Kevin (@kcosr) header
Kevin (@kcosr) profile picture

automation enthusiast | https://teapawt.net/t.co/l1vrFQx6KL

Joined February 2022
Photo by Kevin (@kcosr)Photo by Kevin (@kcosr)Photo by Kevin (@kcosr)Photo by Kevin (@kcosr)Photo by Kevin (@kcosr)Photo by Kevin (@kcosr)Photo by Kevin (@kcosr)Photo by Kevin (@kcosr)Photo by Kevin (@kcosr)Photo by Kevin (@kcosr)
For the second time now, GPT 6.2 Sol has drafted a plan that involves changing a dependent service instead of using the existing interface. There is no capability gap in the interface. It's behaving as if it doesn't understand the boundary between the two codebases.
1
1
4
482
Got a new CI box. 🚀
1
3
302
Qwen 3.8 27B is doing a lot of good work for us, but it still occasionally hallucinates. This is a complete fabrication. Full BF16. ☹️
2
7
444
I briefly tried Dots again and I'm really puzzled about the intended UX. The chat UI looks nice with a typing indicator. But if you ask it to do work that involves a tool, even taking a note, it just pauses for minutes at a time with no working indicator. When I ask again, it has to check status. How is this better than a traditional chat thread?
4
6
709
My unoptimized Rust test suite now runs in 2 minutes, down from 40. - Build Argon2, Blake2, and bundled SQLite at opt-level=3 in tests. - Use cargo-nextest to run tests across binaries in parallel.
1
1
8
719
The speed of 6.1 Terra makes it great for generative UI.
2
293
At my company, we are focusing more on business automations than software factories. While we develop our own software, we are not a software company but a provider of services. We are benefitting far more from having agents triage alerts, Jira and GitLab issues, review code and deployment requests and answer questions for staff by researching across our internal systems than we would from churning out more code.
5
1
3
816
I lost wifi connectivity 5 minutes into a rambling dictation from my deck. Boy am I glad I went with client buffering and incremental uploads for transcription. I just kept talking and it resumed uploads when I went back inside.
5
16
786
I enabled step_model_switching in Codex so that I can disable ultrafast mid-turn. The problem with the default is that if I leave it on during tests and peer reviews, it burns tokens at the ultrafast rate checking status unnecessarily. There is no command completion callback and they limit sleep to 60s max. A watch mechanism like Claude provides would be nice.
2
1
2
432
Where does one even begin to troubleshoot something like this? 🤣
2
4
432
I'm beginning to think that most software engineers will end up not reading the code and maybe not even verifying the work they are doing since it's much easier to just instruct the agents to do that. I also think that each company will have a couple of very experienced staff engineers that will build agentic systems (custom AGENTS.md, custom skills, various plugins, etc) so the code that the rest of the developers generate will match the code quality standards of the company. These people will probably read thousands of lines of code every day, understand the code patterns, and probably perform large-scale refactors. Their big challenge will be to act based on data, not on vibes. This will maybe be controversial, but there's a good chance that the overall output will be higher quality code than two-three years ago as long as the people building the agentic systems know what they are doing.
50
9
2
254
27,161
Introducing tsrs: a Rust port of the TypeScript compiler 🦀 We are not forking TypeScript! ◆ In close collaborating with the TypeScript team ◆ Up to 8x faster than TS 7.0 & ~70% less memory ◆ Runs in WASM, so it works in your browser & on Workers ◆ Soon part of the Oxc toolchain npx tsrs blog.cloudflare.com/typescri…
4
315
Your agent should be able to control the app in any way you can, with proper thread-scoped access controls, of course. Heck, mine can even control client-side settings on individual devices like toggling voice functions. 😀
WTF. You can schedule tasks in T3code just by asking...???? Why isn't this feature blasted absolutely everywhere? This would have saved me so much fucking time.
1
5
968
It's crazy that I'm going to run out of Claude usage before the next reset. A month ago I only used it for code reviews. Now it is the reverse, and most of my Codex usage is just me being impatient and using ultrafast.
7
1
9
1,089
A lot of the work I'm doing involves big, well scoped tasks where Claude will take a full day to implement. I am blown away by its attention to detail and how it will remember small things I drop into the chat, or decisions it needs to inform me about after completion. This may be due to memory features in the harness or training around keeping a log (which you can do yourself), but it just works well out of the box. With 1M context and it's use of subagents, it almost never needs to compact. I am no fan of Anthropic but they are providing the better developer product right now.
1
2
336
Brokered access for AI agents, with resource-scoped permissions, account credential management, audit logging, rate limits, and attribution. Supports Jira, Confluence, GitLab, plus MCP and OpenAPI imports. WIP 👇
7
2
1
18
897
I have some eval and tuning work ahead of me so of course I had to built my own tool. Let's hope it works. pi + smolvm + ghostty
1
1
1
4
556
Staying in the loop while focusing on other things. Works hands-free in the background on my Android device.
2
6
1,122