Redstone, a VC, published an email they received around a year ago.
A lost billion-dollar opportunity for Redstone.
And hopefully, someone at DeepSeek has taken a copywriting course in the meantime😁

Had the pleasure to meet Amjad, CEO of Replit.

Replit is one of the big names in the current AI agents revolution.
It allows users to write, run, and collaborate on code in multiple directly from their browser, without the need for complex setups or installations.

The market is down because DeepSeek, a Chinese company, released a model (R1) comparable to OpenAI's O1, allegedly using less GPUs and spending only 5-6 million dollars (whereas it was previously believed that such models would cost tens or even hundreds of millions of dollars).

This is a knocker-up. Someone once hired to tap on windows to wake you up for work.
Most jobs will either change or disappear in the next years because of AI. And it's fine.

As a society, we should worry about how to help people adapt.
Re-educate them.

AI success isn’t about having all the answers, it’s about embracing experimentation.

✅ Define metrics that truly matter. Focus on metrics aligned with your goals, and let them guide your experiments.

My current goto stack for quick AI development.

Models:
Claude 3.5 Sonnet @AnthropicAI
- Strong at both writing and code generation
- Default in Cursor for writing code

OpenAI o1 @OpenAI
- More complex and nuanced requests

If you want to build a new AI system, remember the following 3 points.

First, are LLMs Really Necessary?
• Complex language understanding? Sure, go for it.

A lot of advice on how to approach 2025 in AI...

The most important trend will be to bring your 2-week PoC into production.

People like novelty (and focus on it). But you'll have to focus on what didn't change in the last 20 years.

In particular data quality and evaluations.

Sora from OpenAI still lacks fundamental understanding of physics and how the world works.

Still impressive, but they overhyped it quite a lot 🤌

Does somebody have strong use cases for the O1 reasoning models?

I find myself using it in chat for specific questions but I still didn't encounter an "API use case" where I felt GPT-4 is noy enough, O1 is.

The narrative of token price going down misses the big picture.
Agentic applications are very tokens angry, and the average number of tokens required for accomplishing an action is growing more than the tokens' price reduction.

Ok, just got here.

Who are the most interesting folks in GenAI, Data, ML?

Any conversation happening here and not on the bird app?