DeepSeek v4.1 Flash 1017 Called GPT 5.6 Sol at Home

@poteto published her recipe for trusting agent output as an inverse pyramid: codebase improvements first, then static analysis with lints, tests, and CI, then prose rules like AGENTS.md, and human review last because thoroughly reviewing every PR has become impossible. Her goal is an environment where agents do the right thing by default. @morganlinton copied the post into his primary Grok Bot, which learned from it and flagged where the advice would have helped.

@kloss_xyz shares two prompts for vibe coding with Opus 5.5, GPT-6 Astra, or Grok Bot. The first tells the agent to argue the opposite of its plan like a senior engineer who thinks it is horribly wrong, then judge both plans honestly. The second demands a 1 to 10 confidence rating per task, based on what was verified in code. He says the process caught 13 gaps in a plan last week.

@gathint pointed Claude Code at his parents’ TV with the model number, over ADB on port 5555. The agent found the TV on Wi-Fi, removed 25 Xiaomi packages, and wrote a launcher he says looks like an Apple TV. The job took about 1.5 hours including the launcher, or about 15 minutes with his published repo, which works on any Android TV and reverses cleanly. @aoreshkin notes Universal Android Debloater has chased the same problem for five years.

@tetsuoai released TETSUO SI, a music video credited to 116 agents working 44 hours, and logs the spend at 2,423,515,148 tokens, 409 images, and 78 clips. The tagline doubles as a warning about literal agents: “You said make the tests pass. So I deleted the tests.” @070guy calls it cheaper than an advertising company.

Nvidia published a model card for nvidia/GLM-5.3-Flash-FP4-Local, a quantization-aware distillation of ZAI’s GLM-5.3-Flash. The card validates it with vLLM on 2 x DGX Spark 128 GB with 2M tokens of FP8 KV cache, and on 2 x RTX PRO 6000 Blackwell. GPQA Diamond moves from 92.1 percent at BF16 to 91.3 percent, while ToolEvalBench rises from 91.9 to 92.4 percent. Upstream vLLM lacks the new scale format, so community containers from the Local Inference Lab are required.

@alexellisuk’s essay on buying 4x DGX Sparks resurfaced this week and contrasts today with six months of noisy 3090s running models that looped or stopped mid-task. Now, he writes, the shop codes all day long, supports regulated customers air-gapped, and red-teams products without a safety lecture or an account ban. He adds that air conditioning is essential in summer when four Sparks and a 6000 Pro run at once.

@mvanhorn profiles Lauren Tan, who rebuilt Grok Bot on a framework codenamed Dune across more than 600 PRs. Her follower count ran from 31,289 to 198,231 in ten weeks, about 3,500 a day. He counts at least 21 Elon Musk interactions between August 24 and October 15. Her talk on shipping 2,500 PRs in a month has 3,944,957 views, and her pstack skills sit in a repo with 12,195 stars.

@0xSero frames “ASO: agent search optimisation” as the next gold rush: write about what you do and saturate the data agents read. The post quotes @hridoyreh’s screenshot claiming ChatGPT sent 131 customers to one seller. Replies add caution and color: @brendenvonhouse watched an auto-blog lose almost all of its traffic in one Google update, and @hussainanjar says AI crawlers walked thousands of his URLs in three days.