[Premium] LLM OS, Turning any LLM into o1, & More

TLDR;

Technology:

😍 Nous’ Exciting Reasoning API

😎 ChatGPT’s MacOS Expansion

🥳 Gemini Takes the Top Spot

Markets:

☹️ ChatGPT’s First Official Victim

😩 NVIDIA’s Blackwell GPUs Could Turn Out a Nightmare

Product:

🤯 Stripe Embraces LLMs

😱 Pixtral Goes Massive

🛍️ Perplexity Goes Shopping

FRONTIER RESEARCH

Nous’ Exciting Reasoning API

Nous Research is quickly escalating in my ‘favorite AI labs’ ranking. Not only are they putting meaningful effort into creating powerful, uncensored LLMs (some frontier products like Claude are pretty obnoxious to use due to the so-called ‘safety measures’), but now they have released a superb new reasoning API called Forge.

The idea is pretty simple. If you think of OpenAI’s o1 models, these are nothing but a GPT-4o-type LLM wrapped around what they named a “Reinforcement Learning algorithm.” That sounds extremely cool to say, but what they are referring to is nothing but a Monte Carlo Tree Search (MCTS) type algorithm that allows this LLM to explore the space of possible solutions.

The model can then generate multiple predictions for the task, explore resolution paths, and backtrack when reaching a dead end until it reaches a ‘budget’ threshold or converges on an answer the model deems worth outputting. Now, with Forge, you can provide these capabilities to other LLMs.

TheWhiteBox’s takeaway:

I think this release is very important. Now, Nous Research allows you to take the following LLMs: Hermes 3, a Llama 3.1 405B fine-tune by them, Claude Sonnet 3.5, Gemini, and GPT 4 and turn them into o1-type models automatically.

Of course, that means you will soon be able to run your open-source small models and turn them into o1-type models (Large Reasoner Models) without dealing with the huge complexities of assembling the search algorithm.

And does this work? Boy, it does.

Using Hermes 70B, a fine-tune of Llama 3.1 70B (not even fully state-of-the-art), combined with Forge takes its performance in AIME, one of two tests used to evaluate whether someone is ready for the Math Olympiads, to 80%, even surpassing the unreleased o1 (superior to o1-preview and o1-mini).

If you thought Large Reasoner Models (LRMs) like o1 were uncharted territory for open-source, think again.

FUTURE

ChatGPT for MacOS Expansion

OpenAI has announced a new step in its integration with the Mac operating system. It now allows integration with other installed developer applications, such as Visual Studio Code or XCode.

As shown in the video, ChatGPT can now access the context of your code, allowing it to synthesize code that is immediately applicable to the code in your repository. You just have to copy and paste the code snippets (a feature they claimed would soon be automatic, even allowing you to code with your voice).

TheWhiteBox’s takeaway:

Although this is a cool feature, there’s much more to it than meets the eye. To me, it’s a clear insight into OpenAI’s plans regarding ChatGPT: they want to turn LLMs into fully-fledged operating systems.

In other words, they envision a future where all applications, software, and business and…

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