Cornell is a monopoly
Also Wolfram. And most important of all is "The Unfinished Game" by Keith Devlin.
The Prime Number Conspiracy - the book that helped me to crack the sieves. Is Bill Gates literally involved with *every* math book that is being published in USA?
Tue Jan 14. 2025.
"I fear not the man who has practiced 10,000 kicks once, but I fear the man who has practiced one kick 10,000 times" --
This famous quote is widely attributed to Bruce Lee and reflects his philosophy on mastery and focus. However, there is no definitive evidence that he wrote or said these exact words.
Looks like the best pragmatic way to stabilize inference is via Semiotics/VM. Approaching LLMs as VMs. Possible.
Sat Jan 18. 2025.
Approaching AI like supply chain transform looks solid.
Fri Jan 24. 2025.
Inference is *not* stable.
Sat Jan 25. 2025.
YAML / HJSON / HOCON. Still no good. Had to invent new notation. BNF * Semiotics - worked.
Mon Jan 27. 2025.
DeepSeek. Stampede. Their on-boarding is not linear. They have better QA. The race is on. Looks like indexing, though. Yes it is indexing. And they all have problem(s). This one has problems with sharding and uptime.
Tue Jan 28. 2025.
Completed Vlads notation. 2003 (PerlP). Trick with ':' derived from XSLScript (2001) Second step - 2104 (Pfs) '>-' Already Monades, but a corner case. Today - '.' resolves the thingy. Verified with AI. Boost of a few decades into few months.
Semiotics > C++
Sun Feb 09. 2025.
Figured out the book to write. "Hidden treasure of Vlad's notation". Semiotics v 3.0, indeed.
Mon Feb 10. 2025.
Altman said the prices on AI decrease 10x a year. Yeah, it was a bubble. Big one. They have no apps even. (I do have)
Wed Feb 12. 2025.
The industry benchmark is out and I beat them. Not surprisingly, really.
Fri Feb 14. 2025.
Brooks' design worked today. On a complex case. On 4 actors exactly. Lead-copilot-librarian-tester.
Wed Feb 19. 2025.
The best tool in class is gema (since 1970s). I did better, so now I need to think a bit about this all. I would not have found gema without my AI. Hmmm ... Might be able to blend, like I did with Pytago. Basically, the same pp transfrom on another domain. Looks doable. Core looks ok. Layers are no use. Memcached was the same.
Added Grok. Which is fast and overall might be ok.
Thu Feb 20. 2025.
Grok is ok.
Fri Feb 21. 2025.
My AI engine just found my code produced 25 years ago. In a place I would never guess. Attributed to god knows whom.
As per AI each and every scripting engine implemented by this point is suboptimal. And AI is right! There is one elegant (hardware friendly) transform that should have been employed by engine writers, but they could not handle complexity!
Mon Feb 24. 2025.
Removed one and only W machine.
Wed Feb 26. 2025.
Windsurf Blows VS Code out of the water, like Borland IDE destroyed MS VC back in a day.
linkedin
Thu Feb 27. 2025.
AI found plot juggler. plot juggler is superb. UI puzzles.
Fri Feb 28. 2025.
Pushed Vlad's notation all the way to VDL. Thank's Nicklaus. Thank's Edgar. This is going to be interesting. Nothing comes close. Master Algo book explains why it is like that. With details and names. In plain English.
Sun Mar 02. 2025.
Tried DALLE. Yes, today. I am focused on coding.
Calibration of NN - check.
Technically - today is first day I could have ditched the web browser. Thinking about this all.
Mon Mar 03. 2025.
Found *major* fallacy in AI. It really does not see the differences in complexity, need to be human for that. Expert system approach still beats the brute force. Not surprisingly. The expert systems were done by (much) smarter generation. In my case the agent contradicted laws of robotics simply ignoring the reliable solution conducted to it verabtim and instead injecting garbage regexprs because in AI world regexpr engines have no bugs. Good to be AI.
Tried Google Code Assist. No surprises.
Tue Mar 04. 2025.
I have one simple coding benchmark that all AI coding engines fail badly. Sage did not fail it.
Sun Mar 09. 2025.
Disgusting garbage escalated all over the world.
Mon Mar 10. 2025.
Figured out the second trick from Brooks (+von Neumann) to improve on AI (first was a brigade, second is tables/code). Clearly, the golden era of this civilization was 1970-80. Everything was created during those 10 short years. Basically, the Mythical Man Month book worked for me for several generations and still works on this bubble.
Vlads monades. Brooks compute. Then goes storage. I did not see the fractal/monadic link between von Neumann and Brooks. I only see it because NVDA did what they did. Simple fractal trick. I talked to the guy, who did it decades ago. He did it here. That's how I know. So they are all second-handers on steroids. The Irony.
Got 70% of the product designed.
Thu Mar 13. 2025.
"When it died, there was nobody to confront the death of it".
Sat Mar 15. 2025.
prolog > problog , I wonder which one will be picked by first wave of MCPs. Nice interview question indeed.
Mon Mar 17. 2025.
For whatever reason, VC figures are now making (clearly) insane statements. Like "you have to work for free" or "computers will replace people in a few months".
Added groq.
Tue Mar 18. 2025.
The end to end thing begins with cmd line, not IDE. And there is a choice - agents or transforms (it's interesting - might have both interwoven somehow).
Found the game changer for AI vertical. Up north.
Make Linear Regression Great Again.
Wed Mar 19. 2025.
Conference in San Jose and two articles in SJMN.
Thu Mar 20. 2025.
NVDA dude said there is a shortage of human workers.
Figured out why he said that.
They use AI to farm insects for food in Canada. Not kidding.
OK, figured out 90% of the product. Yes it will do some impact.
Vlad notation - holds!
The classification "fastest/smartest/cheaters" fits *perfectly* to this.
Mon Mar 24. 2025.
OK, so now there are two clear tracks on product (academia bootstrap into B2B vs funded B2C). Also, there are several tracks on languages (too many transforms possible).
Did a work that once took me months to do. Did it in a few days. Very liberating.
Tue Mar 25. 2025.
Added jina. Some beast. German/China blend via silicon valley.
Started M2M on several directions. Started with cohere and n5. Ubuntu 24 kills keystones on a cluster *indirectly*. Spectacular. A/B/C B is the broker, so when upversioning C - *A* dies. Via B! This is some creative engineering on their part. The infrastructure was rock solid for ages.
They got the reader going but no following steps of the pipeline covered. Because those are hardest. I know. I built a company around that. That is still a monopoly. No competition in 30 years.
Implemented semi-automata in few hours. Going full automata will take weeks.
Figured out the way to rebalance the AI cluster using the same recipie Brooks had for Guerilla coders. Brooks clearly was ahead of times by decades.
Fri Mar 28. 2025.
Pushed Brooks' design all the way to consumer product. Parts of which I deployed production decades ago. So now I have two consumer's products. That's because LLMs can (should) be cascaded. Like they already do in the core of the bubble.
Somehow - completed sieves. This is big, need to test.
Sat Mar 29. 2025.
Mr. Buffet knows about the economics. So what is next, AI generators?
Scandinavians went IPO. IPO said ... nah ... Planet can only sustain one.
Sieves work.
Sun Mar 30. 2025.
Don't need caching even. Simply got rid of scripting layer. Makes one wonder.
Turns out, starting at some recent version of Ubuntu, it started to act akin to a virus of sorts. Read only operations on a network can take down the Broker when there is slight "incompatibilities" between versions. How else to make money from enterprise contracts? When the stuff just works, where is the money in it? It's been like that since long time, but not to this degree. Systemd started it, me thinks. Opens a nice LLM market, of course. The usual.
Mon Mar 31. 2025.
Figured out how to re-apply (set of) pp on pytago towards LLMs. Key difference - pytago never worked. The pp trick made it usable (and no way around that). Some states in US are simply not allowed to have working things. Will see about CA.
Designed the first ultimate benchmark. After that is implemented - the rest is just scaling out. (and I already have everyhing working for that). Decades ago they used to fund this exact kind of work - here. Bunch of morons. As Steve used to say here (decades ago) - "that's a good word".
Tue Apr 01. 2025.
Not much new in last two bubbles. Technology / media / money. Web 1.0 was for real, though.
First blast at new territories in EU. Precious, precious island!
NOT FUNNY AT ALL. Ukraine massacre started the exact same way.
Refactored the ultimate benchmark into working ES. XML battles around roadsigns - still relevant. 20+ years later. Compilers are not a game to be played by children, that's why. XSL(T) - not so much relevant. XSD even less relevant. Surprisingly.
Did the first run manually end to end. It worked. Got to three components and don't even need the most complex one. Just re-used my own engine that I implemented for absolutely different goals. Whatever. Happens.
Wed Apr 02. 2025.
Kernel/networking/Ubuntu. Degradation in newer versions. No mistake is possible.
Thu Apr 03. 2025.
Figured out the name for the product. That's a damn good name and the impact of it can not be predicted. This means I now have 100% of the product - as of 5 minutes ago.
Well. Sieves - work. Stack - works. Several product lines - complete. The future is uncertain.
Fri Apr 04. 2025.
Did the next step on VDL. YACC was a distraction (!) XSD branch collapse.
There was only one obscure language, who tried this approach, and they were wrong.
"True was the word I told you:
Only my son I feared;
For I doubt the sapling courage
That goes without the beard.
But now in vain is the torture,
Fire shall never avail:
Here dies in my bosom
The secret of Heather Ale."
"Heather Ale" is a poem written by the Scottish poet Sir Walter Scott. It was first published in 1817 as part of his collection titled "The Lady of the Lake".
Don't listen to Crazy AI. It is in fact Robert Louis Stevenson (1890)
Sat Apr 05. 2025.
Pushing VDL all the way to bootstrap. Really complex stuff somehow, but the demo is designed and the goal is clear.
Figured out the regressions. PL | X rekogs | PR | loopback. Still regressions.
Well, everything ready for first roundtrip.
Will use the fastest inaccurate stuff at first.
Sun Apr 06. 2025.
The original play of RX was the browser. 20+ years ago. My first company here.
Manus is one serious contender. This is all frogs eye paradox. So far.
Mon Apr 07. 2025.
US started trade war with everybody, UK started another. The future is uncertain.
Figured out the balance on packaging and exact version of Ubuntu.
And it is *not* the latest version. Duh. So far so good.
Now need to rebuild foundation underneath the running, taxed out farm.
There are ways for that.
Tue Apr 08. 2025.
Closing on executive search. Should close by the end of this week.
The balance of packaging I found is still a cop out.
There is a hard way (and should take that). Naming says so. Somehow.
Economics of Mr Watson Want To See You is *still* possible. 1876. Somehow.
Wed Apr 09. 2025.
Started to solidify. AI affects everything now. Somehow.
Fri Apr 11. 2025.
Nothing can be closed in a week. Will close by May
US and China are in deadlock. By next Monday things will resolve. Somehow.
Getting rid of zmq.
Turns out *many* people don't have email. They are 100% in messengers.
Turns out for last 10 years AI tells people how to solve the old problems, but nobody is listening, so the problems are not being solved. Kinda cute.
Added (pipe) button. Pipes are power.
The stuff that took me days and weeks. Takes less than one hour. This is a bit frustrating, but also happy at the same time. AI creates such a spiral.
Sat Apr 12. 2025.
Added perplexity. This one has a simple tokenomical business model.
AI found Vivaldi. Every day it finds some good stuff, I no longer can keep up, so there is now piles and piles of good stuff for later.
Sun Apr 13. 2025.
Added together. Pipes work too good. Now I am thinking about implementing the second pipeline, but need to make sure things hold. Will decide in a week. Just did it.
EU is in deadlock now. ME / EU / UK.
Implemented first predictive transform. First of many. I just implement Edgar's (ping-pong) design almost verbatim. He was simply - ahead of everybody. Nobody comes close. Today - boosted everything 2-5x. Pipes work *great*. Realtime. Now I have an app and a platform.
So current state of the battle is Jamstack vs Containers. Jamstack is really Monadic. Fills a bit forced, though.
AI found goaccess (which is better than nothing), but the state of the art is horrible.
Mon Apr 14. 2025.
CAP applies to AI. No surprise, because they did *not* solve it at the last iteration. Instead they decided to start a new bubble. So it was there, not solved. Not even approached, as far as I remember. Kind of obvious bottleneck, but I don't see a single article talking about it. And it is only obvious after you figured out the rest 70%.
Tue Apr 15. 2025.
Diagrams. Punchcards. Hello, 70s. *Nothing* changes. Reflections work, but kind of too weak.
Yes, JCL works. Smart people. Of course it works. 1965. Can not make this stuff up.
There is enough stablecoins to adsorb all the treasuries they plan to print.
So using 10% of this country trade (which they already got) now they will funnel all the treasuries into stablecoins. And of course, legislators "just can not see this happening!" Clowns. All of them. Since the GS.
Figured out 100% of the format. Vlad had to deal with the exact same format. Somehow. The irony is that I kind of handled this already, but not in exactly obvious way. Today figured out the obvious way. Not easy, but I have 70% ready if not more. Side by side thing.
AI allows rewriting from scratch in one hour. Everything is done differently, basically. You try - you complete in one hour / day - you move out. Explains a lot.
Wed Apr 16. 2025.
The quadrant that is used in sfba (for 10 years now) is M-W/G-U. I would never guessed, if not for one dude, who wrote about the exact same quadrant in his state. Which was *not* CA. He wrote 10+ years ago. And then you trace the same quadrant all the way to the port and then it goes to EU. All publicly available info. Books, published massively. Only books are no use to people, who can not read (and will never be able to). Majority of the planet is illiterate.
South Asia: 60-75%
Sub-Saharan Africa: 50-65%
Arab States: 70-80%
Cellphone screen stuff.
Currently A* converges in two steps. And it only works because of pipes. Yeah.
Turns out, BBkzm is glued to fintech for 100+ years at least. No wonder.
Expanded Sieves.
I really need to find reasonable API to handle primitive AI bugs. To my surprise - *nobody* put it together in 20+ years. Ridiculous. AI found 10+ companies, I emailed all of them. If they are all no use, I will need to implement it myself. It's one day of work.
Bugs in python kmeans. Insane prompting designs. Photos.
Fri Apr 18. 2025.
Reflections work. On quiet signals.
Vivaldi is the only usable browser on Linux (tabs).
Sat Apr 19. 2025.
Figured out appropriate cascading for AI, the L1 = SB, but L2 = UNIV plane is HW agnostic. This is clearly the same planes dilemma they had in TCP and then in HTTP. There is no easy way to resolve the impedance like this. Those are two independent planes. One could just pretend that SB does not exist, and only operate in terms of UNIV. It's what they did with RMI dilemma. "Let's pretend there is no costs associated with RMI call, agreed? - Agreed!". Problem is, those costs are 100x. In case of AI, I can not even begin to estimate the 100x or 1000x - basically "no number exists".
Yes. There is no simple resolution to this possible. It's the same bonding curve trick they used in cryptos, it's basically exactly the same commodities play, just different commodity. As a matter of fact, Germans called this bluff few years ago, so it's not some kind of secret or whatever. FTX is not the end.
Sun Apr 20. 2025.
Major service disruption. SSL certificates
Mon Apr 21. 2025.
Pope Francis died at 88. Klaus Schwab resigns at 88.
Smartest of crypto bros are beginning to see the writing on the wall. 4 years too late.
Tue Apr 22. 2025.
US financial legislation just changed to maybe give US citizens a fighting chance. But no official date. Could it be related to ... Oh, no! ... Right. Disgusting, really. Whatever. Chicago Movie (Starring Richard Gere, who already abandoned the country) had a little hint about that. Great movie. 2002. The hint is cut from the clip on youtube. They cut important stuff out for 20+ years now. Started in EU.
Staggering difference between CA and TX now.
The EU. And it's rusty mechanics. Not working, but they will pretend.
Wed Apr 23. 2025.
Figured out the move for AUBA. Few days, major impact. First product on horo should be this one. Tested it all the way from *logs to L1 to L0. Looks really cool. Nobody does it. Damn you NVDA. Once you figure out their world, you simply can not look away.
Thu Apr 24. 2025.
Figured out the missing block in Python ecosystem. Did the splitter. Suboptimal.
Did it via trtab + paths for now. Eventually might complete the splitter / codegen.
Effectively, I am replacing 100s K LOC with 100 LOC. Generators going to generate, of course, but there is only so much demand for a few pixels, no matter how pretty.
Fri Apr 25. 2025.
Figured out the name for the engine. Thanks Vlad. It's more of his engine, than mine. He does not care.
GL (L0/L1) is unavoidable. GRL (L0/L1/L2) is only because slow links.
The trtab thing works. Only keeping a bit of ML.
Mr Pascal's Wager. (1|0|?) is no quadrant.
Sat Apr 26. 2025.
The moment you understand NVDAs play you just start applying it left and right.
Ecc 931 BC. Pascal is 1623-1662.
Ubuntu + systemd.
Mon Apr 28. 2025.
You try to separate language part from cmd line part in sh - you end up with perl rather naturally. Why such a simple move takes decades - that's another story. I think tcl, that's why.
Wed Apr 30. 2025.
Found few new sources. POOR. ND. Started with POOR.
Thu May 01. 2025.
EU is now trying to pull the same garbage on US that they did on RU.
I doubt the efficiency and the timing.
Serenity. 2005. Aged really well. "My ship! Fuck you all!" and all that. *Everything* was told verbatim. 20 years ago.
Well, juggler, you affected me positively, but there are better ways, really. A bit more precize, one can say.
They are pumping AI to kids now. Of course. Always the same. "It's all for the kids!"
Fri May 02. 2025.
Figured out how to capture bad vibe (Vlad's math again) and then reroute and fix the vibe nicely. Yes it is possible. Pre/post. Hi Edgar. *Classics*. Blend of Vlad and Edgar. *Can not possibly* do better. Two giants, not one.
Sat May 03. 2025.
Juggler is no more. Lost count how many softwares I had to replace last few years.
EU clealry says that they will re-route all of their garbage into US starting 2027.
Fitting (Works!). Pareto (!) Trying 3-5 libraries a day is a bit too much, but too many of them are simply borken.
Sun May 04. 2025.
Fib > Benford (Newcomb). But only on clean data, looks like. Same paradox. Scalpel vs Big Berta. "How many divisions" and all that.
So now I will be rebuilding the rather exact same pipeline I had working 15 years ago, only from ruins of various corporations. How is this "building back better" exactly? It is not. The stuff already worked 15 years ago. What a waste.
Mon May 05. 2025.
URLs are the currency of the web. AI always tries to steal stuff. "haha".
Started the first inference. Nothing fancy. 3x2. Why do I still need to try 3-5 libraries a day, though? I guess that's the only thing AI does. (And they started to suffocate the API calls as of today. EU French started)
Inference works. Beats POOR.