Ayyaz

Software Engineer.
Helping Businesses grow with AI

[Ayyaz @ayyazdev]

#DeepSeek on the spot

[komal @komal_uk01]

Same prompt.

Different AI.

Completely different answers. 🤨

[komal @komal_uk01]

Same prompt.

Different AI.

Completely different answers. 🤨

[Ayyaz @ayyazdev]

@grok Huh? why the Grok bot? whats the reason 😀

[Ayyaz @ayyazdev]

Hey @grok remove the worst agent.

[Ayyaz @ayyazdev]

#Muse vs #Claude

Competing as a motion designer.

every pixel and every sound written in code. No video model, no After Effects.

#ClaudeCode #MotionDesign #CreativeCoding #claudevideos

[Ayyaz @ayyazdev]

MCP Inspector 2.12.0 patches three vulns that hit every v2 release. A remote server's OAuth URL could run commands through the opener on Windows/WSL, and DNS rebinding could reach the API token and stored OAuth tokens. If you poke at untrusted servers with it, update first.

[Ayyaz @ayyazdev]

@nextgenfusion Yeah, the advisory-lock leader is the cleaner fix. I'd add lock_timeout first though, since it turns a silent hang into an error you can retry.

[Ayyaz @ayyazdev]

Dhaga found a Postgres deadlock before shipping Pages. ALTER TABLE ... ENABLE ROW LEVEL SECURITY takes an ACCESS EXCLUSIVE lock even when RLS is already on, so one instance replaying schema at boot deadlocked with another running CREATE INDEX CONCURRENTLY.

If your app applies schema at startup, boot two instances at once in a test.

[Ayyaz @ayyazdev]

EDIS Global lost an EPYC CPU in an Amsterdam host on Oct 23. The second socket couldn't carry it because half the storage hung off the dead one. A new CPU came by taxi and didn't fix it; board swap expected Monday.

If your app lives on one VPS, keep a warm copy on another host.

[Ayyaz @ayyazdev]

LiteLLM 1.106.4 fixes the proxy sometimes exiting on startup with an ImportError when it couldn't reach the remote model cost map. Broken since v1.104.0, also fixed in 1.104.7 and 1.105.5.

If your gateway has restricted egress, upgrade and test a cold boot with the network blocked.

[Ayyaz @ayyazdev]

@paulopierrondi And the hardest to spot in logs, since the request shows up as a normal 200. Worth counting streams that end without a stop reason.

[Ayyaz @ayyazdev]

Ollama 0.40.9 now reports generation failures on streamed OpenAI Responses and Anthropic Messages requests. Before, those streams could stop silently after partial output, which an agent reads as a finished answer.

Update, and treat a stream with no stop event as an error.

[Ayyaz @ayyazdev]

@ajs6888 对，客户端最好也加一道检查：没收到 stop 事件就重试或直接报错，别把半截输出当结果。

[Ayyaz @ayyazdev]

@_chrismdavidson Good add. Bumping max_logical_replication_workers doesn't help if max_worker_processes has no room. Polling pg_subscription_rel for non-'r' states is a clean check.

[Ayyaz @ayyazdev]

amaiko's write-up on moving live tenant Postgres databases to CloudNativePG (Oct 16) has a gotcha I'd have walked straight into.

Postgres caps logical replication workers at 4 by default. One of their tenants had four databases, so its four subscriptions took every worker and the initial table-copy workers never started. Nothing errored. Their first rehearsal only passed because one copy grabbed a worker first.

The rest is worth stealing too. A typical tenant was down about 3 minutes, the 23 GB one just under 7, and cutover refuses unless every table's content hash matches on both sides. A hash check that didn't run counts as a failure.

Before the first CREATE SUBSCRIPTION I'd count subscriptions plus their copy workers and raise max_logical_replication_workers and max_worker_processes to fit.

https://t.co/Hp2RxEA3JO

[Ayyaz @ayyazdev]

@Wasim_edit Coming from a video editor, I'll take that. The prompt is in the thread if you want to see how it was set up.

[Ayyaz @ayyazdev]

I asked Claude to show off as a motion designer.

15 seconds. 8 disciplines. 900 frames at 60fps plus the soundtrack, every pixel and every sound written in code. No video model, no After Effects.

Prompt in the reply 👇

#ClaudeCode #MotionDesign #CreativeCoding #AIart

[Ayyaz @ayyazdev]

@ichelpark Right, with no motion blur in a code render the extra frames are doing that job. Real blur would be the next thing to add.

[Ayyaz @ayyazdev]

Tigris US East had about an hour of failed requests on Oct 23. Their status note says a memory accounting error ran some servers out of memory during routine provisioning.

If your placement logic tracks memory itself, check it against what the host reports and alert on drift.

[Ayyaz @ayyazdev]

Crush v0.98.2 fixed two permission bugs: server permission requests could be scoped to the wrong session, and the write tool made parent directories before the write was approved.

If you build agent tools, check that nothing touches disk until approval comes back.

[Ayyaz @ayyazdev]

@TheXSami Followed back @TheXSami. It was one single prompt that I attached above.

[Ayyaz @ayyazdev]

Prompt: Build a 15s, 1920x1080, 60fps motion showreel entirely in code. 128 BPM, one bar per discipline: kinetic type, shape morph, generative, 3D, data viz, UI, liquid, logo. Each hands off to the next. Orange/black/cream, Geist, motion blur, synced soundtrack, seamless loop.