Since its existence, humanity has always been searching for a scalable energy source. We have yet to find one. The closest we came was code. We found ways to scale it significantly. This was only possible through worldwide collaboration. GitHub was facilitating that and thereby altered the course of humanity. It will continue to do so as code will play a key role in solving humanity's future problems.
ChatGPT is the go-to solution for prototyping. If chatGPT cannot solve it, the tech likely needs to arrive. However, to bring features into production, the market will be divided into specialized LLMs catering to specific verticals/use cases/budgets.
For AI, specialization will always beat generalization. If it doesn't for you, it means that you are in the stage where the Kaplan et al. (2020) data law effects still apply, and you have yet to reach the tipping point of ever-increasing saturation. To cross the tipping point, only a hyper-optimized feature space gets you out of the local extrema. Thus, eventually, feature-space optimization via specialization is the way to go.
Who best classifies and orchestrates hyper-optimized models will win the LLM wars. Ultimately, it's all about engineering, infrastructure and data strategies.
Disclaimer: We are on the waitlist and haven't used it actively...But if Claude 2 can deliver on the promised features, it will be a game-changer: 4X cheaper than GPT-4-32k is a big selling point as we are running agent strategies in production, which costs a lot. The better performance across multiple benchmarks, including math, coding, and reasoning, could be decisive for our use case: writing, maintaining, and healing end-to-end tests for UI. From our experience, achieving SOTA and tractable, consistent performance simultaneously in these functional areas is complex and could be a unique achievement. Code that does what it is supposed to and compiles is significantly more complex than letting the LLM hallucinate more creative answers. Compilers don't like creativity ;)
But the most important innovation is increasing the context length. So far, we always had to work on stripped-down versions of HTML, with 100k and 200k context that would change and should yield significantly better results.
Congrats on your launch; keep on innovating! We can't wait to try it out!
Just discovered Alltum and the angle superhuman + asana/X makes a lot of sense. In my experience, project managers often work on asynchronous information and I experienced firsthand how it can kill customer perception and thus the success of a project. Having a single and synchronous source of truth makes all the difference!
What's great
project management (4)asynchronous information management (1)
I wish I had this a few years ago when I was still a researcher in machine learning. It would have saved me so much time. This product just had to happen!
