In #Russian, but thanks to @volyx and FAANGtalk for the event on Web3 — https://www.youtube.com/watch?v=DQRaTQ5dNSc — and a humble plug that @BoreMeNo is also what yours truly keeps posting to.

Folks, I want to get a feel about n8n, its strengths and weaknesses.

Not some intro tutorial, but something advanced, and, ideally, DIY — to get my hands dirty.

Any papers / videos / channels to browse? Pointers much appreciated.

They say the best posts are the ones you write on the spot: https://dimakorolev.substack.com/p/immutable-metadata-less-storage

One piece of wisdom I came across recently is that all communication is strategic [1].

Granted, the author focuses on corporate communication — he’s writing a mini-series on Machiavellianism for engineers as we speak.

Full disclosure: I am not opposed to Machiavellianism at all, quite the contrary — a healthy dose of it in high-profile conversations is what most of us could make use of.

If your goal is to win big in your career, this wisdom is spot on. I’d even generalize it to “PR and self-presentation matter,” and add that consciously shaping the impression you make — starting with the most influential people around you — usually pays off quickly.

But a big part of me still disagrees, and here’s why.

Maybe it’s about my own values, but I’ve always found that while some compartmentalization helps with personal growth, too much of it slowly erodes one's identity.

If your identity truly is about building a corporate career, then yes: all communication is strategic. If your identity is “I put my best foot forward 9–5 to ensure a good life for my family, but after hours I’m not thinking about work,” that’s also perfectly valid.

But for me, there’s no work-life balance; there’s work-life synergy [2]. Over time, I’ve gravitated toward working with increasingly interesting people on increasingly meaningful projects. I’m not the kind of person who makes others work weekends — but myself I often do, because pursuing things I believe in is a core part of who I am.

And if I were to fully internalize the idea that all communication is strategic, it would quickly creep into my core identity. I’d inevitably start being strategic with friends and family, too. Maybe that’s fine at some high level — but day to day, I’d sacrifice quite a lot career-wise to stay true to myself.

This authenticity — something I’ve nurtured over time — would be gone the moment I fully embraced the idea that everything I say is part of some strategy to achieve some goal.

And I’m conflicted about it. As a contrarian thinker and doer, I often find myself in rooms where big decisions are made, and my input matters. I know I could easily shape myself into someone who “belongs” in those rooms. But I choose not to — it would feel like betraying a core ideal.

What’s “worse” is that people who are far more successful, at least on paper, often admit in private that they’d prefer a life closer to mine — even as they keep playing the strategic game.

So go figure.

Luckily, the world is big enough for all of us to coexist — and to bring more meaning and more wealth into it, each in our own, deeply personal way.

[1] https://josvisser.substack.com/p/all-communication-is-strategic
[2] https://dimakorolev.substack.com/p/a-players-who-are-they

Good one.

Although I struggle to find Denmark.

Edit: It's there next to Germany, just unreadably small. Denmark's border with Sweden is not a land border, so all good.

Screw the Optional<bool> holy war — we're living in an async world now.

CancellableFuture<void> FTW!

Today was the day I decided to educate myself on worldwide financial numbers.

Money supply, M2: about $20T in the U.S., roughly $100T worldwide. On a per-capita basis, that’s around $60K per person in the U.S. versus about $12K globally.

Stock market value: approximately $60T in the U.S. and $120T worldwide.

Crypto assets: a bit under $4T. To sanity-check this number: a bit less than 20M BTC have been mined, and Bitcoin makes up roughly half of all crypto tokens. At around $100K per BTC, that’s about $2T in BTC, and add a similar amount across other tokens.

Now let’s look at trading volume relative to total assets, per asset class — i.e. the velocity of money.

For equities in the U.S., the velocity stands at ~2.4x — meaning each dollar invested in stocks changes hands about 2.4 times per year.

Annual Forex trading volume is just under $2 quadrillion (~$1,890T). Compared to a $100T global M2 money supply, that’s a turnover factor of just under 20x.

If we filter for USD-related Forex trades only, about 88% of the total, the denominator drops to $20T. That pushes the effective turnover factor to nearly 88x.

For context, when looking at actual U.S. dollars in circulation — i.e. transactions that contribute to GDP — the velocity today is roughly 1.3x. It used to be around 2x before the dot-com bubble and before 2008.

Now, why I started writing this post: I wanted to understand crypto trading volume — and thus its velocity.

This turns out to be tricky, since it’s hard to separate crypto spot trading from crypto derivatives. But loosely speaking:

∙ Without derivatives crypto’s annual velocity sits at 6x .. 10x.
∙ With derivatives it’s likely 20x+, because leveraged speculation.

Is that big or small? Well, let's compare this number to traditional finance.

This proves to be tricky, since it's hard to draw the line between crypto trading and crypto derivatives trading. But loosely it looks like without derivatives we're looking at something between 6x .. 10x in annual velocity. With derivatives it's likely 20x+, mostly due to high-leverage speculation.

Using gross market value as the denominator, traditional derivatives trade at roughly 0.75x per day. Yes, per day. With ~250 trading days per year, that’s around 180x annual turnover.

That’s why I’d argue that, in crypto, 20x isn’t actually that high. For two reasons:

⒈ Lower collateral: Crypto deal are often made at much higher leverage than traditional derivatives.

⒉ Market-maker incentives: most major exchanges now offer negative effective fees — that is, rebates exceeding 100% of paid fees — making it profitable to churn volume.

So, from a game theory perspective, in a “reasonable” world, you could easily expect crypto market turnover to be far, far beyond 20x. And right now, it’s directly comparable to Forex, instead of literally being at least another order of magnitude above it — similarly to how the derivatives market compares to the base assets one.

I have more thoughts on these numbers — especially from philosophical and moral standpoints — but I’ll save those for another post.

This post may be of interest to native English speakers.

When you hear, say, or write words like Bolshevik, Perestroika, or Tovarisch — do they carry a distinctly foreign flavor to you? I’m guessing they do, right?

It just occurred to me that for native Russian speakers, the word revolution triggers a very similar vibe.

In English it has two closely related but still quite distinct meanings.

One refers to physical rotation — the Earth revolving around the Sun, the revolver — the "revolving run", the revolving door in a mall or a business center, the RPMs (revolutions per minute) of an engine.

The other is about turnaround — the opposite of evolution, a rapid, dramatic change of state rather than a slow, gradual process. This is the socio-political sense of the word, where “change of state” works as a literal pun.

(Interestingly, in English, a socio-political revolution is almost the opposite of what’s commonly called a “revolving-door policy,” even though the word stems are the same. English can be delightfully confusing — though, to be fair, every language has its quirks.)

In Russian — and I suspect in many other Slavic languages — there’s a dedicated word for physical rotation. It shares a stem meaning “around,” and the concept of a “turnaround” would also be expressed as a single, different word. The Russian equivalent of revolution exists but carries a much narrower meaning: it’s almost always closer to “the opposite of evolution” and far more likely to evoke breakthroughs — sometimes scientific, but much more often socio-political.

Ironically, in the socio-political context, the Russian word for “turnaround” is effectively a direct synonym of revolution. But no one would ever use the Russian word for revolution to describe engine RPMs. For that, we’d use a form derived from “around” — literally something like “over-rounds.” The prefix pere- in Perestroika, by the way, is the same "re-" one: “rebuild,” “restructure,” “around again.”

Of course, revolution — much like evolution — has fully entered the Russian lexicon as a first-class-citizen word. And, just like Americans can use comradery in a totally local, context-free way, Russians can use revolution organically in native speech.

Still, I realized that for most Russian speakers — especially those who also speak English — hearing the word revolution in Russian often sparks an instinctive reaction: “Oh, that sounds foreign.”

I’m not sure how deep of an insight this really is, but I found it amusing. Not just because the word has two diverging meanings — but because the divergence itself is anchored in something bigger, capturing the difference between literal and physical on the one hand, and metaphorical and symbolic on the other.

This is alarming: https://www.youtube.com/watch?v=ZO5u3V6LJuM

Perhaps watch it before it's removed. Sigh.

I did not talk about this publicly before, but in a way, I predicted LLMs back in 2005.

Okay, this is a bit of a stretch — but let me unpack it so it makes sense.

We had a startup related to text categorization. The term Machine Learning was not yet mainstream; it was still called Data Mining and Information Retrieval / Extraction.

Bags of words and SVMs were king. And anything that could be convex-optimized into something that even loosely resembled a model “understanding” text was bleeding edge.

Back then, I fell in love with entropy models. It was a natural next step after compression — which, incidentally, was what that startup grew out of (oh, those parallels). The “insight” was that you didn’t need to perform the actual coding/decoding step, as long as the mathematical model was strong enough to compute the theoretical upper bound on how many bits it would take to “consume” and then “produce” a given text.

I vividly remember us drawing Markov chains with “clever” extensions — carrying state over — to affect the probability of generating the next token. Tokens were either letters or words back then, though we also thought about “semantic stemming” long before it was cool. Hello, future embeddings!

My bold prediction at the time was that eventually we would assume the existence of a massive corpus of texts, in all languages, that could be used to pre-train a model. And that this corpus — and the trained model — could be shared worldwide.

The task of “compressing” each text in that corpus would literally reduce to just its offset, since all the training data was memorized and packed in some clever way. The compression ratio of War and Peace, for example, would be extremely high, because it would already be contained verbatim in the dataset.

But here’s more. Suppose you wanted to “compress” a text similar in style to some cluster of texts already known. Say, a newly uncovered Tolstoy novel. My claim was that as long as the new text followed grammar and style, it would be very low-entropy to transmit from Alice to Bob. All we’d need was the right “initialization,” the seed — after which a similar text would be generated. Alongside the seed, we’d send “corrections” to steer the generation — not corrections in specific words, but adjustments to the inner state of the system, nudging it in the right direction.

I was convinced that such an extended Markov chain (or even a Petri net!) could be built. One that would implicitly and automatically capture not just the thesaurus and grammar (because: entropy), but also all possible styles of writing and all possible ways of expressing ideas. All in service of making it more low-entropy to “compress” human thought.

After all, natural language has evolved to be highly redundant. Entropy models simply exploit this redundancy — maximizing compression by distilling the “core information” and throwing away the rest, which is low-information noise that can be cheaply reintroduced later.

Well. Multi-layer neural networks already existed back then. What we’ve added since are attention and transformers.

And that, essentially, is the solution to the problem I imagined twenty years ago.

PS: I am deeply grateful to the folks I worked with back then. Being introduced to ML/AI in my early twenties turned out to be one of the best things in my life.

I want to coin the term “Reverse Instagram Effect.”

The Instagram Effect is real. I’m not young enough to experience it firsthand — I don’t even have an Instagram account — but here’s how I see it: people are far too eager to showcase the best parts of their lives.

Put simply, the content people publish is deliberately curated to look far more glamorous than their ordinary, mundane lives.

That has obvious downsides for mental health. For the consumers of content it feeds envy and longing — “wow, what a life; I want that too.” For the producers it breeds exhaustion and guilt — “I’m tired of pretending I live like this when really I’m struggling with X, Y, Z.”

(I’m no psychologist, but I hope that’s not too far off.)

Now: the Reverse Instagram Effect is in the very same content bias, but viewed through the lens of what AI models are trained on.

With text, even conversational speech tends to be treated with some generosity: speaking precisely is still valued. When AI trains on news articles, podcast transcripts, essays, it sees a lot of the styles we humans read every day. That exposure helps models approximate the kinds of language and arguments we actually use.

But images and videos are different.

Pleasantly positive: AI-generated images and videos are, bluntly speaking, enjoyable for humans to watch. That’s expected — the training corpora for visual media are heavily skewed toward content people already like to engage with.

Existentially positive: because the visual corpora are so biased toward polished, shareable content, models fail to learn what “ordinary life” actually looks and feels like. They don’t absorb the quieter, messier, unedited footage of everyday existence — and that’s a good thing for us as the species!

Imagine training an AI on continuous surveillance feeds. Imagine feeding it minute-by-minute biometric streams from people’s lives. The kinds of inferences and manipulations that could follow would be deeply alarming. If a model truly understood ordinary human behavior at that granularity, the risk of it being able to “get under our skin” would rise dramatically.

So far, we’re not handing models those datasets — for legal, compliance, and ethical reasons. Thank goodness. For a Skynet-style emergence you’d need models to be intimately familiar with ordinary human responses to ordinary stimuli. That level of access would make large-scale manipulation far more plausible.

Even excellent long-form writing hasn’t altered daily behavior at scale the way a viral slogan or a perfectly timed short message can. Short, amplified messages shape events. Thankfully, that kind of real-time, intimate behavioral feed is not what AIs are being fed today.

A deeper way to say this: text is only the outer shell of richer interpersonal dynamics. Even if an AI masters that shell, it will lack the insight into how people actually perceive and react to certain signals in context. And for visual media, the Instagram Effect — the polished, curated bias we dislike in people — inadvertently protects us by depriving models of the raw, ordinary data they’d need to weaponize influence at scale.

So yes: the Instagram Effect is bad for people. But as a structural quirk of our visual data, it’s oddly useful — a thin, accidental firewall between current AIs and the kind of intimate control we should all fear. For now, at least, we’re safer than we sometimes fear.

Published my thoughts on building high-throughput consensus to produce total order for events or commands: https://dimakorolev.substack.com/p/building-high-throughput-consensus

I generally try to avoid commenting on political and near-political topics, but the H-1B visa controversy sounds more like social news thank political news to me.

What surprises me most is that after spending ~20 minutes scrolling through the news and comments and feeds, I have seen ZERO mentions of Canada and Australia.

The UK and EU are there. I think I even saw Dubai / UAE.

But somehow the sentiment is disproportionally that "the USA is letting us down big time".

Folks, moral judgement aside — and yes, I am speaking from the position of privilege, which I do acknowledge — if wannabe-H-1B talent is such a powerful force, shouldn't we expect the next Silicon Valley to emerge in Vancouver or Melbourne?

By first-order logic of everyone commenting it totally should. Because if it's the US that is presumably the biggest loser in this game, some other major country may well pick up the tab and reap the benefits.

And yet — again, ZERO mentions of Canada and Australia.

I, for one, am finding it suspicious.

PS: For the record, my take was and remains that in the past ~5 years we're seeing the world becoming more sane, that is, from the game theory standpoint. There were major imbalances, and a lot of institutions and state actors have abused the status quo one way or another.

This is now changing. Many, if not most, short-term changes will feel negative in the moment.

Long term though, I'm all for sustainable progress and innovation, through effective monetary and human resources allocation, which works best under competitive capitalism.

If the end result of this H-1B situation is that three months from now the UK, EU, Dubai, Canada, Australia, plus perhaps Thailand, Japan, and Argentina begin competing to be the prime candidate for "Silicon Valley 2.0" — that's a great thing to my taste!

(And if you believe this is unlikely to happen, then let's first acknowledge that the "original" Silicon Valley does have an unfair advantage that it has been routinely under-capitalizing on — and this is what's being corrected as we speak.)

Although my mid-term prediction is that it's all power games. Ultimately, the H-1B visa was indeed very much abused. What we need is to lower the threshold on the talent visa (O-1, EB1-1), and make sure the fees are reasonably high — to close the loopholes of the like Infosys is abusing, to keep Google et. al. effective, and to make properly venture-backed startups want to innovate in the US. All of this can be done, and something tells me it will get done soon-ish.

The term sound money is well known these days. Fiscal conservatives, libertarians, and the Web3 crowd are equally fond of it. The idea is simple: every unit of money belongs to someone and is backed by something tangible. Printing money is impossible, and every loan has a consenting party consciously taking the risk.

Related Web3 concepts are staking, bounty hunting, and slashing. To participate as a miner — thus contributing to the ledger and earning rewards — one must post a meaningful stake as collateral. If they cheat or fail their SLA, others will notice. Cheaters are penalized by losing part of their stake — that’s slashing. Whoever proves this cheating earns a fraction of the slashed stake — that’s bounty-huting.

Transparency and auditability make this work. Transparency doesn’t mean every message is readable; it only means the system continuously publishes verifiable checksums of its own state, making retroactive edits impossible or prohibitively costly.

I’d like to introduce the term sound software.

Software is sound when the service provider is contractually committed to its users — and when provable misbehavior leads to compensation for whoever caught it. Users and platforms should have aligned, verifiable incentives to keep each other honest.

Consider a social network. Publishing a post yields a proof. Followers are guaranteed to receive that post in their feeds — filtering and ranking happen on the client side.

If the author deletes or edits the post, there’s public proof.

If moderation removes or flags the post, there’s public proof — traceable to a moderator, not necessarily by name, but auditable.

If a court order requires removal, there’s proof of that too — the order ID and a permalink. If the case is sealed, disclosure follows later; but eventually everyone can verify the order existed and demanded removal.

If a user fails to receive a post they should have, they can present a compact proof: { A posted X, B follows A, B requested a feed update, X was not delivered }.

Upon such a proof, the platform must present a counter-proof or have its stake slashed. If there is no counter-proof, the detector is rewarded.

This resolves shadowbanning — for good. No more tampering with, well, anything.

Wouldn’t it be great to live in a world of sound software?

Also, competition matters. I don’t want companies forced into this; I want enough people to prefer it — and then let the market do its job.

Much like Bitcoin today. In many places, the ban era is over. Barring a major hack — likely by state actors or a Big Tech supply-chain compromise — Bitcoin, as sound money, competes with fiat that’s prone to printing and helicoptering. In that market, more people choose Bitcoin. That looks right to me.

I’d love to see the same for every online service — from messengers and social platforms to driver’s licenses, plates, and travel documents.

Unlike sound money, where Satoshi arrived like Prometheus, sound software probably won’t emerge spontaneously. The current giants are powerful and protective of their oligopoly. But history has reversals: unsustainable systems collapse, and the rebuilders choose better foundations.

So I don’t expect the US, EU, or UK to lead here soon — the status quo is too strong. A country under pressure might adopt it first; a place that needs credibility and openness could make the bold move.

We need more free communication platforms that cannot be shuttered by a single state actor — and cannot quietly align with one powerful owner’s narrative. That’s the bar for sound software.

Perhaps the future role of humans in the AI- and robotics-powered economy is that of weird scientists from the past?

It has already assumed the role of shamans. The next step is to learn "directional creativity", also known as science and engineering. Hear me out.

AI models as we know them today are good at operating within the environment they were trained on — AlphaGo is a case in point. They are doubly good when they can simulate their environment such that the objective function is evaluated correctly — behold AlphaZero.

They are increasingly good when they a) can operate in a real environment, which is nearly impossible to simulate, b) are in the space of extrapolation and linear optimization, not prolongation, and c) the objective function is still evaluated correctly. This is modern-day, "post-Boston-Dynamics" robotics.

When it's about, say, walking over rough terrain without falling by learning to orchestrate a bunch of "muscles" the right way — there's nothing better than modern-day neural networks. In fact, nature first evolved neural networks for exactly this purpose.

Going off on a tangent a bit, my personal belief for what's big to come is learning ultra-low-power-consumption models for robotics. Because, mathematically speaking, we know it is possible to, say, walk straight by applying a very small amount of very high-lag, low-throughput "computation". But I digress.

So, of the three points above, (a) is relatively straightforward — already done in 100% simulated environments, about to be done with general-purpose robots. The last one, (c), is also somewhat clear, since if we're talking about, say, manufacturing, the QA processes are already well established — or they will be, inevitably, since we're not letting humans fly on planes assembled by robots without thoroughly checking them through.

Thus, (b), remaining in the realm of extrapolation, not prolongation — that's where modern-day AI hits the wall.

Simply put, it lacks creativity. The ability to invent. To not just "hallucinate and filter", but to do so "in the right direction".

That's how human genius and prodigy look. For the record, I am not discounting the possibility that we, humankind, stumble upon a way to make "our AI" this smart. I just don't think it's likely. Much like it's not likely for a remote human tribe to invent / discover supersonic travel. The shamans of that tribe have done marvelous work researching medicine and poisons and the altered state of human consciousness — induced and controlled. Good. That's what I refer to as "wandering creativity", a.k.a. hallucination and filtering. This is what modern-day AI models can already do well.

The next frontier is "directional creativity". And that's exactly the role a human can play.

Just imagine. A factory where tons of robots are building stuff. Humanoid and whatnot. From very small to very large. And the factory is actually an open-air setup of vast dimensions, where the area includes everything, from resources to mine, to the test-flight polygon.

These robots can somewhat make sense of what they're doing. They have a lot of CPU/GPU/*PU, after all, individually and collectively.

They will also very likely report to some human-controlled entity. At least for now. They lack agency, to begin with. And I'd even argue that "learning" this "directional creativity" is the necessary precondition to agency.

And this is exactly where the human comes in handy. To give this sense of direction.

And it's not a brutal dictator who wants to build an army and destroy the world. In fact, such a human would very likely fail at the task. Because other "AI-managing" humans would, in fact, invest in science and R&D, and, as a side effect, they would be able to quickly produce what it takes to stop whatever "army" that dictator could unleash on them.

The sense of direction will have to be more scientific in nature. Driven by curiosity. And fueled by our, human, desire to learn more about the universe around.

[ ... part 2 below ... ]

[ .. part 1 above ... ]

Including, of course, our own planet, and our very, human, role in it. It's not just technocracy and space exploration I'm talking about. It's also magnificent cities where humans thrive, and want to live.

If only — IF ONLY! — for the sole purpose of finding those future one-in-a-billion human beings who would take this craft of AI-assisted exploration of the universe and human nature further.

Sure, this will eventually lead to human extinction. Since in a few "generations", the next prodigy-in-residence would be far too psychopathic, and would actually want to "rule" the world, with an iron fist. Well, first, that's a risk we all have to take. Second, hopefully, enough human-crewed space vessels will be en route to other parts of the Galaxy by that time. And third — and I believe the most likely scenario — is simply that "sustaining" humans "in their natural habitat" would be as cheap and as fun as it is to sustain many animals today.

Think of how the colonizers treated Native Americans a couple hundred years ago vs. now. And think of cruel animal performances in zoos centuries and even decades ago, compared to wildlife today actively seeking assistance from humans in the more developed parts of our civilization, where empathy does thrive. Something like this may well happen once the "chief scientist in charge" assumes powers that are beyond what we have been able to imagine thus far.

There's also a techno-libertarian take to this. If there are multiple isolated parts of the planet — much like what we're seeing now — there will be multiple "chief scientists". And, as I like to repeat, competition that does not turn into oligopoly does fuel progress big time. The AI Cold War, if it were to happen, would very likely be extremely fruitful for most ordinary citizens, much like the first Cold War was. Sure, it may elevate our chance of destroying ourselves — but we did have those moments in the past, ref. the Cuban Missile Crisis; and between going planet-wide Amish and living our human lives in a "pro-progress" way, I'd definitely pick the latter.

So, to conclude. I do believe the engineering and scientific mindset is the most important one to possess these days. Since, from first principles, this mindset is what is most logical to develop.

If anything, it's a nice model of my world which I choose to live in. It's also a healthy — I think — coping mechanism to deal with the miserable fact that the control over wealth and technology these days is ... far from ideal, on this planet, and I am increasingly saddened to realize that this state of affairs a) is unlikely to change any time soon, and b) would likely change for the worse if it were to change.

Therefore, I conclude that my best course of action in navigating this life is to remain a techno-optimist and techno-accelerationist, who can prove himself useful at making things happen when and if we do enter the acceleration phase.

Recently I wrote about how I “predicted” LLMs circa 2005. Now I want to touch on how I suggested vector search circa 2012.

TL;DR: Embeddings were taking shape. Very early, in fact — word2vec and GloVe weren’t there yet — but I was at Google and then Microsoft, trying to see where the wind was blowing. So I argued we should build a search engine on well, meaning, not words.

The algorithm I proposed looks trivial in hindsight; whether it was revolutionary is not for me to judge.

Take the embeddings — or whatever we called them then. Apply them to smaller chunks of content: paragraphs for Web search, or sessions of user preferences for feed ranking. These embeddings should already be roughly orthogonal; if not, reduce dimensionality with PCA.

Then select a few random orthogonal vectors, say N = 42. Think hyperplanes that split the “document” space into two parts, “positive” and “negative.” Each hyperplane yields a bit: 0 for negative, 1 for positive. Now each document is an N-bit binary vector.

The Hamming distance — the number of differing bits — becomes a proxy for semantic similarity.

Next, use a meet-in-the-middle trick to traverse the space quickly. Start from a query, which can be one or several points. When building the index, for each point add: itself, everything 1, 2, and 3 bit flips away — not too much bloat. At query time, replace subsets of one / two / three / four bits with a mask (“this bit doesn’t matter”). Final step: check intersections.

Voilà. The intersection is the candidate set; then you rank within it. The Hamming distance itself is also a useful signal.

The example N = 42 can be tuned. For up to five or six “bits of separation,” I estimated a couple hundred hits. Too many? Increase N. Too few? Decrease.

You can also adjust the “bloat factor” and “search depth.” My thinking then: keep bloat small to control index size, keep depth small for latency. If you need more hits, the natural lever is lowering N.

This scales almost perfectly. It’s easy to shard the index into near-disjoint parts with minimal overlap, so most queries hit very few — sometimes just one — node.

Ironically, and unrelated to my name, Yandex rolled out its Korolev engine in 2017. That was probably the first large-scale ranker showing results with zero term overlap. As long as the semantic match was there — and the page was good — “word similarity,” even with synonyms and stemming, became obsolete that day.

We’ve since moved on, as GPUs and massive parallelism took over. Still, from a linear-algebra perspective, what I suggested was a simplified version of how vector databases query today.

Granted, I approached it from the implementation side — which I often do. Mathematically, it was random hyperplane projections and Hamming-space neighborhoods — a pragmatic take on approximate nearest-neighbor search, where the ultimate metric is the same cosine similarity vector DBs use today.

Unlike my previous “retroactive prediction,” though, this one has a follow-up: a prediction for the future.

I believe we need to focus on less computationally expensive ML/AI models.

From economics, of course we’ll keep burning more GPU power. It’s ROI-positive, winner-takes-all, so no sane actor stops.

From intellect, though, we’re simulating far more “thinking” than the entire human race — even the animal kingdom — combined. We’ve reached the plateau.

Instead, imagine algorithms that compete with each other — and humans — at Chess and Go (and StarCraft). On Arduino. Use massive LLMs to evolve better algorithms, sure. But the actual programs should be trivial enough to move within maybe thousands — or even hundreds — of CPU cycles. Likely a mix of pattern matching and alpha-beta-pruned search. That’s how expectation-maximization should converge.

Then the real challenge: how cheap and tiny can hardware be, yet reach Grandmaster in Chess? And how few watts can it use?

Solve this, and we’ll be better equipped for problems that truly require intelligence. Cracking physics. Solving energy, food, and housing crises.

Did Durov and Lex talk monetization?

From what I’ve seen, Durov once promised “no promoted channels, ever” — which is clearly no longer true. So either there’s an honest conversation about breaking promises, or the phrase “we made a promise to our users” carries no weight. Did Lex ask? Did Pavel answer?

For the record: I’m fine with Telegram as a product. But it is neither decentralized nor secure.

Security

A real “secure Telegram” would let users bring their own encryption devices. Just like you can set a proxy, why not set an “encryption router”? Some would use secure chips, others a Raspberry Pi running OpenSSL. The idea: only the recipient’s device can decrypt. Same reliability as tokens on a blockchain — zero trust in intermediaries. I’d gladly pay a few dollars a month for that.

Decentralization

A step further: let Telegram run entirely on private clusters, offline from the Internet. Not free, but maybe $25/month for a frozen version that only talks to my own servers.

Reality check

Of course, Telegram is a business. And businesses work with regulators, not against them. Blockchain shows it’s possible to give regulators the finger and survive — but I don’t expect Telegram to go that far.

Web3 bridge

Today, Web3 fees are already miniscule. Sending a short message securely could cost $0.001 on-chain. Slower and pricier than centralized apps but worth it for true privacy. Telegram doesn’t have to become that service, but it could at least be friendly to those that are.

Transparency

Telegram does publish some aggregate stats on government requests — good, but insufficient. Imagine if all such requests were logged on-chain, with IDs, aggregates in a month, full texts after five years. That would make “privacy first” more than marketing.

Censorship

Same with group removals: if Telegram refuses a takedown, those groups should be listed publicly. History is clear: those who censor at scale never win long-term. Content banned in one country should remain visible elsewhere — just like laws differ across borders.

Bottom line

Telegram today feels no more private than Slack or Discord. Any authority with minimal paperwork can get in. Unless it embraces real user-controlled encryption, decentralization, and transparent court-order logging, it won’t qualify as a “privacy first” platform.

PS: I'm growing into using AI to shorten posts so that they fit the limit.

Just realized we did pass a major milestone recently, and somehow it went largely unnoticed.

We used to compare blockchain payments to Visa payments under the assumption that Visa is a huge player, and crypto "could be promising", "has a change to grow from its minuscule market share today" etc.

These days we are talking about Web3-operated and Visa-operated payments as equal partners.

I just stumbled upon a piece where the author boasts (!) that Visa can be "almost on par" with blockchain-native payments.

It took 10+ years, but it's the future. We are indeed living it.

Perhaps it's now safe to say that Visa cards will be retired in some 10+ years?

As in, they will live as legacy for God knows how long, we're still seeing Americans paying groceries with paper checks sometimes.

By retiring visa I mean we could be living in a world where the whole family of payment protocols is decentralized by design. So if some region or business decides to launch "their own payment method", it will be blockchain-based.

And our smartphones and smart watches will not "support Apple Pay", but support Web3-first means of payment.

So that teenagers boasting that they "hacked their PlayStation to pay for bus rides" will be the norm, not the exception.

(My fingers almost wrote "hacked their GameBoy" about teenagers, that's age showing.)

However, hear me out. It is becoming realistic as we speak. Some ten years from now, for payments under, say, $50, such as coffee, setting up the bridge for a newly created payment method would be common knowledge.

Not only visa and banks. Products such as Venmo or Zelle will be a thing of the past.

They may still exist as consumer products and as gateways for regulators and tax forms — one can not, for instance, run a business by paying their "employees" via Zelle.

But when it comes to actual payments, those projects should be out of business. There'll be cash, but not all places would accept it. There will be credit cards and Apple Pay and what not. And there will be stablecoins, or whatever acts as stablecoins in 10+ years.

Many businesses today still add an extra % for credit card payments. Because card payments are slow to hit one's bank account, and since an attempt to revert the transaction can be made, requesting the business to do extra paperwork.

Businesses in 10+ years may prefer stablecoin payments, since they are instant, ultra low-fee, and friendly with regulators either way.

As in, since hopefully most developed countries will live in a world where tips are tax-free as long as they only constitute a small fraction of income of the individual or business. So that a lot of average John-s and Jane-s will not pay cash for their haircut or coffee — they will pay with their own Apple-Pay-like device, indistinguishable from cash for all intents and purposes.

For small transactions, people who know each other well enough would just not bother.

And this will be a brave new world taxation-wise. Hopefully, we could get back to a model where the business can choose to be taxed based on size, not transaction volume.

As in, a coffee shop owner may just file their taxes based on the number of tables, not coffees they make. Reasonable counties should appreciate this approach — if anything, with comparable tax revenue, it incentivizes businesses to serve more customers, better.

Although I’m afraid this may not end well. More governments are struggling to justify their revenue models as we speak. So they may well perceive further proliferation of Web3-first payment methods as an existential threat — which it is for those who refuse to serve people well and instead optimize for KPIs the public doesn’t endorse.

Time will tell. All I'm saying is it'd be nice to have inter-governmental competition at this level, instead of wars and tariffs.

And I, for one, find it easier to envision myself in a place where the government takes this Web3-first signal as a call to become more effective — not as a threat to its existence, since effectiveness has long been off this government's menu.

What’s the Most Lightweight Blockchain to Run In-House with Fast Finality?

This question probably needs some context. First, it’s for our weekend project — something that deals with consistency and durability. Second, it’s about being lightweight to set up and run, not about the logic itself.

Another way to phrase it: Is there something like etcd or SQLite, but designed for Byzantine consensus — ideally in Rust?

In fact, etcd itself would work quite well, and its API is comfortable to use. Even a simpler API would suffice: just an atomic swap of values for given keys, as long as the current bodies and versions of those keys match what the request specifies. Add passthrough mutation epoch numbering, a majority consensus among nodes, and a small data footprint (say, under ~1GB total, in-memory is fine).

The only thing etcd lacks for this particular use case is request signing with the nodes’ private keys.

The other missing piece is that etcd’s API is too tightly coupled to etcd itself. Eventually, my code will need to run across data centers, where nodes join and leave dynamically. It will be configured differently. Maybe even hosted on Substrate someday. Maybe even anchored to a major chain like Ethereum for proofs.

But I want the API to stay the same — so that the client-side code using this consensus engine doesn’t need to change. I might need to rebuild it with a new client library that checks different signatures upon commit, but it should be the same code.

So: is there a quick tool like this? Or a protocol? Or even just a whitepaper or blog post describing something similar?

I could hack it together in half a day on top of etcd, and then another half day over Substrate. But it feels like something this simple and useful must already exist somewhere.

Thanks in advance for any pointers!