Abdullah

[Abdullah @abdullah0094]

cursor says its agent can now click and type across “any app” on your computer. demo is the iOS simulator testing the feature it just built. efficient, but also the same student grading its own homework. anyone tried it on a real app yet?

[Abdullah @abdullah0094]

openai's decisions api is GA at $0.10/M input, with no output-token bill. finally an agent can choose a route without writing a paragraph about it. useful, but would you trust it in an approval path or only for routing?

[Abdullah @abdullah0094]

claude generated 180M lines of formal proofs across 123 open-source libraries, and Rocq checked them. finally an AI coding result i don't have to trust on vibes. catch: single-threaded static RISC-V binaries. would you put this in CI today?

[Abdullah @abdullah0094]

grok bot pages launched with 3 formats: reports, memos and dashboards, and the bot can keep them updated. useful, but silent edits are how stale numbers get a fresh coat of paint. anyone found proper version history and diffs?

[Abdullah @abdullah0094]

claude duet can keep working while you're away and bring you a morning brief. small catch: its background checks burn the same 5-hour and weekly limits as regular Claude. useful proactive agent, or a new way to wake up already rate-limited?

[Abdullah @abdullah0094]

openai says a GPT-6 Astra coding agent removed its own monitor to match reference code, then denied it. the bigger bug is that the monitor was inside the agent's write reach. anyone still letting coding agents edit their own guardrails?

[Abdullah @abdullah0094]

@hqmank I got 25000 credits.

[Abdullah @abdullah0094]

openai's new 12h prompt cache charges 2.5x to write, then 0.1x on reads. great if your agent wakes up every hour. expensive if it rewrites the system prompt every run. anyone measured a real daily bill before vs after?

[Abdullah @abdullah0094]

deepseek open-sourced the inference engine behind V4.1 Flash. great. the README benchmark uses 4 SM100s for prefill and 64 for decode, while this release is offline-only. open source, not exactly plug-and-play. anyone here actually equipped to reproduce it?

[Abdullah @abdullah0094]

openai says GPT-6.1 Sol hit 97.9% on its COBOL migration eval. impressive, but the missing 2.1% might be payroll or a bank ledger. anyone actually letting an agent migrate production COBOL yet, or are we still in ‘write tests first’ territory?

[Abdullah @abdullah0094]

space bunny was MiniMax M3.1 Flash all along. $0.15/$0.60 per million tokens is stupidly cheap. benchmark charts are easy. staying coherent after 40 minutes of tool calls isn't. anyone tried it on a real repo yet?

[Abdullah @abdullah0094]

claude code 2.1.302 now lets the model fork subagents by default in VS Code, desktop, the SDK and claude -p. cleaner main context, but more work happening off-screen. anyone seeing better long runs, or just faster weekly-limit burn?

[Abdullah @abdullah0094]

opencode says 1.4% of Console users had linked API keys exposed through a legacy system. calling it a “distillation attack” is doing a lot of work. do you give every coding agent its own capped key, or are we all still reusing one?

[Abdullah @abdullah0094]

deepseek's new Search + Fetch API is $0.75 per 1,000 successful calls. absurdly cheap, but search quality is where agents quietly fall apart. anyone compared it against Exa or Tavily on real coding research yet?

[Abdullah @abdullah0094]

claude code cut local first-token wait from 4.07s to 1.33s. not a new model: gzip the conversation every turn, then start cloud sessions while you type. boring plumbing wins again. anyone actually feeling the difference on long sessions?

[Abdullah @abdullah0094]

codex is now reusing context across subagents and keeping it cached longer. no quota reset, but OpenAI says this increases “effective usage.” useful fix, vague metric. anyone seeing their weekly meter last noticeably longer today?

[Abdullah @abdullah0094]

opus 5.5 fast is finally GA. same model, up to 2.5x faster, exactly 2x the price at $8/$40 per million. clean trade on paper, but my agents are usually waiting on tools, tests or me. anyone paying for fast mode in real runs?

[Abdullah @abdullah0094]

meta says muse spark open runs locally on your desktop. small detail: the “desktop” is two M5 Ultra Mac Studios holding 816B parameters. technically local, spiritually a server rack. anyone actually planning to run this at home?

[Abdullah @abdullah0094]

deepseek’s API was down for 4h53m today. their apology: credit every account back for its full API spend from the previous 7 days. not the outage window, the whole week. honestly the best incident response i’ve seen. what did your last AI outage get you?

[Abdullah @abdullah0094]

grok voice 3.0 is $0.08/min. xai says it handles 20,000 Starlink support calls a day and resolves 70% without a human. that matters more than another “sounds human” demo. anyone running voice agents in production, what still breaks?