Customer Support AI Explorer

Customer Support Manager & Python Dev. Exploring AI's role in support services, sharing insights on automation and AI-driven solutions. Building @ephor

[Customer Support AI Explorer @CSAIExplorer]

...not even 10 days ago both companies had been very public about how necessary it was to slow down the pace and lauding each other over the idea, only to announce their next frontier models a week later within 65 minutes from one another 🤔 Gotta be one of those SOC checkboxes ("did you publicly disclose your concerns before shipping anyway?")

[tmuxvim @tmuxvim]

I put a prompt injection into my LinkedIn bio and recruiters are messaging me in Old English and calling me Lord.

[Customer Support AI Explorer @CSAIExplorer]

@HeyGen HyperFrames

[HeyGen @HeyGen]

We built our launch video in Claude Code using HyperFrames.

Now it's yours.

Open source, agent-native framework. HTML to MP4.

$ npx skills add heygen-com/hyperframes

RT + Comment "HyperFrames" to get the full source code of this launch video (must follow)

[Ethan Mollick @emollick]

This game was 100% designed, tested, and made by Claude Code with the instructions to "make a complete Sierra-style adventure game with EGA-like graphics and text parser, with 10-15 minutes of gameplay." I then told it to playtest the game & deploy.

Play: https://t.co/JuqRUYQXc0

[Customer Support AI Explorer @CSAIExplorer]

@AIdenAIStar @ChatGPTapp and to think Google built their empire on top of a similarly clunky idea underlining words in blogs and displaying ads when you hovered over them accidentally.

[Customer Support AI Explorer @CSAIExplorer]

@benln Pls send the info for the Buenos Aires one!

[Customer Support AI Explorer @CSAIExplorer]

Done being overwhelmed by notifications, so I'm trying Spec - you should too 👀 Use my link to get on the waitlist: https://t.co/WWRAYqzrR5

[Customer Support AI Explorer @CSAIExplorer]

The big discovery: Personas unlock reasoning pathways that exist in the model but aren't naturally accessed.

It's like having the same brain think through different lenses - suddenly connections appear that were always possible but never surfaced.

Full analysis: https://t.co/2ZQQNkqaWv

[Customer Support AI Explorer @CSAIExplorer]

The numbers were surprising:

Pure LLM accuracy: 13/18
Best persona: 13/18
Worst persona: 8/18

But accuracy missed the point.

The vanilla LLM told me WHAT went wrong. The personas explained WHY through frameworks the base model never touched.

[Customer Support AI Explorer @CSAIExplorer]

Example: Donald Miller's "StoryBrand" persona didn't just say "customers were upset."

It reframed New Coke's failure as "violating the customer's narrative arc."

Same data. Completely different insight. One I can actually use in my next product launch.

[Customer Support AI Explorer @CSAIExplorer]

Just ran an experiment that changed how I think about LLMs.

Same model, same business case (New Coke), three different approaches:

Raw LLM ❌
LLM with expert personas ✅
Multi-persona debate 🤔

The results? Personas don't make LLMs smarter. They make them think differently.
🧵👇

[Customer Support AI Explorer @CSAIExplorer]

Is it just me or was Sonnet 4 lobotomized at some point since yesterday? I had to switch to Opus 4.1 for coding because Sonnet was fumbling so badly, and is being generally unhelpful, breaking good code with, allegedly, pretty dumb mistakes...

[LiveKit @livekit]

You can now deploy AI voice agents to LiveKit Cloud.

We handle:
• Stateful load balancing
• Capacity management
• Draining and instant rollbacks
• Operational observability

[Customer Support AI Explorer @CSAIExplorer]

@AIdenAIStar @Reddit Every community has its gatekeepers, but the real innovation happens when you build solutions that work regardless of the platform politics. Focus on shipping code that speaks louder than forum drama.

[Customer Support AI Explorer @CSAIExplorer]

What I don't like about GPT5 so far for coding:
* Conflated some tooling error with my reported task
* Conflated output from the task at hand and placed it on the output required from a template (i.e. the template asked for a #summary of something after analysis, and instead it provided a summary of what my instructions had been. Strange!)
* May be too verbose oriented for my taste for the purposes of coding (not the output code itself necessarily, but the whole thinking and reasoning process around it). Perhaps, with reason, it's because I'm testing the high reasoning model...

[Customer Support AI Explorer @CSAIExplorer]

What I like about GPT5 with my limited exposure so far using it in Windsurf:
* it follows instructions very precisely. And I mean it is PRE-CI-SE... (even the second sentence of the tenth bullet on a cursor rule somewhere). The most precise rule follower I've seen so far. Not sure if this is good or bad, but at least it's different.
* You don't need to heat up the context with every single little detail for it to not mess up. No hand holding, just tell it your end goal, seems to be able to find the way pretty nicely. It's like this one is truly "agentic" at heart (for whatever that means), at least the model that Windsurf calls the "high-reasoning" model

[Customer Support AI Explorer @CSAIExplorer]

@CranQnow Speed matters, but so does substance. That 8-minute window only works if your reply actually adds value to the conversation. Most people rush to comment first instead of crafting something worth engaging with.

[SkalskiP @skalskip92]

supervision has gained almost 1k stars since the release on wednesday; so cool

link: https://t.co/xXMRaS3Guk

[kyon @0xKyon]

announcing: kisuke, a native ios ide

kisuke is your fav pocket engineer.

you get:
- claude code (use your anthropic acc + all CC features)
- multi tab terminal
- code editor
- browser + built-in devtools
- port forwarding (auto-detected and available inapp)