Teodor Zhechev (@tzhechev) header
Teodor Zhechev (@tzhechev) profile picture

Software Dev.

Joined September 2011
Photo by Teodor Zhechev (@tzhechev)Photo by Teodor Zhechev (@tzhechev)Photo by Teodor Zhechev (@tzhechev)Photo by Teodor Zhechev (@tzhechev)Photo by Teodor Zhechev (@tzhechev)Photo by Teodor Zhechev (@tzhechev)Photo by Teodor Zhechev (@tzhechev)Photo by Teodor Zhechev (@tzhechev)Photo by Teodor Zhechev (@tzhechev)Photo by Teodor Zhechev (@tzhechev)
We're all suffering a very distinctly american set of neuroses after the advent of the internet monoculture. This is probably already "invisible" as in "water" to just about everyone though.
gm to all game devs in 2026 I am the man in the cave
2
33
Teodor Zhechev retweeted
If you posted The Economist's estimates of the Russian death toll in the Ukrainian invasion, you'd inevitably get a lot of hate in the comments for being naïve, exaggerating, etc. But the Russia government hack shows the estimates were completely correct.
89
321
25
4,067
149,264
Teodor Zhechev retweeted
we have lost sacred texts if people aren’t immediately recognizing what this is 😔
Grand Theft Auto VI has an insane amount of unique houses and buildings with detailed interiors that you can enter, you never know what you’re going to encounter. In the beginning of the game, I ran into a house to escape the cops and accidentally entered a crime scene. #GTA6
75
227
7
6,757
1,495,249
I'll be honest - I have no idea how anybody has any hype for the most obviously and shamelessly astroturfed game in history. It was the same game since the "third" one. It's going to be just endless astroturf "fan" accounts for weeks now and BS "reviews."
3
20
Teodor Zhechev retweeted
It is hard to do justice to just how bizarre this story is. Last year, the FT cried foul because Nvidia's disclosures attributed 22% of sales to Singapore. The FT and others questioned this figure as suspiciously implausible, given the size of Singapore's market. The explanation was simple. Multinational buyers used Singapore affiliates to centralize invoicing, while physical shipments went elsewhere. Nvidia explicitly disclosed this context in its FY2025 annual report. In Q3, Nvidia switched to reporting revenue by customer headquarters, rather than billing address. This resulted in Singapore ceasing to be a material line item. Nvidia explained that, because billing addresses “often reflected a customer’s centralized invoicing location, even though our products were almost always shipped elsewhere,” reporting by customer headquarters “provides a better representation of the geographic profile of our revenue.” Nvidia also recast prior periods on the same basis. Now, the Financial Times runs the false headline: “Nvidia stopped disclosing Singapore sales as revenues hit $104bn.” The sales still are disclosed but based on where customers are headquartered rather than location of accounts-payable departments. The same FT that faulted Nvidia for doing the latter rather than former, is now faulting Nvidia for doing the former rather than the latter. What is even weirder is that Nvidia continued to disclose the prior Singapore-booked metric in the financial statements filed by its local Singapore subsidiary. That is how the FT got the figure that it claims is being concealed. So, in effect, the public markets are getting both, while the FT is making it sound like they are getting neither. Finally, and perhaps most disingenuously, the article tells readers in paragraph four that Nvidia “previously declined to comment on the reasons for changing its reporting methodology,” only to say to add several paragraphs later “beyond its financial filings", while never quoting the straightforward explanation sitting inside those filings. What a joke.
FT Exclusive: Nvidia stopped disclosing the scale of its business in Singapore last year when the city-state accounted for almost half of the US chip group's worldwide revenues. ft.trib.al/mnaOHL1
Readers added context they thought people might want to know
Nvidia disclosed in SEC filings its switch to reporting revenue by customer headquarters rather than billing location, after which Singapore was no longer a separate geographic line item. sec.gov/Archives/edgar… sec.gov/Archives/edgar…
39
143
18
1,071
73,532
Everybody talking about AI Safety (you know who) seems to conveniently disregard the real dangers of AI as a monopolized service. A duopoly of Anthropic/OpenAI that has absolutely no reason to take security breaches seriously, because what are you going to do - NOT use AI at all - is much worse for everybody and for safety. These same companies were touting how much damage their models could do as examples of the dangers of opensource models "in the wrong hands" would have zero real incentive to treat incidents on their end with actual urgency once open models are banned. Safe infrastructure is democratized infrastructure. It's being able to choose your version of alpine, run on cloudflare or gcp or bare metal and do updates and mitigations in full control. A single multi-trilion dollar company will just treat your security as litigation expense.
35
FYI бубояди Около 18–21% (приблизително една пета) от консолидираната нетна печалба на Kaufland идва от България, според наличните публични данни за 2025/финансовата 2025/26 г. Оборотът на България е около 3,8–4% от общия на Kaufland (1,4 млрд. евро от ~36,7 млрд.). Но маржът в България е значително по-висок (~4,7%) от средния за групата (~1%). grok.com/share/c2hhcmQtMi1jb…
2
39
Teodor Zhechev retweeted
I am actually disgusted by the timing of this drop. And this will get a bit more technical. Let’s do a real technical teardown of this "1,385-sigma Blackwell weight fingerprint" report that OpenAI and Anthropic are breathlessly promoting today, because banking on nobody reading past the headline is how DC narratives get manufactured. ​First, look at the actual CUDA maths: when PyTorch initialises model weights (torch.nn.init) with a fixed PRNG seed, curand/Philox parallelises random number generation across the GPU’s Streaming Multiprocessors (SMs). Because Hopper has 132 SMs and Blackwell has 148 or 152 SMs, the exact same seed lays out initial weights in a different spatial pattern depending on the GPU’s SM count. In huge MoE models, certain low-gradient null-spaces or unrouted slices preserve faint directional traces of those step-0 weights even after training. ​So yes, brute-forcing the seed and SM grid can show that a model’s step_0 weight initialisation script ran on a 148/152-SM Blackwell GPU. Now read what the summary explicitly admits in the fine print: 1. Initialisation is not full training: The report literally states that these signatures show how the weights were created, and "support (but do not prove) the inference that Blackwell also participated in model training". Initialising a model checkpoint takes seconds on a single rented dev node before loading the safetensors onto a standard Hopper cluster. The "2,000 to 3,000 Blackwell GPUs" and "$500 million" cluster numbers do not come from the weight fingerprint and Depue admits in his own thread that he simply "estimated" cluster sizes from release schedules and model details. ​2. Zero geolocation and zero proof of smuggling: The final sentence of the executive summary explicitly admits that the findings "do not establish the physical location of the chips, how they were acquired, or whether export control regulations were violated". Depue himself rented a cloud GB200 instance just to run his verification test. Commercial labs that are not on the Entity List renting bare-metal Blackwell nodes in offshore clouds (Singapore, Japan, or the Middle East) is completely legal under BIS rules and leaves the exact same SM fingerprint. Now ask yourself why a former OpenAI researcher dropped this on November 12, the exact day Jensen Huang was pushed out, with OpenAI and Anthropic immediately stepping in to "independently verify" it. US closed-source labs are watching their paid API moats get eaten alive by free Chinese open-weights models. Conflating a step-0 PRNG initialisation seed with a "$500M smuggled Blackwell cluster" hands DC hawks the exact pretext OpenAI and Anthropic want to regulate global cloud compute and kneecap open-source AI.
I have a major result to share. I’ve found conclusive forensic evidence that the top 4 Chinese AI labs—DeepSeek, Moonshot, MiniMax and Ant Group—have built their newest models on NVIDIA’s most advanced Blackwell chips, which are barred from export to China. This sounds insane, but I discovered that almost every AI model carries a hidden signature of the GPU that trained it. When a model is first set up, the starting values of its weights are laid out in a pattern that depends on the exact GPU. Each combination of GPU and seed number produces a unique pattern—a sort of fingerprint of the hardware used to create the model. Traces of that pattern survive training into the published model. Again, this evidence isn’t based on any insider information or leaking. It’s based on a forensic analysis of the model weights released by the companies themselves. I searched more than six trillion combinations of chip layout and seed number and matched them to open-weights models. In doing so, I found conclusive evidence that six Chinese AI models—including DeepSeek-V4-Pro, Kimi K3, and MiniMax-M3—were initialized on NVIDIA Blackwell systems. Four of the models match Grace Blackwell, the GB200 and GB300 parts in NVIDIA’s flagship liquid-cooled rack-scale systems. I then verified these results on real hardware. Running unmodified PyTorch with the seeds I’d recovered, a Blackwell B300 reproduced DeepSeek-V4-Flash’s untouched starting weights bit for bit, and a rented GB200 reproduced Kimi K3’s to a correlation of 0.999995. A Hopper H100 reproduced none of DeepSeek-V4-Flash’s, and changing the seed by one broke the match. (DeepSeek’s seed is 953049583 by the way.) The method also was verified to match 24 known open-weights models to the hardware that developers said they used. It is exceedingly unlikely that these results are a coincidence. Across six trillion combinations, the best accidental match should reach only about 8 sigma (i.e., eight standard deviations above what random chance would produce). The real matches range from 49 to 1,385 sigma, and even the weakest single test hits 20. These were data-center-scale training runs. With some analysis of release schedules and model details, I estimate that the largest, Kimi K3 and DeepSeek-V4-Pro, used 2,000 to 3,000 Blackwell GPUs each. Clusters of this size cost on the order of $150 to $200 million each to buy, and across these labs alone would add up to more than half a billion dollars of export-controlled GPUs, not counting experimental compute or otherwise. It’s quite likely that we see multi-billion dollar expenditure on export-controlled GPUs. I am highly confident in this result, and it has been independently verified by senior researchers at OpenAI and Anthropic. You can read the full report at unemploymentlabs.com/checkpo…. There is much more I’ve identified about these models through other forms of checkpoint forensics like this, and will have more to share soon.
31
113
15
774
39,446
Problem with using AI to code is you can't just be angry at yourself for being stupid. You now have somebody else to be angry at for being stupid.
1
24
Fingers crossed NVIDIA can land the right CEO once this interim search period is over. Perhaps they could try to get the legendary Hensen Juang.
2
37
Изказването на изп. директор на "Кауфланд България", че ако компанията тръгне надолу, фалитът на КТБ ще ни се стори като "пролетен бриз" е арогантно и неприемливо, както и внушението, че "Кауфланд" е твърде голям, за да бъде санкциониран. Чуждестранните инвестиции са важни, работните места също, но никоя компания не може да използва икономическата си тежест като щит срещу закона. Вместо да плаши с катастрофални сценарии, ръководството на "Кауфланд" е добре да се замисли дали през годините винаги е било коректно към българските производители и дали е имало неприемливи търговски практики. Глобата на КЗК от € 75 млн. е сериозно обвинение, на което трябва да се отговори с аргументи и доказателства, а не с високомерие. От години критикувам самата КЗК за бездействие и политически зависимости и нямам намерение да я защитавам безкритично и сега. Ако санкцията е необоснована, "Кауфланд" има пълното право да я обжалва и да защити интересите си в съда. Но докато печелиш милиарди от продажби в България, най-малкото, което се очаква, е да уважаваш българските институции, потребители и производители. Големият бизнес не е над закона, както и КЗК не е над съдебния контрол.
55
12
7
230
9,589
Радевото управление е трагедия за България, ама още по-голяма трагедия е опозиция, която го обвинява, че репресира една от най-негативните, злоупотребяващи и вредни външни „бизнеси“ защото първите 5 минути от управлението им се застъпва с напълно заслужена глоба (3та по ред.)
4
110
Невежеството, което се излива по кауза Кауфланд в момента напълно разсейва всякакви съмнения, колко са умни умните и красивите. КЗК гони Кауфланд от 10г, не откакто Радев е на власт. Кауфланд и Лидл, и двете Шварц Груп, са топ веригите в страната и държат половината от сегмента. teapaw.org/i/grok/share/c2hhcmQtM…
1
1
3
375
Teodor Zhechev retweeted
I only make serious software
653
5,459
992
48,323
5,121,737
"I'd rather see a France led by a fascist than a France without nuclear power."
Je préfère voir une France dirigée par une fasciste que de voir une France sans nucléaire. Car la France peut survivre au fascisme. Mais elle ne survira pas à la fin du nucléaire. Je pèse mes mots. Votez mieux.
56
1,717
23
31,690
354,512
People used to have to have kids because they would otherwise be destitute in old age. Now the state robs everybody else's kids at gunpoint to take care of the people who never had kids. The people who raised kids did it out of pocket. Not really that hard to see why TFR is ⬇️
1
53
Teodor Zhechev retweeted
In a weird way, everybody is rooting for you. When you go to a job interview, the interviewer wants you to be amazing so they don’t have to look anymore. when you go on a date, the person wants to have an amazing date with you. If you make a piece of content, people want to see amazing pieces of content and they have no problem if it’s yours. the instant you get into the mindset that people want you to do well, your entire world gets better. And you’ll be shocked to know this pattern keeps repeating in different areas of life.
Hit me with the harshest reality truth.
352
9,784
467
85,120
3,003,693
Не знам кой е Гюров, честно казано. Йотова всички знаят коя е. Не беше достатъчно важен президентският пост, че да си струва да излъчим кандидат, за който някой е чувал предполагам. Е щом не е било важно, голяма работа.
1
119
Cut the cables.
how do you take probably 2 hours to put this together, take photos, post them and don’t realize this is reminiscent of something deeply racist
1
165