🍓OpenAI drops o1, a new reasoning model
China's startup ecosystem is on the verge of collapse, Mistral goes multimodal.
Listen to a conversation version of this edition, made by Google’s latest NotebookLM.
OpenAI dropped a preview of o1 yesterday, a reasoning model designed to reason through complex tasks and solve harder problems than previous models in science, coding, and math.
The early feedback has been fascinating, as folks try to dissect the definition of “reasoning”, and what use-cases the o1 excels in, or what it is less good at. Why this is interesting: previously AI models worked more like memorization (trained on vast amounts of data), but o1 applies a new methodology closer to chain-of-thought, which means it’s able break a problem down and work through it step-by-step.
Right off the bat, it feels noticeably faster. The latency between inference and compute seems at lightspeed, and the user experience design had a part to play in that: now you can watch the model “think through” the steps to solve the question.
I threw it a notoriously tough Math Olympiad question, and it solved it after “thinking” for 97 seconds. This is a living thread on Twitter for other similar challenges. It seems to perform less well at fuzzy human-like reasoning, but excels at deterministic challenges.
In the coming months, we’ll likely see breakthroughs in code agents, and applications with heavy mathematical lift, like cell sequencing, algorithms, and game engines.
🍓
The Chinese startup ecosystem seems to have collapsed.
A chart from the Financial Times showed the number of new startups falling from 50,000 in 2018 to just over 1,000 in 2023. It’s not surprising when you realize the stakes are extreme: if your startup fails to IPO, the VCs and government will take steps to bankrupt the founder and seize their house and assets.
Venture is inherently risky and innovation ecosystems like Silicon Valley have a unique “gold rush” ecosystem of forward-thinking, risk-taking builders and funders that is hard to replicate.
Have a great weekend,
Tara
The Latest
Google releases NotebookLM, a tool to turn written content into lively podcasts: Imagine listening to a lively discussion about your latest research project, summarized and analyzed by two AI hosts. For fun, I generated a conversation from today’s edition of The Strange Review:
Mistral releases its first-ever multimodal AI model: Mistral AI is finally venturing into the multimodal arena. The French AI startup taking on the likes of OpenAI and Anthropic released Pixtral 12B, its first ever multimodal model with both language and vision processing capabilities baked in.
Meet Verse, an AI-powered creative app for Gen Z: The iOS app allows users to design and publish multimedia content on an interactive canvas with the help of an AI assistant. Users can create a mini website, called a Verse, for things like moodboards, greeting cards, invitations, storefronts, fan pages, blogs, and more
Klarna Cuts 50% of Workforce, Ends Partnerships with Salesforce and Workday Amid Generative AI Overhaul This move is driven by a strategic overhaul focused on integrating generative AI into their operations.