Why do they die?
Author|Wang Yi
Editor|Wang Bo
There are 738 names on this death list.
Among them, there are some former star AI projects, such as OpenAI's AI speech recognition product Whisper.ai, Stable Diffusion's well-known shell sites FreewayML and StockAI, and the AI search engine Neeva, which was once regarded as a "Google competitor".
"Throughout the process, we've discovered that it's one thing to build a search engine, and quite another to convince the average user to switch to a better option." Neeva co-founders Sridhar Ramaswamy and Vivek Raghunathan wrote in a blog post announcing Neeva's closure.
This list of AI project deaths comes from the AI tool aggregator website "DANG! AI Graveyard. Most of the items on the AI Graveyard page state the background, functionality, technology application, and time of death, like an epitaph engraved in Cyberspace.
AI Graveyard, Image Credit: DANG!
According to the statistics of "Jiazi Lightyear", as of June 2024, this list includes a total of 738 AI projects that have died or stopped working.
- There were 271 literary products such as chatbots and AI writing, accounting for about 37%;
- There are a total of 216 Wensheng picture products such as AI painting and AI design, accounting for about 29%
- There are 73 Wensheng audio and video products such as AI voice and AI video, accounting for about 10%;
- Other products such as AI code tools and SEO optimization tools account for about 33%.
Why do they die?
1. Died not from "shelling", but from "failing to set the shell"
In the AI graveyard, many of them are "casing" products.
For example, AI Pickup Lines (AI pickup lines), users can use it to generate 10 pickup copies per day for free, or they can choose a paid subscription of $9.99/month or $99.99/month, so as to generate an unlimited number of pickup lines and flexibly choose any keyword; In addition, users have the option to purchase a comprehensive database for $499.99, which gives them access to more than 100,000 content pick-ups covering a variety of topics and styles.
However, AI Pickup Lines didn't survive long, going live at the end of 2022 and shutting down in early 2023.
AI Pickup Lines，图片来源：AI Graveyard
The main reason for the closure of AI Pickup Lines is that entertainment is greater than practicality, and with the enhancement of the capabilities of more and more competing large models, it is difficult for such products to connect to a single API to cope with the complex and changeable social scenarios in life, and the barriers will become thinner and thinner. In addition, while such products may generate revenue through advertising or one-time purchases, long-term user retention and profitability are insufficient, and eventually they are shut down beyond their means. The death of "shell" products such as AI weekly report generators and AI coaxing girlfriend copywriting generators is also based on this logic.
However, "casing" is not a pejorative term.
"Jiazi Lightyear" once mentioned in the article "Disenchantment of Large Model Shells: Questioning Shells and Understanding Shells": Non-AI practitioners regard "shells" as beasts; Real AI practitioners are secretive about the "shell". However, due to the fact that there is no clear and accurate definition of "casing" itself, the industry's understanding of "casing" is also a thousand Hamlets for a thousand readers.
Suki, the former designer of Yuque and the co-founder of the current AI assistant Monica, shared the fourfold advancement of "shell" on the spot:
- First order: directly quoting the OpenAI interface, what ChatGPT answers, and what the shell product answers. Volume UI, form, cost.
- Second: Build a Prompt. The larger the model can be compared to R&D, and the Prompt can be compared to the requirements document, the clearer the requirements document, the more accurate the R&D implementation. Casing products can accumulate their own high-quality Prompts, with high quality Prompts and distributed Prompts.
- Order 3: Embedding a specific dataset. Vectorize specific datasets and build your own vector database in some scenarios to answer questions that ChatGPT can't. For example, verticals, private data, etc. Embedding can encode paragraph text into fixed-dimensional vectors, which is convenient for semantic similarity comparison, and can be retrieved more accurately to obtain more professional answers than Prompt.
- Fourth Order: Fine-Tuning. Use high-quality Q&A data for secondary training to better match the model's understanding of a specific task. Compared with Embedding and Prompt, which consume a large number of tokens, fine-tuning is to train the large model itself, which consumes fewer tokens and responds faster.
If you include the pre-training of imitating the Llama2 architecture, it can be regarded as the fifth order. These five-fold advancements basically include every scene of the "casing" of the large model.
Although they are all "shells", the degree of "shells" is different, and now there are many "shell" products that have survived or even survived well because of their clever designs and good pricing strategies.
For example, the AI assistant Monica mentioned above is an upgraded product through the acquisition of ChatGPT for Google. It has built-in GPT-4o, GPT-4, Gemini, Claude Llama 3 and other large models, and has gained millions of users in a few months because of its good dialogue, search, summarization, translation, table processing, picture editing and other functions.
Another example is Perplexity, an AI search product known as the "king of the shell", which has been ranked in the top 10 of a16z's Top 50 Gen Al Web Products for many years due to its extremely fast response speed, accurate question response, and archivable multi-round interaction. As of mid-May 2024, the number of daily visits to its products reached 3 million, an increase of more than five times compared to a year ago.
Aravind Srinivas, co-founder and CEO of Perplexity, said earlier this year: "One can think of Perplexity as an AI 'shell' product, but it obviously makes more sense to be a 'shell' product with 100,000 users than to have your own model without users." ”
Perplexity page, Image Credit: Perplexity
There are also many AI "shell" products made by independent developers that also perform well.
For example, David Bressler, who has many years of experience in market research, built an Excel formula generator called formula bot through the no-code platform Bubble, earning $26,000 in ARR (annual recurring income); There are also independent developers who have made Chatbase, an AI chatbot platform, with an MRR (monthly recurring revenue) of about $64,000. In addition, there are excellent AI products such as Magnific (an image super-resolution and enhancement tool that accumulated 720,000 users in 5 months and was later acquired by Freepik) and PDF.ai (which learned about the content of pdf documents through Q&A, which paid for itself within 6 days of its launch, and successfully exceeded the AAR of $300,000 in September 2023).
Therefore, many AI products die not from "shelling", but from "failing to shell".
2. Sell memberships, sell the number of experiences, and then what?
In addition to "not having a good shell", the second major cause of death of products in the AI cemetery is the single profit model, and there are two main pricing forms for related products: charging members and buying points for the number of experiences.
For example, Wenshengtu Product Purephotos.app and AnimeAI.lol provide points purchase services for enterprise users, while the latter packages products and services into different packages. Perhaps realizing that enterprise users don't make money, since May 2024, Purephotos has begun to experiment with the recently popular "pay-as-you-go" billing model, where the more images generated by users, the cheaper the cost is allocated to a single image.
Purephotos.app定价策略 图源：AI Graveyard
AnimeAI.lol: AI Graveyard
The same goes for Photofix. It is an AI photo editing tool with features like image enhancement, removal of superfluous characters, and literal images. The product is divided into "Basic" and "Premium", with the basic version costing between $0.39 and $5.99 per image and the premium version costing between $0.49 and $9.99 per image.
Photofix定价策略 图源:AI Graveyard
And even if Purephotos.app later added the "pay-as-you-go" charging model, it failed to recover the decline because it changed too late.
After sorting out the pricing strategies of some Wensheng Diagram products in some AI cemeteries, "Jiazi Lightyear" found that most of these products follow the purchase of credits, and the larger the amount of points purchased by users, the cheaper the price will be amortized to each generation task. However, the bottom layer of these products is mostly the API of several mainstream Wensheng graph models, but the pricing is not much lower than the price of its underlying model - take Patience AI as an example, its underlying access to Stable Diffusion, Waifu Diffusion, DALL-E and other models, the product price is $15 1000 points, about $0.015 / point, if it is calculated according to the cost of generating a picture 2 points, Each image costs about $0.03, which is $0.02 for DALLE-2 to generate a single image (the highest specification of 1024 x 1024).
Pricing of DALL-E 2 images with different specifications, image source: OpenAI
With such an uneconomical price, if there is no major breakthrough in the product or the underlying technology, then it is not surprising that these products will eventually die.
Even if the product and design are sophisticated enough, and the pricing mechanism is set up reasonably, if the giants are gone, the startups will also suffer.
This has to say that Neeva, an AI search engine that was once regarded as a "Google competitor", is now lying in the AI graveyard.
Co-founded in 2019 by Dehar Ramaswamy, Google's former vice president of advertising, and Vivek Lagunathan, former vice president of monetization at Youtube, Neeva has gained a lot of attention since its launch because of its focus on ad-free, tracker-free, and prioritizing user privacy.
Unlike many search engine products that choose to plug into Google or Bing's APIs, Neeva chose to build the search stack from scratch and assembled a small team of 50 people. Neeva has introduced shopping pages with larger images and useful comparison information, while prioritizing UGC content from sites like Reddit and Quora, and sports search results have become beautiful full-screen scoreboards that can take users directly to the web when searching for specific keywords.
Compared to Google, Neeva's interface is cleaner and cleaner, such as replacing the blue links of the traditional search results page with a more intuitive page and placing more emphasis on UGC content.
Google (left) and Neeva (right) search results comparison, image source: Medium
Officially launched in the U.S. in June 2021 at $4.95 per month, Neeva quickly attracted a large number of users in a short period of time, growing to 500,000 monthly active users within four months of its launch. By early 2022, Neeva had integrated large language models into its search stack, becoming the first search engine to provide citations for real-time AI answers to most queries. In order to expand the user base in 2022, the basic version of Neeva began to be available to users for free.
Neeva product page, image source: TechCrunch
In order to follow the trend of generative AI and also seek better growth, in January 2023, Neeva launched NeevaAI, a generative AI search product. It was one of the first search engines to integrate AI capabilities to answer queries with summaries and citations, and NeevaAI surpassed Microsoft's New Bing and Google's AI Search beta in the month of its launch.
Neeva was also once the darling of capital, having received investment from well-known VCs such as Sequoia Capital and Greylock Partners, with a cumulative financing amount of $77.5 million.
However, after 4 years of operation, Neeva couldn't hold on: in April 2023, Neeva announced the permanent closure of its search engine. Ramaswamy said in a post that due to the huge challenges in attracting new users and the current difficult economic environment, Neeva will close the web and consumer search products and start exploring the To B business. In May 2023, Snowflake, a cloud database company, acquired Neeva for about $150 million.
It is undeniable that "AI search" is a good product form, from Perplexity in the United States to Mita AI in China, the stability of traffic and the growing number of users have verified the real market demand for such products. However, in the case that giants such as Google and Microsoft have occupied a strong ecological niche, the competition between AI search startups is an extremely capital-heavy game, and in order to make users abandon their original habits and turn to new search products, they not only need to be unique in terms of product power, but also need to spend a lot of money on marketing and promotion, which puts forward high requirements for the financing ability of AI search startups.
At the same time, the ability to find a suitable monetization model is also one of the factors that determine the success or failure of AI search products: advertising alone may be slower to monetize, while other monetization models (such as subscriptions) are difficult to attract a large number of users due to a certain payment threshold, which is also the reason why Neeva's user growth has slowed down after the launch of the paid version.
3. How not to walk into the AI graveyard
In 2006, Y Combinator, a well-known startup incubator, summarized 18 ways to die for startups, including burning too much money, not making money, no computing power, and no product differentiation. From these projects in the AI cemetery, "Jiazi Lightyear" found that the 18 ways to die 18 years ago are still fatal. Even star AI products that have soared in the past will suddenly hit a wall at some point and become the dust of history.
The AI Graveyard is only a few small and medium-sized companies, but some of the larger star AI companies are also dying or gradually falling silent. These companies were valued at hundreds of millions or billions of dollars in their glory days, but they have failed in the last two years – Inflection AI is a prime example.
In May 2023, the company released its first chatbot, Pi, which can have personalized conversations with users through apps or web pages, WhatsApp, Instagram, Facebook.
Pi's page, image credit: TechCrunch
In an interview with Bloomberg News, Inflection AI co-founder Mustafa Suleyman said that while Inflection AI has attracted a lot of interest from investors, including Microsoft, and has 1 million active daily active users, it has yet to find an effective business model.
Pi says it has always been backed by venture capital and has no business model, image source: Pi
The example of Inflection can serve as a wake-up call for entrepreneurs that when an AI application company's core product fails to deliver a convincing enough performance, and the model level is under pressure from an arms race, then the original logic of "model-driven AI applications" may no longer hold.
Fu Sheng, chairman and CEO of Cheetah Mobile and chairman of Orion Star, once told Jiazi Lightyear: "I now firmly believe that the product is looking for the market, and using the market to extrapolate what kind of technology you need." In the past, there was a superstition about the quality of technology, and in the past, many people who did AI came from universities, research institutes or large factories, and those who came out of it may think that the paper is the key and the architecture is the most critical, but in fact, the first demand of the market is the most critical. ”
Fu Sheng, Chairman and CEO of Cheetah Mobile and Chairman of Orion Star, photo source: "Jiazi Gravity"
Kai-Fu Lee, founder and CEO of 0100000 in May this year, put forward a concept of "TC-PMF", he believes that the concept of PMF (product-market fit) can no longer fully define AI-First (AI first) entrepreneurship based on large models, and should introduce Technology (technology) and Cost (cost) to form a four-dimensional concept, that is, "TC-PMF" (Product-Market-Technology-Cost Fit, Technology cost X product-market fit).
In Kai-Fu Lee's view, large models are expensive from training to service, and the shortage of computing power is a collective challenge for the track, and the industry should work together to avoid falling into the irrational ofo-style bloodshed and money-burning play, so that large models can use healthy and benign ROI (return on investment) to store energy for a long-distance run and rush to China's AI 2.0 revolution. "Doing technology cost x product market fit, especially reasoning cost reduction is a 'moving target', which is a hundred times more difficult than traditional PMF." Kai-Fu Lee explained.
Kai-Fu Lee, founder and CEO of Zero One Things, picture source: Zero One Things
All in all, Inflection's failure cannot be simply attributed to the failure of the product, but the failure to find the "TC-PMF", blindly driven by financing, ignoring the free cash flow, technical feasibility and cost controllability, even if the market data of the product is good, it will inevitably be "acquired" because of the shortcomings of commercialization.
Of course, in addition to learning from the underdogs, the question most people are probably more concerned about is: what kind of AI startups can succeed today?
Taken together, there are two types of companies that are more likely to survive:
- The first type is enterprises that truly understand the needs and pain points of B-end or C-end users;
- The second category is enterprises that make functions that cannot be replaced by generative AI products such as ChatGPT and Midjourney, and penetrate and penetrate a certain segment of the scene.
A typical case in the first type of enterprise is AnswerAI.
AnswerAI is an AI Tutor (AI tutor) product for the North American market, and its main function is to take pictures to solve problems + discourse. The founder, Zhou Li, graduated from Peking University with a master's degree in 2007 and has worked as the founder of Tiger Map, Wandoupod, Kika Input Method and LiveIn.
Different from the previous wave of AI Tutor1.0 products, which focused on "photo search", Answer AI is an AI Tutor2.0 product, which can not only search for questions, but also solve problems, and give an argumentation process on the basis of giving answers. The questions are not limited to the questions in the question bank, but can solve new questions that have never been seen before within the scope of ability, which greatly solves the pain points of student users who "have answers but can't understand ideas" and "can't meet new questions". After the product came out, Answer AI received rave reviews on the Internet, with many users saying "This is the best AI product I've ever used".
Answer AI user feedback, image source: Google Play
According to data released by Data.ai on May 21, five of the top 20 educational apps in the U.S. app store are AI agents that help students complete homework, and Answer AI is one of them.
At present, Tutor AI has more than 2 million users worldwide, 80% of whom are from high schools and universities in the United States, temporarily ranking first in the AI Tutor category in North America, and this year's ARR is expected to reach $5 million.
The typical representative of the second type of enterprise is a URL shortening tool company named Bitly, and vidyo.ai with similar ideas.
Headquartered in New York and founded in 2008 by Peter Stern, Bitly offers long chain shortening, dynamic QR codes, and custom link shortening.
Bitly, image source: Bitly official website
This doesn't look like a product made by a company in the era of generative AI, but Bitly has been rated by many as the "best short-chain tool" for its simple interactive operation, stable service capabilities, and built-in statistical functions. Previously, in order to save users 140 characters of space, X used the TinyURL service to quickly and automatically shorten long URLs, and gained a lot of exposure and extra traffic.
Bitly didn't choose the PLG (product-driven growth) route of To C at first, but set its sights on enterprise customers, selling "small screwdrivers" to large enterprises through SLG (sales-driven growth).
Thanks to a strong freemium service, Bitly quickly captured most of the global market, achieving nearly $20 million in ARR in 2018; After adjusting its strategy to PLG in 2020, Bitly has achieved "skyrocketing" growth.
The history of Bitly ARR, image source: Medium
Today, this ancient and low-key company has completely broken the curse of "SaaS can't do To C in the United States" and has broken through the ARR of $100 million; Even the emergence of ChatGPT in 2022 and the fact that many people began to use AI tools such as ChatGPT to shorten the chain did not shake Bitly's growth fundamentals.
The reason is simple: AI tools such as ChatGPT occasionally generate randomly and occasionally use Bitly's domain name generation when faced with users' requirements for "long chain to short chain". Generally speaking, short links generated using Bitly domains are not shortened in real Bitly accounts, so the generated short links are often opened with error pages. Bitly staff also stated on their user service page, "If you're using AI tools to help you with copywriting, be sure to check your links before posting or printing your text." ”
Bitly staff explained the reasons for the short chain errors generated by some AI tools, image source: Bitly
A product with a similar idea to Bitly is the vidyo.ai of AI video editing tool platforms.
vidyo.ai can automatically convert long videos into short videos with one click, users only need to upload the video, or paste the link to the vidyo.ai, and it will automatically edit the highlights of the long video in the cloud, and can also intelligently track faces and add subtitles, and support the format adapted to various short video platforms. vidyo.ai can reduce video editing and processing time by up to 90%, and what used to take 3 people nearly a week to complete can now be done in just 15 minutes with vidyo.ai.
At present, AI-generated video products such as Runway, Pika, PixVerse, and Sora all emphasize the "generation" ability of video, which is closer to the production side, but ignores the needs of the consumer side - after the video is produced, it is to serve users after all, and it is short video that is truly marketable and popular with users. vidyo.ai has seized the field ignored by the giants, and instead of video generation, it has found another way to make AI video editing products that are more "market-oriented", and then occupied a place in the ecological niche of AI video.
After joining Entrepreneur First, an international investment incubator, in 2021, vidyo.ai received $1.1 million in seed funding in 2022, and has accumulated 500,000+ users in 2023, with an ARR of $1.5 million.
Then set your sights back on the country.
"Jiazi Lightyear" once paid attention to a company that "found a new way and made a breakthrough at a single point" - Haina AI. Unlike many HR SaaS companies that do both AI interviews and BPO (business process optimization), Haina AI focuses on the single scenario of "AI interview assessment" and provides customers with quantitative talent assessment methodology and AI automatic evaluation algorithm.
Haina AI disassembles talents into more than 200 dimensions and more than 4,000 behavioral characteristics, and based on the latest open-source large model, uses hundreds of millions of high-quality industry data to refine the industry AI model, and automatically interviews and evaluates the talent's appearance, communication expression, comprehensive quality, professional skills, psychological status, industry experience, etc.
Haina AI product service flow chart, image source: Haina AI
Since its establishment five years ago, most of the top 3 leading groups in China's top 8 industries have used Haina AI, such as SF Express, Wal-Mart, Luckin, etc., and each group interviews 100,000-1 million people every year, all of which are completed by Haina AI, and the customer repurchase rate has reached 100%.
Liang Gongjun, founder and CEO of Haina AI, once told "Jiazi Lightyear" that AI recruitment has developed very slowly in the past five years, and like most to B companies in the past decade, it is difficult to achieve revenue of more than 100 million yuan, because there is no way to scale and the flashpoint cannot come. But now the flashpoint of AI interviews has arrived. In this field, companies that focus on single-point scenarios and have completed PMFs will be the first to stand out in the next six months, and they have gone through the growth process from 0 to 1 and from 1 to 5. This will be followed by a rapid burst of 5 to 10, to 100, and to 1000.
The commonality between Bitly, vidyo.ai, and Haina AI is that they have all found scenarios that generative AI giants cannot reach or cannot do well, and grasp the subdivided needs in this scenario, penetrate and penetrate them; In other words, they have found their own unique foothold outside the range of the big factories.
Wang Xiaochuan, founder and CEO of Baichuan Intelligence, mentioned at a media communication conference in May this year that Baichuan Intelligence hopes to make products that are "beyond the range of large manufacturers". "First of all, the market size of to B in China's business environment is 10 times smaller than that of to C; To B collects RMB and spends USD. The big factories will be involved in this matter, but I didn't expect everyone to be so ruthless, and they all got involved to 0, which must be within the range of the big factories. And we're definitely going to differentiate. Wang Xiaochuan said.
Wu Bingjian, a partner of Heart Capital, once publicly expressed a point of view: the key word of Mobile (mobile Internet) is "competition", large-scale money-burning competition, and only those who win the competition have a chance to run out; The key word of AI is "engulfing", predicting the development of the model, and those who are not engulfed have a chance to escape.
In today's less hot market environment, financing may be a problem faced by every AI project; But since there is no financing, what AI startups can do may be to "not be swallowed up" and try to "run and survive" on their own. As long as you outperform some people first, you may be able to outperform everyone in the future.
"I have always told entrepreneurs, don't pursue technology leadership, don't be entangled in how much of the product is AI, how much is manual, because the technology iteration is too fast, we must pursue whether we can achieve commercial quality, and grab customers, scenarios, and data in our own hands." Zhu Xiaohu, managing partner of GSR Ventures, said at the "AI Creation Era - 2024 Jiazi Gravity X Technology Industry New Trends" conference held by Jiazi Lightyear in May this year.
Zhu Xiaohu also expressed a "particularly obvious feeling" in a sharing this week - this year will be the beginning of AIGC's entrepreneurial return to the essence of business.
This feels right, but Jiazi Lightyear believes that returning to the essence of business is not the same as just doing applications, and a single large-scale model company is also valuable.
Just look at the OpenAI plan to block China's API incident, as well as the "moving plan" quickly launched by various large model companies. Although some large model companies have very little API revenue, they have also joined the battle for customers this time.
The market environment and opportunities are changing rapidly, and for an AI company, a "moat" is not a prerequisite, but only meaningful when the company really has something worth "protecting".
(Cover image source: "Jiazi Lightyear" is generated using AI tools)