AI Graveyard, and 738 Dead AI Projects

The Industry of artificial intelligence has seen a number of “death projects” fail due to technical challenges or lack of market acceptance. There are 738 names on the death list Among them are some former star AI projects, such as Whisper.ai, the AI ​​speech recognition product launched by OpenAI, Stable Diffusion’s well-known shell websites FreewayML and StockAI, and Neeva, the AI ​​search engine that was once regarded as a “Google competitor.” “Throughout the process, we discovered that building a search engine is one thing, but convincing ordinary users to switch to better options is another,” Neeva co-founders Sridhar Ramaswamy and Vivek Raghunathan wrote in a blog post announcing Neeva’s closure. This AI project death list comes from a subpage of the AI ​​tool aggregation website “DANG!” – AI Graveyard. Most of the projects on the AI ​​Graveyard page state the project background, functions, technical applications, and death time, just like epitaphs engraved in cyberspace.
According to the statistics as of June 2024, this list includes 738 AI projects that have died or stopped running. Specifically:

• There are 271 products such as Chatbot and AI writing , accounting for about 37% ;
• There are 216 AI painting, AI design and other cultural products, accounting for about 29%
• There are 73 AI voice, AI video and other cultural audio and video products, accounting for about 10%;
• Other products such as AI code tools and SEO optimization tools account for about 33% .

Why did Ai project die or discontinue?

It’s not that they died from “putting on a shell”, but that they died from “not putting on a shell well”

“putting on a shell” mean implementing security measures or safeguards to protect the system and its data. “Not putting on a shell well” would then imply inadequate protection or vulnerabilities in the AI system, which could lead to various issues or even failures.

• Insufficient Risk Management: Projects that fail to anticipate and mitigate risks can face unforeseen obstacles that stall development or cause abandonment.
• Poor Stakeholder Communication: A project may flounder if there is misalignment between the development team, stakeholders, and end-users. It is important for the success of any project to have clear communication channels where all parties involved are able to express their ideas freely with one another so as to know every person’s perspective towards achieving the goal set.
• Ethical and Regulatory Issues: Ignoring ethical considerations or failing to meet legal standards could land you in court or destroy public confidence hence leading to project failure.
• Financial Mismanagement: Inadequate allocation of funds coupled with unrealistic budgeting might lead into bankruptcy before achieving desired results.
• Technology Integration Issues: Artificial intelligence systems often need to be integrated within other existing technologies or platforms. If this is not done successfully then it means that the intended purpose would not be met thereby affecting its success rate among users.

• Technical Challenges: AI projects often encounter technical hurdles such as insufficient data quality, limitations within the algorithms themselves, or even problems related to computational resources.

In the AI ​​cemetery, many are “shelled” products.
For example, with AI Pickup Lines, users can generate 10 pickup lines for free every day, or choose a paid subscription of $9.99/month or $99.99/month to generate an unlimited number of pickup lines and flexibly choose any keywords. In addition, users can also choose to purchase a comprehensive database for $499.99 to obtain more than 100,000 pieces of content pickup content covering various topics and styles.
However, AI Pickup Lines did not survive long. It went online at the end of 2022 and closed in early 2023.

The main reason for the closure of AI Pickup Lines is that it is more entertaining than practical, and as more and more competing products have enhanced their large model capabilities, such products that access a single API will find it difficult to cope with complex and changeable social scenarios in life, and the barriers will become steadily decreasing. in addition although such products may generate revenue through advertising or one-time purchases, long-term user retention and profitability are insufficient, and they will eventually be closed due to insufficient income. The death of “shell” products such as AI Weekly Report Generator and AI Girlfriend Coaxing Copy Generator also follows this logic.
However, “shelling” is not a derogatory term.
“Disenchantment of Large Model Shells: Questioning Shells, Understanding Shells”: Non-AI practitioners regard “shells” as a scourge; real AI practitioners are very secretive about “shells”. However since “shells” themselves do not have a clear and accurate definition, the industry’s understanding of “shells” is also different for every thousand readers.

What is the five levels of advancement in large AI models cover all scenarios related to their development and deployment?

These five levels of advancement basically cover every scenario of the large model “shelling”.

• First level : directly quote the OpenAI interface, what ChatGPT answers, what the shell product answers. Volume UI, form, cost.
• Second level: build prompts. The big model can be compared to research and development (R&D), and prompts can be compared to requirement documents. The clearer the requirement documents are, the more accurate the R&D can be. Shell products can accumulate their own high-quality prompts, with high-quality and high-volume prompts for distribution.
• Third level: Embedding specific data sets. Vectorize specific data sets and build your own vector database in some scenarios to answer questions that ChatGPT cannot answer. For example, vertical fields, private data, etc. Embedding can encode paragraph text into vectors of fixed dimensions, which facilitates comparison of semantic similarity. Compared with Prompt, it can perform more accurate retrieval and obtain more professional answers.
• Fourth level: Fine-Tuning. Use high-quality question-answering data for secondary training to make the model more suitable for understanding specific tasks. Compared with Embedding and Prompt, which both consume a large number of tokens, fine-tuning is to train the large model itself, consume fewer tokens, and respond faster.
• fifth level is considered as If we also include the pre-training by imitating the Llama2 architecture.

Although they are all “shells”, the degree of “shelling” is different. Now there are many “shell” products that have survived and even thrived because of their sophisticated design and good pricing strategies.
For example, the AI ​​assistant Monica mentioned above is a product upgraded through the acquisition of ChatGPT for Google. It has built-in large models such as GPT-4o, GPT-4, Gemini, and Claude Llama 3. With its excellent dialogue, search, summary, translation, table processing, image editing and other functions, it has gained millions of users in a few months.
Another example is Perplexity, an AI search product known as the “king of shells”. Due to its extremely fast response speed, accurate question responses, and archivable multi-round interactions, it has been ranked in the top ten of a16z’s Top 50 Gen Al Web Products for many years. As of mid-May 2024, the number of daily visits to its products reached 3 million, more than five times the number a year ago.
Perplexity co-founder and CEO Aravind Srinivas said earlier this year: ” People can think of Perplexity as an AI ‘shell’ product, but becoming a ‘shell’ product with 100,000 users is obviously more meaningful than having your own model but no users. “

There are also many AI “shell” products produced by independent developers that have performed well.
For example, David Bressler, who has many years of market research experience, built an Excel formula generator called formula bot through the no-code platform Bubble, and earned $26,000 in ARR (annual recurring revenue); there are also independent developers who have made AI chatbot platform Chatbase by deepening their roots in niche fields, with an MRR (monthly recurring revenue) of approximately $64,000; in addition, there are Magnific (image super-resolution and enhancement tool, which accumulated 720,000 users in 5 months and was later acquired by Freepik), PDF.ai (understand the content of PDF documents through questions and answers, which recovered its costs within 6 days of its launch and successfully exceeded $300,000 in AAR in September 2023) and other excellent AI products.
Therefore, many AI products do not die because of the “shelling”.