AI Workflows, Code Editors, and Platform Disruption | Original
Table of Contents
- AI Thoughts
- AI lacks true intelligence or depth
- Machine learning is advanced applied calculus
- LLMs struggle with structured file formats
- Open source eliminates technological secrecy
- Text-based tools face AI disruption first
- New Platforms Powered by AI Workflows
- AI workflows automate multilingual content generation
- Users submit prompts for format conversion
- Platforms enable content refinement and summarization
- Customizable AI workflows via keyword settings
- AI handles content transformation end-to-end
- The Next Direction of AI Code Editors
- Cloud integration critical for CI/CD workflows
- A/B testing enhances AI-generated content
- RLHF extends to real-world deployment feedback
- Human feedback refines imperfect AI outputs
- Prompt optimization beats output correction
AI Thoughts
Last updated in August 2025
-
Satya Nadella mentioned Jevons paradox. It is worth to learn.
-
Yin Wang: There is no “intelligence” in artificial intelligence, no “neural” in neural network, no “learning” in machine learning, and no “depth” in deep learning. There is no “depth” in deep learning. What really works in this field is called “calculus”. So I prefer to call this field “differentiable computing”, and the process of building models is called “differentiable programming”.
-
Yin Wang: Machine learning is really useful, one might even say beautiful theory, because it is simply calculus after a makeover! It is the old and great theory of Newton, Leibniz, in a simpler, elegant and powerful form. Machine learning is basically the use of calculus to derive and fit some functions, and deep learning is the fitting of more complex functions.
-
Currently, large language models can’t filter by file language like YAML or Python. However, a significant portion of information in the real world is organized this way. This means that we could train large language models using files.
-
For training large language models, we could develop a system that finds exact matches. Perhaps we can combine the KMP (Knuth-Morris-Pratt) search algorithm with transformer architecture to enhance search capabilities.
-
There are no technological secrets. Open source will reveal all the secrets that are closely guarded.
-
AI will affect many tools, including indirect ones. People say they won’t need Figma to draw prototypes, they will directly go to code. I think Postman will be similar; people will directly use Python or other scripts to call or test APIs.
-
One reason we don’t use Postman or Figma in the AI era is that their functionalities can’t be generated through text. They also lack a command + K shortcut to trigger component replacement.
-
User interfaces are becoming a barrier in the AI era. Why upgrade Postman to be AI-powered for testing applications when we can directly use Python’s requests library or other programming languages to test code, as the latter will be powered by AI?
-
Why upgrade Figma to be AI-powered for UI creation when code-based UI generation, enhanced by AI, offers a more direct and potentially powerful approach?
-
LLMs will change text-related applications first, like Google, Search Engines, Text Editors and Writing Tools, Quizlet, Zendesk, DeepL, Medium, WordPress, Trello, Asana, Gmail, GitHub, Goodreads, Duolingo, and Feedly.
-
Conversely, LLMs are unlikely to revolutionize technologies like Git, Linux, ffmpeg, mobile phones, hardware, browsers, operating systems, or voice and video calls. These technologies are code-centric, and their code is not easily generated by AI, unlike API testing tools such as Postman.
-
Technologies with more code are hard to be revolutionized by AI, like OpenOffice, MySQL, Mozilla Firefox, Chromium, VLC Media Player, Qt Framework, LLVM/Clang, and GNOME. If AI could help make these technologies, they wouldn’t be replaced. AI should be helping to make better technologies, and to do that, AI will need more computing power to generate the same magnitude of code.
-
There are two ways LLMs can bring change: first, by altering the content or data within a platform or software, such as content translation in apps like TikTok; second, by directly replacing certain software or platforms, like Postman or Google Search, including Google Translate.
-
There are two ways AI audio tools can bring change: first, by altering the content or data within a platform or software, such as generating audio books for Audiable; second, by directly replacing certain software or platforms, for example, the Sing songs app, as AI can now perform the same tasks humans do, making it easier for people to sing songs as a hobby.
-
There are several ways to measure how AI impacts current software or platforms. One way is to measure how much data or content can be generated or improved by AI, either partly or completely. Another way is to measure how much code can be written or improved by AI, either partly or completely. This means we use what AI generates to improve current platforms. Additionally, AI can help invent new software and platforms.
-
There are three types of products, Generative AI products, the products that use APIs of Generative AI products, other products.
-
One product idea is to use AI to accumulate real-time information, news, or updates from social platforms such as Reddit, GitHub Trending, Twitter Trending, Quora Trending, and Zhihu Trending. Users can use prompts to customize the feed or even add specific social accounts.
-
There are five important types of data: text, image, audio, video, and code.
-
Other important types of data include numerical, geospatial, biometric, sensor, transactional, metadata, time-series, structured, unstructured, semi-structured, health, environmental, log, network, and behavioral data.
-
Google is still better for website indexing, especially if you want to download software or a document from a specific site. It works like a domain search. You’re not using it to find information, but rather to navigate to other sites to perform tasks. An LLM may not have the latest download links.
-
Google works like a domain search; if you want to go to a Maven repository site to check the latest version, you can use it.
-
Google remains useful for image search, whereas LLMs excel at text generation. Still, people often prefer real images to verify hardware details, dimensions, object shapes, or a person’s appearance.
-
AI chatbots are popular because text is harder to process than images. People prefer real images over AI-generated ones since images are easier to understand at a glance. However, AI image generation has untapped potential - users could ask AI to show different angles, zoom in on faces, or magnify circuit board details. Since people primarily work with text rather than images, there’s significant room for growth in AI image tools.
-
AI excels at explaining concepts and facilitating understanding. Furthermore, users can pose questions regarding any specific detail. This is likely the most significant utility of AI tools.
-
I used AI to learn about Large Language Models. The moment it helped me understand K, Q, and V was wonderful.
-
The reason I prefer to use Ubuntu since the release of LLM is that the rich and colorful apps in macOS are less appealing to me. I prefer writing my programs and doing everything through the terminal and text.
-
AI can be evaluated by how well it can update a pom.xml or requirements.txt file to the latest version, update libraries, and perform checks. This process can involve a significant amount of work and can sometimes be complex.
-
In the AI era, programming languages that have better performance and robustness are more important and will be more popular, while syntax is less important. This is because LLM will help generate code, making it less of a hassle as long as the program executes well.
-
People tend to read everything from AI chatbots because it is easy to learn, they can ask questions about any aspect, the format is consistent, and the quality is often among the best found on the Internet.
-
But information is not just about text, you can read most text information from AI chatbots, but you loose the original website and its layout and form, its explaination images and website design.
-
Websites with a lot of interaction are unlikely to be significantly changed by AI, such as web games, Google Docs, Google Sheets, and collaboration tools like Zoom or Slack. They are code-centric and not just focused on text.
-
It is easy to make typos or it requires effort to craft prompts for AI chatbots. That’s why a fully AI-driven digital bank, digital trading app, or AI social media with a simple chat box often doesn’t work. Traditional click buttons, page navigation, and layouts in mobile apps are more convenient.
New Platforms Powered by AI Workflows
2025.01.08
-
Workflows are systems where large language models (LLMs) and tools are orchestrated through predefined code paths.1
-
Imagine a new platform, like TikTok or Quora, X, Threads, Instagram, WhatsApp, Facebook, LinkedIn, Reddit, or YouTube, fully powered by AI translation.
-
Every post or answer created by users can be saved in a single language. The platform will automatically translate content into 20 languages, allowing users to view it in their preferred language.
-
Beyond translation, other AI-powered features, such as summarization, audio generation, and video generation, will play a key role. Essentially, the user submits prompt context, and the platform handles the rest.
-
Users can upload text, images, audio, or videos, and the platform will automatically convert the content into other formats. Users can decide how they wish to receive that content (e.g., as text, images, audio, or video).
-
Platforms can automatically generate summaries, with different types of summarization available in multiple languages.
-
In any text, image, audio, or video on the platform, AI can assist in generating, refining, enhancing, fixing, summarizing, expanding, converting to other formats, or imagining new forms of the content.
-
Users can customize the platform using keywords like “English” or “funny” to adjust the style of AI workflows in platforms like TikTok. Once set, the AI will tailor content accordingly.
The Next Direction of AI Code Editors
2025.01.08
Recently, I was working on adding an xelatex pipeline to GitHub Actions.
I encountered an issue with the fontawesome5 package in the GitHub flow. The solution provided by 4o-mini (installing TeX Live 2021 and using tlmgr install fontawesome5) didn’t work for me. However, 4o suggested a better approach: upgrading to TeX Live 2023 and still using tlmgr to install fontawesome5. While this didn’t completely fix the problem, switching to TeX Live 2023 significantly improved the situation.
I used ChatGPT to help figure out the problem. For more details, check out What ChatGPT O1 Can Do That 4o-mini Cannot.
At this point, I didn’t use editors like Cursor or Windsurf, though I did try them in another project. The issue with these code editors is that they only capture local test output, which limits their functionality in cloud environments.
In workflows like GitHub Actions, Jenkins jobs, or any code deployment or testing flow, code editors need to be better integrated. They should provide seamless interaction with the cloud and CI/CD processes.
This integration also applies to other content creation tools—whether for text, images, audio, or video. These tools should be integrated with A/B testing systems. AI tools could generate content, and A/B testing tools could provide feedback. This dynamic is similar to Reinforcement Learning from Human Feedback (RLHF), where AI models improve over time based on real-world feedback.
This concept of extending RLHF beyond just model outputs—into real-world testing and deployment environments—seems like a promising direction for improvement in both code editors and AI-driven content creation tools.
The test can be either instant or long, and it can be either automated or assisted by humans. If the tests are automated, such as user A/B testing for an AI tool, it still involves human feedback, but the process is automated. For example, we can let the computer check results every day or every hour based on A/B testing outcomes to improve the creation process. Similarly, for Jenkins or GitHub Actions jobs, we can let the computer check after their tasks are completed.
If human assistance is involved, the feedback cannot be fully understood by the machine and is often somewhat vague. For instance, when AI tools create content like images or videos, humans might point out that the content isn’t funny enough, or that a specific detail should be improved. Machines still have a long way to go in making everything perfect, and whether something is “perfect” is often subjective, depending on individual taste. It is human feedback that helps make things better.
In theory, all human-defined rules can be written as prompts. There are user prompts and system prompts. We should focus on improving the prompts instead of fixing the output every time.