The initial GenAI use cases have been interesting but are only a scratch on the surface of what is come. The biggest problem to date, has been people's focus on one shot perfection. The reality is that GenAI is not perfect, it is not going to be perfect, and it will never be perfect. The key is to understand the limitations and work with the limitations.
Limitations are:
- context length is limited, even with 1M+ tokens
- context has a cliff at around 64-128k tokens
- can struggle with ambiguous queries and make assumptions to fill in gaps
- lacks true understanding and reasoning abilities
- very bad at maths
- is poor at novel tasks
- lacks long-term memory
The next generation of GenAI powered tools will be embedded into our systems focused on micro tasks at high speed and with high accuracy in mix modal environments combining logic, multiple specialised agents, and on-demand micro programs/macros.
To overcome the limitations, we will see a combination of:
- handing off tasks to specialised agents / sub agents
- using external tools and APIs to augment capabilities
- implementing multi-step reasoning and planning
- using external memory stores
- employing techniques to quash the context length
- implementing feedback loops for continuous improvement
- leveraging user input to refine outputs
- providing tools to the AI for calculations
I am excited as we transition from wave1 to wave2 as I have built early prototypes of the tools to come.
I witness firsthand the Web2.0 transition and I know this is the most exciting frontier as people start appreciating the limitations and starting working with the limitations, not against them.