Bookshelf

Not book reviews. Just personal notes and mental models from books that changed how I think.

As an Amazon Associate I earn from qualifying purchases made through the "Buy on Amazon" links below, at no extra cost to you. See the About page for the full disclosure and content license.

The Nvidia Way cover
The Nvidia Way
Buy on Amazon

The Nvidia Way — Tae Kim

Read: November 2026

Jensen Huang’s story, told through the company he has run for more than thirty years. Tae Kim’s argument is that Nvidia’s edge is not the chips, the market value or luck. It’s an organizational culture of extreme independence, extreme standards and extreme speed, built around a CEO who believes the company’s worst enemy is its own complacency.

The first third felt like time travel. The PCI bus, the first separate graphics cards, Intel’s integrated graphics, the NV1 flop: all the hardware history behind how a GPU came to exist. The book explains the technical parts in plain terms too, so you don’t need a chip background to follow it.

The CUDA playbook. Mark Harris and the early CUDA team were tiny: a handful of engineers on the driver, the math library and the compilers. Getting it adopted took more than good software. Jensen put CUDA on every GeForce card, not just the expensive ones, so it would be everywhere. Nvidia gave away free cards, recruited academics to teach it (David Kirk flew across the country every other week to lecture), recorded the courses and put them online, and ran summits where scientists could ask for a chip feature and get it. Years later, when AI arrived, developers already knew Nvidia and a decade of bug fixes sat in its libraries. cuDNN, built from the first deep learning experiments, turned that into a product.

The mission is the boss. Jensen’s stock phrases keep people pointed at the right thing: “The mission is the boss.” “Speed of Light.” “How hard can it be?” When someone rambles in a meeting he says “LUA”: listen to the question, understand the question, answer the question.

Whiteboards, not slides. It comes up again and again. A slide deck lets you hide an unfinished thought behind nice formatting, and a whiteboard doesn’t. You also have to wipe it clean when you’re done, however good the idea was, and start over. Nvidia’s culture discourages looking back at either errors or successes.

The Top 5 email. Instead of status reports, which get sanitized until they’re useless, everyone sends their top five things in progress and what they’re seeing in the market, each bullet starting with an action word. Jensen read them as a live feed of the company. After a pile of them mentioned machine learning, he decided Nvidia wasn’t moving fast enough on it.

He looks in the mirror and says “you suck.” That’s his self-improvement method: every morning, no pride in the past, focus on the future. He expects the same from everyone at Nvidia, and the book says he’d rather “torture them into greatness” than give up on someone.

Pain and suffering as advice. Whenever anyone asks Jensen how to be successful, his answer is “I wish upon you ample doses of pain and suffering.” It sounds like a joke but the book treats it as his actual philosophy, and it’s the title of chapter 1. His reasoning: “People with very high expectations have very low resilience. Unfortunately, resilience matters in success.” Greatness comes from character, and character only comes from overcoming setbacks. The hours reflect that view (sixty a week is the bare minimum, and the book doesn’t pretend otherwise).

Engineer CEO vs. business CEO. Jensen walks into a product meeting knowing more about a minor Wi-Fi product than the people running it. The book sets this against Intel, which turned down Apple’s iPhone chip in 2006 because the price was below its forecast cost, and sold off its ARM mobile processor unit the same year. Then it missed AI too. Intel’s CEO couldn’t see where the technology was heading, and Jensen could.

Outrunning, not fencing. “I don’t see Nvidia as some kind of IP fortress. Our strength is just outrunning the other guy.” Three teams on staggered schedules (three chips per design cycle) and “the number one feature of any product is the schedule” are that idea in practice.

Bootstrapping the chicken and the egg. Bryan Catanzaro on ray tracing and AI frame generation: you can’t have AI in hundreds of millions of homes without building it first. It took six years to get the frame-generation model accurate enough, and he points out that most academics can’t spend six years on one project because they need to graduate. The book mentions a lot of PhDs, and this is part of why they end up at Nvidia.

Full stack. Like Apple with the iPhone, Nvidia designs the hardware, software and networking together, where most rivals just make chips. The Mellanox acquisition (InfiniBand) is the networking piece, and the same InfiniBand shows up in the hardware chapter of Inference Engineering, so it was a nice connection to see it from both sides.

“Not everything here is transferable. Most companies can’t ask for sixty-hour weeks and get them, but the whiteboard, the Top 5 email and the habit of answering “I don’t know, but I’ll find out” are cheap to copy.”


Inference Engineering cover
Inference Engineering
Buy on Amazon Free download

Inference Engineering — Philip Kiely

Read: November 2026

A model’s life has two phases. Training is where it learns, and inference is where it gets used to answer real requests. This book is about the second phase: making AI models faster, cheaper and more reliable once people depend on them. It works in three layers: the runtime (one model on one GPU, the graphics processing unit chips these models run on), the infrastructure (scaling across clusters, regions and clouds) and the tooling engineers use across both.

What makes it easy to follow is the order. Each chapter assumes only the ones before it, going from “why would I do this” to models, hardware, software, optimization techniques, other modalities and finally running it all in production. These notes follow the same path.

1. Prerequisites: why and when. Calling a model through an API is the easy start. Running it yourself on dedicated GPUs makes sense when traffic is high enough that paying per GPU beats paying per million tokens, when you run a custom or fine-tuned model, or when your product chains several models and the delays between them add up. This chapter also sets up the measuring stick. Time to first token (TTFT) is how long before the answer starts appearing, and tokens per second (TPS) is how fast it streams after that. Both are reported as percentiles, because an average hides the slow requests: “P99 is 2 seconds” (P99 being the 99th percentile) means 99 out of 100 requests were faster, and the slowest one is what users remember. On picking a model, benchmarks are good for shortlisting but get saturated or gamed (Goodhart’s Law: once a measure becomes a target, it stops being a good measure). Distillation, which trains a small model to imitate a bigger one, is one way to get a cheaper model that behaves similarly.

2. Models: what is actually being computed. An LLM (large language model) writes one token (a word or word fragment) at a time, and each new token is chosen by looking at everything before it. The part that does the looking is attention, which works out which earlier words matter for the next one. Image models work differently: they start from random noise and clean it up step by step, and they’re a pipeline of several models, not one big one. Video is the same idea with far more data. The chapter ends with the key skill, working out what is slowing a model down. A GPU has two limits, compute (calculations per second) and memory bandwidth (how fast it can pull data in). Knowing which one you’re stuck on tells you which fix will help.

3. Hardware: where it runs. GPUs are built for huge numbers of simple tasks in parallel, unlike CPUs (central processing units), which are good at complex sequential work. Their onboard memory (VRAM, video random access memory) caps how big a model can fit, so large models are spread across the eight GPUs in a node, linked by fast connections (NVLink, NVSwitch), and across nodes by InfiniBand. The opposite case exists too: a GPU can be split into up to seven slices for small models. The chapter walks through NVIDIA’s generations (Hopper, Ada Lovelace, Blackwell, Rubin), other accelerators, and local inference on laptops and phones. The author’s view is that the future is local and cloud working together: small, quick queries on your device, heavy ones in the datacenter.

4. Software: what talks to the hardware. CUDA (Compute Unified Device Architecture) is NVIDIA’s platform for writing code that runs on its GPUs, and PyTorch sits above it for describing the math. Model weights are stored in formats like safetensors (weights only) or ONNX (Open Neural Network Exchange, which bundles weights plus the execution graph). On top sit inference engines, which are the programs that actually serve a model: vLLM and SGLang (fast to start) and TensorRT-LLM (the most tunable, but it compiles itself for specific hardware first). NVIDIA Dynamo works above all three to orchestrate large deployments. The chapter closes with how to measure honestly: shadow real production traffic onto the system under test so it sees real load without affecting real users, and profile it. A benchmark tells you how the system performs, and profiling tells you why.

5. Techniques: making it faster. Five families of optimization, and they build on the earlier chapters.

  • Quantization: store the model’s numbers at lower precision so it’s smaller and faster. Risk is low for the weights, moderate for the KV (key-value) cache and high for attention.
  • Speculative decoding: a small, quick model guesses several tokens ahead and the big model checks them all in one pass. When the guesses are right, you get several tokens for the price of one.
  • Caching: while generating, the model keeps notes on what it has already read so it doesn’t redo that work for every token. That’s the KV cache. Reusing it across requests saves a lot of compute, which is why some routers send your request to the server that already holds your context.
  • Parallelism: splitting one model across many GPUs, either for lower latency or for higher throughput.
  • Disaggregation: reading the prompt (prefill) and writing the answer (decode) are different kinds of work, so run them on separate machines. It pays off for something like a coding assistant, where a lot of code goes in and comparatively little comes out.

These don’t always combine. Bigger batches use up the spare compute that speculation needs, for example. The skill is picking a set that works together, and the book keeps returning to one principle: the more constraints you can introduce, the faster the system gets.

6. Modalities: the same playbook on different models. Everything above was mostly about text, and this chapter applies it to the rest. Vision language models read images or video along with text. Embedding models turn text into searchable vectors, which is what powers search, RAG (retrieval-augmented generation) and agent memory. ASR (automatic speech recognition) models like Whisper turn speech into text, and TTS (text-to-speech) models turn text into speech, with streaming and speech-to-speech variants for real-time use. Image generation gets kernel-level tuning, and video generation is the most demanding of all, best run on the newest GPUs with caching tricks and the work split across GPUs. Each modality has its own bottleneck, so the same techniques get applied differently.

7. Production: keeping it running at scale. This is where a fast model becomes a dependable service, and it’s a hard problem of its own, not a footnote. Containers package the model so it runs the same everywhere. Autoscaling adds and removes replicas using both traffic (a leading signal) and utilization (a lagging one), and cold starts matter because a new replica can take minutes to load a big model. A router decides where a request should go, while a load balancer decides where it could go. Beyond one cluster, a control plane handles deployments and global scaling while workload planes handle the traffic. GPU capacity is a blend of reserved instances for the baseline with on-demand and spot for peaks. GPUs also fail often, so the book’s stance is to expect it, cordon bad nodes and keep going. Changes roll out as canaries to a small slice of traffic, because a full duplicate fleet for blue-green deploys costs too much in GPUs. The chapter finishes with observability (latency percentiles, queue depth, utilization), client code and streaming protocols.

“The mental shift from per-token pricing to owning your unit economics is harder than any of the technical parts, and the book says so directly.”


What Got You Here Won't Get You There cover
What Got You Here Won't Get You There
Buy on Amazon

What Got You Here Won't Get You There — Marshall Goldsmith

Read: October 2026

A book about success not scaling the way we assume it does. The habits that got you promoted aren’t the ones that keep working at the next level, and past a certain point, the only real lever left is behavioral.

Keep a “stop doing” list, not just a “to do” list. Goldsmith’s central move: most self-improvement focuses on what to add, but the highest-leverage change is often just stopping something. “If he had simply stopped what he was doing, his reputation and net worth might have remained intact.” Stopping a behavior gets no credit and no attention, but it can matter as much as everything else combined.

We misjudge ourselves in both directions. People overestimate their strengths and overrate their weaknesses. Ask a team to estimate their percentage contribution to a project and the total always runs past 100%. Everyone’s playing a “what’s in it for me” calculus on every choice, including the ones they’d never frame that way out loud.

Truth isn’t the bar for saying something. “The fact that a destructive comment is true is irrelevant. The question is not, ‘is it true?’, but rather, ‘is it worth it?’” Most of what undermines us socially isn’t willful. Being bad at sharing information isn’t the same as withholding it, and paying attention to what we actually say is the clearest signal of what’s turning people off.

Feedback vs. feedforward. Feedback reports on the past, which you can’t change. Feedforward is the same information aimed at the future, ideas you can actually act on. If you’re soliciting feedback for yourself, Goldsmith says there’s really only one question that works: “How can I do better?” Not “what did I do wrong,” not “how was I perceived.” Just that.

It’s not what you say, it’s what they hear. Feedback tells you what to change, not how. The “how” is the real work, and transformations here take time to register with the people assessing you: “You have to change 100% to get 10% credit for it.”

Recognition is underrated and resentment is sticky. Of all the interpersonal slights people carry, not getting credit may be the one that lingers longest. Goldsmith’s fallback advice for handling any suggestion you don’t know how to respond to: just say thank you. When someone offers you an idea, you’re either going to learn more or learn nothing, never less.

“Marshall emphasizes gratitude a lot throughout the book. In this era, you could probably write most of it as LinkedIn recommendations.”

You don’t need all 39 attributes, just the few that are actually broken. No executive embodies every trait on a list of successful behaviors, and that’s fine. The real question for any gap: is it bad enough to merit fixing? If not, leave it alone.

“The same logic applies to professional goals generally: pick where you actually want to excel instead of trying to be well-rounded everywhere.”

Stop blaming others, especially for the choices that worked out. Understanding the past is fine if your problem is accepting it. If your problem is changing the future, understanding doesn’t get you there. We can’t see in ourselves what’s obvious to us in others, which is exactly why this stuff needs someone else pointing it out, and why it stings when they finally do. Goldsmith’s line on that sting: we need the painful, unsolicited feedback moments, because without the discomfort we might never find the motivation to change.

The 20 habits that hold successful people back. These aren’t character flaws, they’re transactional: one person doing something to another. The list, in full, because it’s worth having on hand rather than just a page reference:

  1. Winning too much: needing to win at all costs, even when it doesn’t matter.
  2. Adding too much value: the overwhelming urge to improve on every idea someone brings you.
  3. Passing judgment: rating others and imposing your own standards on them.
  4. Making destructive comments: needless sarcasm dressed up as being sharp or witty.
  5. Starting with “No,” “But,” or “However”: qualifiers that secretly say “I’m right, you’re wrong.”
  6. Telling the world how smart we are: needing to prove you’re smarter than people think.
  7. Speaking when angry: using emotional volatility as a management tool.
  8. Negativity, or “let me explain why that won’t work”: sharing negative thoughts even when nobody asked.
  9. Withholding information: refusing to share in order to maintain an advantage.
  10. Failing to give proper recognition: the inability to praise and reward.
  11. Claiming credit that we don’t deserve: overestimating our contribution to any success.
  12. Making excuses: recasting annoying behavior as a permanent trait so people excuse it.
  13. Clinging to the past: deflecting blame onto past events and people, a subset of blaming everyone else.
  14. Playing favorites: failing to see that you’re treating someone unfairly.
  15. Refusing to express regret: not taking responsibility, admitting fault, or recognizing how your actions land on others.
  16. Not listening: the most passive-aggressive form of disrespect toward colleagues.
  17. Failing to express gratitude: the most basic form of bad manners.
  18. Punishing the messenger: attacking the innocent person who was only trying to help.
  19. Passing the buck: blaming everyone but ourselves.
  20. An excessive need to be “me”: exalting our faults as virtues simply because they’re who we are.

Atomic Habits cover
Atomic Habits
Buy on Amazon

Atomic Habits — James Clear

Read: August 2026

Probably the most recommended book in the habit-building space, and it earns that reputation. The core argument is simple enough to state in a sentence, and the rest of the book is just working out its implications: small, consistent changes compound into outcomes that look dramatic only in hindsight.

You don’t rise to the level of your goals, you fall to the level of your systems. Clear’s whole argument rests on this. Two people can share the same goal and get wildly different results because goals are just the desired outcome, systems are the process that gets you there. Winners and losers often have the same goals, the thing that separates them is what they do between now and then.

Habits are the compound interest of self-improvement. A 1% improvement isn’t noticeable, but a 1% decline isn’t either, and stacked daily over a year, 1% better compounds to roughly 37 times better while 1% worse compounds to nearly zero. The outcomes of any habit are rarely visible in the moment, which is exactly why they’re so easy to neglect.

Identity-based habits beat outcome-based habits. Instead of “I want to run a marathon,” the more durable frame is “I am a runner.” Every action is a vote for the type of person you want to become. The goal isn’t to read a book, it’s to become a reader. The goal isn’t to run, it’s to become a runner. Behavior that’s inconsistent with your self-image won’t stick no matter how good the system is.

The habit loop: cue, craving, response, reward. And the four laws that follow from it for building a good habit: make it obvious, make it attractive, make it easy, make it satisfying. Break a bad habit by inverting each one: make it invisible, unattractive, difficult, unsatisfying.

Make it easy is underrated. The two-minute rule stuck with me the most: scale any new habit down to something that takes two minutes or less. “Read before bed” becomes “read one page.” The point isn’t the two minutes, it’s removing enough friction that showing up becomes automatic, and once you’re showing up consistently, scaling the habit back up is the easy part.

Environment design beats willpower. Clear’s point is that you don’t need more discipline, you need a better-arranged room. Put the guitar in the middle of the living room if you want to practice more, put the phone in another room if you want to stop checking it. Most people try to change habits by fighting their environment, when it’s far easier to just redesign the environment so the right choice is also the obvious one.

The plateau of latent potential. Progress isn’t linear, most of the work happens below a surface that doesn’t show results until a threshold is crossed. An ice cube sitting at 25°F doesn’t visibly change as the temperature rises to 26, 27, 30, 31, it looks the same the whole time, until 32°F, where everything shifts. People quit in the plateau because the lack of visible progress feels like proof it isn’t working, right before the point where it would have.

What’s left isn’t any single tactic, it’s the reflex to ask “what system produced this outcome” instead of just setting another goal.


The Man Behind the Wheel cover
The Man Behind the Wheel
Buy on Amazon

The Man Behind the Wheel — Tim Bouquet

Read: August 2026

Subject: Onkar S. Kanwar (Apollo Tyres)

Core thesis: True industrial scale isn’t built by political networking alone; it requires operational resilience, geographical diversification, and bold global risk-taking during moments of crisis.

The 30-second core insight: Onkar Kanwar transformed Apollo Tyres from a single, union-plagued Indian factory into a $2B+ global giant by systematically decoupling production from local risk, professionalizing family leadership, and acquiring distressed global brands (Vredestein) long before Indian globalization was fashionable.

Key mental models:

Pillar The situation Kanwar’s move Universal lesson
Operational risk Labor strikes paralyzed the primary Kerala plant in the 1980s Built a second factory in Gujarat to break single-point-failure dependency Never let a single operational bottleneck hold the enterprise hostage
Capital allocation Operating under the restrictive pre-1991 License Raj Acquired existing distressed licenses (Nirlon) rather than waiting Find creative workarounds within rigid system constraints
Global expansion Domestic price wars and commodity pressure Cross-border M&A: bought Dunlop South Africa and Dutch brand Vredestein Acquire premium R&D/brand equity in mature markets to upgrade margins
Leadership transition Friction between relationship-driven father and systems-driven son Fought for professionalized strategy over traditional patriarch control Sentimental legacy must give way to institutional longevity

The anti-fragile production principle: Labor union strikes threatened to bankrupt the single-factory Kerala operation. Rather than bowing to pressure, Kanwar used a multi-state footprint to hedge operational risk - “Apollo could not be held to ransom by unions.”

The global value step-up: moving from execution risk, to scale efficiency, to global R&D and margins.

Single factory (Kerala) → Multi-plant domestic (Gujarat/Chennai) → International M&A (Netherlands/Hungary)

The one that got away: Apollo came close to acquiring Cooper Tire for $2.5B in 2013 - bigger than Tata’s Jaguar Land Rover deal - and had already done much of the integration planning before it fell apart. Kanwar’s own framing of the failure: “You cannot regret, because everything happens for the best. Equally you cannot go forward if you are looking backwards. We didn’t look back. We said, ‘Let’s move on.’”

Quotes that stuck:

  • “Onkar Singh Kanwar only knows one direction and that is forward. Whatever life throws in your path, you carry on, you stand on your own two feet and you follow your dream.”
  • “The basic principles do not change when you become big but I remind the HR people that they must listen to staff. As I keep saying, listening is learning.”
  • “A tyre is a unique combination of steel and rubber. The steel withstands the stresses placed on a tyre and the rubber takes care of the strain.”
  • On punctuality: “Onkar is equally wedded to punctuality. To him turning up late shows lack of respect and lack of intent.”