Skip to content
Fortissimo

Futuro Fortissimo Archive

🎼 ff.148 The AI “Myth”

Claude Mythos, autographs and world models

Published · Michele Merelli

  • Is Anthropic's latest model too dangerous?
  • Will Sinner be beaten by a Roomba?
  • The human way out: Twitch and autographs?

🐉 ff.148.1 Claude Mythos: not for everyone

Substack ↗

As you may have heard, there's a new myth out there: Claude Mythos - the 245-page PDF release.

Anthropic's new model is so powerful that, if it were released today, it could cause quite a few problems. It has dug up flaws and vulnerabilities in systems that underpin the internet, including a bug in OpenBSD (an operating system famous for its rock-solid security) that had stayed hidden from human eyes for 27 years.

In short, code is going through what Go went through with “move 37” (♟️ ff.18.4 Artistic chess: the king of chess): even world champions get beaten. But this isn't a board game: these are the foundations - the internet and software - of our economy (and our lives).

So? On the code side, Anthropic (with Apple, Amazon, Cisco and others) has launched Project Glasswing, to give AI wings - yes - but with transparency, as if they were made of glass.

🎾 ff.148.2 Sinner and other sins

Substack ↗

The singularity - in code - seems to have arrived. Other fields will probably follow: biology, economics, genetics. Oh, and robotics, obviously.

AI's first “physical spillover”, as recent developments show:

LATENT trained humanoids to play tennis using imperfect human motion data and noisy videos, moving from hard-coded movements to a dynamic (because AI) “latent action space”.

In China, 100 kWh mobile charging robots turn car parks into charging points: power banks on four wheels.

Figure showed Helix 02 tidying up an entire living room.

🎾 ff.148.2 Sinner and other sins
Sources & links (3)

🌎 ff.148.3 A world out there?

Substack ↗

According to some, getting from a chat to the real world - from ChatGPT to Westworld - takes more than the “next token prediction” of an LLM: physical actions do require understanding concepts, but also placing them in space and time.

So the next OpenAI or Anthropic might be World Models.

Plenty of people are working on them: General Intuition and Meta's former chief AI scientist, Yann LeCun, with AMI Labs (Advanced Machine Intelligence).

And Physical Intelligence (which we met in 🥧 ff.129.2 Pi, FigureAI and 1 hour of “logistics”?) has shown off a use case: cooking a perfect grilled cheese sandwich.

🌎 ff.148.3 A world out there?

💭 ff.148.4 The whole world inside

Substack ↗

But do we really need a different model? Aren't LLMs - perhaps - already able to see a 3D world, maybe by “dreaming” it?

QED? GPT-5.4 recreates Monica's apartment from Friends in two prompts.

💭 ff.148.4 The whole world inside
💭 ff.148.4 The whole world inside
Sources & links (1)

🖊️ ff.148.5 I'll sign for that

Substack ↗

After all this talk of robots, back to humans. And to books. And so, in a perfect loop, to Amazon.

The giant, born precisely to replace bookshops (and always big on automation, 🧝 ff.6.2 Amazon's elves, and stores without staff), seems to have overlooked the human side of things: a bookshop in Alabama sells books signed by the authors who wrote them.

At the same price as the regular ones, mind you, but with a human touch.

They're so popular that the shop has struck a deal with the local hotel to put up buyers coming from far away. Couriers in reverse.

Rachel Karten points out that this same “human” touch is what makes us love Twitch and live-streaming:

Live-streaming offers a refuge from the growing glut of AI-generated content in our feeds.

In a social media landscape where the difference between real and artificial has become almost imperceptible, the unmistakable humanity of real-time video is a refreshing draw.

Rachel Karten on Substack

In short, human imperfection is making a comeback. The same old imperfection filled with gold, even in the cracks of a basketball court.

🖊️ ff.148.5 I'll sign for that

Related issues

Newsletter

Get new issues of Futuro Fortissimo

A free newsletter by Michele Merelli (Bergamo) on technology, AI, nature and society. Written in Italian, delivered by Substack.

📩 Subscribe for free ↗
All results in the archive →Press / to search anywhere