Simrat

[Simrat @simrat11exp]

Reading a business, liking everything until I discover that they have used portion of excess cash for purchasing Bitcoin ETF.

[Simrat @simrat11exp]

Imagine that I am having a conversation with AI where it has to ground answers based on docs on my laptop. So its not static. I can go into a conversation thread with it.

Why I can't use claude or Gemini apps because each session has different context plus voice mode and quality is very ordinary in these apps. So I usually do handsfree conversation during walking or gym.

[Simrat @simrat11exp]

My new AI productivity hack - I often get the itch to dig into a company or a topic while I'm away from my laptop, but that's exactly where all my documents and context live. So I built a bridge: I message a Telegram bot from my phone, code running locally on my Mac picks up the prompt, queries an LLM (Claude/Gemini/Deepseek) against the documents on laptop, and sends the answer back as a voice note, usually within a minute or two. The only cost is the LLM API call; voice generation is free, handled by an open-source TTS model running on the Mac itself.

[Simrat @simrat11exp]

@kaustiwari They hold context, could be books, notes, bunch of past n concalls for a company etc etc.

[Simrat @simrat11exp]

@mehtas7887 Basically, whatever prompt I give. Lets say I ask about summarising some concall, it will send me voice note with a summary. So I can talk about any topic with it over telegram.

[Simrat @simrat11exp]

@mehtas7887 Meaning?

[Kairos @KairosPraxis]

@labubu_trader If true, huge for AT&S who are basically using their entire facility in Malaysia to manufacture substrates for $AMD

[Simrat @simrat11exp]

This weekend has been about spawning dozens of Claude agents to analyse 800+ concalls and help me generate new ideas - new companies which I never read before or old companies which I have not revisited for some time. Lot of names do not come up on technical scans these days due to narrow breadth and crowding in handful of names. Even 1-2 good ideas from this vast pool can drive returns in 2027/2028.

[Simrat @simrat11exp]

Looking at a micro-cap infra solutions business growing 35-40% and can grow PAT 2.5-3x in next 3 years effectively trading at 5-6x fwd. PE. The business is from unsexy sector - infra but its not an EPC or developer business but a kind of an aggregator + platform + services led business model trying to bridge fragmented supply with demand. On first glance it looks like a trading business but a deeper look, its moat is using lot of tech + operational rigour to extract value in a tough sector.

[Simrat @simrat11exp]

"You’ve heard the rumblings. Read the headlines telling us that AI will make developers obsolete. That anyone can code now. Just describe what you want to do, and tools will take care of the rest. That the era of the professional developer is over.

We’ve seen and heard this before. Early assembly programmers were told that compilers would make them redundant. Instead, compilers elevated the level of abstraction and opened software development to far more people. What once required deep hardware expertise became an act of logic and creativity. Entire industries emerged because software became something many could build. Businesses, research labs, and universities suddenly had the ability to create their own tools.

In the 2000s, operations engineers expressed similar concerns when cloud computing arrived. They feared automation would make them obsolete. Instead, it lowered barriers to experimentation and created an explosion of new projects, new companies, and new engineering roles. Every simplification produced greater demand.

Each technological leap forward has followed a similar pattern. Tools evolve, workflows change, and complexity increases, yet the core attributes of great developers remain constant. Creativity, curiosity, and systems thinking have continued to define the craft.

Time and time again we have seen that lowering the barrier for entry doesn’t eliminate the need for human expertise, it amplifies it. Generative AI lets us generate code in seconds, but if you put garbage in, you get really convincing garbage out. The AI doesn’t sit in budget meetings where leadership debates whether to optimize for cost or performance. It doesn’t understand that the customer service system needs five 9s of uptime while the internal reporting dashboard can go down during peak sales periods. It can’t read between the lines when a stakeholder says, “make it fast” but might mean “make it cheap.” The politics, the constraints, the unspoken priorities that shape every technical decision are nuanced and require a developer who understands why it matters to the humans who pay for it and the humans that will use it."

Wonderful read - https://t.co/wUfNAKfotQ

[Simrat @simrat11exp]

@varadprm Yeah I meant it might be sub par roic thing but economics can improve if right atmosphere is created, conducive pitches and rivalry is there.

[Simrat @simrat11exp]

Test cricket is such a lovely sight after boring and overloaded T20s. I wish some billionaire had been a test cricket fan to fund it for the majority of the year.

[Simrat @simrat11exp]

@reselfff It can only improve if people play and prioritise it more.

[Simrat @simrat11exp]

@phreakv6 Well I know for the fact that in my current work exp at Amazon, we are paranoid about even 100ms ttft. So there is big effort to drive this down to open up new usecases.

[Simrat @simrat11exp]

Yes I use deepseek myself and know it is very comparable at fraction of cost. One, it is does not seem like Chinese models would be serving Indian usecases, second, since enterprises pay one of 3 cloud companies, the incentives of those companies is not to serve these open weight models. On consumer AI, product as whole would win, not model alone and we all know network effects would play out there embedding sota models in consumer products again controlled by big tech + new AI labs. So unless India has its own tech stack + models, the story ends at paying big tech one way or the other.

[Simrat @simrat11exp]

Network latency is significant part for a user sitting in India and getting served by data center in US. You are also thinking only about how AI is used currently. Once physical AI is there even 10-20ms hops would count. Also on the open weight models doing 60% of tasks does not mean 60% of the AI monetary value would accrue to them. It would be pareto in nature where sota models take away bulk of the monetary gains from this, so 60% tasks number loses relevance.

[Simrat @simrat11exp]

@phreakv6 They will anyway invest sooner or later as inference usecases are better served near the consumer. On open weight models I doubt they would reach far, until they are really sota. People want high performant reliable models.

[Simrat @simrat11exp]

I have never understood hype around a defence company. Seems like its spinning people on their axis. Always prefer focused managements and businesses. One should never buy into vague guidances. Any growth rate guidance beyond 30-40% CAGR for 3-5 years should be questioned not celebrated.