Historically, we’ve always had to turn data into numbers to work with them. ChartsExcelMachine learning
Numerical representation
is an abstraction.
The very fact that we now have intelligence at the cost of nothing unlock new paradigms. We are no longer bound by the limitations of having to work with numbers. We can work with data in their much more optimal, original forms.
And when paired with burnt in LLM chips, we would have solved not just the cost issue, but also the latency issue.
Every single known AI expert in the space right now is a machine learning expert who reaches for the machine learning solution. As a result, none of them have truly original thoughts outside of machine learning solutions.
We do not need to focus on training LLMs anymore. The LLMs we have today are already enough to automate away 100% of needed human labor.
The key is a well designed system around a well designed knowledge base:
- self improving
- true multimodal
- works with any knowledge source
- infinitely scaling
And it is the only way to 100% automation, training models on its own will never be enough.
The gap is not in the model itself, it’s in the knowledge that it has. We can’t just keep retraining models every year to give it up to date information.
Why not just have a knowledge layer outside of the LLM that is so good at storing exactly the right information?
By itself, the model is never going to know what your personality is like. What you had for lunch yesterday. How it should interact with you based on past interactions.
And every single attempt at tackling AI memory is far too basic (besides my approach), it doesn’t actually solve the issue.
There is also a generational opportunity right now to take trillions in market cap away from Nvidia by developing burnt in LLM chips.
than LLMs run on GPUs
Here is a company doing it already: taalas.com
Here is their demo: chatjimmy.ai
By itself, it is understandable why companies haven’t invested in this direction yet. However, if you pair it with my first point, which is that LLMs are already good enough, and that the issue is in the knowledge layer, suddenly it becomes a ridiculously good play.