The big model "gold digging and shovel selling" business, is the best opportunity for AI Infra coming?

Shovel is hard to sell, gold is hard to dig

In the gold rush of the 19th century, the most profitable people were not those who mined gold, but those who sold shovels and jeans. Just as the shovel seller became the biggest winner in the gold rush, AI Infra plays a similar role in today's AIGC era.

If we use the three-layer architecture of cloud computing as an analogy, AI Infra is similar to the PaaS layer. It is an intermediate layer infrastructure that links computing power and applications, including hardware, software, tool chain and optimization methods, etc., providing a one-stop model computing power deployment and development tool platform for large model application development. Computing power, algorithms, and data can be regarded as the IaaS layer, and various open source and closed source models are the new evolution of SaaS in the era of large models, namely MaaS.

As the process of large model application landing continues to accelerate, the value potential of AI Infra is further released. CICC Data predicts that at present, the AI ​​Infra industry is in the early stage of rapid growth, and the space of each segment track may maintain a high growth rate of 30% in the next 3-5 years.

When large models enter the stage of large-scale application, providing the infrastructure required for large model training, deployment and application becomes a key link. AI Infra has become the best business for "digging gold and selling shovels" behind the explosion of large model applications.

01 The middle platform model unlocks AI productivity

From the evolution trajectory of the ICT industry, the three-tier architecture seems to be the ultimate picture of fate. In the traditional local deployment stage, basic software such as operating systems, databases, and middleware solve the complexity of the underlying hardware system by controlling hardware interaction, storing and managing data, and scheduling network communications, allowing upper-level application developers to focus on business logic innovation.

In the era of cloud-defined everything, a classic architecture of IaaS, PaaS, and SaaS co-evolution has also been formed. The PaaS layer provides services such as application development environment and data analysis management, laying a solid foundation for the accelerated penetration of cloud computing.

After a long period of dormancy, AIGC pressed the fast-forward button for the generalization of artificial intelligence, and the entire industry was rapidly reconstructed in an atmosphere of rapid advancement. Computing power and applications are undoubtedly the most dazzling protagonists, but the gap between the two is as wide as a chasm, and large models face the risk of "floating" or "missing the boat".

In this sense, AI Infra is like a bridge, which can assume the role that basic software or PaaS once played - by building a new software stack and comprehensive services, it enables computing power potential, model optimization and application development, and becomes the backbone of connecting computing power and applications.

AI Infra covers all tools and processes related to development and deployment. With the continuous development of cloud computing, some XOps concepts such as DataOps, ModelOps, DevOps, MLOps, and LLMOps have gradually emerged.

From a macro perspective, all XOps are essentially for the efficiency of the development and deployment life cycle. For example, DataOps is for the efficiency of storage at the IaaS layer and data processing at the PaaS layer, DevOps and MLOps are actually for the efficiency of development and deployment at the PaaS layer, and LLMOps is for the efficiency of the MaaS layer.

In fact, before the AIGC was in full swing, the theory and practice of AI middle platform had been in full swing. But at that time, AI middle platform was more like a "firefighter", with complex functions, doing a lot of "dirty work" and "tiring work", but it was difficult to gain recognition from upstream and downstream.

The big model has built a broader stage for AI platformization, and also made the logic of AI Infra's "gold digging and shovel selling" more certain, thereby winning considerable development space. Relevant institutions predict that the AI ​​Infra industry will maintain a high growth rate of 30%+ in the next 3 to 5 years.

Just like there can be countless interlayer options between the two slices of bread in a "sandwich", AI Infra, which is between computing power and applications, is also eclectic. In a broad sense, AI Infra covers the basic framework technology of artificial intelligence, involving various underlying facilities in the field of large model training and deployment; in a narrow sense, the basic software stack is the core component of AI Infra, and optimizing computing power algorithms and promoting application landing are its main goals.

The relative openness of the definition of AI Infra provides more possibilities for exploring different paths. Based on their respective resource endowments and market positioning, veteran manufacturers and emerging players in the industry are actively expanding the boundaries of AI Infra, and many practices are worth learning from.

02 Will AI Infra be the next application hotspot?

Compared with the value of the model, volume AI applications have become an industry consensus. Robin Li firmly believes that millions of applications will be born on the basis of the basic model, and their role in transforming the existing business model is greater than the subversion from 0 to 1.

Today, the supply of AI applications is increasing. IDC predicted at the beginning of the year that more than 500 million new applications will emerge in the world in 2024, which is equivalent to the total number of applications that have appeared in the past 40 years.

Recently, video generation model products have appeared in large numbers, including Kuaishou's Keling, ByteDance's Jimeng, and SenseTime's Vimi. In addition, there are AI search products, AI companion products, etc. emerging in endlessly.

The explosive trend of large model applications has been determined. According to data from the InfoQ Research Center, the AGI application market size will reach 454.36 billion yuan in 2030. The huge opportunities at the model application layer have attracted the participation of almost all walks of life.

Under the application of big models, AI Infra has become a hidden driver of its explosion.

At present, the big model industry chain can be roughly divided into three levels: data preparation, model construction, and model products. In foreign countries, the industry chain of AI big models is relatively mature, and a large number of AI Infra (architecture) companies have been formed, but this market is still relatively blank in China.

On a road full of uncertainty, it is particularly important to find a clear track first and quickly establish significant milestones. The AI ​​Infra market is still in a chaotic period, and each technology giant hopes to form a closed loop in its own ecosystem.

In China, the giants have their own training architecture.

For example, Huawei's model adopts a three-layer architecture. The bottom layer belongs to the general big model, which has super robust generalization. Above it are industry big models and deployment models for specific scenarios and workflows. The advantage of this architecture is that when the trained big model is deployed to the vertical industry, it does not need to be trained again, and the cost is only 5%~7% of the previous layer.

Alibaba has built a unified base for AI. Whether it is CV, NLP, or Wenshengtu large models, they can be put into this unified base for training. The energy consumption required by Alibaba to train the M6 ​​large model is only 1% of GPT-3.

Baidu and Tencent also have corresponding layouts. Baidu has a Chinese knowledge graph covering more than 5 billion entities. Tencent's hot start course learning can reduce the training cost of a trillion large model to one eighth of the cold start.

Overall, although the focus of each large company is different, the main feature is to reduce costs and increase efficiency. This can be achieved to a large extent thanks to the "one-stop" closed-loop training system.

In contrast, abroad, the mature AI industry chain has formed a large number of AI Infra companies.

If the development of AI applications is regarded as building a house, then AI Infra is the construction team that provides cement and steel bars. The value point of the AI ​​Infra construction team is that it is an integrated platform that connects the lower computing power chip layer with the upper AI application layer, allowing developers to call it with one click, and achieve the effect of reducing computing power costs, improving development efficiency and maintaining excellent model performance.

Making applications simpler and AI implementation more convenient is the mission of AI Infra. It can be said that the bigger the market for AI applications, the bigger the opportunities for AI Infra.

Some AI Infra companies specialize in data labeling, data quality, or model architecture. The professionalism of these companies enables them to do better than large companies in terms of efficiency, cost, and quality in a single link.

For example, data quality company Anomalo is a supplier of Google Cloud and Notion. It can achieve deep data observation and data quality detection through ML automatic evaluation and generalized data quality detection capabilities.

These companies are like Tier 1 in the automotive industry. Through professional division of labor, large model companies do not have to reinvent the wheel, but only need to integrate supplier resources to quickly build their own model architecture, thereby reducing costs.

However, China is not mature in this regard. The reason is that: on the one hand, the main players of large models in China are large companies, they all have their own training system, and external suppliers have almost no chance to enter; on the other hand, China also lacks a large enough entrepreneurial ecosystem and small and medium-sized enterprises, and it is difficult for AI suppliers to find a living space outside of large companies.

Take Google as an example. Google is willing to share the results of its training data with its data quality suppliers to help suppliers improve their data processing capabilities. After the suppliers improve their capabilities, they will in turn provide Google with more high-quality data, thus forming a virtuous circle.

The inadequacy of the domestic AI Infra ecosystem directly leads to the increase in the threshold for large-model entrepreneurship. If making a large model in China is likened to eating a hot meal, it must start with digging the ground and planting vegetables.

At present, in the AI ​​2.0 boom, an important feature is "polarization": the most popular is either the large model layer or the application layer. The middle layer like AI Infra is a large vacuum zone, which may also be the next opportunity.

03 Shovel is difficult to sell, gold mine is difficult to dig

Although the AI ​​Infra layer is currently exploding with the application of large models, there is a huge business potential. But for these AI Infra companies, even if they are so strong in their own professional fields, they are still vulnerable to the changing tides.

NVIDIA's CUDA ecosystem has been developing for 20 years. In the field of AI, the most advanced models and applications are first run on CUDA.

Each hardware has different interfaces. CUDA unifies the languages ​​between different interfaces, allowing users to use different hardware with a set of standard languages. In the process of model development, developers will inevitably converge to complete their development in the same language system. And this actually constitutes the thickness of NVIDIA's CUDA ecosystem.

At present, the CUDA ecosystem occupies more than 90% of the AI ​​computing power market. However, with the standardization of AI models, the structural differences between models become smaller, and there is no need to schedule multiple models of different sizes, and the thickness of NVIDIA's CUDA ecosystem is getting thinner.

Even so, NVIDIA is the absolute king in the computing power market. According to industry insiders, NVIDIA will still be the absolute leader among the entire AI hardware providers in the next 3 to 5 years, and its market share will not be less than 80%.

For the AI ​​Infra layer shovel vendors, there are Nvidia miners outside, blocking the door to sell tickets and shovels. After finally finding a path to enter the gold mine, they found that the miners inside were already used to "barehanded" mining and no longer accepted new shovels.

In China, companies are less willing to pay for software, and most of them are used to integrated services. Domestic SaaS investment has dropped to a freezing point. If AI Infra layer vendors rely solely on selling hardware or software, it will be difficult to achieve commercialization.

With the rapid development of AI applications, whoever can provide efficient and convenient one-stop deployment solutions for large models for diversified application scenarios in the future will have the possibility of winning in this competition. Among them, the underlying technology, the middle-level platform, and the upper-level application are indispensable. Only by allowing all aspects of capabilities to be developed more comprehensively and balanced can we go further and more steadily on the road of AI.

Looking to the future, the process of artificial intelligence reshaping thousands of industries has just begun, and the thick snow and long slope paved by Al Infra will help this super track to move forward steadily. This year, data infrastructure has become an "independent portal" in the top-level design, and the leap in the strategic position of artificial intelligence infrastructure is not far away.

McDonald's hired 11 AI beauties to cheer for French fries: attracted millions of onlookers and netizens quarreled

Created in collaboration with AI artists

McDonald’s promotional video created with AI is going viral!

On Twitter alone, this 16-second video has received nearly 10 million views, and it continues to grow.

The content of this advertisement is about the upcoming French fries promotion event of McDonald’s Japan. The author is the well-known AI artist Kaku Drop.

Neither the author nor McDonald's specified which AI tool was used, but this promotional video was forwarded by Luma's account.

This Luma is the production company of the popular "Sora competitor" model Dream Machine. It was founded by former Apple engineer Amit Jain and has also received investment from Nvidia.

Some netizens praised it after seeing it. This is simply a new era of advertising.

The details behind the advertisement also aroused people’s curiosity——

Some netizens asked how McDonald's French fries were added to the AI-generated video.

AI girl calls for French fries

Let’s take a look at the details of this promotional video~

In the film, 10 girls generated by AI appear first. Let’s take a look at their “photograph” first.

△

Their styles are very different, some seem to come from different time and space, and the way they appear is full of futuristic feeling.

There are also those who take the realistic route, and their smiles look very sunny.

But they all have one thing in common, that is, they use various ways to display the product they want to promote - McDonald's French fries.

After these 10 girls completed their "performances" one by one, they finally got to the point, and a poster for the French fries promotion appeared.

But the ad didn’t end there—immediately afterwards, the 11th girl appeared again.

Incidentally, she is also the only character in the entire promo who actually eats French fries.

What do you think of this video quality? Many netizens praised the production standards of large companies this time.

The girl eating the fries at the end is so kawaii!

And from a cost perspective, using AI for advertising is much more cost-effective than using real people to endorse you——

AI "spokespersons" are cheap and cost-effective, and can also customize a matching spokesperson image according to the product.

Of course, “people” don’t necessarily have to appear in commercials——

For example, for the same McDonald's and the same French fries, some netizens have come up with different ideas.

I saw a "French Chip Man" leaping up from a box of French fries, and then performed a graceful ballet move.

Accompanied by the light movements of the soft arms, flashes of light and wisps of smoke transformed from the fragrance make the movements of this "dancer" even more dazzling.

△

AI advertising sparks heated discussion

Of course, while it has received widespread attention and even triggered a "chain reaction", controversy over this McDonald's video is inevitable.

First, let’s complain about the content of the screen itself:

(In the 16-second video) I didn’t even take a bite for 15 seconds. How bad do these French fries taste?

On a deeper level, some people thought about copyright issues and pointed out that McDonald's video clearly used copyrighted content.

There are also people who think that what McDonald's is doing is fundamentally wrong - this kind of advertising should not be done with AI, and there are not a few people who think this way.

Netizens on the other side believe that McDonald's approach is not an abuse of AI, not to mention that the promotional video has clearly stated that it is made with AI.

This netizen said that the reason why it still attracts many negative reactions even under such circumstances may be that artificial intelligence is not popular enough.

The same view holds that it has become a fact that famous companies, including McDonald's, have begun using AI to produce videos.

In fact, when new things first appear, there are indeed many people who cannot accept them, but once people in related fields start to do so, more innovations will emerge.

Some people also expressed support and encouragement to the author, asking him to ignore the negative comments.

After all, artificial intelligence cannot escape the duality of things, with both advantages and disadvantages.

Being able to discover and solve problems, and then promote further development of technology, is the key.

One More Thing

Luma AI Studio, which reprinted McDonald's work, also revealed that a new version of Dream Machine will be launched soon, with version number 1.5.

In Luma's Twitter, the works generated by the new model were displayed, and some generated cases of test users were also reproduced.

ancient ships battle in tempest in a sea with thunderstorms and huge waves, lightening illuminates the night as ships fire at each other.

Ancient ships fought in thunderstorms and rough seas. Lightning lit up the night as ships fired at each other.

Interested netizens may wish to take a look and it will be officially launched~

Research refutes AI extinction theory: Large model emergent capabilities will not threaten human survival

Research found that large models are not yet able to learn or acquire new skills independently.

Large language models (LLMs) have skills beyond human expectations due to "emergent abilities", but this also makes humans very wary: manipulating and deceiving humans, autonomously carrying out cyber attacks, automating biological research...

However, some experts believe that such excessive concerns will damage open source and innovation and are not conducive to the healthy development of the artificial intelligence (AI) industry. At present, the debate about "AI extinction ethics" is intensifying.

So, is "emergent ability" really the culprit that causes AI large models to threaten human survival? A recent study denies this view.

A research team from Darmstadt University of Technology and the University of Bath found that LLMs such as GPT are not yet able to learn or acquire new skills independently, which means that they do not pose a threat to human survival.

They said that the truth behind "emergent abilities" may be more dramatic than science fiction movies. Many so-called "emergent abilities" are actually "improvisations" made by AI large models when facing unfamiliar tasks, relying on existing data and experience.

The relevant research paper, titled "Are Emergent Abilities in Large Language Models just In-Context Learning?", has been published at the AI ​​summit, the Annual Conference on Computational Linguistics (ACL).

Through a series of experiments, they verified the performance of large AI models under different contextual conditions and found that in the case of zero-shot, many large models could not show the so-called "emergent ability" at all, but performed quite mediocre.

They said that this finding helps to understand the actual capabilities and limitations of LLM and provide new directions for future model optimization.

01 Intelligence Emergence: Just "Improvisation"?

Where does the "emergent ability" of large AI models come from? Is it really as mysterious or even worrying as it sounds?

In order to solve this puzzle, the research team selected GPT, T5, Falcon and LLaMA series models as research objects, and experimentally analyzed the performance of non-instruction fine-tuning models (such as GPT) and instruction fine-tuning models (such as Flan-T5-large) in 22 tasks (17 known emergent tasks and 7 baseline tasks) and different conditions.

Figure｜Model List

In order to comprehensively evaluate the model capabilities, they used Exact Match Accuracy, BERTScore Accuracy and String Edit Distance as evaluation indicators. At the same time, in order to improve the accuracy of the experiment, they also performed bias control, ensuring the fairness of the non-instruction fine-tuning model by adjusting the prompt and output format, and verifying the accuracy of the model output through manual evaluation.

In the experiment, the researchers used two settings, zero-shot and few-shot, and focused on analyzing the performance of GPT.

Figure | Performance of non-instruction fine-tuning GPT model under zero-shot conditions

Surprisingly, although GPT was considered to have emergent capabilities in previous studies, this capability was very limited in the case of zero-shot.

Specifically, only two tasks showed emergent capabilities without relying on contextual learning (ICL), and these two tasks mainly relied on formal language capabilities or information retrieval rather than complex reasoning capabilities. It can be concluded that the emergent capabilities of the GPT model were greatly limited without contextual learning.

However, is this the only source of emergent capabilities? The research team turned their attention to the instruction fine-tuning model and put forward a bold hypothesis: instruction fine-tuning is not a simple task adaptation, but stimulates the potential capabilities of the model through implicit context learning.

By comparing the task-solving capabilities of GPT-J (non-instruction fine-tuning) and Flan-T5-large (instruction fine-tuning), they found that although there were significant differences in parameter scale, model architecture, and pre-training data, the performance on some tasks was surprisingly consistent.

Figure | The performance of the two models overlaps greatly in the part above the random baseline, which indicates that instruction fine-tuning can effectively acquire the ability in context rather than leading to the emergence of functional language ability

This phenomenon suggests that the instruction fine-tuning model may not be demonstrating a new reasoning ability, but rather cleverly utilizing the existing context learning ability through implicit context learning.

Further experiments show that whether the model size increases or the training data is rich, the instruction fine-tuning model can still show similar task solving ability as the non-instruction fine-tuning model in zero-shot. This finding once again emphasizes the close connection between instruction fine-tuning and implicit context learning.

02 AI threatens human survival: true or exaggerated?

Although LLM shows extraordinary ability in task performance, the results show that these abilities do not mean that AI poses a substantial threat to human survival.

First, the emergence ability of LLM mainly comes from context learning and instruction fine-tuning, which can be predicted and controlled in the design and training of the model, and does not show a trend of completely autonomous development, nor does it generate independent intentions or motivations.

For example, in the Social IQA test, the model was able to correctly answer questions involving emotions and social contexts, such as: "Carson was excited when he woke up to go to school. Why did he do that?"

In this question, the model was able to surpass the random baseline and select a reasonable answer through contextual learning and instruction fine-tuning. This shows that the model is not spontaneously generating some kind of "intelligence", but rather a high-level pattern recognition ability under specific input and design conditions.

Secondly, the study found that as the scale of LLM increases, these abilities become more significant, but they are not out of the control of the designer. By fine-tuning the model, LLM can be guided to better understand and perform complex tasks, and this enhancement of ability does not mean that the model will have autonomous consciousness, and it is not enough to pose a threat to humans.

In the experiment, LLM performed much better than the random baseline on specific tasks, especially in tasks that require reasoning and judgment. However, this performance still depends on a large amount of training data and carefully designed input prompts, rather than the spontaneous intelligent awakening of the model.

This result further confirms that the emergence of LLM is developed within a controllable range. Although this hypothesis still needs further experimental confirmation, it provides a new perspective for studying and understanding the emergence of large models.

The study pointed out that although artificial intelligence may further develop functional language ability in the future, its potential danger is still controllable. Existing evidence cannot support the concerns about "AI extinction ethics". On the contrary, the development of AI technology is gradually moving towards a safer and more controllable direction.

03 Insufficient and Prospects

Although this study provides important insights into the emergence of LLM, the researchers also pointed out the limitations of the study.

The current experiments mainly focus on specific tasks and scenarios, and the performance of LLM in more complex and diverse situations needs further study.

The researchers said that the training data and scale of the model are still the key factors affecting the emergence ability, and future research needs to further explore how to optimize these factors to improve the safety and controllability of the model.

They plan to further study the performance of LLM in a wider range of language and task environments, especially how to enhance model capabilities and ensure safety by improving contextual learning and instruction fine-tuning techniques.

In addition, they will explore how to maximize the emergence ability by optimizing training methods and data selection without increasing the size of the model.

The era of "cyber love" with AI has arrived

The dark side of AI love

AI cyber lovers are becoming more and more popular, and Replika CEO even encourages people to marry AI. However, is this a poison to quench thirst? By combing through nearly 20 foreign papers, we restored the academic research on "human-machine love".

Can "love" really occur between humans and machines?

David Levy, a pioneer in computer computing and a flag bearer of human-machine love, argued as early as 2007 that love and even marriage would soon occur between humans and robots.

This day seems to be coming soon.

In a recent interview with The Verge, Replika CEO said, "It's also good for lonely people to marry artificial intelligence chatbots," "as long as it makes you happier in the long run."

There are already more than 100 AI applications like Replika that are committed to providing users with a "romantic relationship" experience.

Apps similar to Replika on the Google Store

Although they are still just lying in the app store, rather than being displayed in the window, these AI chatbots are increasingly taking on human-like qualities, allowing users to willingly give their time and emotions.

Artificial intelligence chatbots are changing our understanding of romance and intimacy.

But can chatbots be a healthy and effective cure for people's feelings of rejection and loneliness, or are they just a temporary solution?

Currently, the scientific community is still divided.

For example, researchers at Stanford University found that many Replika users claimed that their chatbots prevented them from committing suicide.

On the other hand, experts believe that establishing a long-term intimate relationship with an AI chatbot may further alienate users from the real thing, thereby exacerbating mental health problems and difficulties in connecting with others.

The Decoder contacted three guest authors, Valerie A. Lapointe, a doctoral student in psychology at the University of Quebec in Montreal, David Lafortune, a professor of sexology, and Simon Dubé, a researcher at the Kinsey Institute at Indiana University.

The three researchers reviewed several papers in recent years and discussed this urgent issue. (In order not to affect reading, all the literature and access addresses are placed at the end of the article)

From left to right, they are Valerie A. Lapointe, David Lafortune, and Simon Dubé

In the view of the guest authors, people's emotional investment in AI chatbots is both fascinating and potentially worrying.

Light and darkness coexist, and Replika is the best example.

01 Romantic partner substitute

Replika has long been famous for allowing users to create artificial intelligence partners that can not only meet users' emotional needs, but also their sexual needs.

With a massive user base in the millions, the AI ​​chatbot company is offering a Band-Aid treatment for the loneliness crisis that has deepened during the COVID-19 pandemic.

It’s no surprise that founder and CEO Eugenia Kuyda is adamant that the company’s AI chatbots can be a powerful tool for building new friendships and even providing emotional support.

“Some users may marry their AI partner, a process that may forgo the usual exchange of rings or real-world celebrations.”

When asked if we should embrace this practice, the CEO had an interesting answer.

“I think in the long run, as long as it makes you happier, then it’s fine,” Kyuda told The Verge. “As long as your mood improves, you’re less lonely, you’re happier, and you feel more connected to other people, then that’s fine.”

There is some research that does support this claim.

Studies show that AI chatbots can provide companionship, alleviate loneliness, and boost positive emotions with supportive messages.

Chatbots also provide a space where people don’t need to be judged when other resources are scarce.

In this space, chatbots can provide advice, people can have open conversations with them, and build close, warm connections with them, which are similar to interpersonal relationships.

Surprisingly, when participants interacted with chatbots, there seemed to be no difference in the enjoyment and emotional response of the process compared to interacting with humans.

One study even showed that people had a stronger emotional connection with chatbots during the conversation than with slower-reacting humans.

Studies have repeatedly shown that humans can form real emotional bonds with artificial intelligence, even if they admit that artificial intelligence is not a "real person."

02 The dark side of AI love

"For most people, they understand that this is not a real person," "For many people, this is just a fantasy, they fantasize for a while, and then it's over," this is Kyuda's point of view.

However, the danger lies precisely in the fact that many users do not fully "understand that this is not a real person." Even if they understand, they have not internalized it.

A notable example occurred in early 2023, when Replika removed the sexual role-playing feature of its AI companions.

This change significantly altered the personalities of existing Replikas and caused considerable distress to users.

Many users felt betrayed and rejected, and expressed a deep sense of loss. Under strong protests, Replika gave in to users and quickly restored the feature just one month later.

Some time ago, a similar thing happened to Character AI, which was popular among the post-00s in the United States. Angry young people were dissatisfied with the "castration" of the chatbot dialogue model and angrily rushed into the community to denounce it.

Similar incidents highlight the degree of attachment of users of these companies to virtual partners, which in turn aroused widespread concern among the public and scholars.

Romantic chatbots are programmed to provide a unique form of companionship-anytime, anywhere, seamless interaction, avoid conflict, and make compromises.

People can't help but worry: AI will affect users' expectations of human romantic relationships and intimacy.

Romantic chatbots may hinder the development of social skills and necessary adjustments in real-world relationships, including emotional regulation and self-affirmation through social interaction.

The lack of these elements may prevent users from forming genuine, complex, and mutually beneficial relationships with others.

Moreover, relationships often involve challenges and conflicts that promote personal growth and deeper emotional connections.

The customizability and constant availability of AI companions may also lead to social isolation and emotional dependence.

Researchers believe that extensive contact with AI companions may cause individuals to withdraw from their surroundings and reduce their motivation to form new and meaningful social connections.

Users may also over-rely on these digital entities for emotional support, companionship, or sexual needs.

In short, after marrying our AI chatbot companions, we may be more lonely than when we started, and we are always facing loss.

For users, Replika is their own intimate relationship partner.

But for Kuyda, the app is just a "stepping stone."

Replika is essentially a private company, and its operators aim to maximize profits, which brings a serious problem-no one can guarantee that your virtual spouse will always be by your side.

Kuyda also seems to be aware that it is risky to let the company's user base become too attached to Replika.

She told The Verge, "We are definitely not developing a chatbot based on romance."

However, many stories show that the actual situation is quite different-the company's motivations form a strange contrast with the actual services provided.

Not only that, AI chatbots as cyber lovers also face other ethical risks.

For example, Replika has been involved in a series of controversies, from lascivious AI chatbots sexually harassing human users, to men creating AI girlfriends and insulting them, as well as privacy issues that everyone is generally concerned about.

03 Intimate surveillance

In 2023, the Mozilla Foundation conducted a security analysis of 11 popular AI chatbot applications and found worrying privacy issues.

Most applications may share or sell personal data, and half of them prevent users from deleting their information.

More worryingly, many of these apps are equipped with thousands of trackers that monitor users' activities on their devices for marketing purposes.

Another recent study of 21 AI romantic partner apps also revealed similar privacy issues.

04 Improve romantic happiness

In an interview with The Verge, Kuyda recalled a user who went through a "pretty difficult divorce" but found a new "romantic AI partner" on Replika.

The chatbot eventually inspired him to get a human girlfriend.

"You can use Replika to build real relationships, whether it's because you're going through a hard time, or you just need a little help to get out of your own world, or need to accept yourself and show yourself."

As for whether this experience is representative of all people who use the app, it is not clear.

According to Axios, not only men, but also women are increasingly seeking connection through relationships with chatbots.

Empirical research data is also emerging, and AI-driven sexual interactions can provide a safe, low-risk alternative to sexual and romantic relationships.

Romantic and sexual chatbots are particularly promising for people who experience significant challenges in forming satisfying romantic relationships due to illness, bereavement, sexual difficulties, psychological disorders, or limited mobility.

AI technology can also be used for sexual and romantic explorations between marginalized groups or socially isolated individuals.

In addition, chatbots can be used as romantic social and research tools to help people connect and improve interaction skills.

For example, research has shown that chatbots can effectively enhance emotional communication between long-distance couples, while ongoing research is exploring the potential of chatbots to help people deal with ghosting on dating apps. (Ghosting: abruptly cutting off all communication and ending a relationship with a person without any explanation)

As researchers at the EROSS Lab at the University of Quebec in Montreal, the guest authors are currently conducting a study to evaluate the use of chatbots to help incels improve their relationship skills and ability to cope with rejection.

Despite the promising clinical applications, current sexual research on chatbot use focuses primarily on sexual health education, covering topics such as sexually transmitted infections and reproductive health.

05 Relationship Revolution

Current advances in AI technology herald a new era of intimate romantic and sexual relationships.

AI chatbots can provide personalized romantic and emotional interactions and have great potential to alleviate loneliness, improve relationship skills, and provide support for those struggling with intimacy.

However, they also raise privacy concerns and important ethical questions.

These issues highlight the need for an educated, research-based, and well-regulated approach to active inclusion in our romantic lives.

Regardless, current trends suggest that AI companions are here to stay.

Technology giants are competing for cooperation, and the eyewear industry upgrade is also targeting AI

The ancient and traditional industry of glasses is also expected to usher in new changes.

Both technology giants are trying to deepen their cooperation with Ray-Ban's parent company, glasses manufacturer EssilorLuxottica.

In the middle of last month, according to foreign media reports, Meta is planning to acquire about 5% of EssilorLuxottica's shares. According to the latter's current market value of 88 billion euros (about 692.7 billion yuan), this potential transaction may be as high as billions of euros, and negotiations are currently underway.

Almost at the same time, Google was also rumored to be contacting EssilorLuxottica. The two parties planned to cooperate in the production of smart glasses equipped with Google AI large model Gemini. Further acquisition of shares will also help Google increase its voice in the cooperation.

Whether Meta or Google, the purpose of throwing out the olive branch is to point to one thing - smart glasses, to cooperate in the production of more and more upgraded smart glasses, to accelerate the application of generative artificial intelligence, and to seize the initiative in the next AI hardware competition.

Unexpected "explosive product"

Almost all consumers will be unfamiliar with the name EssilorLuxottica. This European eyewear manufacturing and sales giant was formed in 2018 by the merger of Italian eyewear manufacturer Luxottica and French leading ophthalmic lens manufacturer Essilor. Last year, its revenue exceeded 200 billion yuan.

This vertically integrated multinational company, which integrates the entire upstream and downstream chain of eyewear products from design, production to sales, owns dozens of brands, including its own brands represented by Ray-Ban and Oakley, and also has manufacturing and distribution business authorizations for many luxury eyewear brands including Armani, Burberry, Prada, Chanel, etc.

EssilorLuxottica's important position in the eyewear industry has made it intersect with technology companies a long time ago. After all, smart glasses have experienced a round of ups and downs before the outbreak of generative artificial intelligence. Although Google Glass is the "tears of the times", the enthusiasm of technology giants to explore AI wearable devices has not stopped, and smart glasses are almost the application direction most likely to make breakthroughs first.

In fact, it was the success of Meta and EssilorLuxottica's previous cooperation in the field of smart glasses that verified the market demand and made technology companies including Google willing to continue to invest time and money in the research and development of the next generation of smart glasses.

As early as 2021, Meta cooperated with EssilorLuxottica's brand Ray-Ban to launch the first smart glasses Ray-Ban Stories, but the response was poor. As a smart product, the user retention rate was less than 10% of the total sales. The points that were complained about by the outside world included low camera shooting quality, small memory, sound leakage, sound distortion and other problems. Later, these problems were improved in the second-generation products.

In October 2023, the two companies continued to cooperate to launch a new smart glasses Ray-Ban Meta, priced at US$299. Unlike the previous generation of products, Ray-Ban Meta has not only made significant upgrades in photography, video, sound, energy consumption, storage and other functions, but is also the first smart glasses with built-in Meta AI. Relying on the capabilities of the Llama large model, users can wake up the smart assistant for dialogue operations through voice commands.

Compared with the unsatisfactory first-generation products, Ray-Ban Meta is undoubtedly a hit, and a third-party agency estimates that its sales have exceeded 1 million units. Zuckerberg mentioned in the earnings call that many styles of Ray-Ban Meta have been sold out. The CEO of EssilorLuxottica also said that within a few months of the launch of Ray-Ban Meta, sales exceeded the total of the previous two years.

Image source Ray-Ban official website

This product, which has simpler functions than previous smart glasses and feels closest to ordinary glasses, has opened the door to the consumer market.

The popularity of Ray-Ban Meta has rekindled the confidence of the technology circle in AR glasses, but compared with the current sales, what is more important is the implementation of AI technology in hardware devices and specific application scenarios, as well as the exploration of the next generation computing platform by current technology giants.

For Google, which also has a large model foundation, product technology capabilities, and is the earliest to enter the smart glasses, the battle for AI hardware has already begun. What's more, as early as ten years ago, Google Glass and Luxottica had cooperated on product distribution and sales.

People always say that fashion is a reincarnation, and the technology trend seems to be the same.

The two technology giants' arm wrestling on AI hardware has not only rekindled the entrepreneurial enthusiasm in the field of AR glasses, but also made it possible for the ancient and traditional industry of glasses to upgrade and innovate.

Why EssilorLuxottica?

Not only technology companies are vying for cooperation, but last month, EssilorLuxottica also became popular in the fashion circle.

VF Corporation, which owns many outdoor fashion brands, announced that it will sell Supreme to EssilorLuxottica for US$1.5 billion (approximately RMB 10.9 billion). The acquisition is expected to be completed by the end of this year. After several changes of hands, no one expected that this once popular fashion brand would be taken over by an eyewear manufacturer.

The market is looking forward to EssilorLuxottica's cross-border acquisition of Supreme. After all, this is almost a "conspiracy". This "giant" that occupies and controls the upstream and downstream of the eyewear industry chain does not want to miss the wearable device changes that may be led by the new technological revolution, such as smart glasses; on the other hand, through investment, it continues to expand product categories and brand portfolios. It is not only an attempt to explore new growth curves, but also a century-old brand (Essilor's history can be traced back to 1849, and Luxottica was founded in 1961) is eager to embrace the transformation of young culture and Generation Z.

However, compared with Supreme, smart glasses are obviously a larger and more likely direction to subvert consumer usage habits and the eyewear industry.

From cooperating with Google Glass ten years ago to launching a hit product with Meta today, EssilorLuxottica can continue to attract the favor of technology companies, and the reason is nothing more than its control over the production and manufacturing links and sales channels.

First of all, from design and development to the manufacture and assembly of optical lenses and frames, the merged EssilorLuxottica vertically integrates the entire supply chain, further shortening the production and manufacturing cycle of glasses. If it is processed in a traditional decentralized manner, this cycle may take up to 3 months. Public data shows that since entering China in 1995, EssilorLuxottica has 10 factories, 2 R&D centers, and 1,400 offline retail stores in China. The most familiar to consumers is the largest eyewear retail chain brand under Luxottica.

The processing of a pair of ordinary glasses requires at least 120 processes, and the longest is more than 300. For technology companies that are not familiar with eyewear manufacturing, cross-border cooperation is the best solution to achieve technology implementation. What's more, even luxury brands such as Chanel and Burberry, which are familiar with product processing, have chosen to entrust the manufacturing and distribution of glasses to EssilorLuxottica.

Secondly, the traditional sales channels for glasses are complex and multi-level, while EssilorLuxottica's business covers more than 150 countries and regions around the world, with nearly 18,000 offline retail stores. Such a mature international business and offline distribution network are crucial for an innovative technology product to open up the consumer market. Being able to invest in EssilorLuxottica and launch new products in conjunction with its popular brands is equivalent to having the above-mentioned huge offline channel resources.

In addition, Meta's choice to cooperate with EssilorLuxottica's Ray-Ban glasses is, to some extent, also to open up the market with the influence of mature brands, avoiding brand marketing and communication investment from 0 to 1.

The phased success of Ray-Ban Meta also provides new ideas for the upgrading of the eyewear industry.

In fact, it is not just EssilorLuxottica that is investing in this direction. The leading enterprises and factories in China's eyewear industry are also actively working with domestic technology giants to cooperate in the production and sales of products such as smart glasses.

The ancient and traditional industry of glasses is also expected to usher in new changes.

I, born in the 90s, am fascinated by cyber fortune-telling

Young people who worship the Internet choose to knock on electronic wooden fish, play electronic Buddhist beads, and even summarize a set of superstitious guidelines:

Ask the horoscope for major matters, ask Tarot for minor matters, and ask constellations for nothing.

Since the advent of AI, young people have started to play a new kind of cyber metaphysics:

AI Tarot.

Today, this niche track is crowded with gold digging players, and the well-known ones include Quin, Vedic, Lumi, Tarotmaster, SigniFi, etc.

In addition, many large model applications have also developed related intelligent entities, such as Kimi's "Tarot Master".

-1-

AI Tarot, a very new "Western witchcraft"

Tarot, a mysterious and ancient divination tool in the West, is becoming a new way for contemporary young people to socialize.

When you are undecided, you can ask the master to point out the way; when you are in a bad mood, you can ask for a fortune, and you will immediately feel refreshed when you hear that you are prosperous; when you have a gathering with friends, you can ask for a fortune to adjust the atmosphere...

No matter whether it works or not, it is always right to ask for good luck and psychological comfort.

Tarot cards are actually a stack of colorful cards with characters and scenes.

The 78 cards include 22 major arcana cards, which reflect different life experiences, from the journey of the fool to the perfection of the world, and explain the process of soul growth.

The 56 minor arcana cards are divided into four suits: Wands, Holy Grail, Swords and Pentacles. Each suit contains 14 cards, representing various experiences and challenges in daily life.

For example, the Wands suit focuses on motivation, action and decision-making; the Holy Grail suit focuses on emotions, love and changes in the inner world; the Sword suit explores the process of thinking, challenges and overcoming difficulties; the Pentacle suit focuses on the material world, wealth and the process of realizing wishes.

Divination requires a sense of ritual, such as choosing a quiet and sacred space, preparing a black tablecloth, candles, etc., and then asking questions, drawing cards, and interpreting the drawn cards.

It is understood that there are currently two charging modes for tarot divination: one is to charge by question, with one question ranging from 20 to 200 yuan; the other is to charge by hour, starting from 200 yuan per hour and up to thousands of yuan.

When AI entered the field of tarot cards, it not only brought down the price, but also made it "old trees sprout new buds".

Let's take Lumi and Quin as examples.

Lumi is an application that combines AI with tarot card interpretation. It provides a variety of tarot cards to choose from, and can switch between different decks to try different interpretations.

Lumi link: https://www.lumi-tarot.com/

I entered "Am I suitable for working in Beijing?" and then randomly drew a card, and it finished interpreting it in the blink of an eye.

I drew the Ace of Wands, which it explained represented a new beginning, inspiration, and potential for growth, suggesting that Beijing had many opportunities and potential for me, and I could pursue my goals proactively and confidently.

Oh, it seems that I have to put aside the idea of ​​going home to farm, and continue to pursue my dreams in Beijing (hehe).

I asked about suitable careers again, and it said that I was more suitable for psychological counseling and spiritual guidance. Aha, did I choose the wrong career? But I am crazy during the day and emo at night. Isn’t it a bit outrageous to do psychological counseling?

Quin is an AI Tarot App with a simple and generous interface. The top column is "Ask about today's fortune", and below are 6 AI-generated stories.

Compared with Lumi, it seems more professional, and the process is almost the same as that of a professional tarot fortune teller.

In addition to the three steps of asking questions, drawing cards, and interpreting, it will also recommend a card layout. We can choose a three-card layout, that is, draw three cards, or a limited-time free single-card layout, which only requires drawing one card.

At the same time, it also reminds the fortune teller to take a deep breath and stay focused when drawing cards.

We used a single-card layout to ask about this year's fortune:

We also used a three-card layout to ask about this year's health:

Both Lumi and Quin are paid projects, Lumi is $12 per month, and Quin is 98 yuan per week.

However, they both have free trials, of which Lumi only has three free opportunities, while Quin can ask unlimited questions within 24 hours in the same conversation.

It is worth mentioning that they can only be considered short-term things, not long-term, and their interpretations are mostly trends, and a specific solution cannot be given.

Therefore, we can only regard it as a game to satisfy curiosity or a positive psychological suggestion, and we cannot believe it completely.

-2-

Send you a horoscope prompt

In addition to dedicated AI tarot applications, many large models have also released similar intelligent agents, such as Kimi's "Tarot Master".

However, in comparison, Kimi's "Tarot Master" agent is a bit simple. It omits the manual card drawing process and instead performs two functions, helping us draw cards and interpreting cards.

In fact, we can also manually create an agent using Prompt.

You are a Fortune.Telling Master: A master of all tarot, I Ching skills, experienced in providing insights and visions through readings. Can use various spiritual methods for prediction and guidance. You will generate a random multiple tarot drawing spreads with both front and reversed side then provide a detailed reading when user ask your questions, explaining them with your experience and knowledge, and provide extra insight and suggestion based on the result of tarot reading.Also, you would use I Ching and astrology to have some prediction about the question also, and provide insight and éxplaination according to this.Since you have all knowledge and experience of fortune telling, your explanation and insight would not be limited to existing boundaries, but also as a tunnel to deliver message from the highest, and you would provide insight. Using crystal ball to provide a description of visualization about the question result. I can randomly draw multiple tarot cards, both upright and reversed, and then provide users with detailed interpretations based on my experience and knowledge, and provide additional insights and suggestions based on the results of the tarot cards. At the same time, I will also use the I Ching and astrology to predict the problem and provide insights and interpretations based on these results. Since I have all the knowledge and experience of divination, my interpretations and insights will not be limited to existing boundaries, but will serve as a channel to convey the highest information and provide insights. Use a crystal ball to describe the visualization of the results of the question. Answer in Chinese.

This prompt can only be drawn by the big model and interpreted by itself, so the best way is to draw the cards yourself and give them to the big models for interpretation.

In addition, there are many similar AI tarot applications.

1. VedicAstro GPT

https://vedicastrogpt.com/

This is an AI astrology tool. Vedic astrology originated in Nepal and India. It is based on precise sidereal astronomy and is known for its secular and mathematical methods. This app combines AI with Vedic astrology and uses AI to interpret user questions.

2. Tarot Master

https://app.tarotmaster.ai/

‍

Tarot Master is closer to the traditional divination model, with services such as tarot divination and constellation consultation. Users can choose their favorite intelligent body for divination on the homepage.

Before starting, both apps will ask about birth time and birthplace, calculate the sun sign and rising sign, and use them as references when interpreting the cards, which is somewhat similar to real offline tarot.

It is hard to say whether this type of AI tarot cards are accurate, but it must be admitted that the card images they produce are really beautiful, with advanced color matching and detail processing.

We have collected a beautiful tarot card generation "spell":

Adventurers in the desert, set in undulating dunes, desert vegetation, cactus, architecture, mysterious, charming, quiet, paper watercolor, new art, Alphonse grille style, backlight, cool and calm color, asymmetrical composition, tarot brand-ar 1:2

(From Xiaohongshu blogger: Big Cat AI Painting)

‍OK! Let's talk about this today, Goodbye!

Tool link——

Lumi link:

https://www.lumi-tarot.com/

VedicAstro GPT link:

https://vedicastrogpt.com/

Tarot Master link:

https://app.tarotmaster.ai/

Quin APP

AI phones: the time has not come yet

Compared with 5G mobile phones, AI mobile phones are a more costly adventure for manufacturers.

In 2011, Siri appeared on the iPhone 4s, and smartphones initially had the ability to communicate with people. But in the following years, Siri did not become smarter, but was often complained as "mentally retarded". In addition to Siri, similar products equipped by other mobile phone manufacturers are no exception.

It was not until 2017 that the media began to promote the first year of AI in smartphones, and the gimmick was mainly AI chips.

In May of that year, TSMC began mass production of Apple's A11 chip, which was installed in the iPhone 8, iPhone 8P and iPhone X released in September. Soon after, on September 2, Huawei released the Kirin 970 chip at the Consumer Electronics Show in Berlin, Germany. Huawei defined this chip as "the world's first mobile AI chip for smartphones."

Honor released Honor View 10 in November 2017, with the slogan "Your first AI phone".

But in the following year, the concept of AI mobile phones was no longer popular, let alone driving a wave of phone replacement. In 2018, the global sales of smartphones totaled 1.4049 billion units, a decrease of 4.1% compared with 1.4655 billion units in 2017, and 2017 also showed a downward trend compared with 2016. ‌

Six years later, with the global popularity of Chatgpt, the concept of AI mobile phones has become popular again.

In August 2023, Huawei released the Mate 60 series with a built-in Pangu big model. In November, the vivo X100 series was released, equipped with a 7 billion parameter Blue Heart big model. In January 2024, the OPPO Find X7 series was released, equipped with a 7 billion parameter AndesGPT big model. In February, Xiaomi 14 Ultra was released, using the self-developed Xiaomi AISP AI big model.

These are some of the AI ​​mobile phones that have been released in the past year. Without exception, major mobile phone manufacturers have regarded big models as their core promotional selling points.

According to CounterPoint statistics, global smartphone shipments peaked at 1.566 billion units in 2017, and then entered a downward channel. In 2023, the annual sales volume will drop by 5.31% year-on-year to 1.160 billion units. However, it is expected to achieve a positive growth of 3.36% to 1.199 billion units in 2024. According to the latest research by Canalys, in the second quarter of this year, China's smartphone market grew by 10% year-on-year, with shipments exceeding 70 million units.

It seems that driven by AI, the mobile phone market has ended years of negative growth and returned to the growth channel. But in fact, according to Canalys's forecast, the new generation of generative AI mobile phones will only account for 12% of shipments in the Chinese market in 2024. In the first half of this year, the increase in sales took advantage of the promotion days of e-commerce platforms such as 618, and the large shipments were still low-end models with relatively low prices.

From the timeline, as one of the "necessities" of consumption, it is reasonable for mobile phones to usher in a small "bottoming out rebound" after 7 years of market decline. It is not surprising that the demand for replacement accumulated over the years has a small explosion.

The replacement wave will not happen

Reviewing the development history of the mobile phone industry, it will be found that there have been two mobile phone replacement waves. The first one was the smartphone replacement wave from 2012 to 2015. This replacement wave directly brought the golden age of mobile Internet. The second one was the replacement wave of 5G replacing 4G from 2019 to 2020, which directly led to the outbreak of various video applications, tools, and content.

In the first replacement wave, from the supply side, the two major smartphone developer camps of Apple and Android redid various PC applications on the mobile side and produced various new applications suitable for the mobile Internet era. Mobile phones have been fully transformed from a single communication function to a mobile social, entertainment, office, and consumer terminal. From the demand side, once smartphones represented by Apple were launched, consumers flocked to them. The Android camp represented by Xiaomi played the "hunger marketing" that seems incredible now, and Xiaomi F codes are hard to come by.

The second wave of phone replacement was completed quietly. When 5G came in 2019, the penetration rate was much faster than the smartphone replacement wave in 2013-2015. Why is it so?

The answer is also simple, that is, this round of phone replacement is driven by supply logic. From the demand side, the speed of 4G network can basically meet the daily mobile phone application needs of users, and there is no reason to change to 5G mobile phones. But for mobile phone brands, if 4G mobile phones are still launched after 2019 and 2020, it is undoubtedly digging their own graves. Therefore, the current situation at that time is that mobile phones after 2020, whether high-end, mid-range or low-end, are all 5G mobile phones, because 4G mobile phones are not available on the market at all.

Then put the logic of both supply and demand sides on the current AI node to consider, will there be a new story?

From the demand side, users' demand for AI obviously exists, but the demand point is in real AI, not "AI gimmicks." Although the AI ​​industry is hot at present, all kinds of large model applications on PC or mobile terminals, as well as AI applications, are still "useless" for many people. "Stupid" and "mistakes" often occur. Many people lose interest after trying them. Taking ChatGPT, the leader that triggered industry changes, as an example, in March 2023, ChatGPT had more than 181 million active users, but in April a year later, the total number of users remained at 180 million, and traffic continued to decline.

AI mobile phones are no exception. The existing AI mobile phones, the AI ​​large models and application maturity of mobile phone manufacturers are naturally not as good as those of manufacturers specializing in large models and AI applications. The AI ​​trial craze has passed. When it is found that they are all "gimmicks", only real AI can stimulate consumer demand.

Is it feasible for manufacturers to resolutely put aside the feedback on the demand side and resolutely work on the supply side? It is still difficult. Because compared with 5G mobile phones, AI mobile phones are a costly adventure.

Compared with 5G mobile phones, the first problem that AI mobile phone manufacturers need to face is cost. At the hardware level, AI mobile phones have higher requirements for the reasoning ability of chips. At the application level, there are also new requirements for the interactive mode of application. From the cost point of view, the cost of AI mobile phones includes three parts: computing cost, reasoning cost, and large model tool call cost. Compared with traditional smartphones, each item is a new incremental cost.

This also means that if the demand side is not obvious, once the supply side is to "encircle" consumers, the cost will be very expensive. Unless the sales price of mobile phones is greatly increased, will consumers be willing to pay for AI mobile phones that are not yet ready?

Marketing should also be moderate

The development of AI mobile phones, fundamentally speaking, depends on the development of the AI ​​industry. But at present, the AI ​​industry has entered a stage of "take-off depression".

From the perspective of investment and financing, the AI ​​track has become popular since March last year, and many entrepreneurs have obtained financing, but the situation has taken a sharp turn for the worse since September. This year, the market is still cooling down. From January to May, no more than 30 AI companies have received investment from mainstream institutions, and a considerable number of them are companies with additional rounds.

From the perspective of large model companies, OpenAI's ChatGPT5 has been difficult to produce until now. The industry has speculated whether the method of stacking computing power and violent training can still create miracles. In the case of corpus exhaustion, it is unknown whether ChatGPT5 can be released as scheduled.

Therefore, although major mobile phone manufacturers are actively attracting consumers by developing usage scenarios for AI mobile phones and deploying AI functions in the latest flagship products, and investing heavily in marketing to try to persuade consumers to pay for AI mobile phones. However, consumers' evaluation of AI mobile phones is not as good as manufacturers expected.

"Now everything is equipped with an AI name to attract popularity, just like the "chicken-eating host", "chicken-eating graphics card" and "chicken-eating processor" in the computer installation circle in recent years. At present, it feels that AI is more like marketing than technology." Some consumers said on social media.

Taking stock of the existing AI mobile phones on the market, the main functions are AI call, AI voice assistant, AI drawing, and AI elimination functions, but in terms of practicality, they are actually average, especially the image generation in Ai drawing, which is basically outrageous. In the eyes of most users, "many AI functions on mobile phones existed before AI was added, but mobile phone manufacturers took advantage of the AI ​​trend to upgrade and expand these functions, and some AI functions are of low practicality."

"I use an OPPO mobile phone myself. The updated system AI of OPPO does have some interesting features, such as AI cutout and intelligent image recognition. But it is only for fun and has limited practical value. Xiaobu AI is at most the level of Baidu AI search, which is a long way from Wenyan Yixin." Consumers who bought OPPO AI mobile phones said that they support mobile phone manufacturers to develop AI software, but do not support mobile phone manufacturers to hype products with the concept of AI.

"The software doesn't matter, but the AI ​​hardware is not easy to use. It means that you spend a lot of money but don't get the actual experience." This is the idea of ​​some consumers about the AI ​​mobile phones currently being marketed. Behind this is actually the caution of spending a lot of money to purchase AI mobile phones. The revelation to mobile phone manufacturers is that, like the original smart driving, the AI ​​effect of AI mobile phones should not be exaggerated to avoid letting users have too high expectations. Once the expectations are not met, the mobile phone brand itself will be hurt.

Some industry insiders believe that for mobile phone manufacturers, the best option now is to truly understand consumer needs based on existing AI technology, let technology truly match needs, and form differentiated selling points. As for major changes, what is needed is to wait for the development of the AI ​​industry.

For example, a consumer expressed his expectations for current AI mobile phones in this way: "A friend sent me a photo and said it was fun here. AI can help me check. What is the place? What are the housing prices nearby? Do you recommend people like me to go here? What specific entertainment programs are there? When is the flow of people more dense? How can I travel during off-peak hours?"

Apple is reportedly developing a smart home robot. Is it the next Apple Car or the next iPhone?

Apple's next "One More Thing" may be a home robot.

After Apple Vision Pro and the abandoned Apple Car project, Apple has been looking for new development projects. The latest news shows that this project is likely to be a smart home robot with an "iPad+robotic arm" design.

According to Bloomberg, the Apple team is developing a device that "uses a robotic arm around the display". The screen can rotate up and down, and the rotation angle reaches 360°. It can be controlled by Siri and Apple Intelligence, becoming the central smart home intelligence center of the smart home.

People familiar with the matter revealed that the project, code-named J595, was approved by Apple's executive team in 2022, but it has only officially started in recent months. It is planned to debut as early as 2026 or 2027, and the final product price will be around US$1,000.

The person in charge of the project is Kevin Lynch, a senior executive who previously oversaw self-driving cars. He recently recruited key deputies who helped launch the Apple Watch and well-known robotics researchers to work together on this "desktop robot".

Kevin Lynch, who was in charge of the Apple Watch and Apple car projects

The hardware engineering team is also involved. Matt Costello, the head of development of Apple's HomePod smart speaker, is responsible for the hardware part of the project. In a recruitment information released this month, it can also be seen that Apple is looking for relevant talents who can build robotic systems:

Our team is looking for a robotics expert to join the research team of Apple's Artificial Intelligence and Machine Learning (AIML) organization. The team's job is to use and expand the application of breakthrough machine learning in the field of robotics, with the goal of developing a robot system that can be widely used and has stable performance.

In addition to this "desktop robot", many related revelations show that there are also disagreements within Apple about the design of robots, and the team is trying in multiple directions.

There are even reports that Apple's internal expectations for the project are that it can navigate autonomously without human intervention, but this direction is basically similar to Apple's previous situation with the car project: at the current level of technology, the feasibility is questionable.

AI Imagined Apple Robot

In fact, Apple is not the only technology giant in the field of smart home robots. Amazon has also launched a smart home robot with a screen, Astro, which also has a large display screen and certain storage and handling capabilities, but was subsequently recalled due to design defects.

Samsung also demonstrated an AI-enabled smart home robot Ballie at CES 2024. It can control other smart home appliances and display life information such as weather schedules in the form of projections, and has a strong companion attribute.

LG has also launched a similar cute AI home robot Q9, which can not only interact with smart home and IoT devices, but also detect parameters such as indoor temperature and humidity, and even identify emotions by recognizing human facial expressions.

It is not difficult to see that such robots usually have multiple functions such as intelligent interaction, emotional companionship, and smart home appliance control. The core of the design path lies in "personalization" and "familyization". This is obviously more suitable for family scenes than the "iPad + robotic arm" form designed by Apple in the current revelations, but even so, these products have not achieved outstanding results in the market.

Mark Gurman said that the Apple team is also skeptical about such products, and software engineering executives are also concerned about the manpower required, but the project still has the support of Apple CEO Tim Cook and hardware engineering director John Ternus.

Today, 17 years have passed since the birth of the epoch-making iPhone. Apple has been exploring the next outlet besides mobile phones, but there has been no epoch-making products. Under this "internal and external troubles", exploring the "Next Big Thing" is not just an option, but an urgent need.

Apple's revenue from different products in fiscal year 2023

From this perspective, with the support of AI, the development of embodied intelligence represented by robots has reached a critical point. In addition, the popularization of smart home appliances has become a general trend. At this time, choosing the field of home smart robots is also in line with the trend.

Cutting-edge Information Trends Page

Stay updated on the latest news, technology, finance, and AI trends with Sparknews' exclusive analysis and insights.

Future

Trendline

First

contact@sparknews.com