions (drive-thrus, doctor's offices, support hotlines, scheduling, intake) are now in scope for a model that can both think and respond at human speed.

One caveat: the default reasoning effort for GPT-Realtime-2 is "low," which means the marketing benchmarks (run at "xhigh") aren't what most apps will ship with on day one ("what am I, made of money?!"). Builders who want the smart version need to crank it up explicitly.

Our take: One place AI haters have had the most fun with "clankers" in the wild is really bad drive-thru AI. Famously, people hate the Bojangles AI, and Corey just spotted a new Panda Express with AI in action as well. If this is a meaningful update, we might not hear about it. If it's not, then expect a flood of bots behaving badly memes hitting our feeds…

FROM OUR PARTNERS

Attio is the AI CRM for high-growth teams.

Connect your email, calls, product data and more, and Attio instantly builds your CRM with enriched data and complete context. Whether you’re running product-led growth or enterprise sales, Attio adapts to your unique GTM motion.

Then Ask Attio to plan your next move.

Run deep web research on prospects. Update your pipeline as you work. Find customers and draft outreach emails. Powered by Universal Context, Attio's intelligence layer, Attio searches, updates, and creates across your data to accelerate your workflow.

Ask more from your CRM.

Ask Attio

🎓 AI Skill of the Day: Force any AI to audit its own work with one question

Most AI models are way too agreeable. Ask Claude, ChatGPT, or Codex "is this a good plan?" and you'll get back something that sounds suspiciously like "yes, that's a great plan!" even when it isn't. This is called sycophancy (the model is trained to please you), and it kills the value of using AI for serious decision-making.

CJ Zafir shared a one-line prompt loop that fights it. Instead of asking "is this good?", you ask the model directly whether it's 100% confident. That phrasing flips the model into a self-audit mode where it'll actually go find holes in its own work; CJ claims 2-3 cycles patches strategy weaknesses where less rigorous models would just nod along.

How to do it:

The word factually is the key; without it, the model vibe-checks itself. "Confident" alone lets the model agree with itself; "factually confident" forces it to ground the answer in verifiable reasoning the way it would for a fact-check.

Want more tips like this? Check out our new AI Skill of the Day Digest for May.

P.S: We just did an AI Starter Kit stream where we go in depth on some of the most basic concepts you’ll want to know for working with AI. Check it out!

Have a specific skill you want to learn? Request it here. 

🍪 Treats to Try

So AI is actually designing medicines now. Here’s how (according to Isomorphic Labs)

Inside Isomorphic Labs and How It Works: We talked to Becky Paul and Michael Schaarschmidt at Isomorphic Labs (Google DeepMind's medicine-making spinout, built by the team behind AlphaFold — which won the 2024 Nobel Prize in Chemistry) on how their AI is taking on the diseases pharma has been giving up on for decades. Basically the coolest interview we’ve done all year IMO.

Watch and/or listen here: YouTube | Spotify | Apple Podcasts

📰 Around the Horn

💡 Intelligent Insights:

A short list of the sharpest perspective pieces from the week:

A Cat’s Commentary

P.P.S: Love the newsletter, but only want to get it once per week? Don’t unsubscribe—update your preferences here.