The Problem with Search
The way that we search for information is inefficient, noisy and is not aligned with our objectives. Can we use AI to address these problems and change the way that we search?
The web gave us new ways to search and access information; First through website directories and static webpages and later through online forums and message boards. The next innovation brought search engines which indexed web pages allowing keyword search. Web 2.0 followed, and social media platforms allowed creators to create and upload their own content. At this point search had became passive and driven by recommendation and engagement algorithms.
These types of search algorithms have a number of problems - they are inefficient, noisy and not aligned with our objectives. In addition current search is explicit which means that we need to know what we are searching for and then set out to find it.
How do we search for solutions today?
Actively searching through web & document search engines
The user sets out to enter keyword search queries, spends time reading through results and determines which are the best quality relevant results.
Passively consuming streams of information through social media feeds / streaming platforms
Passive news feeds / notifications / photo streams / videos etc. are shown to the user passively to keep them engaged on a social media or content platform and monetise their attention in order to sell advertising. Here the user’s attention is the product.
This type of passive information stream is generated by social media platforms, online advertising, streaming sites, product recommendations and sponsored search results.
The problem with today’s search solutions
Explicit
The user needs to know that they need to search for a solution, they need to know what they are searching for and finally they need to set out to find it and verify the results.
Inefficient
The user needs to browse through, evaluate and verify the results.
There might be multiple search engines / data silos that need to be searched to find the correct information. (eg. Web search and mailbox search)
The information required may form part of a larger collection of information requiring extra work to locate the relevant details. (eg. A large webpage or document).
Non-Collaborative
Other users may be searching for the same solution and may have already found relevant information.
The information required might be in someone else's personal mailbox or IMs. It might be in a different department’s workflow (eg. In the case inbox of a technical support team member).
If one person has the solution to someone else’s search, then collaborative search should allow for that information to be shared (if it is mutually beneficial).
Collaborative search should scale to large numbers of teams or users, and should only surface information that is not private or restricted.
Non-Aligned
A search ranking algorithm or recommendation algorithm that is optimised to keep a user engaged, buy a product or click on a sponsored link is not aligned with the user.
The information returned by these biased algorithms will never fully be in the user’s best interests.
A user-aligned information search should be optimised to provide the user with the most relevant information at the right time without any compromises.
Noisy
Social media feeds and web search results are interspersed with adverts and sponsored articles or links.
This creates noise, making it more difficult to only consume information that is relevant to the user.
This comes at the expense of focus and causes distraction.
Not context aware
Current search tools have limited knowledge of our current context. This might be limited to our location and a user profile.
They are not aware of our current objectives and projects. They do not know what the user is currently looking at or what they are currently doing.
What should search look like?
How do we autonomously surface the most salient information at the right time and deliver it to the right people based on their context and requirements?
How do we retrieve this information from multiple data silos including ones that are not on the public web while respecting privacy.
How do we reduce noise and make collaboration extremely efficient?
An example of an efficient, aligned and collaborative search:
When driving, a satnav / maps application will only raise alerts and important information depending on the driver’s speed, location, direction and destination. (Context aware and not noisy)
The driver does not search for this information before driving down a road. It is delivered implicitly when needed. (Efficient and Implicit)
The driver is not subjected to an engagement algorithm attempting to display information with an objective of maximising engagement and monetisation. The application is aligned with the driver’s current objectives (their destination & safety) and raises notifications accordingly. (User Aligned)
Information is crowdsourced from other drivers and users of the system implicitly while respecting privacy. For example if multiple drivers are moving slowly - then the app will deduce that there is a slow moving traffic queue and raise alerts to any other drivers or reroute them. (Collaborative)
Dialogue as Search
The first disruption to our incumbent search tools and information platforms is taking the form of dialogue.
Large Language Models (LLMs) have enabled useful chatbots. It is clear that in many use cases these chatbots allow us to search for information more efficiently and effectively using a dialogue interface instead of a web search engine.
The LLMs are able to engage in dialogue and return information from an extremely large body of text sources; including forum posts, social media posts, web pages and books. Vision models (LVMs) are also able to return information from images and videos. These chatbots can also search the web and use APIs on your behalf to find and collate information that is outside of their training data.
This makes searching via these Chatbots much more efficient and also less noisy.
One to Many, Many to One & Many to Many Conversations
We are also getting an idea on how chatbots enable collaborative search by allowing a user to ‘chat’ with a large amount of data (or even large number of people).
LLMs are also starting to unlock new powerful ways to allow a single person to converse with the collective knowledge of thousands of other people simultaneously.
Inject a book into the prompt of an LLM, and you can ‘chat' with the knowledge that that particular author has written (one to one).
Inject all posts in a web forum on a particular topic, and you can ‘chat’ with the collective knowledge of the thousands of authors that posted to that forum (one to many).
Inject the transcriptions of a collection of YouTube videos or podcasts, and you can ‘chat’ with the collective contributions of all speakers within those videos (one to many).
It doesn’t matter where these contributors were geographically, what language they used, what time zone they were in, or at which point in time they contributed. The knowledge is available to you in your preferred language and is available at any time.
I use the word ‘chat' loosely, because you are not conversing with the actual persons but with the collective knowledge that they had contributed up to that point in time.
Finally, we can easily envision technology, where thousands of users can chat in real time with a chatbot, and LLMs (or a future derivative model) are able to distil this collective realtime knowledge into a relevant conversation with a single user (many to one).
The responses from this user, are then fed back into the collective knowledge, which may in turn be used in conversations between chatbots and other users (many to many).
Are Search Agents next?
The real disruption to search, will happen when users no longer need to actively search / chat, and important information is surfaced implicitly in the background.
This could take the form of search AI agents. Perhaps we would each have a personal search agent, or have access to multiple specialised search agents. (eg. a specialised health information agent, a social media search agent and a work information search agent).
We can view information as a huge distributed network of nodes, each node containing specific information. The nodes of this network are made up of people, devices, sensors, databases, information stores and eventually other agents.
Through the use of APIs and Search Tools, our agents will then interact with and traverse through this network to locate the solutions we require.
Therefore to solve the problems with current search, next generation search agents should be able to:
Reduce noise in the network
Our agents should only surface information that is relevant to our current objectives at the right time. Currently we must read all emails in our inboxes, join all meetings and endlessly scroll newsfeeds to ensure that we do not miss relevant information. Instead our agents should be engaged to filter and only return required information. The rest can be discarded as noise.
Reduce friction in the network
Different languages, timezones, different periods in time create friction which makes it more difficult for us to find relevant information and communicate. Again our agents, should be able to retrieve information written in different languages and at different points in time, collate it into our preferred format and present it to us when required.
Reduce latency in the network
Currently it takes time for published solutions to be discovered by persons looking to solve similar problems. Collaborative search agents should be designed to reduce this latency, by sharing information with all relevant persons as soon as it is published.
Ensure that information is always contextually relevant
Information should only be retrieved and surfaced if it is important for the user to know at that time. These notifications should be time sensitive based on current objectives, actions and context. Our agents should prompt us using our devices or wearables on a need-to-know basis, instead of us having to scroll through endless data streams to keep up to date.
Automatically share data that is relevant between people
Similar to crowdsourced information in the driving app example, non-private information gained by particular nodes, should be discoverable by our agents if we would benefit from knowing it. Perhaps different agents can collaborate and negotiate with each other on our behalf to procure important information.
We should look forward to a time where we no longer need to spend so much time searching the web and scrolling through social media.
Perhaps in the future our smartphones will no longer command our attention. Instead we will wear a combination of wearables such as a smartwatch, normal glasses (not AR/VR) with cameras and an ear pod for audio. The multiple sensors in our wearables will monitor our context, our health and our current activities. They will ‘hear’ what we are hearing and ‘see’ what we are seeing through their cameras and microphones.
Our personal search agents will then use our context to decide when they need search for information and raise any important information in realtime. This information can then be fed to us through audio (synthetic speech) or by displaying text, images and video on our smartphone’s screen.
The experience would be similar to the analogy of the driver and the car navigation app, which ensures that we are notified implicitly about any time sensitive information we need to know as we drive down the street.
In conclusion, as AI algorithms improve they will unlock the development of personal agents. We must use this opportunity to design agents that are truly aligned with our objectives, and are able to search on our behalf in the background. If built properly, perhaps we will move away from addictive algorithms designed to keep us looking at our screens and engaged on social media.
Instead our smartphone disappears into our pockets, and we refocus our attention to the real world - augmented with all the information we need to know, when we need to know it, spoken into our ears.