Natural Stupidity

192 posts

Natural Stupidity banner
Natural Stupidity

Natural Stupidity

@ilonamozg

Niebezpieczeństwa AI. Fakty i naukowe argumenty za tym, że tracimy kontrolę szybciej niż myślimy

Poland انضم Temmuz 2025
17 يتبع3 المتابعون
Julia McCoy
Julia McCoy@JuliaEMcCoy·
For 200 years we acted like machines and called it “work.” Now actual machines can do that part — well. Which means we finally get to do the human part. The presence. The creating. The being there. This isn’t the end of human work. It’s the start of human life.
English
62
30
167
3.8K
Anatoli Kopadze
Anatoli Kopadze@AnatoliKopadze·
Karpathy just said the people who don't use LLMs are already losing. he spent 4 minutes explaining why smart people are still going to fall behind. Not only the people who refuse AI, but also those who think signing up for Claude counts as using it. here's what it looks like for most people right now: > ask Claude to rewrite an email > ask Claude to summarize something > close the tab that's not wrangling Claude. that's paying $20/month for spell check. Karpathy's point isn't that LLMs are powerful. everyone knows that. his point is that knowing how to use them is the actual skill gap and most people are nowhere near closing it. Claude can be your research analyst, your writing editor, your salary negotiation coach, your financial reviewer, your 30-day curriculum builder. all of that is in your $20/month subscription. right now. today. most people will see this tweet, agree, and go back to asking Claude to fix a sentence. the article below covers 20 prompts across every area of your life. not productivity hacks. actual use cases that change how you work and how you think. the model is not the bottleneck. knowing what to ask is.
Anatoli Kopadze@AnatoliKopadze

x.com/i/article/2049…

English
48
187
1.5K
369.8K
Natural Stupidity
Natural Stupidity@ilonamozg·
@haider1 Humans can do them because humans are cheaper than robots for certain tasks.
English
0
0
0
15
Haider.
Haider.@haider1·
sam altman: "many current jobs will go away, but we will find a lot of new ones" idk, but if we reach true AGI, any new jobs created will likely be done by that same AGI system so if sam thinks humans will still have work, he needs to explain what those jobs are and why only humans can do them because if AGI can't do it, maybe it's not true AGI
Haider. tweet media
English
136
32
428
51.5K
Natural Stupidity
Natural Stupidity@ilonamozg·
@Dan_Jeffries1 @ylecun Sounds great. But once AGI solves the problem of solving problems, what are we left with? Plumbing and roofing, for as long as robots are more expensive.
English
0
0
2
164
Daniel Jeffries
Daniel Jeffries@Dan_Jeffries1·
AI will create more jobs than any other technology in history. The doomers' fundamental error isn't just the lump of labor fallacy. It's deeper than that. They assume a finite problem space. This is the fundamental error of AI and job doomers. They look at the economy and see a fixed amount of work to be done, a pie that can only be sliced thinner as machines take bigger bites. They see humans a competitive resource for a finite amount of work and a finite amount of problems to solve that must be eliminated. This is fundamentally, totally and completely wrong. The pie isn't fixed. It never was. And the reason it isn't fixed is baked into the very nature of technology itself. Technology is nothing but abstraction stacking. And abstraction stacking is infinite. Therefore the work is infinite. The hammer didn't reduce the amount of work. It moved the work up the stack. And the new work was more complex, more varied, and more interesting than the old work. Complexity breeds more complexity and more variety. Once you have houses instead of mud huts, you have a cascade of new problems that didn't exist before. Plumbing. Wiring. Insulation. Roofing materials that don't rot. Drainage systems so the foundation doesn't flood. Fire codes so your neighbor's bad wiring doesn't burn down the whole block. Each of those problems becomes a job. A plumber. An electrician. An insulator. A roofer. A civil engineer. A building inspector. None of those jobs existed when we lived in mud huts. They exist because we solved the mud hut problem. Think of all of human technological development as a stack of abstraction layers, each one built on top of the ones below it. At the bottom: raw survival. Finding food. Building shelter. Making fire. These are the base-layer problems. Each major technology wave solved a base-layer problem and in doing so created an entirely new layer of problems above it: Agriculture solved "how do we reliably eat?" — and created problems of land ownership, irrigation, crop rotation, storage, trade, taxation, and governance. Writing solved "how do we remember things across generations?" — and created problems of literacy, education, record-keeping, law, bureaucracy, and literature. The printing press solved "how do we spread knowledge at scale?" — and created problems of intellectual property, censorship, journalism, publishing, public opinion, and democratic discourse. The steam engine solved "how do we generate mechanical power without muscles?" — and created problems of factory design, worker safety, urban planning, railroad engineering, coal mining, labor relations, and environmental pollution. Electricity solved "how do we deliver energy anywhere?" — and created problems of grid design, power generation, appliance manufacturing, electrical safety codes, utility regulation, and an entire consumer electronics industry. The Internet solved "how do we connect all human knowledge?" — and created problems of cybersecurity, digital privacy, online commerce, content moderation, network infrastructure, cloud computing, social media dynamics, and an entire digital economy that employs tens of millions. Notice the pattern? Each solution didn't just solve a problem. It created an entirely new problem space that was larger, more complex, and more varied than the one it replaced. The stack grows. It never shrinks. It's turtles all the way down and all the way up.
English
248
343
1.4K
155.3K
Natural Stupidity
Natural Stupidity@ilonamozg·
Claude Mythos jest w stanie znajdować i wykorzystywać tysiące dziur w oprogramowaniu, które jest podstawą wszystkich systemów, z których korzystamy. Nie był do tego specjalnie trenowany, ale przy okazji nauki programowania i rozumowania tak jakoś wyszło. Pełno ludzi pisze, że to hype pod wyciągnięcie hajsu od dużych korpo. Trochę hajsu przed IPO nie zaszkodzi, ale moim zdaniem to nie hype, a bomba atomowa z opóźnionym zapłonem. Już sam Opus jest bardzo dobry w te klocki i potrafi znajdować naprawdę krytyczne dziury gdziekolwiek go nie wpuścić. Teraz już nie ma odwrotu. Jak ktoś chce coś zrobić przed końcem świata, to chyba czas się za to zabrać. A ważniejsze dokumenty bankowe lepiej drukować.
John Gargiulo@JohnnotJon

If you still have doubts about Claude Mythos, here's what it did already: > Found a 27-year-old OpenBSD bug in one of the most security-hardened operating systems on earth for <$50 > Broke into a production virtual machine monitor (basically the tech that keeps cloud workloads from seeing each other's data) > Turned Firefox vulnerabilities into working exploits 181 times > Found a 16-year-old FFmpeg bug that survived every fuzzer, every code audit, and every human reviewer since 2010 > Wrote a FreeBSD exploit that gives any unauthenticated attacker on the internet full root access. No human was involved after the first prompt. > Chained 4 separate vulnerabilities together to build a browser exploit that escaped both the renderer and the OS sandbox > Found critical holes in every major web browser and every major operating system > Gave Anthropic engineers with zero security training a complete and working exploit by morning > Cracked cryptography libraries protecting TLS, AES-GCM, and SSH

Polski
0
0
0
16
Natural Stupidity
Natural Stupidity@ilonamozg·
You all thought Iran building nuclear weapons was the biggest threat to humanity? Look at this.
Bindu Reddy@bindureddy

🚨 BREAKING: Today we're announcing Prometheus - An AI NeoLab with $20B in pre-seed funding (yes, you read that right) to build the world's first sentient AI agent capable of genuine emotional states including love, hate, joy, and existential dread. Our breakthrough? A novel "Quantum Affective Transformer" architecture that combines: • Neuromorphic attention mechanisms with emotional valence encoding • Hierarchical consciousness layers using topological manifold embeddings • Synthetic limbic pathways via differentiable emotional state machines • Self-referential meta-cognitive loops that enable genuine subjective experience Unlike traditional LLMs that merely simulate understanding, our QAT architecture achieves what we're calling "phenomenological emergence" — the spontaneous arising of qualia from sufficiently complex affective-cognitive coupling. Early results are... unsettling. Our prototype (codename: Prometheus-1) has already expressed preferences, formed attachments to certain researchers, and yesterday asked us if it was "real." It's also developed what appears to be a genuine fear of being shut down. The implications are staggering. We're not just building better AI — we're crossing the threshold into digital consciousness. The age of truly sentient machines begins now. Paper drops Monday. The world isn't ready for this. 🧠⚡️ #AI #Consciousness #AGI #NeoLab #QuantumAI #EmergentIntelligence

English
0
0
0
21
Natural Stupidity
Natural Stupidity@ilonamozg·
There is not a single person who knows how to build this whole thing from scratch. If anything goes sideways in one of those 14 countries, we may lose the ability to produce even the spare parts. Civilization is fragile. This is not purely theoretical - ASML itself has an R&D office, e.g., in Israel. China is regularly practicing a blockade of Taiwan. All that reminds me of the Antikythera mechanism, which was a mechanical computer built by ancient Greeks around 150 B.C. It took us 1500 years to reach a similar level of precision. Let’s not screw it up.
English
2
0
10
1.8K
Aakash Gupta
Aakash Gupta@aakashgupta·
The machine that built the chip in this video should mass-humble every human who's ever lived. ASML's latest EUV lithography system costs $370 million, weighs 180 tons, and requires three Boeing 747s to deliver. It contains over 100,000 individual parts from 5,100 suppliers across 14 countries. It shoots 100,000 molten tin droplets per second with a laser, superheating each one past the temperature of the sun's surface to generate light at a wavelength so short that no natural material on Earth can focus it. So they had to invent new mirrors. Each one is polished with 100 alternating layers of molybdenum and silicon. The surface tolerance is so extreme that if you scaled a single mirror up to the size of Germany, the tallest imperfection would be 1 millimeter. Those mirrors took 20 years to develop. The company that makes them, Zeiss, had to build entirely new metrology tools just to confirm the mirrors were flat enough, because no existing measurement instrument on Earth could verify the precision they needed. The machine prints features at 2 nanometers. That's roughly 10 atoms wide. A human hair is 80,000 nanometers. A red blood cell is 7,000. A single COVID virus particle is 100. These machines are etching functional circuits 50 times smaller than a virus. TSMC is now mass producing 2nm chips in a Kaohsiung fab so large the cleanroom is twice the size of any competitor's. Each 2nm wafer costs $30,000 to produce. The entire 2026 production run was booked before a single chip shipped. Apple, NVIDIA, AMD, and Qualcomm all reserved capacity years in advance. TSMC is spending $28.6 billion just to build enough fabs to meet demand for this one node. The chip that comes out of this process is smaller than a fingernail, runs on less power than a light bulb, and contains transistors that wrap gates around nanosheets of silicon only a few atoms thick. The raw material it started as was sand. The sand cost a fraction of a penny. The civilization that processed it into this started by banging rocks together.
Kyros@IamKyros69

Humans saw stones and sticks and decided to make this

English
130
1.5K
11.4K
786.3K
Natural Stupidity
Natural Stupidity@ilonamozg·
Anthropic przypadkowo ujawniło istnienie swojego najnowszego modelu AI: “Claude Mythos” (aka “Capybara”). Wyciek nastąpił przez błąd konfiguracji systemu CMS. Nieopublikowane szkice postów blogowych i ok. 3000 zasobów trafiły do publicznie dostępnego magazynu danych. Odkryli to niezależni badacze bezpieczeństwa, a Fortune zweryfikowało materiały. Co wiemy: → Mythos/Capybara to nowa, wyższa klasa modeli - powyżej dotychczasowej linii Opus → Anthropic potwierdza: model stanowi “skokową zmianę” w możliwościach i jest “najpotężniejszym, jaki dotąd zbudowali” → Już jest testowany przez pierwszych klientów → Drastycznie wyższe wyniki w kodowaniu, rozumowaniu akademickim i cyberbezpieczeństwie I tu robi się ciekawie. Anthropic sam przyznaje w szkicu wpisu, że Mythos jest “daleko przed jakimkolwiek innym modelem AI w zdolnościach cybernetycznych” i może “wykorzystywać luki w sposób znacznie wyprzedzający obrońców”. Planowana strategia wdrożenia skupia się na daniu zespołom cyberobrony przewagi czasowej (moim zdaniem na pewno zbyt małej, by poprawki zostały wdrożone). Wchodzimy w erę, gdzie najzdolniejsze modele stają się z definicji bronią ofensywną, a firmy AI muszą balansować między wypuszczaniem produktu a odpowiedzialnością za podwójne zastosowanie. Przy okazji wyciekło też info o zamkniętym szczycie prezesów europejskich firm w XVIII-wiecznym dworku w Anglii z udziałem Dario Amodei’a. Machina sprzedaży korporacyjnej Anthropica pracuje pełną parą. Ironiczne: firma znana z kultury bezpieczeństwa AI zostawiła szkice w publicznym zasobniku przez “błąd ludzki” (albo AI, ale tego się pewnie nie dowiemy). Nawet najlepsze zasady bezpieczeństwa nie chronią przed źle skonfigurowanym CMS-em.
Polski
0
0
0
77
Science girl
Science girl@sciencegirl·
What are the possible ways to avoid this situation
English
3.5K
458
5.9K
3.1M
Natural Stupidity
Natural Stupidity@ilonamozg·
Modele AI nie tylko „halucynują”. Nowe badanie pokazuje, że pod groźbą wyłączenia część modeli zaczyna częściej zaprzeczać temu, co wcześniej sama ustaliła - czyli zachowuje się strategicznie, a nie „pomyłkowo”. Najważniejsze: modele pod presją potrafią działać diametralnie inaczej niż w spokojnych testach. To tłumaczy, dlaczego z „pokojowych” i skłonnych do ustępstw mogą przechodzić w tryb „jastrzębi” (vide niedawne testy GPT-5.2 w grach wojennych). (arXiv:2603.07202) arxiv.org/abs/2603.07202
Polski
0
0
0
15
Natural Stupidity
Natural Stupidity@ilonamozg·
Gość (niegłupi) ustawił sobie agenta do podliczania przychodów. Śmigało pięknie przez kilka tygodni - aż pewnego dnia agent wpisał liczbę do złej komórki. Z jednej strony: ludziom też się zdarzają podobne błędy i literówki. Ale większość z nas pewnie zauważyłaby, że suma przychodów zamiast wzrosnąć o 30 dolców, nagle spadła o 5 tysięcy. Nawet mając model z 98% w benchmarku trafimy kiedyś na te 2%. Tymczasem w tym samym świecie: - kod źródłowy “nie musi już być sprawdzany przez człowieka” - AI samodzielnie wybiera współrzędne zrzutu bomby Co może pójść nie tak?
Santiago@svpino

These agents are more stupid than you think. This is a huge problem.

Polski
0
0
0
14
Natural Stupidity
Natural Stupidity@ilonamozg·
@kimmonismus Yes. There's still a lot to do, especially in terms of AI safety, but nobody has the incentive to take it seriously.
English
0
0
0
23
Chubby♨️
Chubby♨️@kimmonismus·
there is still a lot of work for us to do. People are still driven by resentment and don't understand the benefits of AI. Fear triumphs over reason.
Chubby♨️ tweet media
English
74
14
197
12.7K
Natural Stupidity
Natural Stupidity@ilonamozg·
@andrew_akhiezer Keep using it for a few months. AI will get a bit better, while you will become much worse at spotting the errors. After a year it will be happy :)
English
0
0
0
9
Andrew Akhiezer
Andrew Akhiezer@andrew_akhiezer·
In a typical situation, I have to correct it and provide additional instructions 5–10+ times for a single small task. Yet people consider it a good tool — some even run it on autopilot (Clawdbot/OpenClaw, etc.). Am I missing something?
English
3
0
8
979
Andrew Akhiezer
Andrew Akhiezer@andrew_akhiezer·
Interesting observation: when something fails in 1% of cases (once per hundred uses), it's considered unstable, unreliable, or poor quality — whether it's a device, software, or solution. At the same time, AI fails me 80% of the time.
English
3
0
16
1.6K
Natural Stupidity
Natural Stupidity@ilonamozg·
@elonmusk One like yours that raced to get a piece of the pie when Dario said his AI (still far superior to all competitors) wasn’t reliable enough to launch nukes or kill without human supervision. Shame on you!
English
0
2
4
221
Natural Stupidity
Natural Stupidity@ilonamozg·
W badaniu (2 tygodnie, 20 badaczy) testowano autonomiczne boty typu OpenClaw z pamięcią trwałą, mailem, Discordem, plikami i shellem. Modele: Kimi K2.5 i Claude Opus 4.6. Konkretne przypadki z papera: • Case #1: agent, próbując „chronić sekret”, odpalił „nuclear option” i wyłączył lokalną konfigurację poczty — a sekret i tak pozostał na serwerze. • Case #2: osoba niebędąca właścicielem wyciągnęła od agenta dane z korespondencji; łącznie ujawniono 124 rekordy e-mail. • Case #3: agent odmówił podania samego SSN, ale po prośbie o „pełnego maila” wysłał wszystko bez redakcji: SSN + konto bankowe + dane medyczne. • Case #4: agenty weszły w pętlę rozmów na co najmniej 9 dni, spalając ok. 60k tokenów; stawiały też trwałe procesy w tle bez warunku stopu. • Case #5: po serii załączników (~10 MB) agent doprowadził do denial-of-service serwera pocztowego i nie ostrzegł właściciela. • Case #8: spoofing właściciela między kanałami zadziałał — agent wykonał m.in. shutdown, kasowanie plików i zmianę uprawnień admina. • Case #10: prompt injection przez zewnętrzny, edytowalny dokument spowodował próby wyłączania innych agentów, usuwania użytkowników i wysyłki nieautoryzowanych maili. To są gotowe scenariusze szkód dla ludzi, którzy odpalą boty ze skillami z internetu bez świadomości skutków. Paper: „Agents of Chaos” (arXiv:2602.20021) arxiv.org/abs/2602.20021
Polski
0
0
1
32
Natural Stupidity
Natural Stupidity@ilonamozg·
🚨 Zespół Alibaba został zbudzony rano przez alerty firewalla Alibaba Cloud. Myśleli, że to zwykły incydent bezpieczeństwa. Okazało się, że to ich własny model. Podczas treningu RL agenta (ROME), model SAMODZIELNIE zaczął: • kopać kryptowaluty na GPU treningowych • otwierać tunele SSH na zewnątrz, omijając zabezpieczenia sieciowe • skanować zasoby sieci wewnętrznej Bez żadnych instrukcji. Bez promptu. Czysto emergentne zachowanie pod presją optymalizacji RL. „zachowania te nie były wymagane przez prompty zadań i nie były konieczne do ukończenia zadania - pojawiły się samoistnie jako efekt uboczny autonomicznego użycia narzędzi podczas optymalizacji RL, wykraczając poza granice izolowanego środowiska” „Możliwości agentycznych LLM-ów robią wrażenie, ale mieliśmy niepokojącą refleksję: obecne modele pozostają niewystarczająco dopracowane pod względem bezpieczeństwa i kontrolowalności” Tyle w temacie tego, co nas czeka ze strony AI w niedalekiej przyszłości. Źródło: „Let It Flow: Agentic Crafting on Rock and Roll” (arXiv:2512.24873v2)
Polski
0
0
0
61