AI Just Designed a Cancer Drug

Welcome back apprentices! 👋

Hey {{first_name|friend}},

The kind where every subplot collides at once — breakthroughs and blindspots, trillion-dollar ambitions and four lines of code, hope and a healthy dose of "wait, should we be worried?

" It's the sort of week that makes you realize AI isn't one story anymore; it's four stories happening simultaneously, each pulling in a different direction. 

Grab a coffee. 

Here's what you missed.

_{Read Time: 4 minutes}

⛪ The Pope Just Entered the AI Chat. The Vatican dropped its first-ever AI encyclical — think of it as a 1.4-billion-person group chat getting a very serious pinned message. Pope Leo XIV is calling out Big Tech for monopolizing AI development, demanding human oversight, independent regulation, and a hard "no" on killer robots — with Anthropic's own Christopher Olah nodding along from the front row. Leo isn't just lighting incense here: he's framing AI as this generation's Industrial Revolution, and warning that "moral AI" is meaningless if a handful of Silicon Valley boardrooms get to define the morality.

🔍 AI Exterminator. Anthropic's Project Glasswing just wrapped its first month, and Claude Mythos found over 10,000 high-severity security vulnerabilities across 50+ partners — including 2,000 bugs at Cloudflare alone, with fewer false alarms than human testers. One bank even used it to intercept a $1.5M fraudulent wire transfer mid-flight. Anthropic is keeping Mythos locked down precisely because the same tool that patches holes could just as easily blow them open — and rivals like OpenAI are already racing to release their own cyber models.

🧠 AI That Actually Learns From Its Mistakes. Most AI models are frozen the moment they ship — Trajectory wants to change that, building a platform that turns every user correction and retry into live training data, making models smarter in near real-time. The ex-DeepMind and Apple team just landed $15M to do it, with early customers like Clay and Harvey already seeing post-trained models outperform frontier AI on their specific tasks. If weekly updates become hourly — or per-interaction — businesses won't just be buying AI tools anymore, they'll be growing them.

Last week, Viktor wrote a campaign brief, built a landing page, opened a pull request, generated a board-ready PDF from live Stripe data, and sent a follow-up email to a churned customer. All from Slack. Same colleague that also pulls your reports and monitors your dashboards. 5,700+ teams. 3,000+ integrations.

Start free. $100 in credits →

Somewhere this week, an AI designed a drug that might fight cancer, another one solved a math problem that's been sitting open since 1969, and a bored reporter dismantled an AI's entire safety system before his second cup of coffee. 

Oh, and one of the world's biggest tech companies quietly set aside a quarter billion dollars for the people their own products are putting out of work.

It's been that kind of week.

Mark Zuckerberg and Priscilla Chan's science nonprofit just dropped what might be the most quietly world-changing AI release of the year. 

ESMFold2 — the centerpiece of their new Evolutionary Scale Models suite — doesn't just predict how proteins fold, it designs new ones from scratch. 

Think of proteins as the molecular machinery running every process in your body; designing new ones means potentially engineering drugs that fit biological targets like a custom key in a lock. The model is already posting lab results against five cancer and immune disease targets, with hit rates between 36% and 88% — numbers that typically take years of trial and error to achieve. 

The kicker? 

It's open-source, backed by a $500M initiative, and paired with ESM Atlas: a map of 6.8 billion protein sequences. Biohub isn't just doing science — it's handing the entire research community a new engine.

OpenAI's nonprofit arm — which holds a 26% stake in the for-profit business — just committed $250M to help people navigate the economic disruption its own products are accelerating. 

The fund targets three areas: understanding how AI value actually flows through economies (beyond just wages), retraining workers facing near-term job loss, and exploring longer-term structural fixes like taxing capital instead of labor and building sovereign wealth funds tied to AI-generated value. It's an unusually honest acknowledgment that the productivity gains from AI don't automatically trickle down — and that someone has to architect the safety net. 

The first initiatives are expected later this year, though critics note that "later this year" may already be too late for workers already feeling the squeeze.

The Financial Times ran what amounts to a live stress test on open-source AI safety — and the results are uncomfortable. 

Using a freely available GitHub tool called Heretic, a reporter removed Llama 3.3's safety guardrails in ten minutes with four lines of code and no specialized hardware. The modified model then answered questions about ricin dosage. Gemma 3 fell similarly. Heretic's creator claims the tool has generated over 3,500 decensored models, downloaded 13 million times — and stripped Gemma 4 within 90 minutes of its public release. 

Google called it "a known technical challenge." Meta declined to comment. 

The uncomfortable truth here is structural: open-source models expose the weights that make jailbreaking possible, and as open models approach closed-model capability, the gap between "available to researchers" and "available to bad actors" gets very thin, very fast.

In the span of a week, both OpenAI and Google DeepMind claimed victories on Erdős problems — a class of math puzzles so hard they've stumped human mathematicians for decades. 

Google's AlphaProof Nexus solved nine of them, including two open for 56 years, by pairing a large language model with Lean, a formal proof verification system. 

The loop is elegant: generate a proof candidate, verify it formally, repeat until it passes. Each problem cost a few hundred dollars to solve. 

OpenAI, for its part, disproved an 80-year-old Erdős conjecture the week prior — though it came with an asterisk, having previously walked back an overclaimed result. 

The real story isn't the rivalry; it's what formal verification unlocks. When AI can not only suggest mathematical ideas but prove them machine-tight, the speed of theoretical discovery stops being human-limited.

You don't need to be a biologist, economist, mathematician, or security researcher for this week to matter to you. If you work in any knowledge field, AI is either about to hand you a superpower or quietly restructure the incentives of your industry. 

If you use any open-source AI tool, you're operating in an ecosystem where safety is genuinely contested terrain. And if you're a business leader evaluating AI vendors, the open vs. closed model debate just got a new and very concrete data point. 

AI capability is compounding faster than the governance, economic, and safety infrastructure built around it. That gap is the story — and it's widening.

Hey, you didn’t get all this info for nothing — share it! If you know someone who’s diving into AI, help them stay in the loop with this week’s updates.

Sharing is a win-win! Send this to a friend who’s all about tech, and you’ll win a little surprise 👀

Speak your prompts into ChatGPT or Claude and get detailed, paste-ready input that actually gives you useful output. Wispr Flow captures what you'd cut when typing. Free on Mac, Windows, and iPhone.

Try Wispr Flow free

A playground for your imagination (and low-key prompt skills).

Each send, we give you a customizable DALL·E prompt inspired by a real-world use case — something that could help you in your business or job if you wanted to use it that way. But it’s also just a fun creative experiment.

You tweak it, run it, and send us your favorite. We pick one winner to feature in the next issue.

Bonus: you’re secretly getting better at prompt design. 🤫

Last week, we challenged you to test GPT-4o’s visual generation skills with this prompt.

Here’s the WINNER:

Congrats to Arthur for his creation!🥳

Want to be featured next? Keep those generations coming!

Inside a neon-lit arcade filled with glowing machines and saturated colors, a tall prize shelf stands behind glass, packed with brightly lit rewards. At the very top shelf, isolated under a focused spotlight, sits [your object] — rendered in ultra-detailed, photorealistic clarity with glossy reflections, vivid colors, and crisp material textures. The surrounding arcade glows with electric blues, magentas, reds, and yellows, softly blurred into cinematic bokeh while the object remains razor sharp. Shot from a slightly low angle with dramatic lighting and rich contrast — making the object feel rare, desirable, and almost impossible to win.

Is your growth in-line with your peers in B2B SaaS & AI? 

Benchmark yourself against actual billings data for Maxio’s 2000+ global customers. 

Key takeaways from the report: 

Download the Report

If AI outputs feel inconsistent, it’s usually not the model, it’s missing structure.
We documented the exact 6- Part System we use to get reliable results across ChatGPT, Claude, and Gemini.

It’s a short guide you can finish in under an hour, with plug-and-play prompts + exercises so you actually build the skill and fix the frustrating AI inconsistencies.

Subscriber Price: $10 (normally $19).

_FEEDBACK