Browser Use blocked 100/100, and batch your Jev calls
sahir says one bot gate blocked Browser Use on 100 of 100 separate runs, and Gregor Zunic agreed it got them 5/5 on his own attempt. The cheaper lesson of the week: Jev charges your state tokens once per call, so batching questions you might not need usually wins.
Built with it
ILIAS ISM replaced a Postiz self-host with one Bun process
ILIAS ISM looked at self-hosting Postiz so his coding agents could run distribution, then built his own 0-dependency MCP server in an afternoon: one Bun process and bun:sqlite WAL against Postiz's five services. It exposes 20 tools across 27 providers, including get_analytics and a pre-flight validator, and runs in a few MB of RAM on his existing Hetzner box for $0 a month.
1GB of RAM replaced by 5 lines of code
Someone measured it
Allie the Icon: state tokens are charged once, so batch the questions
Allie the Icon laid out Jev's input pricing as state plus the sum of all questions, with the state charged only once no matter how many questions run in parallel. That makes speculative questions, batching questions before you know you need the answers, usually faster and often cheaper than sequential calls that re-send the same state.
surprisingly few people realize state is shared and asking a lot of Qs is very efficient
Hamilton Ulmer says Jev still beats Gemini Decisions on HotdogBench
Hamilton Ulmer added Google's new Gemini Decisions API to HotdogBench and says it landed expensive next to Perplexity Decider, with Jev still ahead. Sundar Pichai priced it at $0.10 per million input tokens with $0 billed for output, and claims it leads several benchmark evals with the lowest calibration error. On Ulmer's run, Fable 5.2 had the fewest mistakes at the highest cost.
surprised at how expensive it is vs Perplexity Decider, and still ~ Jev beats it.
Writing criteria
Dan Shipper tuned a summary skill to 80% of his own taste
Dan Shipper built a weekend skill for writing context window summaries and iterated with autoresearch until it hit 80% of his taste against a no-skill baseline. He credits Mike Taylor, who spent about 14 billion tokens over four weeks testing how to tune skills and published the whole log.
Tuning beats writing: tuned skills won up to 95% of held-out posts, while a skill Claude wrote in one shot lost to having no skill at all.
The case against
Browser Use's captcha solver confirmed failing on gated sites
Browser Use's anti-bot and captcha handling failed in public this week, and Gregor Zunic confirmed each report. Tim Smykov says two sites blocked 5 of 5 fresh Browser Use Cloud sessions on US residential proxies with solveCaptchas=true, with rotate-image captchas still up after 120 seconds.
sick gate lol, got us 5/5
Also on the timeline
Kunal Abichandani puts Jev in a voice ordering hot path
His voice coffee ordering system renders the drink and any changes while he is still mid-sentence, about 200 ms from word to screen.
Mike Taylor dropped Descript and Loom for Kieran Klaassen's editor
Mike Taylor says he now does all his video recording and meetings on cut instead of Descript, Loom and Zoom. His favourite feature is the Jev-based multimodal scene detection.
Frequently asked questions
What is an MCP server and why did ILIAS ISM build one?
An MCP server is a tool that exposes capabilities (called tools) to AI agents. ILIAS ISM built one to replace a self-hosted Postiz setup, reducing five services down to a single Bun process (a lightweight JavaScript runtime) that runs in just a few MB of RAM.
Why is batching questions cheaper than asking them one at a time?
Jev charges once for the state (context information) no matter how many questions you ask in parallel, so batching questions before you need the answers only pays that charge once instead of paying it again for each follow-up question.
What is a context window summary and how did Dan Shipper tune one?
A context window summary is a compressed version of information for an AI model to read. Dan Shipper iteratively refined his summary skill by judging drafts blind against his actual work until it matched his taste 80% of the time.
Does Browser Use work with captcha-protected websites?
No, Browser Use's captcha solver failed in recent tests on gated and captcha-heavy sites, with rotate-image captchas staying up even after 120 seconds, and the fixes have not shipped yet.
What is multimodal scene detection in a video editor?
Multimodal scene detection uses AI to automatically identify and separate different scenes in video based on multiple types of information (like visual changes, audio shifts, or speaker changes), which is a feature in the cut video editor.




