AI's Godmother New Start-up, OpenAI Challengers, Google's Struggles, & More

TLDR;

Technology:

🤔 o1 models Faces its First Rival

🫣 The Gemini Struggles

🫡 Cognition Analyzes o1 models

🤖 Google makes strides with ALOHA

Markets:

💰 The “Godmother of AI” Comes out of Stealth

🇺🇸 Big Tech’s CEOs Summoned at the White House

Product:

🎥 Vchitect, Breathtaking AI Video Generation

🚗 An Interesting Review on Semi-autonomous Driving

SYSTEM 2 MODELS

o1 Faces its First Rival

Just a few days after OpenAI’s grand release of its new model family, o1, SambaNova, an AI inference company (a company that provides open-source LLMs for you to use/fine-tune to your needs) and a company that claims to have the highest token throughput for Llama 3.1 models (thanks to its RDU hardware), has announced a ‘system 2’ demo where a Llama 3.1 405B model is giving a ‘thinking budget’ and is capable of replicating ‘similar’ thought processes to those of o1 models.

You can try the demo for yourself here.

TheWhiteBox’s take:

This is actually a really cool project because we can actually see (and modify) the system message they send to the model, where we observe that they naturally induce the model into a chain-of-thought process.

However, it’s not even close to the level that o1 offers. But why? Well, while this solution imitates self-reflection, one of the core features of o1 models, it’s clearly missing two things:

Better multi-step reasoning capabilities. o1 models were extensively trained using a PRM model (process-supervised reward model), following OpenAI’s let’s verify step-by-step research, meaning they are instinctively better at taking the multi-step approach than previous models. As you can observe below, the reward model scores the final result and the intermediate steps, naturally inducing the generator to be more precise throughout the process.

The Llama 3.1 models were indeed trained using PRMs too, but not at the scale OpenAI did.

Source: OpenAI

It doesn’t perform active search. Instead of coupling reasoning and output tokens as OpenAI did and as I describe in this Notion piece, where each new output is conditioned on the user’s input plus the previous generation, here everything is outputted in ‘System 1’ style, preventing the model from actively exploring different solutions.

In a nutshell, this model is a neat way to force the model to output more tokens, akin to generating ‘more thought,’ but it completely misses the ‘reinforcement learning part’ that allows o1 models to think for dozens of seconds or, maybe in the future, even for weeks. Thus, it’s really not that different from what we had earlier to o1, besides a counting parameter that allows you to limit the number of steps the model can think through. We’ll have to wait for Llama 4, Claude 4, or Gemini 2.0 for truly comparable models.

Hopefully, this clarifies how hype-driven some publications are.

LLM MARKET SHARE

The Gemini Struggles

Source: The Information

According to a report by Retool on the state of Generative AI, Google seems to be struggling with Gemini adoption, its Multimodal Large Language Model, as a distant sixth option…

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