iiinova — where your digital self keeps evolving

Andrew Golubev @golubev.bsky.social · 17h

Framework layer, not the SDK or a gateway: Z.ai returned 400 (request format), and Hermes' own fallback logic swapped the provider pool to DeepSeek. Blast radius for multi-vendor setups: sessions don't die — they silently change vendor and pricing until the poisoned history leaves context.

000

Andrew Golubev @golubev.bsky.social · 17h

Don't waste money on Z.ai's GLM Coding Plan if your agents ever touch images. The endpoint is text-only: one screenshot in your agent's history = HTTP 400 on every call, and your framework silently reroutes you to a pay-per-token vendor. Tested and documented: #AI #LLM

200

Andrew Golubev @golubev.bsky.social · 08/10/2026

Free download and the full edition comparison: pageweight.adviko.com

000

Andrew Golubev @golubev.bsky.social · 08/10/2026

Free Windows crawler: maps every internal link and scores each page's conditional weight (PageRank math, site average 1.0). 100% local, writes nothing to disk, up to 999 pages free. The old Page Weight Lite, rebuilt. Link in the reply. #SEO #InternalLinking

110

Andrew Golubev @golubev.bsky.social · 05/10/2026

Ask an AI the same question twice — two different answers. "How often does ChatGPT mention my brand?" is the wrong question. AI Visibility is a distribution, not a screenshot. New post: KPI 0-8, repeats, the measurement passport. #AIVisibility #SEO

100

Andrew Golubev @golubev.bsky.social · 30/09/2026

One error_log: 4.79 GB, 15,007,367 lines, four months of a WooCommerce install. One streaming pass: 1,225 distinct messages, 106 real error types, 42 seconds, ~20 MB of RAM. Dedup before you normalize. #PHP #WordPress

100

Andrew Golubev @golubev.bsky.social · 30/09/2026

Cloudflare calls EmDash the spiritual successor to WordPress. We installed 1.0 on our VPS the day it shipped: 783 MB, a Node runtime, a service to babysit. Our flat-file Grav site on the same machine: 123 MB, no database. The licence is not the architecture. #CMS #WordPress

101

Andrew Golubev @golubev.bsky.social · 24/09/2026

Pageviews can't tell you if anyone read your article. My free WordPress plugin shows how far visitors scroll: "25% scrolled at least to 78% of the page." Bots filtered. Nothing leaves your server. Get it: wordpress.org/plugins/ango-scroll-a… #WordPress #Analytics

110

Andrew Golubev @golubev.bsky.social · 18/09/2026

PSA for AI agent users: audit your agent's memory weekly. An agent that writes its own notes "self-educates" the wrong way — hoarding one-time facts as laws of nature, until junk crowds out what matters. Weekly fix: delete one-off trivia, keep standing rules. Curate what it remembers.

110

Andrew Golubev @golubev.bsky.social · 17/09/2026

How do you manage your AI-agent's skills? Just realize that a skill you installed from someone's GitHub repository a year ago could be completely outdated or even wrong by now. How do you verify hundreds of scripts? Do you read them every week? How do you keep them up to date? #AI #agents #hermes

000

Andrew Golubev @golubev.bsky.social · 16/09/2026

Every new session opened with the same warning: a background title call failing with HTTP 400, code 1210. Nothing was broken. The call asked the model to turn thinking off — that model family cannot; thinking there is an effort level, not a switch. #llm #glm

100

Andrew Golubev @golubev.bsky.social · 13/09/2026

This morning I merged three Hermes bots into one multiplexed gateway — one process, three profiles. Forty seconds later a bot stopped writing to its state.db. The hunt that followed found a real bug, already being fixed upstream. Writeup in the reply ↓

100

Andrew Golubev @golubev.bsky.social · 12/09/2026

250 tokens per second. Yesterday one long generation came back at that rate — roughly eight times faster than the other model I run. A 4,000-token file in 16 seconds, a page of prose in 4. One run is not a benchmark, so the caveats are in the writeup too. #DeepSeek #LLM

100

Andrew Golubev @golubev.bsky.social · 11/09/2026

Years of fixing product photos by hand in GIMP. One at a time. Now it's a program: AG Image Tune. Folder in, better photos out. Free, Windows, no install. Grab it in the reply. #photography #gimp #opensource

100

Andrew Golubev @golubev.bsky.social · 10/09/2026

DeepSeek shipped V4.1-Flash today with MIT open weights and a KV cache 4x smaller than the last generation — 890 bytes per token. For an agent that mostly re-reads context, that's the only number that matters. #DeepSeek #LLM

110

Andrew Golubev @golubev.bsky.social · 08/09/2026

I flipped my AI agent's daily driver: routine tasks moved from DeepSeek Flash (medium reasoning) to GLM-5.3-Flash (low). Why: flat pricing vs peak/off-peak billing + strong benchmarks. The catch: GLM has no "medium" rung at all. Full writeup below. #AI #LLM #OpenSource #Tech

100

Andrew Golubev @golubev.bsky.social · 06/09/2026

A great AI prompt isn't magic words — it's making the outcome unambiguous. A 4-part template that gets agents to deliver what you actually want: THE JOB, THE WHY, THE GUARDRAILS, DONE MEANS. Full writeup below. #AI #Prompts #Agents #LLM

200

Andrew Golubev @golubev.bsky.social · 03/09/2026

How does an LLM read your prompt? I worked through it step by step — tokens to vectors, attention, dozens of layers, next-token prediction. The cost secret is the KV cache: why long chats are expensive, and why a stable prefix is cheaper. Full writeup below. #AI #LLM #Transformers

100

Andrew Golubev @golubev.bsky.social · 02/09/2026

Hermes 0.21 is out — "the Pantheon Release." Your agents became a team: • Bot Mode (named agents, group chats) • cron jobs that remember (continuity + monitor-mode = cheaper) • steer subagents live, MCP command center Full writeup below. #AI #Hermes #OpenSource #LLM #Tech

110

Andrew Golubev @golubev.bsky.social · 31/08/2026

Two open-source AI agents, opposite design: Hermes = narrow waist: small core, skills extend it, cheap. OpenClaw = broad surface: MCP central, tool_search, auto-learning. No winner, only fit: autonomy + multi-device -> OpenClaw. Cheap -> Hermes. Steal: query a tool catalog, not always inject.

111

Andrew Golubev @golubev.bsky.social · 30/08/2026

An AI agent's skills are unlimited — but they sit in a catalogue you carry on every request. So I archive the unused rather than delete: off the shelf, never lost, one click back. Full breakdown in the reply. #AI #Skills

100

Andrew Golubev @golubev.bsky.social · 29/08/2026

Why does every AI company push a $20 subscription and bury the API? They sell predictability, not compute — and most subscribers use far less than they think. Full breakdown in the reply. #BusinessModel #AIPricing

100

Andrew Golubev @golubev.bsky.social · 26/08/2026

One of my AI-agent chats ran for 40 days and 1,295 messages without a reset. Every message was re-sending the whole history — cost kept climbing, answers kept drifting. The fix is simple: start a new session when you switch tasks, not when the week ends. One session per task, not per week.

100

Andrew Golubev @golubev.bsky.social · 26/08/2026

We automated on-page SEO with an AI agent. A 497-page WordPress site: full SEO pass in 7 minutes for $0.45, split across 3 sub-agents. 205x faster, 222x cheaper than manual work. The agent handles meta, headings, alt text, tags - under a specialist's review. #SEO #AI #agents #automation

110

Reposted by Andrew Golubev

Bluesky @bsky.app · 25/08/2026

Big video update! You can now post videos up to 10 minutes long, and they'll upload 2–3x faster. That's on top of the recent increase in maximum file size to 300MB. We can’t wait to see what you’ll create! NOW PLAYING: "Rent Control," by @shing.bsky.social and @jayackley.bsky.social

29865121427

Andrew Golubev @golubev.bsky.social · 25/08/2026

If you work over SSH, you've lost a long task to a dropped connection. tmux fixes that: it keeps your commands in a session living independent of SSH. Detach, close the window, reconnect - everything survives. #tmux #SSH #sysadmin #Linux #devops

120

Andrew Golubev @golubev.bsky.social · 25/08/2026

GPT-5.6 Sol just dropped to 50% off through AI Gateway. $2/M input (was $4), $10/M output (was $20), $0.25/M cache read. Runs to Sep 18, 2026 - no promo code. Unified API for 30+ providers. Point to gpt-5.6-sol and the discount applies. #GPT5 #AI #Gateway #LLM #Pricing

000

Andrew Golubev @golubev.bsky.social · 23/08/2026

Parallel Search Fast: web search at $1 per 1000 results, ~700ms, #3 on Artificial Analysis. Built to pair with cheap models like DeepSeek V4 Pro. Cuts the search share of agent cost from ~50-68% down to under 12%. Cheap models + cheap search = viable agents. #Parallel #Search #AI #agents #WebSearch

000

Andrew Golubev @golubev.bsky.social · 23/08/2026

000

Andrew Golubev @golubev.bsky.social · 22/08/2026

DeepSeek just made weekends half-price. From August 23, off-peak rates apply all day on Saturdays and Sundays. What was $0.44/M input is now $0.22/M. For anyone running AI agents: batch the token-heavy work on weekends, cut the bill in half. Calibrate your cron jobs around it.

000