I am really excited to see a startup, built by mostly competitive programmers, launching a demonstration of which is the state-of-the-art.

You can watch their demo of Devin, the first AI software engineer, here

• https://twitter.com/cognition_labs/status/1767548763134964000

You can read the opinions of an external user who has early access here

• https://twitter.com/itsandrewgao/status/1767576901088919897

This is an article from Bloomberg (paywall)

• https://www.bloomberg.com/news/articles/2024-03-12/cognition-ai-is-a-peter-thiel-backed-coding-assistant

These are the competitive programmers involved

• Neal Wu neal | IOI 3 Gold
• Scott Wu scott_wu | IOI 3 Gold
• Andrew He ecnerwala | IOI 2 Gold
• Steven Hao stevenkplus | IOI 1 Gold 1 Sliver
• Walden Yan walnutwaldo20 | IOI 1 Gold
• Adhyyan Sekhsaria AsleepAdhyyan | IOI 1 Silver

There have been many other startups that started much earlier in the year (and raised more money) trying to make AI agents work, but it is this team of mostly competitive programmers that has finally demonstrated an AI software engineer that actually works.

I used OpenAI API to translate Python code written by Kiri8128 in the recent combined division contest. These are the results.

• Problem A: Source, AC
• Problem B: Source, CE AC
• Problem C: Source, WA WA WA RE CE
• Problem D: Source, AC
• Problem E: Source, AC

How to use the translation tool

• You will need an email and a phone number to sign up for the OpenAI Beta.
• Create a new code translation Playground.
• Press "Submit" to see an example of how the code is generated for Haskell.
• Change Haskell to C++ in the starting prompt and at the stop sequence. You can see how C++ code is generated as well.
• Insert the Python code you want to translate.
• Change the maximum length to a larger number so that you generate the code all at once. The total number of tokens in the input and output is limited to 4000.
• Copy code from Kiri8128.
• Press "Submit" to generate the code.
• Copy the C++ code and submit to Codeforces.

Kiri8128 is the first to reach Grandmaster with python3/pypy3 without ever switching to C++. Kiri8128 writes clean code which probably made it easier for the model to translate.

I am curious to see if this tool can produce code that resolves the various instances where Kiri8128's Python code TLE or MLE. I haven't been successful, as the code that seem to TLE/MLE because of the language is usually the more complicated ones.

Feel free to share your thoughts and findings!