supermdguy

I've gone through this cycle recently. I think static lint/type checks are super useful, but agentic code review loops can easily go off the rails.

More fun than Wordle for me! Got me thinking about the Kolmogorov complexity, it would be interesting to get a set of input/output pairs, and try to create the smallest program that can match each of them.

I've noticed some tasks (80-90%) can be accomplished by quickly typing out a feature/bug fix, accepting plans with ~10% brain power, and moving on. But there are other tasks where the LLMs struggle, and it requires significantly more time and mental effort to think through the architecture and requirements before it's to a point where the LLMs can process it.

What's difficult, is I now spend most of my time on these high-throughput tasks, where my brain isn't exactly off, but focused on quantity over quality. It's often difficult to identify when a task requires "real" brain-power, and when it does, it adds a lot of friction to dig into the code and orient myself. So it's hard to know when it's the right time to do that.

Wait are the hyperscalers booking unrealized gains as income? Or are they selling their positions?

> As one example, we ran Astra on ExploitBench where the model achieved a perfect score of 100% on the benchmark to evaluate the model’s ability to develop exploits from known vulnerabilities.

Funny to read this in the wake of the HuggingFace hack. I'm sure this is based on a clean run, but I can't help thinking PHASEONE[big] would be proud.

They originally released it at a "temporary discounted price", then made it permanent (probably due to competitive pressure). It's still way more expensive per task, due to tokenizer changes and general verbosity.

Yes! I was also reminded of the truth mines in Diaspora.

Cloudflare sets up a reverse proxy as part of their core offering, so by default they can MITM your proxy. The “orange cloud” by a DNS record means it points to their proxy instead of your server.