Ashhar

[Renjit Philip Thomas @renjitphilipe]

@krishnabgowda Sir, requesting urgent intervention on BBMP’s repeated waste dumping near Divyajyothi Lakepoint Tower Apartments, Begur Lake, Singasandra. Despite resident complaints, garbage piles continue to threaten public health. Kindly escalate for immediate action.

[Ashhar @aproli90]

@DavidBurkett38 @sama Why is everyone talking about their revenue and not their burn? 😬

[Ashhar @aproli90]

@ClaudiuEnacheS @AnthropicAI Isn't technology (including basic need like internet) lagging behind in Europe due to regulatory bodies who still have ancient mindset?

[Ashhar @aproli90]

@boltdotnew GIMME TOKENS

[Ashhar @aproli90]

Perplexity just killed it with their new AI Assistant on Android 😎

It's everything in Gemini + web search + better voice interaction, etc etc.

After using it for a while, I had zero hesitation to switch my default assistant from Gemini to @perplexity_ai 😄

[Ashhar @aproli90]

Made more improvements to https://t.co/UdTPKKTVQu hosted on my HF Cloud.

1. Very clear cache hit and miss messaging to give transparency to user and educate them
2. External images can now be easily embedded without CORS errors 🎉

[Ashhar @aproli90]

Tried https://t.co/badBmVop8g by creating a personal portfolio. It's very different from bolt & vercel.
It has pre-built sections that you can insert and have complete control to edit the code.

I think it's great for creating minimalistic websites really fast 😃

[Ashhar @aproli90]

My experience with the new hot screen sharing feature of Google AI Studio

It was fast and able to interpret screen contents quite well 😊

The only issue was that it's a bad listener, leading to unnecessary to and fro 😬

Link: https://t.co/W0VSTbznZG

[Ashhar @aproli90]

While I'm still waiting for my PR on Claude Prompt Caching to get approved in the official https://t.co/UdTPKKTVQu repo, it was encouraging to see 43 people cloning my forked repo and saving tokens 😅

My Repo: https://t.co/ocOQVSlkof

[Ashhar @aproli90]

Tried https://t.co/oqsMlI1ec0, a simple, chat-based way to design your database schema and deploy it to @supabase

You can even make changes iteratively after initial setup, but note that once you redeploy, table data is lost (I made that mistake 😓)

[Ashhar @aproli90]

I've been playing with the buzzing DeepSeek for few days. Here is my few cents:
1. It's cheap and great at content writing and medium-complexity coding
2. It's still quite behind Sonnet & Haiku when it comes to high complexity coding
3. It didn't do well with large system prompts

[Ashhar @aproli90]

If you don't want to setup https://t.co/UdTPKKTVQu on your local machine, I have set it up for you on HF cloud
Link: https://t.co/ZLOQZMumSq

Just enter your Anthropic API key and enjoy.
This has prompt caching baked in with models & prompts optimized for best cost & performance

[Ashhar @aproli90]

I believe this can be massive! I tinkered with the source code of https://t.co/UdTPKKTVQu and added Claude's prompt caching feature.

This means, upto 90% cost savings on tokens when iterating fast 😃

PR in main repo still open.
You can use my fork: https://t.co/ocOQVSlkof

[Ashhar @aproli90]

Tried the experimental Google Whisk. I thought it could serve a good use case case of creating high quality product/food photoshots.

The outputs were quite impressive, but it's extremely slow at the moment (> 5 mins per output)

Link: https://t.co/B0Kr6gIuy3 (required US VPN)

[Ashhar @aproli90]

It's unbelievable that I could create 50+ logos in few just few mins for our "Chart Genie" project using tools like Recraft AI and Google ImageFX.

@recraftai can create vectors in SVG (means unlimited scaling 🤯)

Link: https://t.co/8qQS4KzaMu
☝ Both you and I get 200 credits😃

[Ashhar @aproli90]

Compiled all my trials and analysis in this short doc: https://t.co/EZZTDYj9RW

[Ashhar @aproli90]

https://t.co/9G6GNRL0Ob is awesome.
https://t.co/mvKGaa8Nqs is awesome + cheap 😃

For new folks, https://t.co/UdTPKKTVQu is a way to run https://t.co/w3qKYXlgZl locally using the LLMs of your choice :)

My goal: Can I get the same outputs at a lower cost?

<contd in comments>

[Ashhar @aproli90]

After trying multiple cost-effective models, the winner emerged as Claude Haiku 3.5 🥇🎉

[Ashhar @aproli90]

In Claude Sonnet 3.5 (new), I accidentally discovered that "solve" step-by-step reasoned better than "think" step-by-step.

Interestingly, "think" step-by-step worked good in the old Sonnet 3.5.

If you're solving a reasoning problem using Claude, do try this & share your view