---
title: "«Вастрик.Ай» digest for the previous 24 hours — 2026-09-22"
description: "Channel discussion of AI tool economics and limits, GLM 5.3 Flash, subagents, a Dota 2 prediction experiment, and code comments."
date: 2026-09-22
---

# «Вастрик.Ай» digest for the previous 24 hours

The main topic was the practical economics and limitations of AI tools. Participants discussed Gemini Desktop/Spark with the Obsidian plugin: access to the vault works, but the quality of its work and separate quotas remain open questions—the Google AI Pro subscription is consumed separately from the API and Antigravity. In parallel, people complained about opaque Codex limit consumption and ambiguous refusals on grounds of “cyber abuse”; they found no shared explanation of the boundaries or how corporate subscriptions respond.

GLM 5.3 Flash drew much attention: participants praised its low price and limits, especially the free 19:00–04:00 Moscow time window. According to reports, the Z.ai promotion was extended through October 7. Resource: https://docs.z.ai/devpack/notice/event-glm-5.3-flash

They discussed when subagents help: mainly on large tasks that can truly run in parallel, rather than on cheap plans. One live example is an agent learning to control a game's built-in AI through macros and reverse engineering of an old game. The prototype is still rough, but its author promises to publish it once refined.

In one experiment, 12 LLMs predicted TI2026 Dota 2 matches. DeepSeek performed best; agents lost to a fixed context, and Astra showed no advantage. The virtual ROI was negative; the experiment cost $25.5. Post: https://t.me/token13th/13

Participants also debated comments in code. One radical position was to ban them with a linter for agent-written code; a compromise was to retain only explanations of non-obvious platform/SDK bugs.