Shelf.

Pets-vs-cattle framing for cloud systems: durable state should live outside replaceable hosts, so a failed server or sandbox can be discarded and replaced.

Geoffrey Huntley on Ralph: a coding-agent Bash loop that works one item at a time, uses specs and fix plans as context, fans out subagents, and relies on tests/builds as backpressure.

Six accessible chapters from Feynman's Lectures on Physics, adapted for general readers from the Caltech lectures with the classroom feel still intact.

Anthropic on Managed Agents as a meta-harness: split durable session logs, stateless harnesses, and sandboxed tools so long-horizon agents can scale securely as model capabilities change.

Anthropic on pushing agentic coding with harness design: frontend design evaluators, generator/evaluator loops, and a planner-generator-evaluator architecture for multi-hour full-stack app builds.

Harness pattern for long-running coding agents: prevent one-shotting the app and leaving half-implemented work across context windows by using an initializer, feature requirements, progress logs, git history, and end-to-end verification.

Field guide to agentic systems: simple LLM calls, workflows like routing/parallelization/evaluator-optimizer when useful, and reserve autonomous agents for open-ended tool-using tasks.

70 years of AI show that scaling computation (search + learning) beats human-knowledge-based approaches—in chess, Go, speech recognition, and vision. You can just scale compute bro.

Turns out agents are hard: VM-level isolation over containers, hypervisor-level state snapshotting for async work persistence, etc.

Five evolutionary breakthroughs - steering (bilaterians), reinforcing (vertebrates / RL & basal ganglia), simulating (mammals / neocortex & world models), mentalizing (primates / theory of mind), speaking (language & culture) - each paired with where AI matches or still falls short.

Book 3 closes the first Red Rising arc with the main dude turning infiltration into full revolution against the caste system.

Book 2 turns Red Rising into military sci-fi/space opera, scaling the world building. Main character does seem like a self-obsessed caricature who loves to hear himself speak—or think.

Acquired on Google's 20-year AI arc: Brain, Transformers, TPUs, Gemini, Waymo, and whether Search can survive the company's own disruption.

Acquired on how Trader Joe's became America's favorite grocery chain by refusing e-commerce, delivery, coupons, SKU bloat, and the supermarket-CPG playbook.

Network effects: cold start to moat, atomic networks and the hard side, with interviews from teams at LinkedIn, Zoom, Uber, Airbnb, and similar companies.

Silicon valley bible pretty much. This book is a lot more useful now that I've personally seen a lot of the dynamics play out from work as oppose to first reading it in highschool.

Dystopian science fiction novel. Space racism seems to be a common thread in critically acclaimed scifi books.

Practical book on product positioning. Feels no shit Sherlock till you apply it. I enjoyed the read.

I was joking with a friend recently about how good of a COO Deng would be.

Jensen says Nvidia wins because it can make the whole AI supply chain believe in, build for, and depend on Nvidia-scale demand before anyone else can.

The winner is whoever locks up the real bottlenecks first: advanced logic, HBM memory, packaging, and power.

He says AI capabilities are still scaling predictably and will soon produce AGI-level "geniuses in a data center," but the real-world impact depends on how quickly we can safely deploy, verify, and absorb that intelligence.

Space datacenters??? Talked about: orbital solar vs ground power, grid & turbine bottlenecks, Starship cadence, Grok alignment & xAI, TeraFab & memory, Optimus vs China, DOGE.

Microsoft wins if every AI agent needs Azure/GitHub/M365 as its computer, identity system, storage layer, and control plane.

Agents are still bad employees. Useful as autocomplete and narrow tools, but not autonomous workers. They lack memory, continual learning, robust reasoning, and reliable execution.

Walter Isaacson is the GOAT of biographies.

Apo B burden, LDL oxidation = inflammation = risk of ASCVD at young age.

Consumer health went from vitamin D and lifting to GLP-1s and compounding-pharmacy peptides - retatrutide, BPC-157, melanotan risks, GH secretagogues. WashU work: stimulants mostly match good sleep for alertness, not true focus; Huberman wants sleep/cortisol “write” tech, focus wearables, exercise-blood banking, and LLMs fed real health data - plus AI decoding octopus camouflage.

Every tool that makes creation cheaper makes verification more expensive. AI collapses the cost of faking a resume to seconds while making authenticity harder to confirm. Crypto handles inter-tribal commerce; AI optimizes within trusted groups.

Hesitation kills companies more than bad decisions do: make the call, even when you're not sure.

Was good. Don't remember much.

You might be building an AI wrapper, but AI is just an Nvidia wrapper which is a TSMC wrapper which is an ASML wrapper.

Thesis is that the real bottleneck in AI is not chips but reliable electricity. I disagree.

Don't be a thinkcel.

"Fighting is not the only thing of value a man can do." Setting in a medieval, bloody, unfair world that gets frequently decimated by storms.

I admire when authors George R.R. Martin their main characters.

"I'm wondering if every person I pass has similar depths, and if there's any way to avoid the mistake of judging them so shallowly that I'm rocked when they show their true complexity."

Second book: a fusion of Sherlock Holmes and Mistborn.

Steampunk fantasy set 300 years after the events of the first trilogy.

Two failure modes (compliance vs passivity). One graph.

@balajis take on Canada/China in the coming years.

Evaluated AI predictions from past years, and collect a convenience sample of AI predictions made this year.

There is always a shortcut.

It’s easier to imitate wholeness than it is to assemble it from parts.

It’s the rate of experimentation, the number of shots on goal, that provides the magic, not the percentage of successes, which might be very low at first.

Lots of variance in opinion about whether an idea is good means there’s a good chance the idea is good; lots of variance in opinion about whether a person is good is a warning sign.

A counterbalance to Angela Duckworth's Grit.

The “elephant and rider” is a very useful framework for thinking about motivation.

Stop self-justifying after a reasonable point; decide, treat the decision as “good enough for now,” and relax. Chronic self-doubt impairs judgment and flexibility more than provisional confidence does.

Reasonableness is public, recursive, and has no ground truth.

How to attract what you want by aligning your inner world with your desires. Clean up your inner world, identify what you truly want, and give yourself permission to pursue and receive it.

Guilt is fear of disapproval, desires are good, so is positive visualization. Don't seek the gods eye view, rather the first person perspective. It is possible to self evaluate accurately.

Mistborn series: still reading.

Mistborn series: Great world building. Overthrowing a 1000 year empire with allomancy and propaganda.

Taking a look at neoclouds in the lense of banks and real estate companies.

Trends show the hardware cost of frontier AI clusters more than doubles each year. 100B by 2030.

OpenAI trying to reach 10GW of compute.

Zvi wrapped, 2025. Mostly AI related stuff.

Guide on Claude Code 2.0.

Becoming more willing to stare into the abyss is one of the most important things you can do to become a better thinker and make better decisions about how to spend your life.

The Expanse series: eh. I mean, its alright.

The Expanse series: I respect the author for going George R. R. Martin on the main cast.

The Expanse series: The rise of an authoritarian stratocracy.

The Expanse series: Space terrorist insurgency with naval hardware. And the political powers are outnumbered and outgunned hazbins.

The Expanse series: Looking back through history, there are a lot more men who thought they were Alexander the Great than men who actually were. Also, I really like Avasarala.

The Expanse series: Politics while the planet is trying to eat the humans.

The Expanse series: Weaker main characters than previous two entries. A classic revenge tail- I cannot deny the schadenfreude I felt at Mao’s plan going to shit from the start.

The Expanse series: Lots of POVs. Good personalization and relationship building.

Bootstrapping a software consulting company.

ECE experience in 2014 uWaterloo.

The Expanse series: A Space Opera. i CaN fiX hEr.

The critical distinction between critiquing others' views and believing your own position is correct.

You're not trying your best if you're not happy.

We currently lack a mathematical definition of what real-world intelligence even is.

Extraordinary claims require extraordinary evidence, but extraordinary evidence might be more common than you think.

When external forces shape your motivations.

Blind yourself to information and be unpredicatable.

Pick a North Star and go for it.

B2B sales...

If you are going to eat shit, don't nibble.

Life wasn't about avoiding suffering.

Society profits from the convexity of entrepreneurial risk; individuals bear the concavity.

How internet communities can temporarily meet in the physical world as popups, then evolve into permanent societies.

Good brainrot book - don't have to think.

Overly anthropomorphized ASI with incompetent personality - Skippy misses obvious ideas despite billions of years of processing.

Trauma, faster than light travel, and gambling on suicide trips.

There’s a time and a place for everything. It used to be called college.

The new economics of consumer apps.

Banks and crypto.

11 thoughts on AI.

Zuckerberg's vision for AI friends and therapists to ad-max.

Circumvent the generalized efficient markets and reach the Pareto frontier through dimensionality.

You want to think reasonably about things that are uncomfortable to contemplate.

Recovery day = binge Youtube, rest day = do what you want when you're motivated and able to get out and about.

There is always something to learn.

If it hurts, you're probably doing it wrong.

One of the most effective things you can do is to change your situation.

Extroversion is apperently strongly correlated with happiness.

Motivation = (expectancy * value) / (impulsiveness * delay).

Scientifically-backed advice for having a better life.

Game theory and abysmal security of Tea app.

YC's startup ideas they'd like to see founders tackle.

Weekly AI news roundup and analysis.

Still reading.

Continual learning is a bottleneck for AI to move out of self contained, short horizon tasks.

Sentiment on OpenAI's Operator.

Short blogs on: Abundance, The Thiel-Andreessen Axis, How DC Missed Deepseek, Shanghai, Effective Altruism, Egalitarians, and Democratic Oligarchy.

Do what you’ll regret the least when you’re 80.

The story of BackRub.

Impact of superhuman AI over the next decade.

Maintaining Party control, prioritizing security vs economic reforms.

China prioritizing hardware tech (e.g., semiconductors, renewables) over demand-side reforms stifles services and consumption, risking deflation. [article]

China's stagnation stems from overinvestment in infrastructure and housing post-2010 stimulus. US will face debt-driven inflation crisis. [article]

Still reading.

Still reading.

Still reading.

Still reading.

What you aim at determines what you see.

Still reading.

Death ends a life, not a relationship.

Still reading.

Still reading- super dense book.

The Office a satire of real corporate dynamics- Sociopaths exploit overperforming Losers by promoting them into Clueless middle management to shield themselves.

How Deng and party 10x-ed the GDP of China in two decades.

Practical book on negotiating.

Establish comptency, make an ask, be transparent with self intrests, show extra effort.

Here's what it taught me about B2B enterprise sales.

Steve Ballmer got way more shit from everyone than he deserved.

By recognizing Moore's Law would make computing universal, Microsoft used IBM's distribution to establish DOS as the standard before licensing it to every PC clone manufacturer.

NVIDIA bet everything on CUDA for six years with no market until AlexNet proved deep learning worked on GPUs, creating the $500B infrastructure powering all AI.

NVIDIA survived 90 graphics competitors by inventing programmable shaders- something which will power ML.

Ethereum made code and money identical-smart contracts with their own funds-creating a Raspberry Pi-speed world computer powering a $200 billion economy.

Enron collapsed by booking 20 years of future revenues as immediate profits through mark-to-market accounting while hiding losses off-balance-sheet, leaving no future value left to steal.

AI is shifting from single large models to compound systems that combine multiple components (models, retrievers, tools) because these multi-part approaches achieve better results than just scaling up individual models, as seen in systems like AlphaCode 2 and AlphaGeometry.

AI agents need event-driven architecture (EDA) rather than just better models to scale effectively, because they require real-time access to data, tools, and other agents across distributed systems-making.

Faster rounds mask risk, inflating a venture bubble that bursts when the pace slows.

My personal startup bible. In unforgiving environments where failure modes compound with time, the truly prudent play is to sprint-because every hour you save shrinks an exponentially expanding risk surface that no amount of redundancy or planning can offset.

If Wall Street is in the business of spatial arbitrage, Silicon Valley is in the business of temporal arbitrage.

The hidden trap isn’t caring what others think-it’s outsourcing your identity to their presumed judgments.

Most of our growth comes from internal measurement, while most problems stem from external comparison.

AI apps: pick a single industry and nail it; horizontal plays belong to developer middleware, not end‑user products.

AI cheapens compute so radically that trusted, proprietary data becomes the bottleneck; the real upside now lies in owning, curating, and tooling that data.

Commoditizing a complement doesn’t just cheapen another layer-it creates a noisy facade of “competition” that diverts antitrust scrutiny while the real monopolist quietly tightens its grip on the profitable chokepoint.

The Rich Barton Playbook for winning markets through Data Content Loop.

My friend Ragnor sent this to me after I got accidently nicked by a (highly unlikely rabied) cat.

Despite rising benchmark scores, recent AI model progress has yielded little practical improvement in real-world tasks.

Boosts RAG accuracy by prepending chunk-specific context to document embeddings. You'd want to merge semantic and lexical matching.

Graph RAG = property graphs + traversal paths + explainable retrieval.

Data is defensible when it is both unique and valuable (core systems or processes) and has feedback loops.

Wisdom on how to live a good life. Points I agree, points I disagree. Could be verbose at times, but overall, was great for the plane ride.

By David Deutsch. Still reading.

Data and behavioral science to help you find love.

Trillion dollar clusters, AGI will arrive by 2027-2028 through scaling and unhobbling existing models, the geopolitical implications, the importance of protecting AI research secrets, his criticisms of building AI infrastructure in Middle Eastern countries, his controversial departure from OpenAI, and his new AI investment firm.

Gwern Branwen, an anonymous internet researcher who was among the first to recognize AI scaling laws, discusses his writing process revolving around intellectual rabbit holes, his minimalist lifestyle living on $12K a year while influencing AI development through his essays, and his belief that AGI is just a few years away.

Dylan Patel: Complexity of the semiconductor industry, China's potential to centralize compute resources, and how massive investment in AI hardware could enable training runs with 10,000x more compute than GPT-4 by the end of the decade.

Professor Sarah Paine: How Mao conquered China by combining propaganda, peasant mobilization, and warfare tactics.

Professor Sarah Paine: How Japanese military culture, particularly the samurai code of Bushido, influenced their strategic decisions and battlefield conduct. Analyzes how Japanese values of honor, death, and willpower led to poor strategic choices, severe inter-service rivalry between the Army and Navy, and a refusal to surrender despite inevitable defeat.

Professor Sarah Paine: Geopolitical dynamics between Russia, China, India, and Pakistan during the Cold War, exploring how their shifting alliances, territorial disputes, and interventions shaped modern South Asia.

A clinical psychologist uses a (controversial(?) manosphere/red pill content) economic model to explain decision making in relationships.

Reflection on philosophy of 1) accelerate technology 2) prioritizing defensive capabilities and decentralization of power so people maintain maintain agency and freedom even as technological progress.

The optimal rate of failure is rarely zero, and if you never fail, you're likely being too risk-averse, which can be costly both personally and ethically.

1/3 whitepaper from angle.money. Technicals on how they manage a basket of assets with exposure limits and handle black swan events without requiring governance intervention.

AI competition between the U.S. and China parallels Cold War dynamics, with Chinese teams like R1 rapidly deploying models optimized for memory-efficient reasoning while American export controls and manufacturing limitations may prove ineffective as the bottlenecks shift from chip capabilities to engineering talent and energy infrastructure.

Multi-agent AI system built on Gemini 2.0 that generate novel, testable research hypotheses, as validated through experimental confirmation of its predictions in drug repurposing for acute myeloid leukemia. The elo system for eval is the same as the one that gens the hypotheses which I would intuit cause a ircular reasoning problem.

By David Deutsch

On average, you spend 80k hours in your career. This book is a guide to choosing a career with high social impact: strategic career choices can significantly improve the world. Some Effective Altruism arguments.

More von Neumann probes shenanigans. Still reading.

A Von Neumann probe's search for a massive megastructure built by an alien civilization, while dealing with conflicts within AI clones of himself.

Software engineer who dies and is resurrected as an AI in a von Neumann probe, leading to a galaxy-spanning adventure of exploration, self-replication, and existential dilemmas. Becoming a become a Von Neumann probe added to my bucket list.

Blockchain network for registering, tracking, and monetizing IP where users can set usage terms and receive royalties as their work is used across applications and AI systems.

Vitalik argues for defensive acceleration (d/acc) approach to technological development that prioritizes technologies that enhance defense, decentralization, and democracy while being more cautious about potentially dangerous technologies like AI, contrasting this with both anti-tech views and unrestricted technological acceleration.

Fornite battle royale with alien politicians held hostage by an AI system losing its shit. There's a fine line between gritty determination and melodramatic edgelord monologues, more and more of this series is crawling (get what I did here?) towards the latter.

More of an edgy AI system in this book.

Series has been going downhill- but I'm still invested in the story. There's dinosaurs in this book.

Personal comment: Stablecoins have created a $160B market-driven system that strengthens USD by extending its reach globally and serving as dollar reserves in emerging markets - achieving currency dominance possibly more effectively than proposed government Bitcoin purchases.

Response to Marc's claim of Consumer Financial Protection Bureau (CFPB) systematically debanking crypto entrepreneurs and politically disfavored individuals, sparking debate about whether these are accurate and what role the CFPB actually plays in addressing debanking.

A floating fortress occupied by warrior gnomes. A castle made of sand. A derelict submarine guarded by malfunctioning machines. A haunted crypt.

I'm binge-reading the series now. This level takes place in an impossibly-complicated subway system built out of the world's subterranean railway systems.

Carl and cat descends deeper into more dangerous floors where they face off against a murderous groups of people.

Carl and his ex's talking cat are forced into a Fortnite battle royal style dungeon-crawler gameshow broadcast across the galaxy.

Adlerian psychology: free you from seeking others' approval, helping you take charge of your own life by focusing on what you can control while accepting that not everyone will like your choices

By Viktor Frankl. Even in the darkest circumstances - drawn from his experience in concentration camps - one can find purpose and meaning in life through choosing one's attitude toward suffering.

They didn't even use Cuda- rather parallel thread execution. The Chinese Quants are cracked. Anyways key take aways: Achieved strong performance with just 33B parameters and $5M training cost through an efficient reinforcement learning approach that skipped supervised fine-tuning and used group-based optimization instead of a critic model. Focused on deterministic tasks (math, coding) using rule-based rewards rather than expensive neural reward models or human annotation, significantly reducing costs. Rather than relying on massive scale, they achieved efficiency through iterative fine-tuning, high-quality data curation, and model distillation to create smaller versions (1.5B-14B parameters) while maintaining reasoning capabilities.

Gen Y young professionals are unhappy because their reality can't match their inflated expectations, which were shaped by their parents' optimism, being told they were special, and constant exposure to peers' curated social media success.

Still reading: a self-help book for societies.

Life is finite, and when you measure it in discrete units - whether days, activities, or relationships - you realize just how limited our remaining experiences truly are.

Guide that reframes career decisions from rigid, high-stakes tunnel choices into a series of experimental dots to be connected over time, emphasizing the importance of understanding both your authentic desires and the reality of what's possible while being willing to adapt as you and the world change.

Re-reading book my prev manager wrote. Essential principles covering risk management, strategy clarity, technological advantage, and continuous improvement in the context of decision making.

Collection of philosophical aphorisms by Nassim Nicholas Taleb that examines human nature and modern life.

Exploration of how AI could improve human well-being across health, neuroscience, poverty, governance, and meaning in the next 5-10 years, arguing that with careful development and global cooperation, AI can usher in an era of unprecedented progress.

Analysis of the stablecoin landscape, examining diverse viewpoints from policymakers and academics. Stablecoins are a transformative financial technology with significant implications for the dollar, global payments, and individual liberty.

Stablecoins offer small businesses lower transaction costs, faster access to funds, and more predictable cash flow by enabling real-time domestic and cross-border payments while reducing reliance on expensive intermediaries.

Prediction markets, like Polymarket, are not just for betting on elections-they're tools for gathering information by aligning incentives. This idea, called 'info finance,' uses markets to extract useful predictions and insights, with applications in governance, science, and public goods funding, and it's now practical thanks to blockchain and AI.

The most intresting thing I'm thinking about here: Unlike traditional game AIs with straightforward victory conditions, modern AI agents often have ambiguous success metrics, but combining real-world economic incentives (e.g., increasing a crypto wallet's balance) with the ability to visually interact with interfaces via an API moves closer toward more tangible and practical win conditions. Web3 companies will begin with the end (-user experience), instead of letting the infrastructure determine the UX. As Nassim Taleb observed, 'Over-engineering breeds fragility. Simplicity scales.'

How distributed systems reliable communicate with malicious actors using cryptographic signatures. [Technical]

Still reading

I was born poor and miserable. I'm now pretty well-off, and I'm very happy. I worked at those. I've learned a few things, and some principles. I try to lay them out in a timeless manner, where you can figure it out for yourself. Because at the end of the day, I can't quite teach anything. I can only inspire you and maybe give you a few hooks so you can remember.

Scaling laws, AI Safety Levels, AGI timeline, meaning of life, programming advice for non-technical people.

Scaling large language models can lead to AGI, debating data limitations, generalization, and synthetic data/self-play: 70% likelihood of transformative AI by 2040 if scaling succeeds.

Pretty political podcast on Trump, Crypto, DOGE, and startups.

Companies are constantly calculating the lifetime value of a customer. We can increase the L. The longer someone lives, the greater your LTV will be, no matter what.

Lots of smart people go into VC, investment banking, and product driven by 1) status 2) lack of conviction from endless optionality.

Developing good taste requires intentional engagement, balancing instinct and aspiration, to resist the flattening effect of algorithmic curation and cultivate deeper meaning and joy in a world of overabundant, passive consumption. Reminds me of the PG article on good taste.

'Today, however, capital seeks to exploit our sociality in all spheres of work. When we all become "human capital" we not only have a job, or perform a job. We are the job.'

Looking for a better understanding of the tech, business, and startup worlds? A primer on the history and future of the internet? A crisp explanation of the crypto and blockchain movements, and why they matter?

How companies use demand curves and proxies for willingness-to-pay to maximize revenue-sometimes in ways that defy logic.

Bundling flattens demand curves, allowing sellers to maximize revenue while benefiting buyers with lower prices and more surplus, making it a win-win strategy for taste-based goods like media.

Marc Andreessen champions technology as humanity's greatest tool for progress, abundance, and freedom, rallying against stagnation, bureaucracy, and pessimism.

Vinod Khosla envisions AI as humanity's ultimate tool for abundance, reshaping work, creativity, and governance while emphasizing the need to outpace China.