Alexander Pucher

[Alexander Pucher @alexanderpucher]

@osanseviero Awesome. Really looking forward to stepping things up with audio and video vs gemma4 e4b

[Alexander Pucher @alexanderpucher]

@ClaudeDevs The animation was well worth the extra tokens

[Alexander Pucher @alexanderpucher]

Gemma4 E4B is perfect for voice agents - except it's not. Fused TTS+LLM, good knowledge - what's the issue?

It ain't obvious at first.

It takes a two-turn conversation to see the problem: that second turn needs user context from the first turn. Unless we want to build up (and blow out) all user context via voice, we'll need to transcribe the first turn's voice input. No big deal.

OK, so should we run the model twice - first to respond fast, then to transcribe? Unfortunately, the model sometimes hears things differently. It's hella confusing when the response and transcript don't fit together.

So transcribe first, then run it again to respond? Now we're losing all those fused-voice speed benefits.

Maybe transcribe and respond within the same request? But now we're wasting time waiting for the transcription stream before getting to the response.

OK, so maybe ask Gemma to first respond and then transcribe what it actually heard within the same request? Now we're aggregating and chunking streams, but yeah, that works!

... most of the time. Unless Gemma messes up the output format, which it eventually will, since we keep rewriting the history, and after a few turns Gemma's output starts to align with the assistant output it sees, which only contains the response but no transcript (because the transcript, by definition, gets assigned to the user).

System prompt? Stop kidding me. You didn't actually read that last run-on sentence either, did you?

The solution? Back to chained STT before LLM. And now there's no need for Gemma4 E4B in the first place.

[Alexander Pucher @alexanderpucher]

@picoito start with superpowers for codex. omo w/ opencode if you want automation

[Alexander Pucher @alexanderpucher]

@AIshaqui81766 @Gradio congrats! do you have any plans to support 5090s or other blackwell?

[Alexander Pucher @alexanderpucher]

Which coding cli works best for you?
- claude code
- codex cli
- opencode / aider / oss
- I only care about skills and models

[Alexander Pucher @alexanderpucher]

@IvSegal @suno yup, would've been my first chocie too

[Alexander Pucher @alexanderpucher]

@ScriptedAlchemy for coding kimi 2.6 or glm 5.1. for planning kimi > glm but it ain't a gpt 5 or opus. DSv4 is still a wildcard

[Alexander Pucher @alexanderpucher]

@vaaselene nice free tier

[Alexander Pucher @alexanderpucher]

@aryanlabde $200 claude max is a utility bill for larger projects. Use for planning, codex or anything on OR or even local models work great for execution

[Alexander Pucher @alexanderpucher]

Claude went down and all work in progress was halted. Yet, the session limit got maxed out in the meantime somehow

[Alexander Pucher @alexanderpucher]

That @ChatGPTapp codex + @claudeai code dual setup is coming in handy. The finishing touch would be a local qwen 3.6

[Alexander Pucher @alexanderpucher]

@ClaudeDevs any way to pass claude designs to claude code cli without manually creating and downloading a "handoff package" ?

[Alexander Pucher @alexanderpucher]

One of those AI moments. @claudeai and @ChatGPTapp going on a lunch break together.

[Alexander Pucher @alexanderpucher]

claude max double quota hack - make sure your quotas reset on mon or tue. whenever something's obviously broken, @claudeai resets the limits on thu. /s

[Alexander Pucher @alexanderpucher]

Life lesson - It's easier to adapt your workflows to the AI agent than the other way around. If something wasn't part of their training, it's impossible to make it work consistently. My notes from one year of agent-first coding:

1/ Choose your stack and language wisely. @claudeai , @ChatGPTapp , @GeminiApp & Co know everything on the surface level, but they have different strengths. We initially used Claude with Java, but switched to TS and Python - felt like productivity doubled.

2/ Agents do 90% of the work, but you don't know with 10% is missing. Use Superpowers, agent frameworks, etc - doesn't matter. There's always something that falls off the plate once the project grows. Checklists, hooks, and CICD are all great, but context-engineering remains an art.

3/ AI models change all the time. Just because you you finally made it work tday doesn't mean it keeps working tomorrow. @AnthropicAI made headlines recently, but this holds true in general. Thinking budgets, alignment, caches, etc change all the time. Feels like parenting a toddler, but without the reward of growth.

It's an awesome tech when it works, but you can't rely on it without guardrails and common sense yet.

[Alexander Pucher @alexanderpucher]

I'm back to managing claude agent memory manually. Auto mode is great to get started, but the memory becomes bloated on unpredictable over time

[Alexander Pucher @alexanderpucher]

@JakubSmal78647 Yup. Let's vibe-code another app for that too.

[Alexander Pucher @alexanderpucher]

A year ago I was running 1-2 build servers for an entire team - now that's 1-2 per developer. Coding agents can be a big productivity boost, but sometimes with unexpected side-effects. The massive growth of CI usage is one of these.