Ibad Desmukh

Building infrastructure for open-source mechanistic interpretability research

[Ibad Desmukh @ibaddesmukh]

@AsfiShaheen @AsfiShaheen you might enjoy studying Linear Algebra https://t.co/1Y0r1OJDqT

[Ibad Desmukh @ibaddesmukh]

@00000sol0 Yes https://t.co/ySMOihJaZ8

[Ibad Desmukh @ibaddesmukh]

Electronic Design Automation (EDA) software companies like Synopsys $SNPS and Cadence Design Systems $CDNS may be the silent winners of the current AI infrastructure buildout. While memory and chip manufacturers are the key focus right now, the expected demand growth for their products is driven by the currently dominant Mixture-of-Experts (MoE) based Transformer AI architecture. However, any future paradigm change, such as the introduction of post-Transformer AI models like Joint-Embedding Predictive Architecture (JEPA), or next generation State Space Models (SSMs), would require massive redesign of AI chips. This constant evolution will generate continuous, sticky demand for EDA software subscriptions. Nvidia recently made a $2 Billion investment in Synopsys, while Cadence Design Systems announced its collaboration with TSMC to accelerate design of next-generation AI silicon.

THIS INFORMATION IS FOR INFORMATIONAL/EDUCATIONAL PURPOSES ONLY AND DOES NOT CONSTITUTE FINANCIAL OR INVESTMENT ADVICE.

[Ibad Desmukh @ibaddesmukh]

One of the reasons why WhatsApp efficiently scales to over 3 billion monthly users [1] is tail-recursion optimization [2]. Mainstream programming languages like Python cannot execute iterative processes efficiently through recursion [3]. A typical recursive call in Python uses a call stack that grows linearly with each call i.e. O(n) space.

Conversely, WhatsApp runs on Erlang, where each user connection is an extended tail-recursive process that uses O(1) space instead of growing O(n) with every call. This is more memory efficient.

An analogy of this is an assembly line, where one machine fully completes its job, passes the product on to the next machine, and forgets about it. If there is no tail recursion, each machine passes the product on, but it must wait to finish one last step after the next machine finishes, so every machine needs to hold the product in its memory until the product reaches its finished stage. This causes the memory requirement to grow with the size of the entire job i.e. O(n) space.

References: [1] I. Mehta, "WhatsApp now has more than 3 billion users a month," TechCrunch, May 1, 2025. [2] P. D'Incau, "An interview with Eugene Fooksman hashtag#erlang," Paolo D'Incau's Blog, Mar. 27, 2013. [3] H. Abelson, G.J. Sussman, and J. Sussman, Structure and Interpretation of Computer Programs, 2nd ed. Cambridge, MA, USA. MIT Press, 1996.

[Ibad Desmukh @ibaddesmukh]

I'm excited to release a fault-tolerant machine learning training pipeline that automatically recovers from cloud interruptions. Press Ctrl+C anytime to simulate an unexpected interruption during training, restart the system, and see it automatically resume from the checkpoint where it stopped. Training machine learning models on spot instances allows for up to 90% cost savings [1]. However, spot instances may be interrupted at any time making fault-tolerance essential.

The ML training pipeline is built to forecast solar power generation using an LSTM with synthetic data for demonstration. The entire implementation is less than 500 lines of clean, testable code. It uses Python and PyTorch for model training, and pytest for test-driven development. System design is based on a philosophy of simple interfaces and deep modules [2]. To ensure data reliability, atomic writes are used for checkpointing.

Browse the code and try it yourself at https://t.co/IQFjmVxhpb

References:
[1] Cast AI, "Reduce cloud costs with spot instances," Cast AI Blog. [Online]. Available: https://t.co/QJQys2BDYW. [Accessed: Oct. 20, 2025].
[2] J. Ousterhout, A Philosophy of Software Design. Palo Alto, CA: Yaknyam Press, 2018.

[Ibad Desmukh @ibaddesmukh]

@goodreads The Rise and Fall of Athens - Plutarch

[Ibad Desmukh @ibaddesmukh]

@_safiamahmood Mocha 😀

[Ibad Desmukh @ibaddesmukh]

@Sabbandkardo Lol I'm guessing less than 10%

[Ibad Desmukh @ibaddesmukh]

@Sabbandkardo Still think there's a long way to go up.

[Ibad Desmukh @ibaddesmukh]

@Sabbandkardo Whats SYS % allocation in your portfolio

[Ibad Desmukh @ibaddesmukh]

@najam_ali @shahzaibis @SajidSomi @narejo @rogueonomist Growth cagr of revenue or net profit?

[Ibad Desmukh @ibaddesmukh]

@AzamAKhan2 @AzamAKhan2

[Ibad Desmukh @ibaddesmukh]

@narejo @pakstockexgltd yes please

[Ibad Desmukh @ibaddesmukh]

@Sabbandkardo @serialinvestor @sohailkarachi @najam_ali Engro profit doubled but stock is down

[Ibad Desmukh @ibaddesmukh]

@AzamAKhan2 certainly, the undervaluation is psychological rather than fundamental

[Ibad Desmukh @ibaddesmukh]

@Suhail sorting by importance (favorited sites/frequently visited) etc.

[Ibad Desmukh @ibaddesmukh]

@lexfridman hamlet by shakespeare

[Ibad Desmukh @ibaddesmukh]

@AzamAKhan2 thanks!!