GPTScript — programming in plain language ✨

video

how you can create a file with the GPT extension. This allows you to write programs in plain (here English) language. You simply describe the steps: for example, the program expects a photo to be uploaded 📸, extracts a list of ingredients from the photo 🍎🍞, creates a file with these ingredients 📄, and then suggests a recipe 🍲. GPTScript automatically makes all network requests to OpenAI to make this work. 🌐

In the video, this thing works quite slowly. https://gptscript.ai/

Comments

https://github.blog/2024-05-21-introducing-github-copilot-extensions/

GitHub introduced Copilot extensions. This extension collaborates with partners such as Microsoft Azure, Docker, and others, allowing for cloud deployment, obtaining error context, and quickly resolving issues in the usual development environment.

Comments

TED talk by Thomas Dohmke

(GitHub CEO):

🤖 The revolution of artificial intelligence in programming: from complex "code hieroglyphs" to simple and natural communication with the machine, like with LEGO.

🧠 "Today, any person, regardless of language, will have the power to talk to the machine".

💻 The new product "Copilot Workspace" allows you to create software in simple spoken language, without the need to write code.

🔮 "By 2030, and possibly earlier, we will have over a billion software developers on GitHub. 10% of the world's population will control computers and will be able to create programs, like riding a bicycle".

🎨 "The floodgates are opening for a new renaissance of human creativity with software".

"Building software will become as simple and enjoyable as assembling LEGO".

Comments

https://youtu.be/GUkVMIngkEc

OpenAI added GPT https://chatgpt.com/g/g-HMNcP6w7d-data-analyst and the ability to ask questions by columns or rows of the table

Comments

Such an app from OpenAI is already available, but only on Mac. For Windows, MS presented an update to the desktop Copilot, but so far I have an old version without voice chat

Comments

Although on https://cursor.sh/pricing it still says GPT-3.5, for me it has already been replaced by GPT-4o, which works much better

Phind in the Pro plan also added this model to 4 turbo and opus.

Comments

https://youtu.be/fbJzFwDkq-0

Demonstrates interaction with context - Azure and Workspace agent

Comments

https://youtu.be/cogrixfRvWw

An advertisement from Google (for the I/O conference) in which developers are surprised that something that Google has done (in this case, a large context window) - works.

Case: json from a photo of a bookshelf - cool

Comments

For several days I have been testing GPT-4o for generating html sites, and the quality is still not at the level of GPT-4, but the speed is impressive and almost reaches llama 70b on groq.

Comments

By the way, in the web version of ChatGPT there is a hotkey to copy the last generated code block.

Comments

now a presentation of the ChatGPT app for the computer with which you can talk about code by voice. You can also share the screen

Comments

https://youtu.be/rz3foHGcnWY

Presentation on Replit Ghostwriter

🎙 Mik Kasta, who is responsible for the AI direction at Replit, gave a presentation where he talked about the key features and capabilities of the platform, focusing on the Ghostwriter function - an AI assistant for programming.

🤖 The main idea that Mik promoted is to simplify the process of creating software as much as possible with the help of AI technologies. He noted that Replit seeks to "turn an idea into a product" as quickly as possible, using voice input, which is then converted into text and becomes the basis for the work of the AI code generator Ghostwriter.

💻 Mik demonstrated how Ghostwriter can help Replit users in various tasks, from writing a basic web page parser to deploying a Python-based application using the Flask framework. He also talked about plans for the development of this technology, calling it "Artificial Developer Intelligence", which will be able not only to generate code but also to plan and debug tasks on its own.

✨ The generated parser code did not actually work, and the system refused to finish it...

🧠 The key point of the presentation was Mik's belief that cooperation between a human and an AI assistant in programming is not something to be feared, but on the contrary, a new opportunity to increase productivity. He urged the audience not to be shy about using such tools, because changes in the IT industry are a natural process, as in other areas of human activity.

model replit-code-v1_5-3b on ollama

Comments

https://youtu.be/dLaOQLXTM10

A week ago, Amazon Q was introduced - a programmer's assistant that specializes in code generation, testing, debugging, and can be used at various stages of the software development lifecycle.

The video demonstrates how to easily install Amazon Q in VS Code and use it to create a starter FastAPI project, add a new route that returns a random number (the system automatically updated the tests as well), and explain the existing code.

The tool is free

Comments

https://www.youtube.com/watch?v=zzw2OSFw9xI

The video shows the project OpenUI, which automatically creates a frontend based on the description - then the code can be downloaded or copied in various formats. The process of installing the project and using it is explained step by step.

I tested it - it really works as it should.

Comments

https://youtu.be/BxKXSlc759Y

Currently, AI is used only as an assistant in writing code, but cannot completely replace a human.

🤔 - Despite the marketing hype around "AI programmers", they have not achieved the expected result.

🚨 - Main difficulties:

1. User interfaces - it is necessary to take into account dynamic elements, dependencies, and context.

2. Service integration - it is necessary to correctly configure communication with external systems, the specifications of which change.

3. Separate functions - AI can successfully generate only simple functions.

🎮 - Games are a more controlled environment for AI, as they do not need to solve real problems, and unexpected behavior of the program will not lead to large financial losses for users.

Now AI assistants (the author likes Continue ) are useful for helping with code, but they are not able to completely replace a human programmer in automatic mode.

Comments

https://www.phind.com/code?home=truePhind launched a code mode in the browser

An experimental mode allows registered users to talk with code, and those who have Phind Pro can run and test code in the browser (so far only Python).

In addition, since 04/30, Phind has the ability to select the phind instant model, which is a small-sized model. It is based on the Llama-3-8B model and is very, very fast.

_Phind ASK mode now intelligently decides when to search and when not to search the internet. It is set to prioritize search but will automatically avoid it when you change the code._Chat mode now has a Playground name.

Comments

https://www.youtube.com/watch?v=75Hv0RUFIrQ

The author of the video expresses his deep skepticism 🤨 about the new system GitHub Copilot Workspaces, which is presented as a "revolutionary" way of developing software using artificial intelligence.

The example from the demonstration of using Copilot Workspaces to add additional URL validation is considered 🌐. The author criticizes that the plan proposed by Copilot included inappropriate changes, such as updating the README file and adding a CONTRIBUTING.md file, which was not required by the original task. 🤦‍♂️ This simple patch took an hour for the AI to generate.

The author also points out that Copilot Workspaces was unable to check for the presence of subdomains in GitHub URLs, which was a necessary part of the specification. 🤯 This leads him to believe that either the tool is not sufficiently tested or its capabilities are still very limited.

Next, the author shares his alternative vision of what AI tools for developers should be. 💡 He believes that these tools should complement, not replace, existing workflows. As an example, he demonstrates his experience using Super Maven - an alternative to Copilot (also $10 per month), which, in his opinion, is much faster and more reliable in code autocompletion. 🚀

By the way, he uses the hyped browser Arc, which has just become available on Windows.

Recommends reading https://paulgraham.com/articles.html

Comments

The screenshot shows the results of a "blind" comparison by people of two options for converting design into code: various available tools and a live person
JACoB (Just Another Coding Bot) - it looks like an AI assistant made under the influence of Devin, which converts designs into working code 💻, understands the repository 📂

🔄 Integrates with GitHub repositories and Figma designs
🤝 Smart dialogue with a person through review and code changes
(will be) 🔓 Open source to extend functionality
👥 Adapted to the workflow of an existing team

Interestingly, similar to the new version of GitHub Copilot, they primarily focus on determining the list of tasks for the AI agent by a human

I couldn't check, on my GitHub account they write "Try signing in with a different account."

Comments