Sam Altman 详述 AI 安全竞赛的三种终局

AISam Altman’s three endings for the AI safety race

PLUS: Make Claude Code check its work after each edit

Zach Mink

Good morning, AI enthusiasts, and welcome to our 4,761 new readers. OpenAI’s current AI safety playbook — shipping models once they pass safety tests and flagging any concerns — is what Sam Altman reportedly calls “pacing 1.0.” The sequel options are a bit more extreme.

Altman reportedly gave staff three possible outcomes: a global watchdog, a single super-lab merger, or one lab racing ahead to “melt everyone else’s GPUs,” an outcome he called “terrible.” OpenAI now has a year to choose its own path.

Reminder: Our next live workshop is today at 12 PM EST — join and learn how to get the most out of Microsoft Copilot and leverage its newest features to get real work done. RSVP here.

In today’s AI rundown:

Sam Altman details the AI industry’s endgame

Ex-OpenAI research chief’s Arda debuts at $1.5B

Make Claude Code check its work after each edit

Anthropic’s Claude takes on scorpion antivenom

LATEST DEVELOPMENTSOPENAI

🔮 Sam Altman details the AI industry’s endgame

Image source: Images 2.5 / The Rundown

The Rundown: OpenAI CEO Sam Altman gave an internal talk to staff last week that covered three possible endgames for safe AI, according to The Information — ranging from a global watchdog combining the top labs to one company racing ahead.

The details:

The report claims Altman favored plan 1: a global enforcer built like the International Atomic Energy Agency, which OAI has lobbied for since 2023.

The “Demis plan” (named after DeepMind co-founder Demis Hassabis) folds top labs into one lab serving partner countries, similar to CERN’s particle research.

Altman called plan three (named for AI researcher Eliezer Yudkowsky) “terrible,” with one lab riding self-improving AI to “melt everyone else’s GPUs.”

He called picking a path, due within a year, “pacing 2.0,” and said it must include a “speed limit” or stopping point for self-improving AI.

Why it matters: Anthropic’s Jack Clark just talked about an intelligence cap and said the world has “a small number of years” to act on safety — and Altman’s scenarios seem aligned with his No. 1 rival. An Avengers-style team-up of top labs under one roof would make for a fun plot, but as with every pact so far, getting the rest of the field is not easy. TOGETHER WITH PAVE

🏗️ Build customized, connected apps in minutesVisit quickbase.com

The Rundown: Pave is an AI app builder that turns the way your team works into a custom app, connects the tools you already use, and runs the process so you’re no longer the glue holding it together.

With Pave, you can:

Describe your process in plain language

Connect the systems you already use

Go live with a working app in minutes

Try Pave today. ARDA

🏭 Ex-OpenAI research chief’s Arda debuts at $1.5B

Image source: Arda

The Rundown: Former OpenAI Chief Research Officer Bob McGrew just came out of stealth with Arda, a startup building AI that helps factories set up assembly lines for new products far faster, launching at a $1.5B valuation after a $160M Series A round.

The details:

McGrew pitched Arda as being able to do “for manufacturing engineers what Claude or Cursor does for programmers,” planning builds from design files.

The comparison is Asia’s factories, where big crews relearn a new design in a day, while reprogramming one automated station can cost nearly $1M.

Defense firm Anduril, Arda’s first customer, says basic assembly instructions now take 15 minutes vs. 10 hours, and it’d be “behind” without the tech.

McGrew wants robots on the line next, reprogrammed by Arda’s agents in days, calling the goal “a fully automated version of Shenzhen in West Texas.”

Why it matters: McGrew joins Mira Murati, Ilya Sutskever, and others on the list of OAI departures now running billion-dollar startups, but Arda is the rare one building factories instead of frontier models. It also comes months after Jeff Bezos’ Prometheus raised $12B for an “artificial general engineer,” focusing on designing physical products. AI TRAINING

🪝 Make Claude Code check its work after each edit

The Rundown: In this guide, you’ll make Claude Code format a settings file and run checks after selected edits. You’ll introduce a small error, see which test catches it, and fix the code without changing the test.

Step-by-step:

Open a new practice folder and ask Claude to build the guide’s small shipping calculator. Run the five tests to check that it works before adding a hook

Add a hook that formats the settings file after Claude writes it. Save the settings on a line and check the hook makes them easier to read without changing values

Add the shipping tests to the same hook. Have it finish formatting before running the checks, then confirm all five pass

Ask Claude to introduce the guide’s deliberate error. An order worth exactly $50 should now fail the free-shipping test. Read the failure before fixing anything

Restore the original rule and run the same tests again. Check that all five pass, then follow the guide’s steps to turn the hook off

💰 Bad AI architecture ate your token budgetVisit glean.com

The Rundown: For enterprise leaders scaling AI, Glean delivers more useful work per token — because better context, smart model routing, and efficient agent design cut waste before the model ever spends a token.

How a context layer reduces wasted processing

Why model routing sends tasks to the right model

How Glean saves 81% fewer tokens per task vs. Claude Coworker

Explore the full story to calculate potential savings. AI & DRUG DISCOVERY

🦂 Anthropic’s Claude takes on scorpion antivenom

Image source: Citrini

The Rundown: Anthropic published new research in which an internal AI designed proteins to lock onto scorpion venom toxins and disarm them, run with the investment firm Citrini Research, with two outside labs confirming that 17 of 211 designs worked.

The details:

Citrini’s five staffers worked at Anthropic’s offices after they couldn’t get the safety block lifted on their own computers, with an estimated budget of $200,000.

Claude ran for 64 hours off five human messages, drawing up 318,484 candidate designs and sending 211 to Twist Bioscience and Adaptyv Bio for lab work.

17 designs passed the binding tests, but each grabs only one toxin, and the proteins haven’t been tested to see if they can actually stop the venom.

Citrini declared its moonshot experiment of a universal antivenom ‘failed’, but its takeaway was that the “frontier has moved” for AI and drug discovery.

▸ James built a 3D-printed barrel nut with Claude

Today’s workflow comes from reader James Miller:

“I 3D-printed a custom threaded barrel nut without ever opening AutoCAD. I use my 3D printer to make parts for various projects, including a pinball machine and a boat, but I struggle with AutoCAD.

As an alternative, I described the part to Claude in plain English and asked it to create a .3mf file. I opened the file in Bambu Studio and printed the part.”

See the workflow on our Community Workflow Hub →

What’s AI still not doing for you? We want to hear the task that’s still eating your time, or the thing AI keeps getting wrong. Share yours Visit The Rundown University.

🛠️ Trending AI Tools

🏠 Town - Personal AI assistants that team up across your whole company

📞 Voxtral 3 - Mistral’s voice AI suite for building multilingual phone agents

🗣️ Phonon - Gradium’s open speech models small enough to run on phones

*Sponsored Listing

📰 Everything else in AI today

Radical Numerics claimed a world first in AI-designed medicine, saying the FDA cleared a virus its Omnii-1 AI designed for a patient whose infection resisted treatment.

Ex-DeepMind researcher Thore Graepel debuted Metis Reasoning, an AI startup applying the AlphaGo tricks to chatbots so they can check their work against evidence.

Chinese AI lab DeepSeek released the code behind DeepInfer, the engine that runs its V4.1-Flash AI, with Huawei chip support that lets firms serve it without Nvidia hardware.

Skyfall AI unveiled SKY-001, an AI that drafts 3D designs of parts and buildings more than 51 times faster and significantly cheaper than frontier models.

OpenAI announced the acquisition of Orchestra Bio, a startup whose AI agents track and monitor studies, adding it to the company’s Rosalind AI science tools.

🎓 Highlights: News, Guides & Events

RSVP to workshop today @12 PM EST: Get real work out of Microsoft Copilot

Anthropic co-founder says the singularity is underway

Anthropic designs a pay package for the public

Zach Mink • 8 minutesAI

Rowan Cheung • 4 minutes

Stay Ahead on AI.