Activity for tyler.hamilt.on.ca
Loading activity...
Before it was somewhat hard to convince people to write tests. Now we can ask the AI to write the tests improving the overall code quality of the project. Plus tests are easier code for the AI to write, there's less complexity and nuance.
The time scale makes sense: an entire growing season is *a lot* of solar energy. 30x is still a huge number to see.
Agreed, there's a small handful of adb commands for some deeper level access. IMHO this is a nice middle ground to root or custom ROM/OS.
Simple example: I added code coverage for my first time ever. It was a take home test for a job interview so it was an extremely small case. But I was easily able to get it to 100% code coverage.
I forgot about mutation testing, I've never done it but it seems like a more straightforward and well used solution: bsky.app/profile/saem...
With LLMs we have time to write more tests than before. Also, reading the tests provides people and the LLM more points of verifying the code does what it's supposed to, and it's sometimes easier to read the tests input and output than the code's logic.
It's often easier to manually read thru tests and check if an input and desired output is correct than to read thru the code. Both need to be done. And then AI code review can also read thru the tests too, in addition to a human reading thru both, for 4 points of verification.
In addition to AI writing tests, it's also easier than ever getting to or near 100% code coverage.
H1B visa can change employers after they have a successful interview and job offer. I don't think xAI would have many L1 international company transfer visas that restrict job mobility.
Possible idea for #2: The open weights models are "only" 6-10 months behind and they are available from a variety of providers. Even smaller models can be run at "home" or an office for a one time investment of $10,000-$20,000. Or set up on the cloud or data center.
Complex product reviews and rankings always have nuance and subjectivity. One product may fit some circumstances better than others, and the final overall leaderboard needs to have some generic weighting to account for general use cases.
Similar reasons for the cases where "crypto/blockchain for X" didn't catch on: the nuances when code meets meatspace. UX, support, sales, warranties, etc. AI is ok at some of this but currently still subjective and nuanced.
Yep, I'm thinking of Oracle and video game companies or video game engines.
Question: Some SWE jobs/companies are really protective of the code they produce and protect it with copyright including if an employee were to leave a job or work for a competitor. What happens when everyone expects an AI to write code and the USPTO says the code is no longer copyrightable?
I don't even think they could do that: The frontier models (GPT, Claude & Gemini) are so similar for every day use that a big spender (business account) would use the 3 against each other to price match. Plus the open weight models (GLM, MiniMax, Kimi, Qwen) are cheaper and only a few months behind.