TikTok Executives Start Businesses, and There Are Still New Stories to Tell in AI Coding | Focus Analysis

Text | Zhou Xinyu

Editor | Su Jianxun

Multimodal, hardware, and AI Coding constitute the hottest landscape in the AI track in the second half of 2024.

The former two, where many powerful players are entrenched, are familiar to everyone. While AI Coding, which was unknown last year, has suddenly become one of the few new stories in the AI application layer.

"Intelligent Emergence" has learned that Chen Zhijie, the head of TikTok's algorithm, plans to leave and start a business recently, focusing on the direction of AI Coding. At the same time, Chen Zhijie joined ByteDance in 2020 with a job level of 5 - 1.

Chen Zhijie. Source: Chen Zhijie's LinkedIn

AI Coding, A New Trend for Globalization

Chen Zhijie, who comes from Baidu and ByteDance, has many years of team-leading experience in the field of recommendation algorithms. From 2010 to 2019, he served as the chief technical architect at Baidu, mainly responsible for advertising algorithms.

After joining ByteDance in 2020, he was first responsible for the algorithms of domestic products such as Douyin, Toutiao, and Xigua Video. After 2022, Chen Zhijie began to serve as the head of TikTok's algorithm, mainly responsible for the recommendation system and data science team, supporting TikTok's businesses such as videos, live broadcasts, e-commerce, and content distribution.

In 2024, with the explosion of AI Coding products such as "AI Programmer" Devin, Cursor, OpenAI o1, etc. globally, the AI Coding track has also been rediscovered and become the new top trend in AI applications.

For example, Li Yanhong once mentioned at the Q3 director meeting that the AI Coding application Comate "is still a good direction in the future". Alibaba Cloud and ByteDance have also launched AI Coding products such as Tongyi Lingma and MarsCode, respectively.

An algorithm engineer from Dark Side of the Moon once told "Intelligent Emergence" that compared to long texts, programming is currently the field with the least "hallucinations" and the most stable output quality. Currently, it has been commercially implemented in scenarios such as code completion and generation, and the "assisted programming" of Kimi is a favored function internally.

In addition to the entry of old players, some AI Coding startup projects have also emerged in this track since this year. For example, AIGCode, founded by former investor Su Wen of Hua Chuang Capital in January 2024, has already received 2 rounds of financing.

Previously, some media reported that Ming Chaoping, the former product head of Dark Side of the Moon's overseas product Noisee, started an AI Coding startup project "New Yan Yi Ma", and started financing with an angel round valuation of 50 million US dollars, almost on par with the first-round valuations of domestic AI unicorns such as Dark Side of the Moon.

However, the consensus of many AI Coding startups is still to go global.

Several insiders told "Intelligent Emergence" that although Chen Zhijie's AI Coding project plans to be implemented in China first, the ultimate goal is to go global.

"The advantage of Coding is that there are no cultural and language restrictions. Unlike language models and language-related applications, it does not require a lot of adjustments according to the cultures of different regions and different users." An investor told "Intelligent Emergence", "AI Coding products are very easy to be standardized, and standardization means lowering the threshold for large-scale expansion."

As two key factors in the development of AI applications, model capabilities and the willingness of downstream customers to pay, currently AI Coding has a more fertile growth soil overseas.

"The models with the strongest Coding capabilities are in the United States, o1 and Claude, especially Claude, which can basically reliably execute code-related tasks with only natural language input." An AI practitioner commented.

He believes that most domestic models cannot directly and reliably understand and generate code, especially long code segments, and developers often need to assist with complex systems such as RAG.

An overseas case that successfully "changed its fate" with Claude is: The US model unicorn Anthropic behind Claude recently disclosed that after integrating the model Claude 3.5 Sonnet, the AI Coding product Bolt of the US enterprise StackBlitz saw its ARR (Annual Recurring Revenue) increase from 0 to 4 million US dollars in just 4 weeks, with more than 100,000 weekly users.

The US AI Coding product Bolt. Source: Screenshot from Bolt's official website

Bolt co-founder and CEO Eric Simons said: "Claude 3.5 Sonnet is the core technology that makes this product a reality."

The "Easiest" to Replicate Million-Dollar Annual Revenue

The entrepreneurial trend and investment fever in AI Coding, the reason for this is that this is one of the few AI application scenarios that can make money at present.

In any industry, algorithm engineers are a necessary reserve, and the improvement of programming efficiency by large models is clearly visible. A recent survey report by Menlo Ventures shows that in 2024, the adoption rate of AI Coding-related products ranked first in the main scenarios, reaching 51%.

The downstream demand directly makes AI Coding a good business with an annual revenue of up to tens of millions of dollars.

According to a report by The Information in December 2024, in the past 3 months, the annualized income of Anthropic's software development and code generation business has achieved a 10 - fold growth. Microsoft also revealed in its earnings conference in July this year that the ARR of GitHub Copilot is about 300 million US dollars, contributing a 40% growth to Github's annual performance.

An even more significant growth occurred in Cursor. According to the estimate of the research institution Sacra, Cursor's ARR reached 65 million US dollars in November 2024, with a year-on-year growth of 6400%.

The operation interface of Cursor. Source: Cursor's official website

Many AI Coding startup companies want to replicate the same growth myth. Several investors commented that the market for AI Coding is large enough, and there are many细分 scenarios, and the homogeneous competition is not that fierce yet.

But an AI practitioner also commented to "Intelligent Emergence" that what AI Coding projects often compete on is not the model capability, but the productization capability.

At present, most AI Coding entrants either benchmark Cursor to create Copilot products that can have natural language conversations. For this type of product, in addition to the technology itself, it also tests how the product form integrates with the customer's IDE (Integrated Development Environment) and naturally integrates into the programmer's workflow.

"Currently, a strong productization capability is the core competitiveness of many Chinese AI startups with a voice overseas." An AI practitioner commented to "Intelligent Emergence", "This is also an important reason why investors are optimistic about the overseas expansion of domestic AI Coding projects."

Welcome to communicate!

Welcome to follow!