in opencode when using astra, I delegate much coding to sol, then Astra reviews, and scores for me how many rounds of rewriting it had to do of Sol's work. Sol 6 is doing well by this benchmark, so far better than 5.6.

uranium is not from our sun.

Astra consumes too much usage. I've come to a compromise: Astra high/xhigh delegates to subagents and then reviews the result. It has cut usage a reasonable amount. I use opencode2. I've been doing this for a few days. Astra helped design it: gist.github.com/timrichardso…

why do people think "drawing lines in the sand" is a definitive statement of non-negotiability? It's a line in sand.

Activists. 2024: Corporate CEOs are paid way too much: limit their money! 2026: We must keep (human) Corporate CEOs running companies at all costs!

I found someone who can advise the government on sensible policy settings around productivity growth and a structurally sustainable budget. youtube.com/watch?v=tN8akF8_… #auspol

what I don't understand is: (a) rezoning increases value of land; (b) at inheritance, descendants want to maximise value of inheritance. So (b) should create support for (a) over time. Why is this not a thing?

Gnome's flathub no longer bans Gen AI outright. When did that happen? docs.flathub.org/docs/for-ap…

I got my openai banked reset. Thanks. In terms of openai incentive this means I get pushed now into tomorrow's queue for roll out (OpenAI is better off racing to provision people who don't have their daily reset yet; I'm now a sunk cost). @thsottiaux ?

right now, no astra and "Couldn't load usage limit resets. Please try again"

Omarchy foundation has, or will have, lots of money. I hope some of it is used in the spirit of competitions as used to spur the invention of naval chronometers (book "Longitude"), or X Prizes @dhh

in opencode2 tui you can have a vertical tab bar of sessions, and there are animations when one is active. And tonight, I saw a little '?' when a session is stopped with a question, and I smiled. Maybe I just never noticed since I get terminal notifications too.

this was awesome @dhh

My niece has transferred her computer science degree from UTS to UNSW (Sydney Universities). She got credit for most studies but has to do the first year programming course because UNSW needs you to know C.

My X feed is giving me FONAT: Fear of not automating things

Now my agents.md has an instruction to avoid metaphorical and other figurative language when naming things. Wtf is a _sealed_plan()?

With AI security (hacking) and AI watermarking, we are in a golden era ... of arms races.

𝙊𝙥𝙚𝙣𝘾𝙤𝙙𝙚 v1.18.17 released. TL;DR: App defaults stay current; session retries are bounded and compaction retains better context; Go exposes quota usage; stats move to R2 SQL. 𝗔𝗽𝗽 • Fixed model selection to follow the server’s current default instead of retaining stale legacy configuration when no default is set. 𝗔𝗴𝗲𝗻𝘁 • Improved compaction for smaller models with explicit conversation and prior-summary boundaries, stronger carry-forward rules, and up to 15K tokens of recent context instead of a fixed two-turn tail. • Fixed runaway provider retries by limiting them to five and adding jitter to exponential backoff. 𝗭𝗲𝗻 Added • Added an authenticated OpenCode Go usage endpoint reporting rolling, weekly, and monthly status, percentage, and reset time. Changed • Applied Go’s China-region eligibility check to DeepSeek V4 Pro as well as V4 Flash. • Changed blocked-account errors to identify fraud or terms violations and provide a support contact. 𝗣𝗿𝗼𝘃𝗶𝗱𝗲𝗿𝘀 • Fixed GitHub Copilot model discovery to expose PDF input when both vision and application/pdf support are advertised. • Fixed MERGE Gateway reasoning variants and provider-option routing. • Fixed DeepSeek V4 Flash sampling defaults by applying top_p 0.95 only to official routes and explicitly versioned 0731 IDs. • Fixed all Muse-family models, including Muse Glimmer, to receive the correct Meta model-specific system prompt. 𝗗𝗮𝘁𝗮 • Moved Go analytics aggregation from Athena to Cloudflare R2 SQL using bounded day/week usage and geo queries. • Fixed daily totals with UTC-aligned windows and an exclusive handoff between legacy and live inference streams, preventing duplicate counts. No noticeable bundle change Compare: github.com/anomalyco/opencod…

I switched from using Linux to Windows, so I could use the Chat GPT Desktop app. Great news! I dual boot from 2 separate SAD. So new I can start using Linux again. =)