[Premium] I Owe You an Apology, Canva's Nasty Surprise, and Giga Data Centers

TLDR;

Technology:

🤖 1X Neo Lives with You… Kind of

🙏🏾 I Owe You an Apology

🧐 A SuperHuman AI Forecaster

Markets:

😱 A 10GW Data Center in The Horizon?

😎 In AI, The Future is For the Generalist

Product:

🎇 Replit’s AI Agent is Amazing

😣 Canva Gives Customers a Nasty Surprise

🤨 The New iPhone’s AI Features Disclosed

ROBOTICS

1X Neo Lives With You… Kind Of

1X has released a video with the S3 YouTube channel that shows the robot ‘living’ with the YouTuber and doing exciting stuff like pouring coffee and other chores.

TheWhiteBox’s take:

While the video is impressive, one specific quote caught my interest. On delivery, Neo won’t be autonomous but teleoperated.

In other words, it seems that the interactions are unique for each household, thus requiring extensive data gathering during an unspecified time until they have enough data to fine-tune the model to your home and the chores it has to perform.

Now, this already feels werid to me. Am I about to have a remote human doing and seeing my private life for an undetermined amount of time? It's not my dream introduction to chore robots, that’s for sure.

If this is a short time, that might work, but deep learning is notoriously data inefficient, so this teleoperated phase could be longer than desired. Another issue I have with these models is that they will be inevitably connected to the Internet, as the base intelligence, the LLM, won’t be ‘on-device’ (it’s a GPT-4-level model, which means TeraBytes of RAM required, or at least in the multiple hundreds; no way that’s stored physically in the actual robot).

As you probably know, a door to the Internet is a cybersecurity risk, so what prevents someone from hacking my robot and teleoperating it… in my home? Hopefully, these questions will be addressed, right? Right?

THE DARK SIDE OF RESEARCH

I Owe You an Apology

As I want to be someone you can trust to understand and navigate the world of AI, I must acknowledge when I make bad claims. And this time, I’ve been fooled.

On Sunday, I enthusiastically presented Reflection 70B, a model that seemed too good to be true. And, well, it was indeed too good to be true.

Although it has not been confirmed, the allegations that the model is nothing but a Claude 3.5 Sonnet wrapper fine-tuned on reflection-tuning (that’s legit), while claiming to be a Llama 3.0 model, are serious.

As you can see below, the model has some back-door modification that prevents it from generating the world ‘Claude.’

In other words, it seems that the creators of Reflection 70B are desperate to prevent the model from generating strings of text, acknowledging that it’s, in fact, Claude (in another tweet, the chatbot acknowledged Anthropic had trained it).

"Reflection API" is a sonnet 3.5 wrapper with prompt. And they are currently disguising it by filtering out the string 'claude'.reddit.com/r/LocalLLaMA/c… — Joseph (@RealJosephus) 10:10 PM • Sep 8, 2024

Again, we need full…

Enjoyed the introduction? Subscribe or sign in to continue with this edition.