Written by | Neocortex Group
Edit | Jeff Wang
This week's AI scene has been relatively uneventful, with neither blockbuster model updates nor huge funding, but the noteworthy news revolves around a single theme - how to make generative AI more useful.
First of all, ChatGPT is now available without registration. As the most popular generative AI product, it is free to use without registration and would have been unimaginable half a year ago. At that time, OpenAI CEO Sam Altman even said that because of the shortage of computing power, he hoped that users would reduce the use of ChatGPT. It seems that the lack of computing power has been alleviated, partly because OpenAI and Microsoft have purchased enough GPUs, but also because users are finding generative AI not that useful. Since reaching a peak of 1.8 billion traffic in May 2023, ChatGPT's user growth has been slowing.
To this day, most of the people around me have experienced generative AI to some degree, chatting with it, asking it to answer a few tricky questions, or asking it to help generate a few pictures, and feeling the imagination of AI, but it doesn't stop there, it has not become an integral part of most people's work. In my case, I tried to get Claude and ChatGPT to put together the earnings data, but they gave a lot of errors, and I couldn't get them to correct them, because I didn't even know if the wrong numbers were because they were citing the wrong data or if they were just making it up. For me, the most commonly used function of chatbots is translation, and it's only slightly better than the free Google Translate or Youdao Translate, but at least in this use case, the output of the AI is stable.
In addition to translation, another stable scenario for chatbots should be writing code. This week's news that Alibaba Cloud is fully implementing AI programming internally has a lot of heat, and Alibaba Cloud even gave this AI official employee the number AI001, and said that it hopes that the company's next 20% of the code will be completed by it. But its practical AI to help write code is not a new thing for a long time, just as the news of Alibaba Cloud swiped the screen, Baidu's official account also issued a special article to introduce Baidu's version of the AI assistant Comate, and secretly poked and poked that it has been working for more than a year, but did not get a work card, because it is okay to brush your face in Baidu (implying that Baidu's working environment is more intelligent). In fact, AI coding and translation have something in common, both of which are to convert one language into another, and this kind of work content is highly closed, with relatively fixed standard answers, and AI does not need to understand what is happening today to give high-quality results.
However, when ChatGPT came out, we didn't expect it to help us translate files or write code, we were looking for a really useful personal assistant, a super smart Siri, Alexa, Xiaodu or Xiaoai. Putting generative AI on mobile phones is nothing new, and almost every domestic Android manufacturer has launched its own large models, but none of them are smart enough to be indispensable.
The good news is that Apple hasn't given up on its efforts to productize generative AI. A paper published by Apple this week shows that it is developing a model that can understand what's on the current phone screen and answer questions based on that information. A few weeks ago, it was reported that Apple had given up on developing its own large model and chose to cooperate with a third-party company, but it seems that Apple has not given up on R&D investment in artificial intelligence, but unlike other technology companies that focus on R&D capabilities of the basic model, Apple focuses on how to better apply the basic model to the iPhone. At the same time, Samsung also said this week that it has plans to continue to upgrade its voice assistant Bixby, although it fully adopted Google's Gemini model when it released the new Galaxy flagship at the beginning of the year, so much so that there are voices that Samsung has abandoned its self-developed Bixby assistant.
However, while most companies are trying to apply more AI to their businesses, Amazon announced this week that it is abandoning Just Walk Out, a technology that was first developed by Amazon in 2016. At that time, Amazon arranged a large number of cameras and sensors in the unmanned store, customers scan the code to enter, and after selecting the goods, they can leave the store directly, and the system will automatically track and identify the goods purchased by customers, and complete the deduction in the background. However, the reality is skinny, and there is news that in order to ensure the accuracy of the results, Amazon has hired a team of 1,000 people in India to watch and tag the videos. Judging from the data in 2022, 700 out of every 1,000 transactions need to be reviewed by the team, while Amazon originally hoped that the number would be 20 to 50, and artificial intelligence has really become artificial + intelligent here.
The following is a summary of the smart news worth watching in the past week, produced by the Neocortex team.
Key Points
ToC Applications
ChatGPT can be used without registration;
Apple wants the big model to read your iPhone screen;
Samsung plans to upgrade voice assistant Bixby;
Job number AI001, Alibaba Cloud fully implements AI code writing internally;
ToB Applications
OpenAI's new speech engine can synthesize simulated human voice speech;
Amazon's offline retail stores will abandon "check-out-and-go" technology;
Models and chips
Perplexity plans to sell advertising;
Microsoft and OpenAI plan to invest $100 billion to develop AI supercomputers;
Intel's foundry business lost $7 billion in 2023.
ToC Applications
ChatGPT can be used without registration
On April 1, OpenAI announced in a blog post that users can use ChatGPT directly without registration. There is little change to the ChatGPT interface after it is opened, but the chat history will be cleared when the interface is closed, and registered users can save and view the historical chat history. This opening is only available for the free version of GPT-3.5, and the paid products GPT-4, DALL· E 3 still requires a login to your account.
ChatGPT, which once set the record for the fastest growth in users, has been slowing since its peak of 1.8 billion traffic in May 2023, according to analytics firm SimilarWeb. At the same time, Anthropic's Claude and Google's Gemini have both made breakthroughs in user growth. OpenAI's opening up of ChatGPT may attract more users.
OpenAI said it will collect more usage data to improve the model, and users can turn off the feature in the settings if they will. OpenAI has also introduced additional content protections for users to access ChatGPT directly without registering, "blocking prompts and generated content in a broader category," though it doesn't specify what those categories are.
Apple wants the big model to read your iPhone screen
According to a paper published March 29, Apple researchers have developed a new model, ReALM, that can understand the information on the screen, as well as the context in which the conversation and the conversation is generated, allowing it to interact more naturally with voice assistants.
ReALM stands for Reference Resolution As Language Modeling, and the study focuses on how to get large models to understand visual elements on mobile phone screens. Compared with the existing solutions, ReALM has achieved substantial performance improvements. In the paper, Apple's research team writes that it's critical for chat assistants to understand context. Reading what the user is seeing on the screen is a critical step in ensuring that voice action is truly implemented.
Recently, it was reported that Apple is considering introducing a third-party model to implement smart features on the iPhone, and Google's Gemini and Baidu's Wenxin Yiyan are both potential partners. However, judging from this paper, Apple has not given up on R&D investment in artificial intelligence, but unlike other technology companies that focus on R&D capabilities of the basic model, Apple is more focused on how to better apply the basic model to the iPhone.
Samsung plans to upgrade its voice assistant Bixby
On April 1, Choi Won-joon, executive vice president of Samsung's mobile business, revealed in an interview with CNBC that Samsung is actively considering adding ChatGPT-like generative artificial intelligence functions to its virtual assistant Bixby to improve user experience and service quality.
When it previously announced several AI features, Samsung didn't advertise Bixby too much. For the Galaxy S24 to be released in January 2024, Samsung has chosen to work with Google to integrate Google's Gemini series models into the flagship model. As a result, the industry speculates that Samsung may abandon Bixby. But Samsung said in this interview that it has not given up on Bixby. Instead, the company will continue to invest resources to strengthen Bixby's AI capabilities to remain competitive in the market.
Bixby was launched in 2017 with the Samsung Galaxy S8 smartphone and now covers a wide range of devices such as smartphones, smartwatches, and home appliances. The software offers a variety of features, including real-time translation or restaurant recommendations. But voice assistants are often less conversational and rely more on the user asking questions and giving answers.
Samsung has not announced a specific timeline for Bixby to get new AI features, but said it is working hard to move forward.
The job number is AI001, and Alibaba Cloud fully implements AI code writing internally
According to a report by the National Business Daily on April 2, Alibaba Cloud is fully implementing AI programming internally, including using Tongyi Lingcode to assist programmers in writing code, reading code, checking bugs, optimizing code, etc., and assigning a formal employee ID AI001 to Tongyi Lingcode. According to Alibaba Cloud sources, 20% of the company's code will be written by Tongyi Lingcode in the future, but programmers are still the core of R&D, and they will have more time to focus on system design and core business development. Within Alibaba Cloud, Tongyi Lingcode has served as a code assistant in various development links, taking API development and testing work as an example, it can shorten dozens of minutes of manual writing and testing to seconds. Next, Alibaba Cloud will fully configure the Tongyi Lingcode plug-in in multiple internal development tools for all employees to use. Jointly developed by Alibaba Cloud and Tongyi Lab, and released at the Apsara Conference in 2023, Tongyi Lingcode supports more than 200 programming languages such as Java, Python, Go, JavaScript, TypeScript, C/C++, and C#, and has been downloaded more than 2 million times.
ToB Applications
OpenAI's new speech engine can synthesize simulated human voices
On March 29, OpenAI shared its progress in AI speech synthesis tools. In the sample demonstration session, OpenAI used a 15-second audio sample and a text sample, and the speech engine then used the voice of the speaker in the audio sample to naturally read out the content in the text sample. OpenAI said that the speech synthesis tool can not only create emotional, realistic voices, but also switch between different languages and read text samples in Chinese, German, French, etc. with the same accent.
In fact, OpenAI launched a speech synthesis tool as early as the end of 2022, which has been applied to ChatGPT's voice function and can be obtained through APIs. However, the features of this update will not be available to the public at this time. OpenAI said there are serious risks associated with AI-synthesized voices, especially in U.S. election years, and government agencies have expressed concerns.
As a result, the update is currently only open to a small range of trusted partners, including companies in education, healthtech, and other fields, who can only reproduce their voices with the explicit permission of the sound owner, and must label the resulting audio as AI-synthesized. OpenAI will do more testing and discussion based on the results of customer trials to decide whether and how to deploy speech synthesis technology at scale. The company also said that the large-scale deployment of speech synthesis technology is premised on voice owners knowing that their voice is being used by speech synthesis services and that the system is able to detect and block the generation of celebrity-sounding voices.
Amazon's offline retail stores will abandon check-out-and-go technology
On April 2, it was reported that Amazon's next-generation Amazon Fresh store will abandon the "checkout and go" Just Walk Out technology and instead adopt the "smart shopping cart" Dash Carts system, and the existing stores will also be transformed accordingly. At present, this change only affects Amazon Fresh, and does not affect the relatively small convenience store Amazon Go, some small Amazon Fresh stores in the UK, and third-party retailers.
Just Walk Out, which stands for "checkout and go," is a technology first developed by Amazon in 2016. At that time, Amazon arranged a large number of cameras and sensors in the unmanned store, and after customers scanned the code to enter, they could leave directly after selecting the goods, and the system would automatically track and identify the goods purchased by the customer, and complete the deduction in the background. The technology was officially launched in 2018 in Amazon's "new retail business" Amazon Go store, and has since been applied to some Amazon Fresh stores.
Dash Carts is a checkout feature launched by Amazon in the Amazon Go store after 2020. Amazon can automatically charge customers to their Amazon account when they leave the store by scanning the items they place in their shopping carts, which also have touchscreen displays real-time receipts, product promotions, and more.
Both technologies make the shopping experience smoother by automating the checkout process, but Just Walk Out requires a large number of cameras and sensors to be placed throughout the store, while Dash Carts only loads sensing elements and lens equipment in the shopping cart, avoiding the "surveillance" caused by too many cameras and reducing the difficulty of technology deployment.
Amazon says the decision to move away from the Just Walk Out technology was primarily based on customer feedback. They learned that consumers enjoyed the benefits of skip-the-line checkouts with Just Walk Out, but also wanted to be able to see their total spend in real time. In addition, Amazon is also facing the high cost of Just Walk Out technology.
Models and chips
Perplexity plans to sell ads
On April 2, AI search company Perplexity said it would start selling ads. Currently, Perplexity answers user questions based on web resources and incorporates videos, images, and data from partners such as Yelp in the answers generated. The answer also provides a link to the source and suggests "relevant questions that may be of interest" to the user. These "may be of interest" questions account for 40% of all Perplexity searches. The company plans to introduce native advertising in this section, allowing brands to influence the setting of these issues.
Ads will be in the form of "native units". "Native unit" is a form of advertising that integrates with the content style and form of the platform, aiming to provide an effective delivery of advertising information without disrupting the user experience. After introducing brand advertising, Perplexity also needed to prove itself to brands in terms of the platform's user scale, brand safety strategy, access to audience insights, and the effectiveness of targeting.
Perplexity was established in August 2022 with the aim of using AI technology to create an ad-free "Google search". In January, Perplexity closed a $73.6 million Series B funding round led by Institutional Venture Partners at a valuation of $520 million. In early March, it was reported that Perplexity was raising a new capital round at a valuation of about $1 billion, but it has not yet been officially announced. It currently has more than 10 million monthly active users, and the main revenue comes from the user's monthly subscription fee of $20. Despite Perplexity's slogan that "search should stay away from advertising," CEO Dmitry Shevelenko said advertising has always been an important part of the company's strategy.
Microsoft and OpenAI plan to invest $100 billion to develop an AI supercomputer
On March 29, media reported that according to people familiar with the matter, Microsoft and OpenAI are closely planning a data center project. The project, which aims to build an AI supercomputer called Stargate, is expected to cost up to $100 billion, 100 times the cost of some of the largest data centers today. The computer will be equipped with millions of dedicated server chips to power OpenAI's AI technology. It is reported that the company's executives plan to launch the Stargate project as early as 2028 and hope to gradually expand the scale and scope of the project by 2030. Microsoft will fund the project, most of which will be used to buy the chips that will be used to build the supercomputer. Participants revealed that Microsoft plans to use AI chips made by Nvidia in the project.
Intel's foundry business lost $7 billion in 2023
In an April 2 filing with the U.S. Securities and Exchange Commission (SEC), Intel disclosed for the first time the financials of its semiconductor manufacturing business, commonly referred to as its "foundry business."
According to the report, in 2023, Intel's foundry business revenue will be $18.9 billion and the loss will be $7 billion, compared to the revenue of $27.5 billion and the loss of $5.2 billion in 2022, Intel has made less and lost more in the past year.
In response to the widening losses, Intel CEO Pat Gelsinger said that the company's foundry business losses are expected to peak in 2024 and are expected to break even between the current quarter and the end of 2030.
He noted that some poor decisions have placed a significant burden on the foundry business, especially the previous decision not to adopt extreme ultraviolet (EUV) technology. Because of this, Intel has to outsource about 30% of its wafer production to chip foundries such as TSMC. Currently, Intel is focusing on reducing the proportion of outsourced production to about 20%, and Intel will turn to extreme ultraviolet (EUV) technology to meet the growing production demand.
At the same time, in order to recover the gap with major competitors such as TSMC and Samsung, Intel plans to build or expand chip factories worth $100 billion in four US states. Previously, the company also announced that Microsoft would be a customer of its foundry services, with $15 billion in orders associated with foundry services.
-END-
![After ChatGPT became popular, a valuable operation needs to have 6 abilities.[fig]](data:image/gif;base64,R0lGODlhAQABAIAAAP///wAAACwAAAAAAQABAAACAkQBADs=)