• 5 min read## What is the future of AI?
What is the future of AI? I don’t know! But let’s look at what thought leaders say.
Recently, Richard Socher gave an interesting interview about the future of AI. His perspective is particularly valuable since he was the first person who applied Deep Learning to natural language processing (in other words, we owe him the current AI revolution). He is now the CEO of You.com, a startup that has developed what many consider one of the best LLM tools in the market.
You can watch the entire 1.5-hour interview here:
<a href="https://www.youtube.com/watch?v=YK2Be7LcaK8" target="_blank"></a>
The interview offers fascinating insights for anyone interested in AI, working in the field, or developing products using AI. Let me share what I found most interesting.
To understand the background: You.com is an AI-powered search engine (recently re-branded as a productivity engine). It works similarly to Perplexity.ai, which you might be familiar with — both are AI-powered search engines that provide conversational answers while citing their sources.
Here’s a summary of the key points from the interview:
### On the Future of Work
AI will transform the way we work, similar to how the steam engine did. Here’s a thought-provoking perspective: 150 years ago, 90% of people were farmers, and most human resources were used in agriculture. Today, only 5% of people work for food production, with the remaining human resources being used to be productive in other domains. This led to a huge productivity gain for society. Back then, it was impossible for those people to imagine such a future.
Just like automation in farming freed up people to create entirely new industries, AI will likely free knowledge workers from repetitive tasks. What will they do instead? That’s the exciting part - we probably can’t even imagine the new types of work that will emerge!
Key observations:
* Currently, we observe that below-average performers are those who benefit the most from these technologies, while these machines are trained with data from top-performers!
* AI will bring people to work at a new abstraction level. Architecting and orchestrating will be the new tasks of humans.
* Use cases that are coming: personal tutors, personal health coach, personal assistants. Why? Currently wealthy people benefit from it; AI will democratize it.
* As in other hypes, we see inflated expectations. But this time, we passed the way of no return: LLMs are very useful and are here to stay. As an analogy we can look at self-driving cars: people promised self-driving cars within short time because it worked in very simple use cases — highways. Same for LLMs, they work very easily on simple use cases, so people assume it will work for all use cases — this creates the hype.
* Personal assistants as agents will transform the internet. They will act for us on the internet (like searching information or booking a flight). This means, all websites that monetize based on traffic will need to find a new business model, since people will not go to the internet on their own — their agent will do — and the agent will not care about advertising. This means we will need to find new commercial models to finance the internet.
### On the Future of Productivity Tools
The integration of AI into productivity tools seems to be taking two paths:
1. Improving existing tools (think of Microsoft Office). Only large companies will have enough data for this training
2. Creating entirely new categories of software (for example a productivity engine — and many more which still need to be discovered)
For existing tools, having control over the workflows and data will be key. This is why the companies owning these tools will likely have an advantage — they can integrate AI directly into their tools, making it difficult for competitors to catch up. In other words, it might be risky to create a product that augments Microsoft Office, as Microsoft themselves might be much better positioned.
Where new players might shine is in the new categories of tools that will emerge.
### How AI-Productivity Tools Can Be Improved
You.com’s experience as an early creator of a conversational search engine has revealed interesting patterns in how people use AI:
* Simple questions (e.g., “what is the age of the president of France?”) are still better handled by traditional search engines. These represent a huge part of the queries.
* Complex analyses requiring extensive internet search and information synthesis are where LLMs can really add value
* Their approach involves many specialized tools in the back-end that activate depending on the question type. For instance, writing a poem doesn’t require web searching, while finding the latest paper on LLMs does. This system, where one model guides requests to the right tool, is what’s known as an agent.
* Focus on quality over costs for the moment, since reliability is not yet there and costs go down and are getting commoditized.
### About AI-Agents
Agents are currently a “hot topic,” but they come with interesting challenges. Take a simple example: finding the age of the president of France requires multiple steps:
1. Generate a web-query “President France”
2. Extract the name “Emanuel Macron” from results
3. Generate another query for “Emanuel Macron age”
4. Extract the age from results
5. Generate a final answer
Each step needs to be accurate. Even with 95% accuracy per step (which is quite good), a 15-step process ends up with overall accuracy below 50% (0.9¹⁵). For 50 steps it will almost always fail. This significant drop in accuracy could frustrate users, suggesting that agent systems won’t be the solution to every problem.
### Your Thoughts?
I hope you found these insights interesting. I’d be curious to hear your thoughts — do you agree with these perspectives? Is there anything else you’d like to know about LLMs and their future impact?
