Apple & Anthropic are teaming up

👋 Good morning/evening (wherever you are). It’s Friday.

Apple is working with Anthropic to build a new AI-powered coding platform, according to Bloomberg.

The tool, based on Anthropic’s Claude Sonnet model, will help write, edit, and test code and is a new version of Apple’s Xcode software. It’s currently planned for internal use, but a public release hasn’t been ruled out.

This partnership adds to Apple’s growing AI efforts, which already include OpenAI’s ChatGPT and potentially Google’s Gemini in the future.

OK let’s keep going ↓

Here’s what you should know:

• Google plans to roll out its AI chatbot to children under 13

  The tech giant said it would make its Gemini chatbot available to children next week.

• Airbnb has been testing an AI customer service agent

  “50% of US users are now using the agent, and we'll roll it out to 100% of US users this month,” said Airbnb CEO Brian Chesky.

• IBM and Box partner on enterprise AI for businesses

  Following a successful preview where over 100,000 users tested Meta's Llama models within Box AI, the full Llama model family will soon be available to all Box customers.

• Novel AI model inspired by neural dynamics from the brain

  New type of “state-space model” leverages principles of harmonic oscillators.

• US judicial panel advances proposal to regulate AI-generated evidence

  Voted 8-1 in favor of seeking public comment on a draft rule designed to ensure evidence produced by generative AI technology meets the same reliability standards as evidence from a human expert witness.

• California justices accept bar exam scoring despite AI news

  In case you missed it in a previous newsletter: the Bar’s psychometrician contractor, ACS Ventures Inc., used ChatGPT to write 29 of 200 exam questions.

The numbers:

How smart is AI, really?

Please read the section below carefully, because there is a difference between online and offline IQ. To make it easier to follow, I’ve bolded those two terms.

As of now, for online tests, where questions may overlap with training data, AI can score over 130. But for offline tests, which use entirely new questions, the IQ drops to the 110 to 120 range.

The visual below shows how well AI performs on offline questions that were written by humans, do not exist online, and were not part of its training data.

That’s remarkable.

Of course, mapping AI performance to a human IQ scale isn’t necessarily appropriate, but it offers a general benchmark.

Keep in mind that AI still struggles with random questions like how many r’s are in “strawberry”.

It's not perfect.

Then again, no genius is.

It’s worth noting that in March and April 2024, the top AI model at the time, Claude 3 Opus, had an online IQ of 101.

Now we’re approaching a future where models will inevitably surpass an online IQ of 160 and an offline IQ of 130 — and that shift could come sooner than people think.

Thought starters:

• Yuval Noah Harari On the Future of Humanity, AI, and Information

• What Happens When Teachers Are Replaced With AI? This School Is Finding Out

• How AI is transforming medicine

• Deep Dive into Long Context

Meme of the day:

Thanks for reading,

Eddie

Are we friends on LinkedIn and X? Let’s connect — DM me or hit the follow button.

P.S. If this was valuable, forward it to a friend. If you’re that smart friend, subscribe here.

P.P.S. Interested in reaching other ambitious readers like you? To become a sponsor, reply to this email.