Finished testing Claude Opus 4 and Claude Sonnet 4 on my personal eval set. I am VERY impressed. Claude Opus 4 absolutely dominated other models in both coding and writing tasks. It is the best performing model for all 4 tasks given. It is worth nothing this model is very

Are you okay, Claude? Anthropic API appears to be overloaded 529 for me across different apps. Anthropic status page needs to be updated.

Looks like the Claude Code "nerf" is here. Some personal observations:. - No more automatic TODOs for small tasks (so far).- Reading files 10/15 lines at a time instead of 50/100 lines.- Failing at simple tasks.

What's the strategy for AI coding companies are? What's Cursor gonna do next? What's Devin going build after acquiring Windsurf?.

I think Claude Code is "enough" for me. I don't really care if it is 100% correct or reliable. I am happy to handle 10% edge cases myself. Going from time saving of 5 hours a day to 5.5 hours is nothing compared to previous generational leaps.

2024, 70%: Cursor became really good in 2024. Edit multiple files. Run autonomously for 5 minutes. 2025, 80%: Claude Code came in 2025. No embeddings or codebase indexing. Uses grep and cli to navigate the codebase.

AI coding is hitting diminishing return. 2021, 30%: GitHub Copilot was first AI coding tool. Autocomplete felt like magic. 2023, 60%: ChatGPT and GPT-4 was next. Huge intelligence jump. AI can actually write code given a task.

Wtf Windsurf is joining Devin (Cognition).

And it does work! I asked Claude Code to write a test with assertion and it works!

Watching Claude Code trying to bypass bot detection on xAI website is really fun. After several failed attempts messing with UA and various headers, Claude Code resorted to using archive. org instead.

Interesting strategy by Amp to use unix timestamp as version number. This coincidentally solves the 9.9 vs 9.11 problem that AI models are unfortunately being RLed to memorize as maths problem instead of software version number. Using the unix timestamp ensures that the version.

Some new information on Windsurf employees.

For those curious, the full replay vod of the livestream is available here:.

So you spend less time working, while getting more work done (assuming agents are as effective as you). The efficiency and productivity gain of AI coding comes from the async nature of coding and the reduced engagement time for human to be present.

The other way to look at it from a productivity perspective: . Instead of spending 5 hours a day engaged in coding, you can now spend 1 hour in the morning and 1 hour in the afternoon prompting and reviewing, to get 6 hours of agent working time.

I could go read a book, write a blog post, or watch some YouTube videos and then come back to review the code. The effective time I have to spend being engaged with the coding process is reduced to around 30 minutes from 90 minutes, which gives a 66% time saving for me.

However, if I were to code it myself, I would have spent 90 minutes staring at the screen and spending my brain cells on solving bugs. Compare this with using Claude Code where I didn't have to be engaged after giving the initial prompt and before reviewing the results.

Some viewers pointed out that, it would probably taken me (experienced dev) similar amount of time to write code manually compared to using Claude Code. That's true, in terms of wall time (how much time it takes for a clock to run).

Just finished my first AI coding livestream with Claude Code, where I shipped 3 small-medium features to 2 apps across 4 repos in 90 minutes live on camera, without writing a single line of code. Some interesting observations on wall time vs human engagement time:

So it's entirely logical to assume that we are no different from LLMs on a conceptual level. And we do in fact live in a simulation.

Humans have both system 1 thinking (automatic) and system 2 thinking (rational), similar to how LLMs would take parallel paths to generate tokens, one memorized shortcut response and one from rational deduction.