With total cumulative financing exceeding 200 million US dollars, Today AI will understand you better and better starting from today.

Recently, Today AI, the developer of Personal AI Assistant, announced that it has completed cumulative financing of over 200 million US dollars, with investors including IDG Capital, Tencent, 5Y Capital, Ant Group, HSG, Transsion Holdings, Tongge Venture Capital and Hanyuan Asset of Shanghai Jiao Tong University.

This financing comes at a time when the personal AI assistant track is continuing to heat up. Over the past year, AI has evolved from simple question-and-answer interactions to executing tasks per instructions, with functions such as searching emails, accessing calendars, and operating computers gradually becoming its basic capabilities. In just the past month, personal AI assistants have become another highly sought-after application direction overseas following AI Coding, with tech giants entering the market one after another and capital accelerating its deployment.

The trend aims to solve a long-standing unsolved proposition: Whether it is question-and-answer interactions or task execution per instructions, the starting point is that users first state their needs and background information. Today AI wants to take a different approach: it remembers the goals and recent status shared by users, and proactively proposes the next step at the appropriate time.

In China's AI industry, Today AI chooses to build an independent product centering on personal AI assistants, and regards this as the core positioning of the company. Today AI is one of the earliest entrepreneurial teams in China to explore this direction. The team is committed to building a personalized proactive AI system, providing every user with a more understanding and more proactive AI assistant.

"The capability is here, but the bond between AI and users has not been established yet", the Today AI team once explained their original intention in this way. 

AI can do more and more things, but it still takes a lot of effort to make good use of it: you have to figure out what to ask first, then explain the context again, and most of the time you have to start all over again next time. Today AI aims to help users save this part of energy, and offer help proactively at the right time.

To achieve this, continuous adaptation during daily use is required. That's why Today AI chose to deliver the product to users first: it launched the public beta overseas in September, and the domestic version went online on September 24, until this time it officially disclosed the financing. Previously, a large number of users have used it to plan work, arrange travels, organize fitness plans and manage schedules, and the product has been tested in these real life scenarios.

Remember your past, and keep up with your present

What kind of help a person needs often depends on the situation he is in at the moment.

When traveling, people may not start to consider where to go the next day until night; after the vacation ends, the plan to resume exercise also needs to be adjusted according to the recent status. Many demands emerge gradually as life changes, and it is difficult to turn them into a clear AI instruction in advance.

This puts higher requirements on personal AI assistants: It should not only remember the experiences and preferences shared by users, but also keep up with their recent arrangements and status. By connecting the two, the AI can bring up matters worthy of attention in time.

A user who lives in Tibet once asked Today AI to sort out materials such as travel photos. Not long ago, he was considering the itinerary for the next day. That night, Today AI proactively recommended a list of niche temples that local Tibetan residents often visit in its briefing, and also attached the departure time of shuttle buses, opening hours and the best light periods suitable for shooting.

Screenshot from user @ahazal, used with authorization

This recommendation surprised him a lot, because these temples perfectly matched his interests. The photos and travel experiences sorted out last time were put to good use on this trip.

According to the Today AI team, what the product mainly remembers is the basic information actively shared by users, as well as the preferences manifested in past interactions. These memories are also managed by users. Users can view the personal information and preferences recorded by Today AI, modify inaccurate parts, and delete content they do not want to retain.

But to offer help, it also needs to combine the user's current situation.

A user who has practiced boxing for a long time usually analyzes sports records with Today AI, and also connects her health data to the product. On the first day back to work after the holiday, Today AI noticed that her exercise volume during the National Day holiday decreased significantly, so it reminded her in the morning briefing: do not rush to increase the intensity when resuming training.

This reminder is not complicated, but it takes into account the change of the user's recent life status. Even for the same person, the appropriate suggestions may vary with time and context.

But proactive service puts higher test on the judgment of AI: What information is worth remembering for a long time, and what needs to be updated in time? When is the right time to remind, and what matters do not need to disturb users?

These tiny judgments determine whether the AI can offer help at the right time, and also affect how much daily burden it can reduce for users.

Reduce the delegation cost, so that more things are worthy of being handed over to AI

For many tasks delegated to AI, you still need to repeatedly explain the background and confirm the progress, and the energy you spend may be more than doing it yourself. The Today AI team calls this cost the "delegation cost", and hopes to reduce it, so that users can delegate tasks to AI more naturally.

Schedule arrangement is a typical example, plans often cannot keep up with changes. A user who is preparing for the CET-6 exam usually uses Today AI to arrange exam preparation plans and daily schedules.

Once, she received an admission email for volunteer work, knowing that she had to attend a temporary training the next morning. Combining her previous exam preparation arrangement and the newly added schedule, Today AI suggested that the original word memorization time be adjusted to the evening.

She does not need to restate her exam preparation arrangement, nor does she need to rearrange the schedule by herself, so the delegation cost is reduced.

This is also the reason why Today AI organizes information around the same person: arrangements in study and life are often interrelated, and a temporary change may disrupt the original plan.

This idea is also reflected in the interaction design. Today AI adopts a continuous single-session design, users can successively explain different tasks in the same conversation, and specific tasks are carried out in parallel in the background, and the progress can be viewed separately. Users do not need to switch sessions frequently, and they can continue to talk about the things they discussed before.

In the words of the Today AI team: "You hand over the tasks, then go on with your life. When you come back, it is still there."

The proactivity of personal AI assistant also requires clear execution boundaries. In the CET-6 preparation example, Today AI only gives adjustment suggestions, and it is up to the user to decide whether to modify it or not. For operations involving payment and important decisions, the product will also confirm with the user first before executing related operations.

Reducing the delegation cost is to let more people make good use of AI, and leave energy to more important things. To achieve this, it is far from enough for AI to help once.

Being helpful in the long run relies on unseen efforts

To make such help happen continuously, the AI needs to accurately record the user's situation, provide timely reminders, and complete tasks smoothly.

The Today AI team is continuing to invest in this. In terms of memory management, it reduces the interference of outdated information on judgment through the Dreaming mechanism, compression and update; in terms of proactive reminders, it adjusts the reminder strategy according to users' reading and response to pushes, so as to reduce invalid disturbances.

At the same time, the team is also polishing the system infrastructure, task queue and cross-device communication, to ensure that long-running tasks will not get stuck or lost, and the progress can be checked at any time.

In the view of Today AI, personal AI assistants are supported by excellent products, appropriate models and solid infrastructure together. It is not difficult to make an amazing demonstration, but the difficulty lies in being helpful every day, and most of these efforts are not easy to be noticed by the outside world.

For personal AI assistants that need to understand users in the long run, there is another question worth answering: The longer you use it, will the AI really understand you better?

One-time AI tasks are usually evaluated by the delivery results, but personal AI assistants also need to examine the changes brought by time: as historical information continues to accumulate, whether the suggestions provided next time are more appropriate, and whether it can more accurately discover demands and complete tasks?

To this end, Today AI has built a set of simulated user evaluation system: the simulated users have continuous life experiences, the team compares the performance of AI in three scenarios: no historical information, only personal portrait, and 30 days of accumulated interaction history, to observe how its help effect changes.

Today AI hopes to verify through this method whether the long-term accumulated memory can be transformed into more accurate and effective help. The team said that relevant research methods, test scenarios and phased results will be gradually disclosed, and more researchers, developers and users are welcome to jointly test.

Today AI still has many parts to be polished. Whether the operation is smooth, whether the conversation is fluent, and whether the function prompts are clear, all are the parts that the team continues to improve according to user feedback.

Today AI admits that this is just the starting point. To make AI understand a person in the long run and provide help at the right time, there is still a lot of work to do in product, model and engineering.

The cumulative financing of over 200 million US dollars enables Today AI to continue to promote this long-term investment, and allows the team to move step by step towards the vision of "Intelligence for All" — to make AI affordable and usable for more people, even people who are not familiar with technology can get their own exclusive help, and leave more time for the life they care about.

At present, this forward-moving team is also recruiting more partners. If you are also excited about such a future, willing to invest in the long term and turn it into reality with your own hands, you are welcome to join Today AI.