Nazir's website

Open source as a business model

2026-07-29T09:40:08+00:00

At first glance an open-source business sounds strange.

You spend money on development, publish the code for free, competitors can copy it, users may not pay. Where is the business?

But in reality open source is almost never a business model in itself. It is a go-to-market strategy. A way to get distribution, trust and standardization.

A company opens up not because it is kind, but because in some markets openness gives you more than secrecy.

Why companies choose open source:

Distribution. A developer can just grab it and try it without a sales call, procurement, demos and all that corporate bureaucracy. For developer tools this is a superpower.
Trust. If you are putting a database, an infrastructure tool or an AI model inside your system, you need to understand what is going on in there. Especially when it comes to security, data, compliance or on-premise installs.
Standard. If your project becomes the default choice, plugins, integrations, documentation, specialists and content grow up around it. And from then on it is no longer you selling to the market; the market starts selling you.
Hiring and development. Strong engineers are more willing to go where their work is visible to the world. Plus the community finds bugs, writes integrations and tests edge cases.
Making the adjacent layer cheaper. If you make money on cloud, support, enterprise features or applications, it is in your interest for the base layer to become more accessible and cheaper.

The last point is especially important for AI.

Many Chinese AI companies are now actively releasing open-weight models: DeepSeek, Qwen, Kimi and so on. This is not quite classic open source, because often the weights are open but the full training data, methodology and infrastructure are not. But the logic is similar.

If you cannot beat OpenAI with a closed premium model and consumer distribution, you can play differently: make the model so accessible and so good that developers, cloud providers, startups, research labs and corporations start using it. You buy reach at the price of disclosing part of the technology.

This makes particular sense if your real business is not selling a file with weights. The real business can be:

inference as a service
enterprise deployments
fine-tuning and customization
cloud resource consumption
closed products on top of the model
government and corporate contracts
the ecosystem around your standard

In other words, open source in AI often works as a land grab. You give away the base layer in order to become part of someone else’s stack.

How open-source companies make money

Managed cloud

The most straightforward model. You can run the code yourself, but sensible companies pay for a ready-made cloud version, because nobody wants to debug Kubernetes at night to save $800.

MongoDB Atlas, Elastic Cloud, Confluent Cloud all come roughly from here.

Open core

The base version is open, the enterprise features are paid: SSO, access rights, audit logs, compliance, advanced security, team management, support.

This is the GitLab model. A developer starts for free, the company pays once organizational complexity appears.

Support and subscriptions

The Red Hat classic. Linux is free, but corporations pay for stability, security patches, certification, support and long-term maintenance. IBM did not buy Red Hat for $34B out of charity.

Source-available instead of true open source

Many companies started out as open source and then closed their license to shut out cloud competitors. MongoDB moved to SSPL, Elastic also experimented with licenses, HashiCorp switched Terraform to BSL, after which the OpenTofu fork appeared.

This is an important trade-off: you protect your monetization, but part of the community starts to feel you have broken the social contract.

Services and consulting

It works, but scales worse. Good as a start, bad as the final model if you want software margins.

How much do they make on this?

A lot, if the project becomes an infrastructure standard.

Red Hat was doing ~$3.4B in annual revenue before the IBM acquisition. GitLab is already at almost ~$1B in annual revenue. Elastic is around ~$1.7B. MongoDB is around ~$2.5B. HashiCorp was at about ~$600M in revenue before the sale to IBM. Confluent has also already crossed ~$1B in subscription revenue.

But here is an important caveat: almost none of these companies are “pure open source”. They are open core, managed cloud, source-available, enterprise subscriptions. Community romance on top, very pragmatic monetization underneath.

How to decide: open or closed?

I would look at it not ideologically, but through one question: where is your real moat?

If the moat is in the code as a secret, keep it closed.

If the moat is in distribution, cloud, data, brand, relationships with corporations, UX, ecosystem or speed of execution, you can open up part of the stack.

Open source works well when:

the product is for developers, infrastructure or AI products
reach matters more than fast monetization
the market works on the principle of “the winner becomes the standard”
the user cares about trust, self-hosting or on-premise installs
plugins, integrations and a community can grow up around the product
you know how to monetize the adjacent layer: cloud, enterprise features, support, applications

Closed code is better when:

the code and the model are the main secret
the product is easy to copy without your distribution
there is no clear paid layer
security, abuse and regulation matter too much
consumer UX and brand matter more than developer adoption

My current conclusion: open source is not about “being kind” and not about “not knowing how to make money”.

It is a way of telling the market: “let the base layer become a commodity, because I am going to make money on what sits above it, is more reliable, more convenient or closer to the corporate budget”.

And if that works, the business can be enormous.

But if you opened up the only thing that was your moat and then did not build a cloud, an enterprise layer, distribution or a product on top, you have not built an open-source business.

You have just given the world a nice repository.

The team should be so small that it's uncomfortable

2026-07-10T10:46:47+00:00

There’s a common mistake in startups: hiring people a bit earlier than you actually need to.

It seems logical: there’s a lot of work, everyone’s overloaded, you need to plug the gaps, speed up, add management, more features, more parallelism.

But in practice almost always the opposite happens. The team gets bigger and the speed drops.

Why?

Because a small team forces clarity.

When there are few of you, you can’t run ten directions at once. You can’t hide a bad product behind processes. You can’t spend months discussing the roadmap. You can’t keep people who are “sort of useful”. Everything is visible immediately.

A small team forces you to answer uncomfortable questions:

what’s the most important thing right now?
who is actually moving the product forward?
what work can we simply not do at all?
which feature isn’t worth a dedicated person?
where are we trying to cure a lack of focus with hiring?

A big team often gives a feeling of progress. More people, more meetings, more status updates, more of everything going on.

But that can be just organizational noise.

I’d say a good early-stage team should be so small that it’s a little scary. So that it constantly feels like “we don’t have enough people”.

That’s a normal feeling. It keeps the system sharp.

Being short on people is also unpleasant because it forces you to deal with the most tedious stuff.

Not what’s easy to show in a demo. Not the small features that give a sense of motion. But the big fundamental things: understanding the real problem, simplifying the product, fixing the core flow, removing the excess, making it so the user comes back on their own.

When there are a lot of people, it’s very easy to hide unimportant features behind busy people.

Everyone has a task. Everyone has progress. Everyone has something to show at the sync. And gradually the team starts serving its own busyness instead of the main risk of the business.

A small team doesn’t allow that. If there are few of you, you can’t keep a person on a feature that’s “sort of useful”. Too expensive. You have to keep coming back to the boring question: which of this actually moves the company?

If the team is too comfortable, most likely you’ve already started buying comfort at the cost of speed. And in a startup comfort is almost always expensive.

Not because people aren’t needed. They are.

But every new person has to add more speed than complexity. And that’s a very high bar.

So the best question before hiring isn’t “do we have work for them?”.

There’s always work.

The best question is: “if we don’t hire this person, what will we stop doing?”

If the answer is “nothing, we’ll just be a bit less loaded”, I wouldn’t hire.

If the answer is “we’ll stop doing the secondary stuff and finally get to what matters”, all the more so.

A small team isn’t garage romance. It’s a forcing function. It makes you choose, cut the excess, and do only what really matters.

How to read LLM benchmarks

2026-02-17T14:44:57+00:00

Every week someone releases a new model, and every one of them “wins on the benchmarks.” OpenAI publishes a table where they are ahead of the whole planet, Anthropic its own table where they lead, Google a third one where Gemini wins. Marketing wars in their purest form. But if you dig deeper, benchmarks are a genuinely useful thing, you just need to understand what they measure and what to trust.

What a benchmark is

A benchmark is a standardized exam for AI models. A set of tasks with known correct answers that lets you compare different models against each other. The idea is simple: give it 1000 questions, count the percentage of correct answers, and there is your result.

The problem is that there are now dozens of benchmarks, each measures something of its own, and not all of them are equally reliable. Let’s sort out what is what.

The main categories

All benchmarks can be split into a few big groups:

General knowledge: how “well-read” the model is across different fields
Reasoning: the ability to reason logically and draw conclusions
Code: the ability to write working code
Math: solving problems from school level to olympiad level
Multimodality: understanding images, diagrams, charts
Human preferences: whether real users like the answers

Benchmarks worth knowing

MMLU-Pro (Massive Multitask Language Understanding)

A test of general erudition: 57 subjects from elementary math to law, medicine, and philosophy. The format is multiple choice, but the Pro version has 10 answer options instead of 4, which makes guessing much harder. Top models currently score around 89-90% (Gemini 3 Pro leads with 90.1%). Useful as a baseline quality filter, but already close to saturation: all strong models show similar results in a narrow range.

GPQA Diamond (Graduate-Level Google-Proof Q&A)

Graduate-level questions in physics, chemistry, and biology. The name “Google-proof” means the answers can’t simply be googled; you need to genuinely understand the material at a deep level. Important context: experts with a PhD in the relevant fields score around 65%, and ordinary well-educated people only 34%. When the benchmark first came out, GPT-4 scored 39%. Now top models have passed 78%. One of the best tests of depth of understanding rather than simple memorization of facts.

SWE-bench Verified

Probably the most practical benchmark for those who work with code. These are real GitHub issues from popular open-source projects: the model has to read the bug description, figure out the codebase, and generate a patch that fixes it. Not abstract algorithmic puzzles, but a programmer’s real work with context, dependencies, and legacy code. The current leaders are Claude Opus 4.5 with 80.9% and GPT-5.2 with 80%. Interestingly, on the harder SWE-bench Pro version the same models score only ~23%, so there is room to grow.

MATH

Math problems from school level to olympiad level. Importantly, this isn’t just “compute 2+2” but problems that require chains of reasoning: algebra, geometry, probability theory, combinatorics. A good indicator of a model’s reasoning abilities.

Chatbot Arena (LMSYS)

A separate story and possibly the most honest benchmark of all. It works simply: real users ask a question to two anonymous models, see both answers, and pick the better one. From millions of such votes an ELO rating is built, like in chess. Right now Gemini 3 Pro with a rating of 1492 and Claude Opus 4.6 are at the top. The main advantage is that this benchmark can’t be “gamed,” because it is literally the preferences of live people on real tasks.

ARC-AGI

A benchmark that AI still can’t properly solve. These are tests of abstract thinking: you are shown a few examples of a grid transformation (input → output), and you need to figure out the pattern and apply it to a new input. Sounds simple, but it requires a generalization ability that modern LLMs are still weak at. Important as an indicator of the fundamental limitations of current architectures.

Humanity’s Last Exam (HLE)

The newest frontier benchmark for genuinely hard problems. The questions were collected from world-class experts in different fields, from quantum physics to the linguistics of ancient languages. The idea is that if a model can’t answer these questions, it definitely can’t be trusted with serious expert work without human oversight.

Why you can’t just trust the numbers

Data contamination

The industry’s main headache. If test questions accidentally end up in the training data, the model simply “remembers” them rather than solving them. The older and more popular a benchmark is, the higher the chance it has leaked into the training data. That is why updated versions appear: MMLU → MMLU-Pro, and LiveBench, which is refreshed every month with new questions.

Optimizing for the metrics

Companies actively tune models for specific benchmarks; that is a fact of life. 95% on some test may mean the model was drilled on exactly that task format, and on your real cases it will perform worse.

The gap between numbers and practice

A model can brilliantly solve benchmark problems and at the same time be bad at following complex instructions, generate odd text, or hallucinate facts. Standard benchmarks don’t measure the “vibe,” the ease of working with it, or practical usefulness for specific tasks.

What to do in practice

Don’t fixate on a single number. Look at several benchmarks in the category that matters to you specifically. Need code? SWE-bench matters more than MMLU. Need general knowledge? Look at MMLU-Pro and GPQA together.

Chatbot Arena is a good reference point for overall quality. If a model ranks high there, it will most likely be good for the majority of ordinary tasks.

Fresh benchmarks are more reliable than old ones. GPQA, MMLU-Pro, LiveBench, HLE: there is less chance of contamination there than in classic tests from five years ago.

The best benchmark is your own tasks. Take 20-30 real examples from your work and run them through several models. That will tell you more than any public tables, because it measures exactly what you need.

Useful resources

Chatbot Arena: live rating based on human preferences
Artificial Analysis: model comparison by quality, speed, and price
LLM Stats: aggregator of results across different benchmarks
SWE-bench: coding leaderboard

World models: the next big step

2026-02-13T09:30:10+00:00

LLMs understand text. You write - they answer. But ask a language model what happens if you push a glass off a table - and it will give a statistically likely answer. Not because it understands physics, but because it has read a lot of texts about falling glasses.

This is a fundamental limitation. Language models work with tokens, not with reality.

World models are a different approach. This is AI that builds an internal model of the world and can simulate it. It doesn’t generate text about physics, it understands physics. It can predict what happens if you take action X in context Y.

Who’s working on this:

AMI Labs - Yann LeCun left Meta and is raising €500M for a startup in Paris. His thesis: LLMs are a dead end for real AI.
World Labs - Fei-Fei Li, the “godmother of AI” from Stanford. $230M from Andreessen Horowitz and Nvidia. Their product Marble generates interactive 3D worlds.
DeepMind - Genie 3 creates interactive environments in which you can train AI agents.
Decart - an Israeli startup, $21M from Sequoia. Their Oasis is a fully AI-generated Minecraft. No code, just a neural network.
Odyssey - founders from the self-driving industry. They make “interactive video” - 3D worlds from text and images.

Practical applications are already visible. OpenAI’s Sora is learning to understand that objects don’t disappear when they leave the frame. Tesla FSD builds a model of its surroundings and predicts the behavior of other cars. Robots need to understand how objects interact in the physical world.

The path to AGI may lie through a combination: LLMs for communicating in natural language, world models for understanding the physical world. Layers, not one universal algorithm.

2026 is the year world models stop being a research topic and become products.

Renaissance person

2026-02-12T10:29:01+00:00

I’ve noticed that the most interesting people around are the ones who are hard to describe in one word. You ask “what do you do?” - and they hesitate, because an honest answer would take about five minutes.

That used to be considered a problem. You can’t focus. You’re spreading yourself thin. You’ve got ants in your pants.

And now I look at it and think - maybe that’s actually the advantage?

The education system is built for specialization: pick one direction, dig deep, become an expert. The logic is clear - if you know one thing better than everyone else, you’ll get hired. But in reality the most valuable solutions appear not in the depths of one field, but at the intersection of several.

A person who understands both code and design builds products differently from a pure developer. Someone with experience in sales and psychology sees the customer more fully than just a good salesperson. Strange combinations of skills are not a bug, they’re a feature.

Da Vinci did everything at once - painting, anatomy, engineering, architecture. Not because he couldn’t choose, but because one fed the other. Understanding anatomy made his paintings come alive. An artist’s eye helped him design machines that were not only functional but beautiful.

Every interest adds connections in your head. The more connections - the more often you notice patterns that others miss. This isn’t about “knowing a bit of everything” - it’s about the ability to see how things affect each other.

If you have many different interests and all your life you’ve felt there was something wrong with that - maybe it’s the other way around. Maybe that’s your main asset.

Zero-sum game

2026-02-10T10:12:11+00:00

Everything in the world already belongs to someone. Land, resources, markets: it’s all divided up. That’s how it was a thousand years ago, and that’s how it is now.

But back then that meant one thing: if you want more, take it from your neighbor.

The Romans built an empire not because they were evil. There was simply no other way to grow. No technology, no productivity growth, so wealth could only be redistributed. By the sword.

Historians call this the Malthusian trap. Population grows, resources are finite, periodically a war or an epidemic happens, population falls, the cycle starts over. Thousands of years of the same thing.

And then something changed.

In 18th-century England people started building machines. Looms, steam engines, railways. For the first time, labor productivity began to grow faster than population.

That was the moment a different path appeared: not dividing what exists, but creating something new.

Your great-grandfather and his great-grandfather lived roughly the same way. And you live radically differently. Not because you’re smarter, but because there are several technological revolutions between you. Each one added new pieces to the shared pie.

We’re used to thinking of the world as an arena of competition. My win is your loss. But that’s not a law of nature. It’s a consequence of stagnation. When nothing new appears, all that’s left is to divide the old.

The problem is that our brain evolved precisely in that kind of world. Tribe vs tribe. Us vs them. Resources are finite, the stranger is dangerous. Those instincts haven’t gone anywhere.

Look at the rhetoric of recent years. Trade wars. Protectionism. Sanctions. “They’re stealing our jobs.” That’s the language of scarcity. The language of dividing up. When growth slows, we automatically fall back on old patterns.

But it’s a temporary lapse.

AI, fusion, space, biotech: the next wave of technology is already here. Every breakthrough creates new pieces of the pie. Every jump in productivity makes conflicts less attractive: it’s easier to create than to take.

Humanity always returns to growth. Sometimes through crises, sometimes through wars, but it returns. The alternative, endless dividing up, is too exhausting.

Reinforcement learning in humans

2026-02-08T16:04:32+00:00

Can you learn something just by imagining it in your head?

In 1983 the psychologists Feltz and Landers gathered data from 60 studies on the subject. The answer: yes. Mental practice genuinely improves skills, from free throws to playing the piano.

And it is not magic. When you imagine a movement, the brain activates the same areas as during the real action. The signal just does not reach the muscles. But the neural connections get stronger.

Essentially the brain is a simulator with built-in reinforcement learning:

It builds a model of the world
It runs scenarios
It gets virtual feedback
It adjusts behavior

I used this myself when I was doing wrestling. You lie there in the evening, eyes closed, and replay throws in your head and how they would play out under different conditions. I did not know back then that this is called motor imagery, I just felt that it helped. My classmate Borya did the same thing with kicking a ball: he would run through the kicking technique before going to sleep.

Chess players analyze games in their heads, without a board. Pianists “rehearse” pieces without touching the keys.

And lucid dreaming is a whole other level. It is essentially a world model you can interact with: a full-blown simulator of reality where you can train anything without consequences.

Why does this matter now? Because AI is going down the same path.

In robotics they figured out long ago that training a robot in the real world is expensive, slow and dangerous. So first they run thousands of simulations in Isaac Sim or MuJoCo, then transfer to hardware. It is called sim-to-real.

World models are one of the main directions in AI: teaching a model to predict “what happens if” without real interaction with the environment. Essentially, building an internal simulator of the world.

For many years we thought neural networks needed terabytes of real experience. It turns out a good model of the world and imagination are enough.

The first simulator of reality did not appear at Nvidia. It appeared inside the skull about 300,000 years ago.

Aggregation of uncertainties, or iteration beats planning

2026-02-07T09:06:08+00:00

There are two kinds of systems.

The first kind tries to predict, plan and control everything. Gosplan. Five-year strategies. Detailed business plans.

The second kind does not pretend to know the future. Markets. Evolution. Startups. The immune system.

The first kind looks smarter. The second kind wins.

Why?

Every link in a chain of decisions carries uncertainty. Uncertainties do not add up, they multiply. 10 steps at 90% confidence each = 35% at the end. The longer the plan, the more useless it is.

The economist Hayek wrote back in his day: knowledge is distributed. It cannot be gathered in one place without losses. Every time information is passed up the hierarchy, it gets distorted. That is why Gosplan lost to the market: not because the market is smarter, but because it does not lie to itself about what it knows.

Startups beat corporations for the same reason. Not because founders are more brilliant than managers. But because a startup says “we don’t know, let’s test it”, while a corporation says “we did the research and we know”.

Evolution works without any single plan. Just variation + selection. The immune system does not know which virus is coming, so it prepares for everything at once.

The practical conclusion: iteration > planning. Not because planning is bad, but because reality is more complex than any model. Systems that embrace uncertainty beat systems that deny it.

And one more corollary: the higher the uncertainty of the outcome, the less sense there is in long planning. When you don’t know what you will get, it is better to build it and see.

Agent Teams in Claude Code: what has actually changed

2026-02-06T16:27:04+00:00

Claude Code has long been able to run subagents: separate agents for specific tasks. It worked simply: the main agent hands out a task, the subagent does it and returns the result. Like a boss and a worker.

Agent Teams is a different model. Now agents can talk to each other directly, argue, and coordinate on their own without constant supervision from above. In effect it’s a shift from “I hand out tasks” to “the team figures it out itself”.

What exactly is the difference

Subagents work in isolation: do the task, return the result, done. They don’t talk to each other. If one finds something important for another, there’s no way to pass it on directly.

In Agent Teams each teammate has its own context, but they see a shared task list and can message each other. If the security agent finds a problem that affects the performance agent’s work, it will message it itself, without the lead getting involved.

Where this is actually useful

Debugging with an unclear cause. You launch 5 agents, each investigating its own hypothesis. They don’t just work in parallel: they argue with each other and try to disprove each other’s theories. This matters, because a single agent tends to latch onto the first plausible version and stop looking.

Code review from different angles. Three agents look at one PR: security, performance, test coverage. Each focuses on its own area, but if the security agent sees that some code affects performance, it pings its colleague right away.

How to use it

First enable it in settings.json:

{
  "env": {
    "CLAUDE_CODE_EXPERIMENTAL_AGENT_TEAMS": "1"
  }
}

Then describe the task and the team:

Users are complaining that the app crashes after the first message.
Create a team of 5 agents and have each one investigate its own hypothesis.
They should argue with each other and try to disprove each other's theories.

Or for review:

Create a team to review PR #142.
One agent for security, one for performance, one for tests.
Have each one do a review and share its findings with the others.

You can require a plan before any work:

Create an architect teammate to refactor the auth module.
Require approval of the plan before any changes.

What to keep in mind

The feature is experimental and burns tokens: each teammate is a separate session. For simple sequential tasks the regular mode is more efficient. Agent Teams make sense when parallel work and communication between agents genuinely add value.

Optimize risk, not return

2026-02-06T15:31:03+00:00

Most investors look at potential return. “How much can I make?” is the first question. I’d say that’s the wrong focus.

Over a long horizon, risk/reward in the markets is leveled out. There’s no fundamental difference between investing in real estate or bitcoin. The market evens everything out. The real question is different: how well do you understand particular risks, and what edge do you have in dealing with them.

The right approach is to invest where you have an edge in understanding the risks. Where you see and know more than the other market participants.

A simple example: investing in Russian companies carries less risk for a Russian than for an American. You understand the context, read the news in the original language, feel the mood. Plus you have no infrastructure risks: sanctions, frozen assets, problems withdrawing funds. For an American investor those are real risks; for you they simply don’t exist. That’s an edge you should use.

The same works in any field:

If you’ve been in an industry for 10 years, you see risks a newcomer doesn’t
If you understand a technology deeply, you can assess the real probability of failure
If you have insider knowledge of a market, you know what others don’t

The formula is simple: look for asymmetry in the perception of risk, not in expected return. Invest where your assessment of risk is more accurate than the market’s.

When Karpathy feels behind

2025-12-27T12:31:37+00:00

If even Karpathy says he feels like he is falling behind as a programmer, that is a good indicator of the times.

The profession is being rebuilt. We no longer just write code; we try to wire together a pile of stochastic, semi-transparent systems: agents, prompts, contexts, tools, integrations. The new layer appeared very quickly, and nobody has a proper mental model of it yet.

The feeling of “I could be doing several times more if I just wired all of this up properly” seems to be universal right now. It is not about intelligence, it is about the speed of change.

https://x.com/karpathy/status/2004607146781278521

Two types of destruction

2025-10-20T09:53:05+00:00

There are two types of destruction.

Destruction by forces coming from inside the subject, and destruction of the subject by external forces.

The first happens much more slowly and more rarely than the second. But the first is a form of sustainable development, and the second of unsustainable.

With external destruction the subject almost always returns to its previous state, since the internal structure still remains the same. A real break and transition happen only after internal work on reformatting.

That’s why it’s important not to force changes from outside. You need to help them happen from inside.

This applies to everything - from simple microorganisms to complex social structures like states.

Venture capital in Russia

2025-10-08T13:34:45+00:00

For a short time, Russia had a venture capital market, which was certainly a good option for some extremely high-risk types of business.

That time has passed now, at least for the mass entrepreneur. Maybe venture capital will come back to Russia, but for now local entrepreneurs need to learn to build technology companies without it.

Because Russia lacks a large layer of infrastructure services that have already shown their effectiveness and return on investment in other countries, there seem to be good chances of finding financing at the early stages even in today’s Russian realities.

At later stages, the missing late-stage venture investors could well be replaced by the stock market. As it was, for example, not so long ago in the US in the 70s, 80s and 90s.

Microsoft and Amazon went public very early by the standards of today’s American companies. No need to look far: even the relatively modern Tesla went public at a fairly early stage of development. Microsoft at a market cap of ~$800M (~$2B in today’s dollars) at the time of its IPO, Amazon ~$450M (~$850M in today’s dollars), Tesla ~$1.6B (~$2B in today’s dollars).

The reason was precisely the relatively weak state of the venture industry.

And that allowed ordinary retail investors to make very good money. Which, by the way, almost never happens now in that same US. Look at the valuations Airbnb or Uber went public at and what growth in market cap they’ve been able to deliver over these years.

I’d like to believe Russia can repeat that success. Retail investors in Russia have money. What’s missing is an environment that doesn’t create as much uncertainty and distrust as we have now.

In Russia, the president has set a significant goal of increasing the stock market’s share of GDP, from the current ~25% to 66% by 2030. I hope the government and the relevant agencies will nevertheless create the preconditions for achieving it. Because the potential is really there. The country needs its own developed space industry, electronics, robotics and AI.

Some materials for self-study

2025-09-21T10:40:54+00:00

I consider all the materials below excellent as an introduction to the subject.

On LLMs, the best source is https://karpathy.ai/ He has a series of lectures on YouTube there about how they work.

Read the OpenAI and Anthropic blogs; there is a lot of good information there on the latest progress of frontier models.

Some videos I highly recommend watching:

The moment we stopped understanding AI [AlexNet]
Ilya Sutskever: “Sequence to sequence learning with neural networks: what a decade”

[Visualizing transformers and attention

Talk for TNG Big Tech Day ‘24](https://www.youtube.com/watch?v=KJtZARuO3JY&ab_channel=GrantSanderson)

[Attention in transformers, step-by-step DL6](https://www.youtube.com/watch?v=eMlx5fFNoYc)
A Survey of Techniques for Maximizing LLM Performance
The New Stack and Ops for AI
How do Graphics Cards Work? Exploring GPU Architecture
[AI Semiconductor Landscape feat. Dylan Patel BG2 w/ Bill Gurley & Brad Gerstner](https://www.youtube.com/watch?v=QVcSBHhcFbg)

Agents

2025-09-21T10:30:22+00:00

People describe agents in different ways, but I would say an agent is an AI application that can make decisions autonomously and work “for a long time” without human intervention. That is, both generating a decision and validating it against the built-in rules happen inside the application.

Popular scenarios for agents:

Customer support agent
SDR (sales rep)
Programmer

In principle, you can take a list of professions with repetitive tasks, a large number of workers, and a relatively expensive hour of work, and that will be a good list of professions/scenarios for automation with agents. All three examples above would be at the top of such a list.

For the most part, agents are still written directly against foundation model APIs, with complex logic layered on top to make them work reliably. But this is a frontier area, so best practices are being formed right now. I will write about this separately.

Take a look at Replit Agent (programming), Claude Code (programming), and Sierra (customer support) to get a sense of where the market is heading.

I also want to point out another area: automating routine tasks in the browser. There are tools for browser automation (which is in fact also an autonomous agent), like Browser Use or Computer Use from Anthropic. These are tools that let you literally say: go to that site, enter this login/password, then find this button, click it, download the results as a PDF, and so on. Still a bit raw for mass use, but the scale of future applications is already visible. And it is huge.

Second layer apps

2025-09-21T10:11:49+00:00

The most interesting thing happening right now, in the first half of 2025, is Layer 2 applications built on top of the foundation model API products from the giants above. There are a lot of very cool things out there.

There are thousands of them, but I’ll just single out a couple that I use myself:

ChatGPT or Claude: well, this one is obvious. Classic second layer products, since on top of their own basic API the teams bolted on a pile of logic that turns a not very human-friendly API into a product you use every day for different purposes.
Perplexity: partly a competitor to ChatGPT, but more about finding information and research. Simplifying, they are rethinking Google’s entire product stack (from search to shopping) with AI. Successfully in some places, not in others. But their search has definitely found its niche and deserves attention.
Cursor: an AI IDE based on VS Code. Writes code for you from your instructions. An autonomous agent is planned soon.
Supermaven: a VS Code plugin that does roughly the same thing as Cursor, but inside VS Code. Its main strength is more in code autocompletion.
Cline: another developer product. Also works as a VS Code plugin, but is better at working with the full project documentation, so it understands context better.
Ollama: an SDK that lets you install open-source model weights locally and effectively run LLMs on your own GPU (essentially for free) for your own purposes on your laptop.
OpenAI Operator: still closed, I think, but this is a move into agent territory.
Browser Use: a framework that lets you automate work in the browser with a prompt. A top tool for scraping data, automating simple browser tasks, and so on.
Browserbase: third-party infrastructure for browser automation. The next step compared to Browser Use, where you can run something similar on third-party servers and not deal with infrastructure problems.
Replit Agent: they built an agent that creates a full project from your description (not a single code file, but all the files the project needs). You write what you need and get a working app. A very cool tool. Usually easier to pick up than Cursor.
ElevenLabs: an API for working with voice, a super tool; we build a lot of things on it. For example, an AI sales rep that calls offices following our script and takes into account the real-time data we feed it.
Claude Code: an agent that runs locally on your machine and writes code autonomously. In general it can be used for more than just code. I use it rarely so far, but it looks very promising.

Again, there is a sea of products; I’ve listed only a few here that I’ve used recently or see value in. The Valley is having a renaissance right now; the techies have come back from Texas to build AI applications. Keeping up with everything is impossible.

Right now most applications are around text GenAI. But very soon we will see mass-market products in image generation, video generation, and even 3D worlds.

A brief history of AI applications

2025-06-03T20:28:42+00:00

Intro

People have been doing machine learning for a long time. Progress was always happening in different corners of computer science. Artificial neural networks have been known since, I think, the 50s, and 30 years ago there were already cool applications people had written - for example here Yann LeCun (now a director at Meta), one of the founders of CNN architectures in neural networks, shows a neural net recognizing digits. But there were no big breakthroughs in those applications, more like cool demos that were hard to apply to real problems.

Everything developed more or less evenly until, in the late 2000s, Geoffrey Hinton (the one who recently, unexpectedly, got a Nobel Prize in physics for his work in AI), who had long worked on psychology, how the brain works, and later neural networks, realized that graphics cards (GPUs) turn out to be better suited than CPUs for the math that neural networks need to do in enormous quantities (linear algebra, matrix operations, tensors and so on). And that’s when the breakthrough happened. Specifically in image recognition.

It’s important to understand the context. By 2010 researchers had one big problem - there was no proper dataset for training and testing image recognition models. There were various small sets like MNIST (handwritten digits) or CIFAR-10 (32x32 images in 10 categories), but that was like learning to drive on a toy track.

In those same years Fei-Fei Li from Stanford did something fundamental - ImageNet. She and her team collected a dataset of 14 million images in 22 thousand categories. But the coolest part - they launched the ImageNet Large Scale Visual Recognition Challenge (ILSVRC) competition in 2010. The task is simple: here’s a million training images in 1000 categories, show how well your model can classify them.

For the first couple of years all the solutions were based on classical computer vision - features were extracted by hand, then fed into an SVM or something similar. The error rate was around 25-30%. And then in 2012 Hinton and his team showed up with their model and simply destroyed everyone - 15.3% error! That was a gap of almost 10 percentage points from the nearest competitor. That’s a lot.

AlexNet

Thanks to improved compute and new approaches to architecture, Hinton and his team managed to train a very cool (for its time) neural net - AlexNet. One of Hinton’s students, by the way, was Ilya Sutskever, who would later be one of the co-founders of OpenAI.

You could say the clock on deep neural networks started with AlexNet, in 2012 - when people realized that the architectures described 30 years earlier actually work, they just had to be made much bigger (more neurons) and you had to spend much more compute for the network to learn.

Then came another period of steady progress across all the disciplines within machine learning over the next 5 years - good image recognition, speech recognition and synthesis, image style transfer and so on. Very hyped apps appeared - like Prisma or MSQRD. OpenAI was founded, somewhere around the turn of 2015/2016.

But with all that, no fundamental breakthrough happened, since neural nets back then (quite small by today’s standards) required a lot of GPUs for training and inference (that’s what the actual work of a trained network is called). Everyone sort of wanted even bigger networks, understood that bigger would mean better results, but there wasn’t really a way to go bigger.

That was the case until in 2017 the folks at Google wrote the now historic paper - Attention Is All You Need. In it they presented several breakthrough ideas on neural network architectures at once, namely the architecture now called the Transformer. One of the main ideas was “attention”, which let the network understand context better, and another was how to technically work better at the level of operations on the GPU.

It was a breakthrough paper. Nevertheless, even though the paper was written at Google, the main beneficiary soon turned out to be OpenAI. And the person who helped them succeed at that was Ilya Sutskever. I listened to an interview where he says that when he saw the transformer paper he immediately understood that this was what would let them scale their neural nets at OpenAI.

The transformer era

After transformers appeared, a real race began. In 2018 Google rolled out BERT (Bidirectional Encoder Representations from Transformers) - a model that learned to understand the context of words from both sides of a sentence. It was a breakthrough for language understanding tasks - classification, question answering and so on. BERT broke all the records on benchmarks and showed that pre-training on huge corpora of text gives incredible results.

OpenAI had long been working on text generation, or more precisely on predicting which word comes next in a set of words fed to the model as input. They applied “transformers” there and the quality of predictions grew substantially, it was a breakthrough. That’s how the GPT models appeared - Generative Pre-trained Transformer. OpenAI first offered GPT models through an API for developers, and then they made a consumer product, ChatGPT.

GPT-1 was released in 2018. It was a relatively small model with on the order of 100 million parameters, but it already showed that you could take a transformer, train it to predict the next word on a pile of text from the internet, and then fine-tune it for specific tasks.

In 2019 they released GPT-2 with 1.5 billion parameters. And here something interesting happened - OpenAI at first refused to publish the full model, saying it was too dangerous, could generate fake news and so on. People laughed, but when the model was eventually released, it turned out it really could generate very convincing text. That was the first warning bell that we were approaching something serious.

Everything turned out just as Sutskever predicted. Today “transformers” are in fact under the hood of ML applications everywhere, which are now called AI applications. Almost all the improvements in video, image, text generation and so on today are applications of the architecture the folks at Google published in 2017.

What’s more. Over the past 3-4 years almost all the major hardware vendors have started designing GPUs/TPUs/AI chips specifically for the needs of transformers.

The ChatGPT revolution

In 2020 OpenAI released GPT-3 with 175 billion parameters - a monster compared to the previous versions. But the real revolution happened when OpenAI figured out how to make these models useful for ordinary people.

The thing is, if you just take a pre-trained model, like GPT-3 was, and try to talk to it the way you’re used to talking to modern chatbots, you won’t get any long dialogue out of it. It will be more like a set of logical but not very connected texts. Not bad for getting a short answer, but definitely not an AI conversation partner. Passing the Turing test was still a long way off. The secret sauce was RLHF - Reinforcement Learning from Human Feedback.

The idea of RLHF is simple: take a bunch of people, they rate which of the model’s answers are good and which are bad. Based on those ratings we train another model (a reward model), which learns to predict what people will like. And then we use reinforcement learning so that the main model generates answers this reward model will like. Essentially, we teach the AI to be helpful, harmless and honest through human feedback.

It was RLHF that turned the raw GPT-3.5 into ChatGPT, which blew up the internet in November 2022. It was the fastest-growing consumer product in history.

This breakthrough, which OpenAI was the first to pull off, around 2021/2022, spurred everyone else - startups, venture investors, the big tech giants - to invest money and time in this industry. In particular, Nvidia’s stock grew 10x on the frantic demand for its chips, since they are the default choice for anyone who wants to train and run inference on neural nets.

LLM

LLM stands for Large Language Models, in general the name of a class, but in practice today it’s the subset of transformers that work with text.

If we talk purely about text, what’s happening today can be split into two stages:

First everyone improved quality by increasing the number of GPUs and the amount of data (for example OpenAI’s GPT-1/2/3/4/4o models)
Then everyone hit a certain ceiling on compute/data and started doing active post-training optimization, so-called reasoning - to simplify, improving answers by having the network run its generated answer through itself again and evaluate its quality. Then it returns the improved answer (for example OpenAI’s o1, o3, o3-mini models)

Now every LLM provider has a bunch of different models for different purposes - some faster and cheaper, some pricier and thinking longer, some better for general tasks, some better for coding and so on.

There’s a huge number of different LLM models on the market right now - proprietary and open source. There’s plenty to choose from. I’ll list just some of the ones everyone is talking about as of spring-summer 2025:

OpenAI

Many different models, good quality on average across many tasks
They have ChatGPT - by a huge margin the most popular consumer AI product, they have hundreds of millions of users

Anthropic

Can probably officially be called number 2 after OpenAI on the sum of all factors
My favorite, we use them a lot
Their model Claude Sonnet 3.5/3.7 is excellent for almost all tasks, in particular for code generation and copywriting
Update: the 4 models have now already come out

Gemini

Google’s product
The latest models are very good
The main feature is the huge context window, meaning you can load a lot of material into the chat or just keep a conversation going for a very long time without interruption
Google in general should be the leader here considering they do almost everything, invented those same transformers and make their own chips (TPUs, Nvidia’s competitors), but so far they’re behind in the race for developers and consumers. But I think they have a chance to catch up

Gemma

Also Google, but open source
The latest models are quite decent on the benchmarks

Galitsky

2025-05-27T12:34:36+00:00

Sergey Nikolaevich is a remarkable man. It is impossible not to be happy both for his team’s victory and for his personal success, even if you are not a Krasnodar fan. An incredible person and an example for all of us.

Krasnodar is a very young club, founded in the 21st century, with only 17 years of history. Yet from the very beginning it was clear that Galitsky was moving toward the championship thoroughly and purposefully, building the leader of Russian football step by step. For many years now Krasnodar has had the best training infrastructure and football academy of any Russian club, and that alone is a huge victory. And now the thing the team was relentlessly heading toward all these years has finally happened. I am not surprised in the least. It was only a matter of time.

Galitsky has proved more than once that professionalism, resilience, self-belief and persistence pay off. He proved it in business, in his personal life, and now, once again, on the football pitch.

Galitsky once said he wanted to see Krasnodar become champions with a squad made up entirely of graduates of the club’s academy. I think that is exactly what will happen.

On Tesla's market cap

2025-01-12T20:36:44+00:00

Tesla’s market valuation is around $1.2 trillion. I got curious, based on some high-level fundamental estimates, how realistic that figure is given the company’s ambitions. Tesla’s main current business is cars, but the company also has three other potentially large lines that analysts often factor in: the energy business, a ridesharing service (eventually based on autonomous vehicles), and robots.

Cars

There are about 1.6 billion cars in the world. Roughly 80–90 million new cars are sold every year.

Annual sales of some manufacturers:

Toyota: ~10 million cars (1st place)
Volkswagen: ~9 million (2nd place)
Hyundai–Kia: ~8 million (3rd place)
BYD: ~4 million
Tesla: ~2 million

Tesla currently ranks roughly 11th–14th among the largest automakers by output. Its closest competitor in EVs is BYD, which sells more than 4 million cars a year (including roughly as many EVs as Tesla), yet its valuation is around $100 billion. (Yes, BYD has a lower margin per car, but let’s set that aside for now.) For comparison, Toyota’s market cap is around $300 billion.

I don’t believe that in the long run any automaker can maintain a radically different margin from its competitors, since the market is highly competitive. So it seems reasonable to assume that Tesla will eventually converge to the margins of other mass-market manufacturers. If Tesla became the largest manufacturer, like Toyota, selling roughly 10 million cars a year (about 10–12% market share), then at today’s prices its automotive business could be valued at around $300 billion, by analogy with Toyota.

Energy business

This is the hardest part of Tesla’s business to value, because entering the energy market effectively means competing with a huge swath of industries. I’ll use a very simple analogy: ExxonMobil. Exxon has an Upstream segment (oil and gas exploration and production) and a Downstream segment (refining, distribution and retail sale of fuel, including gas stations). Downstream is more relevant for us, since Tesla doesn’t yet produce a meaningful amount of energy on a global scale (solar is still small) but does distribute energy products (charging, batteries, etc.).

I admit this is a crude comparison, just for a high-level sanity check. According to public data, the Downstream segment brings Exxon about 30% of the company’s total revenue. Exxon’s total market cap is about $470 billion, so the Downstream share could be valued at roughly $150 billion. Let’s assume Tesla’s energy business can eventually reach the same scale and replace oil with electricity. Then, at today’s prices, Tesla’s energy line could be worth about $150 billion.

Ridesharing

This one is simpler. Uber’s market cap is about $150 billion. Suppose Tesla, with its autonomous cars, could “kill Uber” and take over that market entirely. That could add another $150 billion to Tesla’s market cap.

Robots

This is the most interesting part. In theory, the long-term price of a humanoid robot could be comparable to the cost of a small car (given the materials cost), say around $20,000. If so, there’s no fundamental reason why people would buy more robots than cars (provided the robots’ functionality is sufficient). That suggests we can take the current car fleet (1.6 billion vehicles) as a reference point.

Look at how personal computers reached 80% penetration in ~30 years. For robots the process may go slower, since they’re more expensive than computers. But if we assume mass adoption takes 30 years, with an annual replacement rate of 10%, then during the active adoption period that’s about 100 million new robots a year. If Tesla can take a large 30% share of the market, i.e. about 30 million units a year, that would exceed Toyota’s current annual car sales. Toyota’s market cap is about $300 billion for 10 million cars sold, so by analogy Tesla’s robotics business could support a valuation of $900 billion.

Summary

Cars: $300 billion
Energy: $150 billion
Ridesharing: $150 billion
Robots: $900 billion

That adds up to about $1.5 trillion. Given the already fairly optimistic assumptions, and assuming 100% success in reaching such ambitious goals, that sum justifies Tesla’s current valuation of $1.2 trillion. But this is an extremely optimistic scenario. More realistically, the final figure should be at least two to three times lower.

What fundamental factors might I be missing?

Sutskever at NeurIPS

2024-12-21T17:00:32+00:00

I only just got around to Ilya Sutskever’s recent talk at the latest NeurIPS.

Nothing new idea-wise, but as a retrospective on the past 10 years of deep learning history, from a person who in many ways shaped this field of research, it definitely deserves attention.

The talk itself is only 16 minutes.

https://www.youtube.com/watch?v=1yvBqasHLZs&t=460s

OpenAI founders' correspondence

2024-12-13T21:24:10+00:00

The main read of the day - the correspondence of OpenAI’s founders from 2015 to 2019. Made public in connection with the case currently going through the courts.

https://www.lesswrong.com/posts/5jjk4CDnj9tA7ugxr/openai-email-archives-from-musk-v-altman

Not all the emails are there…

Continuing this hot topic, OpenAI is releasing material today refuting many of Musk’s claims, or at least aimed at that.

https://openai.com/index/elon-musk-wanted-an-openai-for-profit/

How to measure the level of a civilization?

2024-12-10T08:52:01+00:00

In 1964 the Soviet astrophysicist Nikolai Kardashev proposed an elegant scale: by the amount of energy a civilization is able to use.

Type I: a civilization able to use all the energy reaching its planet from its star. For Earth that’s on the order of 10^16–10^17 watts.

Type II: a civilization that has harnessed all the energy of its star. For the Sun that’s on the order of 10^26 watts (ten orders of magnitude more than Earth receives). At this level it becomes possible to build a Dyson sphere, a giant structure around the star to collect all of its energy.

Type III: a civilization that controls the energy of an entire galaxy, on the order of 10^36 watts for the Milky Way.

Where are we now? According to the BP Statistical Review for 2022, humanity consumes about 19.1 terawatts (1.91×10^13 watts).

Carl Sagan proposed a formula for measuring a civilization’s progress: K = (log₁₀ P - 6) / 10, where P is the energy consumed in watts.

By this formula a Type I civilization has a value of 1.0, Type II 2.0, and Type III 3.0. Humanity is currently at 0.73, meaning we’ve covered almost three quarters of the way to Type I on the logarithmic scale. In absolute numbers, though, the picture is different: we’d need to increase our energy consumption almost 10,000 times to reach Type I.

Progress

2024-10-21T09:32:01+00:00

For quite a long time we have been living in a world where technological progress is almost entirely associated with progress in IT: the world of computers, the internet, mobile apps, machine learning, telecommunications.

But technology is far more than just IT. Mining, construction, energy infrastructure, mechanical engineering, engineering systems, medicine, and many other fields are all technology, and progress in them has, unfortunately, genuinely slowed, more in some places and less in others.

One could speculate about the reasons for this slowdown. Was the rivalry between the US and the USSR, for example, an engine of progress? And is a multipolar world, in this sense, a positive that we don’t talk about? That probably deserves a separate note.

The example of the “caught” Super Heavy booster shows two things once again:

Technology doesn’t develop on its own; it is the result of hard work
People are capable of achieving the highest goals they set for themselves

This success reminds us of the importance of ambitious projects in various areas of engineering. Breakthroughs in controlled thermonuclear fusion, quantum computing, or bioengineering could change the world no less radically than the arrival of the internet.

Thanks to them for the reminder. We absolutely need technology to be more than just a symbol of IT.

The value of human life

2024-10-20T08:56:01+00:00

Part 1

In the kaleidoscope of cultures and civilizations that inhabit our planet there is a fault line that becomes more visible every day. This invisible but tangible boundary divides the world into two parts: “Western” and “traditionalist.” And at the very center of this divide lies a fundamental question: what is the value of a human life?

In the Western world, shaped by centuries of humanist philosophy, the Enlightenment, and the struggle for human rights, the life of the individual has been raised onto the pedestal of the highest value. Here every life is considered sacred, inviolable, and priceless. Laws, social norms, and ethical principles are built around protecting and preserving human life at any cost.

This paradigm shows up in everything: from medical ethics, where doctors fight for a patient’s life no matter what, to the legal system, where murder is considered the gravest crime.

In contrast, the traditionalist world views human life through the prism of other values. Here the individual is part of a complex system in which tradition, religious norms, family honor, or the interests of the community can stand above the life of a single person.

In this worldview, death is not always perceived as an absolute evil. On the contrary, it can be a path to higher glory, redemption, or the fulfillment of one’s duty. Here the hero who sacrifices himself for a higher cause evokes not only grief but deep respect.

The collision of these two worldviews produces deep mutual incomprehension. To a Westerner, the idea of sacrificing a life for abstract concepts seems barbaric and inhuman. To a traditionalist, the Western cult of individualism looks selfish and short-sighted, unable to see a picture any bigger than a single human life.

This incomprehension shows up in international conflicts, culture wars, and even in everyday clashes between migrants and the native population in multicultural societies.

The question of whether the gulf between these worldviews can be bridged seems natural, but isn’t it itself a product of Western thinking? The pursuit of “mutual understanding” and “dialogue between cultures” is a typically Western concept, based on a belief in the universality of human experience and the possibility of rational consensus.

But what if the very idea that mutual understanding is necessary is just one more manifestation of cultural imperialism? Perhaps true respect for differences lies in acknowledging that some worldview barriers cannot be overcome.

Part 2

Karl Popper, in “The Open Society and Its Enemies,” formulated the paradox of tolerance: unlimited tolerance must lead to the disappearance of tolerance. Applied to our dilemma, this means that Western society, in striving to understand and accept traditionalist values, risks undermining the very foundations on which it is built.

On the other hand, traditionalist societies, when confronted with Western values, see in them an existential threat to their way of life. In this context their resistance to “Westernization” can be seen not as backwardness but as a form of cultural self-preservation.

Instead of an illusory “mutual understanding,” perhaps we should strive for a state of dynamic equilibrium. In this model both worldviews exist in constant tension, mutually restraining each other’s extremes.

Western individualism taken to the extreme can lead to the atomization of society and the loss of meaning in life. Traditionalism, without the restraining influence of ideas about human rights, risks sliding into totalitarianism and cruelty.

Perhaps it is precisely in this confrontation that the potential for humanity’s development lies. Just as physics has Bohr’s complementarity principle, where mutually contradictory concepts are necessary for a complete description of reality, so in the cultural sphere these opposing approaches to the value of human life may be the necessary poles between which the whole diversity of human societies develops.

In this context the task is not to reach a compromise or a synthesis, but to maintain a productive dialectical tension between these worldviews. This tension can become a source of creativity, innovation, and constant rethinking of our values.

Thus, instead of seeking an illusory “mutual understanding,” we must learn to live in a world where fundamentally incompatible views on the value of human life coexist. It is precisely this contradiction that makes us human in all our complex, paradoxical, and astonishing diversity.

The goal comes first

2024-10-19T19:07:31+00:00

The goal comes before everything else. Once it has been defined, progress becomes possible, in the form of movement toward it.

On the same theme, I recommend reading “The Goal” by Eliyahu Goldratt. The book is interesting both for its idea and for its delivery.

Notes to self

2024-09-26T12:10:29+00:00

One of the most important things I keep reminding myself of - you can’t spend too long thinking about a problem, you have to act faster. This is true almost always.

A book can’t be treated as a source of truth. A book should be treated as a conversation partner.

What if you were in the Matrix

2024-09-25T20:47:40+00:00

Imagine if you were told today that you are in the Matrix.

Everything you see around you, all the people you know, all your past experience, none of it is real.

But there is a pill, and if you take it you will leave the Matrix, losing everything you know. What is out there, you don’t know.

Would you take it?

Sigmoid and exponential

2024-09-20T08:02:27+00:00

A sigmoid and an exponential look pretty much the same up to a certain point. And many people often mistake one trend for the other in their forecasts. What is more, we have never seen infinite exponential growth in processes tied to the physical world. So, properly speaking, the question to ask is: where exactly will the inflection point be?

Blockchain and a banking system

2022-02-01T14:28:14+00:00

Let’s look at banking systems in the majority of modern countries. Basically, it has 3 layers of data exchange:

(1) the Central Bank (CB) database, exchange between the CB and commercial banks, to keep comm banks level ledger;
(2) interbanking data exchange, exchange directly between banks, like SWIFT;
(3) comm bank database, exchange between users of one bank, to keep users level ledger.

Interesting that if we’re looking only at these data layers, w/o looking at other non-data mechanisms within this system, all layers are similar to layers we have at modern blockchain systems, for example, in Bitcoin/Lightning:

L(1) - Bitcoin blockchain, exchange between main blockchain and lightning nodes;
L(2) - Lightning payment channel, similar to SWIFT, exchange directly between lightning nodes;
L(3) - Lightning nodes, similar to bank database, exchange between users of one node.

History rhymes. Or not?

What is a gas fee

2022-01-11T18:16:00+00:00

Ethereum basics

The concept of gas fees, how they are calculated and how affect your transactions is fundamental to understanding how the Ethereum blockchain works.

Basically, Ethereum is a decentralized computer that runs different types of transactions (txns) which could change the blockchain state.

Some examples of txns are:

Send ETH from one account to another
Interact with a smart contract, like swap a token on Uniswap

All txns are initiated by accounts of the Ethereum network and basically are cryptographically signed instructions for the computer, the so-called Ethereum Virtual Machine (EVM).

EVM is what defines the rules for computing a new valid state from block to block.

There’re two types of accounts in Ethereum:

Account managed by a human
Account managed by a program (i.e. smart contract)

Both types of accounts can:

Receive, hold and send ETH and tokens
Interact with deployed smart contracts

A smart contract is a program that “lives inside” a blockchain and can be triggered by txn from another account and execute many different actions, such as transferring tokens or even creating a new contract.

Since each Ethereum txn requires computational resources to execute, each txn requires a fee and must be mined to become valid.

Gas basics

“Gas” refers to the fee required to conduct a transaction on Ethereum successfully.

The simplest and most common txn type is a transfer of ETH from one account to another. Such txn requires 21000 gas to be executed.

Complex txns, like smart contract creation or smart contract call can cost substantially higher.

On a high level, the gas fee can be split into two different multipliers:

[Gas fee] = [Gas units] * [Gas price per unit].

[Gas units] is an amount of computational effort required to execute specific operations on the Ethereum network.

[Gas price per unit] is a price depending on the network busyness and denoted in gwei.

Gwei is 10^-9 ETH.
Gwei itself means giga-wei, and it’s equal to 10^9 wei.
Wei is the smallest unit of ETH, and it’s equal to 10^-18 ETH.
Wei itself is named after Wei Dai, creator of Bitcoin “predecessor” b-money, distributed electronic cash system.

Calculating gas used to be very complicated, but it was simplified after EIP 1559 implementation (as part of the London Upgrade) in the middle of 2021.

If your wallet supports EIP 1559 (like Metamask), gas for your txn will be calculated as:

[Gas price] = [Gas units] * ( [Base fee] + [Priority fee] )

It this particular example: [Gas price] = 248688 * (168 + 2) * 10^-9 * ETH_price = $128

[Base fee] is set by the network. It increases and decreases automatically based on network busyness. This fee gets burned when it’s paid.

[Priority fee] is optional and not burning. It goes to a miner as an incentive to process your txn sooner.

For example, If you put 0 gwei for [Priority fee] more likely there will be no reason for the miner to include your txn into the block.
Without this fee, miners would find it economically viable to mine empty blocks, as they would receive the same block reward.

You can find actual block base fee and priority fees for different types of txn execution speed (e.g. low, average, high) on various websites. For example here: https://etherscan.io/gastracker

[Max fee] includes both [Base fee] and [Priority fee] and it’s a maximum amount of fee you’d be willing to pay to a txn.

Metamask initially sets this amount based on the previous block’s history.

The higher [Max fee] is, the more chances txn will be successful in case of a surge in blockchain use (i.e. [Base fee] increase) after txn is submitted.

What is a gas limit and why it exists

So, now we know how fees per gas unit are produced.

Let’s move to the [Gas limit] and how it differs from actual gas usage.

[Gas limit] refers to the maximum amount of [Gas units] you are willing to consume on a txn.

If your gas limit will be lower than the actual amount of gas units needed to execute a txn, txn will be unsuccessful. The EVM then reverts any changes, but the gas will be consumed.

If we can exactly estimate a [Gas limit] for a txn it will be equal to actual gas usage.

But, usually, due to various reasons, it’s very hard to estimate exactly how many gas units are needed to execute a particular txn.

Even two absolutely the same txns can have different gas usage because of different EVM states while executing. Here’s an example:
https://ethereum.stackexchange.com/questions/44643/why-gas-used-are-different-for-same-transfer-tx

Final words

Txns are basically a list of instructions for EVM. And any instruction has a universally agreed cost in terms of gas.

More details about these ‘costs’ and other implementations of the EVM can be found in an Ethereum Yellowpaper:
https://ethereum.github.io/yellowpaper/paper.pdf

If you want to go deeper on this topic you can start with Ethereum documentation:
https://ethereum.org/en/developers/docs/gas/

Gas fees is definitely a great concept that showed credibility over many years, but the next big step in Ethereum adoption will be a new UI/UX layer where a regular user doesn’t even know about its existence.