Consistently found Abhishek Murarka 💹🐂 to be a clear thinker - especially after listening to his thoughts on a long form podcast recently

I've found LLMs to be quite good at converting imprecise user instructions to precise in-code actions - I think this will turn out to be a very fruitful application design pattern for enabling new use-cases by reusing existing interfaces to reduce user-onboarding friction

Getting recommended by LLMs - with proper context and reasoning - is going to be the next evolution in SEO - in effect, this is compressing all the wisdom of the crowd and giving you true crowd sourced review on a topic - Claude recommending twilio alternatives here

The time required to get something to work on Telegram Messenger v/s WhatsApp API is astounding - WhatsApp still has much bigger distribution but Telegram devX beats it everyday of the week - approvals on WhatsApp can be an indefinite wait with no feedback

Contrarian view - Custom UIs for every use case add complexity & training for users. Chat/voice interfaces are familiar, simple & effective for most AI apps - of course, caveats apply - let users talk, type or show if possible - its much easier

Knowing the limits of what you can do with AI is going to be crucial in designing useful systems - technical knowhow is going to be as important as domain expertise - its context engineering and not just context

💯 Agree - and there is a related point to this - you can vibe code your way to a demo but much more difficult to create a cross-functional, maintainable, deployable system that integrates into existing workflows and provides sustainable value - this is the opportunity I think

While Claude has become quite good at following examples and generating structured output, extracting text from images is still not its forte - had to resort to Amazon Textract Amazon Web Services for OCR which has delivered much better accuracy - an example where AI is not better than

openai/gpt-oss-120b is surprisingly strong at converting natural language instruction to structured JSON compared to tool calling accuracy especially for deeply nested input schemas

Chat or messenger type UXs are an incredible form factor for getting started with agentic systems - there is almost no learning curve involved - domain experts can simply level up as power users without the friction of mastering a new and constantly updating interface

I think Starknet (BTCFi arc) 🥷’s infra is top-tier—fast, scalable, secure. But it needs its “Excel moment”: a game-changing app that’s sticky & iconic, driving adoption like Excel did for Windows. What app could do this?

Not every "guess" in data processing is a hallucination. In raw images from any digital camera - be it a DSLR or smartphone - each pixel captures just one color through a filter (like in Bayer arrays), so about 2/3 of the color data is interpolated from neighbors. I would call

Prompt optimization is a fragile process, but GEPA and DSPy continue to amaze me with remarkable results - turning a manual brittle process into robust engineering through reflection based prompt optimization - got almost ~30% latency reduction by teaching

Compute isn't the real bottleneck - shortage of talent at the intersection of domain expertise, knowing what can be built and knowing how to build it is - this is where magical experiences get created

Success of coding agents like Claude Code and Cursor like IDEs just proves this point - these products have all the 3 ingredients to make it work - it just turns out that people with the domain expertise here are programmers themselves who not only know what can be built but also

I am no thought leader, but what vitalik.eth has been saying recently as well as Eli Ben-Sasson | Starknet.io is imo a good reset of the north star - its asking the same question as the quoted tweet but of the base blockchain/L2 itself instead of the application teams - a good number of

Counter-intuitive at first, but now obvious in practice: Asking Claude Code to write code for an analysis task usually beats asking it to just do the analysis directly - often by a lot. Real example: Client had transaction ledgers split by location in a dozen separate PDFs.

Retrofitting existing ERPs and legacy systems with modern interfaces and tools is more likely to get a buy-in from institutions than a full rewrite from scratch ever will - irrespective of how easy it is with agent swarms. While there is an astounding hype around how enterprises