Waqqas Mir

AI Enthusiast

[Waqqas Mir @w_mir01]

@TheRobCollett @typesafeai @typesafe_ai Sonnet, discrete level: SD 2.39
Jev, discrete level (argmax): SD 1.39
Jev, as shipped (probability-weighted mean): SD 0.36

[Waqqas Mir @w_mir01]

Same resume, same job description, 5 runs

Sonnet 5 score: swung up to 16 points. temp=0 barely helped and the thinking pass has its own variance.

@typesafeai Jev: ±1 point

Jev vs Sonnet
0.3s vs 19s
$0.00013 vs $0.044
1 pt variance vs 16
Ranking quality: 0.76 vs 0.81 Spearman

[Waqqas Mir @w_mir01]

@altryne This is how I spent my Saturday
https://t.co/jQwabnyG36

[Waqqas Mir @w_mir01]

Same resume, same job description, 5 runs

Sonnet 5 score: swung up to 16 points. temp=0 barely helped and the thinking pass has its own variance.

@typesafeai Jev: ±1 point

Jev vs Sonnet
0.3s vs 19s
$0.00013 vs $0.044
1 pt variance vs 16
Ranking quality: 0.76 vs 0.81 Spearman

[Waqqas Mir @w_mir01]

@typesafeai @typesafe_ai Takeaway: if a number is the product, don't let a text model sample it. Decompose the judgement into scored questions and do the arithmetic in code.

[Waqqas Mir @w_mir01]

@typesafeai @typesafe_ai The catch:
Jev can't replace the call. It generates no text, and the report still needs Sonnet for skills gaps, strengths and rewrite suggestions. So it's +0.3s on top, not a complete replacement.

What you get with Jev is a score that doesn't wobble.

[Waqqas Mir @w_mir01]

@tldraw and delivered 🚀

[Waqqas Mir @w_mir01]

@Joshua_WD Thanks.
I'm trying paseo these days with my vps. it seems to be working well with my workflow of shifting between mini and phone.

[Waqqas Mir @w_mir01]

My recent unlock has been running claude code on a hetzner vps with tailscale.

on my mac mini, i run it through cmux and on my mobile I use termius.

Not being tied to keeping my mac awake and my workstation is a gift.

Thanks to @levelsio for the detailed set up guide.

[Waqqas Mir @w_mir01]

@pggalaviz The set up was easy. Asked grok and chatgpt for a detailed step by step guide by asking them to analyze @levelsio tweets.
Buying the vps took time as hetzner machines aren't readily available. other than that, its a 30 minutes activity.

[Waqqas Mir @w_mir01]

@khushiirl 1. Gitgent
2. Githands
3. Agently
4. Botools (read like bottles, tools for bots)
5. GentZ
6. Gitizen

[Waqqas Mir @w_mir01]

@TheViableEdge @sama @thsottiaux
Please hook up Adam for the next early access 🤝

[Chris Brownridge @chrisbrownridge]

Finally figured out how to get unique outputs from Claude with Seedance 2.5 for this @NordVPN ad.

I gave it a simple ask
"i want you to plan out a wild, wacky 30s ad for the product NordVPN dont make it look like anythiung else"

24m later, here is the output. One shot.

[Theo - t3.gg @theo]

If you don't want to keep up with all the Astra vs Fable 5.1 content that is incoming, here's a quick summary

[Waqqas Mir @w_mir01]

I'm seeing this in my meetings with CEOs running small to medium sized tech companies.

CEOs who don't use AI themselves are still debating which LLM to choose for the team, and they're still worried about a secret world wide trap of getting people hooked on AI before raising the AI prices.

We're nearing the end of 2026 and companies are still having these discussions while they sit and wait.

[Waqqas Mir @w_mir01]

@madarco if you're waiting to buy the ticket, this is your sign 😇