Recently, investing in "perceptive and socially savvy" AI has become a popular trend among investors.

When intelligence starts to take physical forms of different kinds, how they can live alongside humans is becoming a new question that the AI industry has to answer.

A robot notices you are thirsty, walks over and hands you a glass of water; a pair of AI glasses identifies the building in front of you and actively tells you its history; a smart earphone notices your packed schedule today and reminds you to take a break. Technically speaking, none of these behaviors have flaws. The machine correctly processes the information, completes the reasoning, and gives a response that meets expectations.

But what if you are in a negotiation with a client, arguing with a friend, or just want to stay alone for a while? The machine may understand what you said, see the world in front of you, and even infer your mood, yet still have no idea whether it should show up at this moment.

Over the past few years, the development of artificial intelligence has revolved around one core question: how to make machines smarter? Now, another question is emerging: as more and more intelligent agents enter human life, how should they get along with people?

This is not just a problem for the robotics industry.

From AI glasses and smart earphones to home smart terminals and vehicle cockpits, to embodied robots with motion control capabilities, intelligence is entering human daily spaces in different physical forms. The models they can call are getting more powerful, the information they can perceive is getting richer, and their physical distance from humans is getting closer.

Yet an increasingly obvious fact is that the improvement of intelligence level will not automatically lead to the improvement of experience quality.

"After enabling machines to understand language and the world, the industry may be moving towards the next test of intelligence: interaction," an investor told Thecapital.

From Creating Intelligence to Finding Bodies for Intelligence

Looking back at the development of this round of the artificial intelligence industry, we can see an increasingly clear evolution path.

The starting point is the breakthrough of intelligent capabilities themselves.

Large language models have endowed machines with unprecedented capabilities in language understanding, knowledge processing and reasoning. Subsequently, multimodal models began to incorporate images, sounds and videos beyond text into a unified understanding framework, and AI gradually moved from processing symbols to understanding richer real-world information.

But at this stage, the main interaction method between humans and AI is still the dialog box on the screen: people input questions, and machines generate answers; people give instructions, and machines complete tasks. No matter how the internal capabilities of the model evolve, the relationship between humans and intelligent systems still follows a set of simple and straightforward logic: humans initiate actively, and machines respond passively.

Next, the industry began to find new carriers for these intelligent capabilities.

The first is embodied intelligence. Researchers are trying to enable machines not only to understand the physical world, but also to directly participate in it through movement, operation and environmental feedback.

The second is a wider range of AI hardware.

Smart glasses give AI visual input close to the human perspective; earphones provide an audio channel that accompanies people for a long time; smart home terminals enter relatively fixed living spaces; vehicle cockpits gain the possibility of understanding driving status and occupant behaviors.

Different terminals have different perception capabilities, feedback channels and physical forms, but behind them is a similar trend: AI is gradually evolving from software that needs to be actively opened to an intelligent presence that can continuously appear in daily life.

In 2026, Google further announced the direction of smart glasses equipped with Gemini at the I/O conference, including audio glasses and display glasses. Meta launched the Ray-Ban Display priced at $799 last autumn, which is also equipped with a wristband that can read wrist myoelectric signals, and this year it has continuously expanded its capabilities in navigation, communication and contextual interaction. Domestic manufacturers including Xiaomi, Thunderbird and Rokid have also intensively launched AI glasses over the past year.

These products do not necessarily have mechanical arms, nor do they need to walk like humans, but they are changing a basic fact: AI no longer only exists when we ask it questions. When the device can perceive the user's environment, remember historical preferences, and actively provide help at the right time, the traditional instruction-centered interaction method begins to appear insufficient.

"In the first few days after users take the product home, they will keep calling and testing it. Once the novelty wears off, most of the time they won't even remember its existence," an entrepreneur who develops AI glasses told Thecapital, which is why more and more teams are starting to explore active interaction. "But once the active interaction goes too far, the first reaction of users is to take it off."

Traditional intelligent systems mainly deal with the relationship between one instruction and one result. Future intelligent terminals need to deal with a more complex problem: the relationship between themselves and people, as well as the relationship between people and other people around. This means that the industry cannot just keep adding more capabilities to intelligence, but must also redefine how these capabilities should be used.

Machines Have Gained Capabilities, But Lack "Social Awareness"

There is a Beijing dialect term referring to the quality of being socially aware. A socially aware person not only knows what others need, but also knows when to offer help, when not to disturb, how to express kindness, and how to exit gracefully when the other party is unwilling to accept help.

Machines just lack this capability. "Active interaction is a double-edged sword. If you fail to grasp the proper measure, it will become harassment," said an investor.

Take the robot handing water as an example.

An entrepreneur in the field of interactive intelligence told Thecapital such a scenario: the robot notices that the user has not drunk water for a long time, so it actively pours water, hands it over and sends a voice reminder. From the perspective of task execution, almost every step is correct. But if the user is in an important meeting, this behavior may constitute an unnecessary disturbance.

This kind of awkward situation is not new. At the end of the 1990s, Microsoft put a blinking paperclip assistant called Clippy in Office. As long as it detected that the user was writing a letter, it would pop up and ask: "It looks like you're writing a letter. Do you need help?" It can be counted as one of the earliest batches of "active AI", and also one of the most complained software images. After several adjustments, it was completely removed from the 2007 version of Office.

The past task-oriented intelligence mainly answers three questions: what is the goal, how to execute it, and whether the execution result is correct.

The next-stage interactive intelligence needs to further answer: should I act now, in what way should I act, and at what moment should I stop. It needs to understand not only the physical environment, but also the context and rules of human society.

Communication between people has never been a simple exchange of information. A pause, a look, a change in body orientation, or even a deliberately kept silence, may carry complex social meanings, yet these pieces of information cannot always be reduced to a clear verbal instruction.

Therefore, when intelligent terminals begin to enter homes, workplaces and public spaces for a long time, the real difficult thing may no longer be just identifying what users want, but judging whether users want to interact with them at this moment. That is, to make robots socially aware.

Interaction Is Evolving From a Product Feature to an Independent Intelligent Capability

This change has already begun to appear in the technical routes of leading AI labs.

In May 2026, Thinking Machines Lab founded by former OpenAI CTO Mira Murati released the research preview of Interaction Models. The team put forward a noteworthy judgment: most of today's AI interactions are still turn-based. The machine does not start to answer until the user finishes speaking; when the machine generates answers, its continuous perception of the outside world is restricted.

However, real human communication is not such a simple back-and-forth. We observe others when they are speaking, judge whether we should interrupt according to their tone, and constantly adjust the communication rhythm of both sides through pauses, responses and gestures.

The research of Thinking Machines Lab tries to enable the model to process continuous audio, video and text input at the same time, and carry out real-time interaction with micro-turns of about 200 milliseconds. When explaining this direction, the team cited concepts in communication studies such as co-presence and simultaneity.

This means that interaction is no longer just a set of voice interfaces attached outside the model, but a problem that the model architecture itself needs to solve.

Capital is also pouring into this direction. Sesame, co-founded by Oculus co-founder Brendan Iribe, secured $250 million in Series B financing led by HSG and Spark Capital in October 2025. It develops a conversational voice model that reduces the response time to two to three hundred milliseconds, plus a pair of lightweight glasses that only support voice interaction.

"This round of dollar funds looking at interaction no longer focus on whether the speech recognition is accurate, but on whether the AI knows how to respond and leave proper blank space in communication," said a dollar fund investor.

There are also domestic startups exploring on this track. Different from making robots judge human emotions as accurately as possible, the HKUST-originated SYNAPUFF focuses its research on human-machine symbiotic interaction, which means enabling intelligent agents to gradually establish an interaction mode suitable for the current context through the closed loop of perception, decision-making and feedback. One noteworthy idea is to use low-social-cost, retractable behaviors to conduct tentative interactions.

For example, a robot with body movement capabilities can test whether the user is willing to communicate with it by slightly turning its head, changing orientation, or a short gaze. For terminals without mechanical bodies, similar behaviors can be presented as a slight tactile prompt, a non-intrusive sound effect, or a visual signal that does not require an immediate response. If the user shows willingness to interact, the system can further interact; if no response is received, the system should know when to stop. What this type of mechanism really needs to study is how to make machines learn to test, judge, negotiate and exercise restraint, rather than take active actions more frequently. More importantly, this capability should not be completely dependent on a certain terminal form.

The same interaction decision may be expressed through actions on a robot, through visual prompts on glasses, through a short sound on earphones, and through a slight tactile feedback on the steering wheel in an intelligent cockpit.

Although the physical carriers are different, the underlying problems to be solved share commonalities: understand the current context, judge people's willingness to interact, choose the right intervention timing, and complete communication through appropriate feedback methods.

Next-Generation Interaction Requires a New Evaluation System

If we take interaction as an independent intelligent capability, we will face a new challenge: how to judge whether it has really improved? In the field of task execution, the evaluation criteria are relatively clear. Whether the robot successfully grasps the object, whether the AI correctly completes code generation, and whether the autonomous driving system safely reaches the destination can all be measured by specific indicators.

But social interaction does not always have a single correct answer. The same user may have completely different reactions to the same interaction behavior at different times. This makes it difficult to fully evaluate the system's capability simply through static Q&A, emotion recognition accuracy, or one-time interaction satisfaction. A possible direction is to understand interaction as a continuous closed-loop decision-making process, and evaluate the key links in it.

For example, can the system accurately judge the timing of active interaction? How many unnecessary disturbances have occurred? After the user refuses, can the system stop in time? Is the user willing to continue the interaction? Can the same intelligent agent form a stable and predictable behavior pattern in long-term coexistence? These problems are not completely unquantifiable, but they require evaluation methods different from traditional task-oriented intelligence.

In long-term interaction, there is another easily underestimated capability: personality consistency. An intelligent agent does not necessarily need to be always gentle and compliant, and even does not necessarily need to take serving users as its only goal.

For example, in entertainment and IP scenarios, a naughty and rebellious character may be more attractive than an always-correct assistant.

The characters in Disneyland are very good examples. Many characters barely communicate with tourists through language, but they can make people quickly understand their respective personalities through eyes, movements, postures and reaction rhythms.

The playfulness of LinaBell and the enthusiasm of Mickey Mouse are not conveyed through a self-introduction, but are repeatedly perceived through a large number of consistent non-verbal behaviors. This experience also inspires intelligent interaction: the character personality perceived by people is not only the text generated by the language model, but the result jointly shaped by language, movements, sounds, touch and interaction timing.

Disney itself is also applying this experience to robots. The BDX bipedal robot developed by its research team does not say a word, and can let tourists tell whether it is curious or nervous just by tilting its head, pacing and jumping. Last November, the walking Olaf robot also made its debut at Disneyland Paris.

SYNAPUFF and other entrepreneurial teams are trying to combine the expression of non-verbal behaviors with the contextual decision-making of intelligent agents, and further study the behavior performance under different personality constraints.

From this perspective, future interaction systems may need to distinguish two different types of capabilities.

One layer is the relatively general symbiotic interaction capability, which solves the problems of how intelligent agents understand social contexts, respect human boundaries, and choose the right timing and way to communicate.

The other layer is the personalized personality expression capability, which determines what kind of character, language and behavior style the intelligent agent uses to get along with people.

The former solves how to get along, while the latter determines who you get along with.

A rebellious character does not always have to please the user, but it still needs to understand when the joke should stop; a rigorous AI assistant does not need to have a distinct entertainment personality, but it also needs to master the proper measure of communication with people.

For more and more intelligent terminals that exist in life for a long time, this may be a more complete interaction framework than simply improving the accuracy of emotion recognition.

Hardware Forms Will Continue to Diversify, While Interaction Capabilities May Gradually Become Universal

In the past consumer electronics industry, competition mainly revolved around hardware performance, functional configuration and cost. When more and more devices can access powerful foundational models, the differences between terminals will no longer only depend on model parameters or task execution capabilities. A new question arises: what kind of intelligent agents are users willing to let participate in their lives for a long time?

Different hardware may have completely different value propositions.

AI glasses emphasize portable perception and instant assistance; smart earphones emphasize low disturbance and continuous connection; robots can provide movements, spatial behaviors and physical contact; home terminals may pay more attention to long-term memory and multi-person collaboration.

These products do not necessarily need to follow exactly the same interaction strategy, but they all need to solve the problems of active intervention, contextual understanding, expression methods and interaction boundaries. As a result, interaction may gradually evolve from product features scattered in various terminals to a type of technical capability that can be independently researched, trained and deployed.

For model companies, this means that interaction should not just be a set of prompt words or interface rules outside the foundational model. Hardware manufacturers will find that the upper limit of product experience may no longer be determined only by sensors, screens, motors and computing power. New opportunities may lie precisely between the two: building a system that can adapt to different terminals, understand social contexts and continuously optimize interaction behaviors.

"It has become increasingly difficult to widen the gap at the hardware level. The circuit board from the same solution provider, with a different shell, becomes a new brand," said an investor who has long focused on consumer electronics. "Now when we invest in AI terminals, we all focus on what the interaction definition logic is. If this layer is not done well, the product has no soul."

But this path is not easy.

It is necessary to accumulate multi-modal interaction data in real contexts, deal with the differences between different hardware feedback capabilities, and solve problems such as long-term memory and personality consistency. More critically, the industry still needs to prove that these capabilities can stably improve user experience, rather than just create some seemingly humanized demo moments. No matter whether these capabilities are finally realized by foundational model companies, terminal manufacturers, or independent interaction technology providers, they will all face a common technical proposition: how to make intelligent systems not only understand human instructions, but also understand human social life.

An entrepreneur admitted frankly that the hardest part of interaction data is not collection, but labeling. "One person frowns, is he annoyed by the robot, or just thinking about something? For the same video, five ordinary labelers can give three different answers, and only experienced interaction experts can give the accurate direction."

This is exactly why the human-machine symbiotic interaction paradigm is worth being redefined.

In the past, humans needed to learn to use machines. Touch screens, graphical interfaces, voice assistants, each generation of interaction technology has lowered the threshold for people to use computing power. But as AI begins to have the capabilities of continuous perception, active decision-making and autonomous action, the relationship between humans and machines is undergoing new changes. We no longer just operate a tool occasionally, but may need to share time, space and attention with an intelligent system that exists around us for a long time. It does not have to look like a human, nor does it have to imitate human emotions, but it must understand people's rhythm, boundaries and willingness, and learn to optimize its own behavior in complex real contexts.

The previous stage of intelligence was to enable machines to understand language, understand the world, and have action capabilities.

The next test is to truly integrate these capabilities into people's daily life.

Making machines smart is the starting point of intelligence. Making machines learn to get along with people is the beginning of intelligence truly integrating into life.

This article