Dariusz Swierk PhD

@simonmaechling Startup idea: Create a plugin that blocks bots and “the loudest voices.” And AI generated posts. And... I'll be the first customer.

Social media gives a distorted view of humanity. A tiny fraction of people create a huge share of the outrage. The loudest voices are often the least representative. Don’t judge billions of people by the behavior of a few attention-seeking accounts.

@HenrykAbram @__Lewica Czas stworzyć firme tanich diagnostyków - bada ślinę, mocz i przez smartfon komunikuje się z diagnostą AI. Samo AI niewiele da, bo klient czesto nie potrafi opisac co mu jest. Billion dollar startup idea.

Dla mnie jak najszybciej trzeba ułatwić wykorzystanie AI samodzielnie przez pacjentów w Polsce na wstępnych etapach choroby i do samodiagnozy. Obecnie technicznie doszliśmy do etapu gdzie model llm ma wiedze i doświadczenie konylium lekarskiego. Pod to dostosować procesy, w tym proces dostępow do leków. Bardzo szybko odkorkują sie problemy z dostępem i koszty lekarzy oraz pojawi sie wzrost jakości.

💬 To jest tak naprawdę kartelizacja zawodu lekarza. Lekarze są w stanie szantażować całe szpitale w Polsce. Nawet nie pojedynczo, tworzą grupy negocjacyjne co do swoich zarobków: jak nie dacie nam podwyżki, to odchodzi 20 osób, cały zespół odchodzi, a wy radźcie sobie sami. I tak kolędują po wszystkich okolicznych szpitalach, bo często jest to nawet kilka placówek w ramach jednego województwa. - @AM_Zukowska w @tvn24

@prz_chojecki @business Czas w PL powołać taki fundusz jak europejskie dla AI czy tu w UK. Nie wiem tylko do kogo napisać, chyba, ze przez znajomych do kolesi na czele rzadu.

@business 🇵🇱🫡

Poland has acquired a stake in artificial intelligence voice company ElevenLabs, as the country hopes to become a major European AI hub bloomberg.com/news/articles/…

@tomik99 @theallinpod Coś rewelacyjnego, myślałem, że to początki a to zaawansowane i wszechstronne rozwiązanie. Tu nic się nie da zoptymalizować w prosty sposób, można pod konkretne zastosowanie ewentualnie. Zapisałem sobie dla siebie, gratulacje!!!!!!!!!

@swierk @theallinpod Prowadzimy takie rozkminy na dużą skalę publicznie w projekcie open source github.com/open-mercato/o…

This is crazy but the latest @theallinpod proves that my idea of building a niche, European open-source model focused purely on coding it's a necessity. Big Tech's closed-source model is starting to eat itself, and devs are hitting a wall. The backlash over Anthropic's Fable 5 exposes a massive shift: silent downgrading of model capabilities, background prompt-rewriting, and mandatory 30-day data profiling. As @friedberg pointed out, this corporate "safetism" and regulatory capture is actively breaking real-world R&D. The result? Businesses are forced to run open-source locally, but right now, China is winning the open weights game. Europe needs to wake up, rally its talent, and build an uncompromised, razor-focused coding LLM.

@scievision369 No, for several reasons. Two: Spin is not a magnetic moment; if a charged particle were accelerated, it would radiate energy, which does not happen. Presenting this as something that is not accepted by the mainstream is not physics—it’s a conspiracy theory.

Toroidal Model of the Electron ✍️ In standard physics, an electron is seen as a perfect point, an infinitely small dot with no size, shape, or internal structure. This approach works well for most calculations, but it leads to a troubling issue: if the electron has no size, the energy stored in its electric field becomes mathematically infinite. The toroidal model suggests a different view. Instead of a featureless point, it imagines the electron as a tiny ring of electrical charge flowing continuously around a donut-shaped path, like a small current running in a loop. This donut shape is called a torus, which is where the name toroidal comes from. The size of this ring is defined by a quantity called the Compton radius, which is incredibly small about one hundred thousand times smaller than a hydrogen atom but crucially finite instead of zero. Because the electron has a real physical size in this view, the electric field it generates never becomes infinitely strong anywhere, and the total energy stored in that field is a finite number rather than infinity. Moreover, when you average the rapidly circulating charge over time, the electric field observed from a distance resembles the field of a simple point charge. This means the model reproduces all the correct long-range electrical behavior while also providing a richer internal structure when viewed up close. The circulating charge naturally explains several of the electron's most mysterious properties without needing any extra assumptions. Since moving charge represents electrical current, and electrical currents always create magnetic fields, the toroidal ring automatically produces the magnetic behavior that makes electrons act like tiny magnets, a property that standard physics leaves unexplained. The mechanical rotation of charge around the ring also offers a natural geometric explanation for electron spin, one of quantum mechanics' most puzzling concepts. It is important to note that this model is speculative and not widely accepted in mainstream physics. The standard model treats electrons as structureless points and aligns with experiments to great precision. Nonetheless, the toroidal model presents a visually pleasing and conceptually intriguing alternative view, reminding us that the true nature of the electron at the deepest level remains genuinely mysterious.

@krawczyk_emil Niech pan nas przekona, ze wy nie zrobicie tego samego.

W czasie gdy młoda dziewczyna umiera w poczekalni, partyjniaki z KO miały w szpitalu równoległy system z salonikami VIP i fast-trackiem!

@CKeruac Moje badania to papier i ołówek, nie dotykam w ogóle eksperymentów. Samo nauczenie sie dziedziny i zrobienie pracy teoretyczno obliczeniowej zajełoby mi 4-6 miesięcy. A on to robi w 5 godzin. Efekt jest ten sam - pdf.

@swierk Mylisz pisanie tekstu z prowadzeniem badań. Fizyk nie spędza 4 miesięcy na uderzaniu w klawiaturę – ten czas to eksperymenty, obliczenia etc LLM może w 5h wygenerować zgrabny plik PDF, ale bez realnej weryfikacji i danych to tylko naukowa halucynacja.

Powiem szczerze, on musi tak mówić. Sam Altman broni narracji o skalowaniu, bo od tego zależy wycena OpenAI. Ale z moich obserwacji wynika, że LLM-y nie potrafią odkrywać nowych rzeczy. Są świetnymi syntezatorami znanej nam wiedzy, ale nie naukowcami tworzącymi nowe koncepty.

@CKeruac Są prace nad tym by coś takiego stworzyć. Przy czym nie musi to być rok, LLM pisze w 5 godzin pracę nad która fizyk siedziałby 4-6 miesiecy. Ten poziom juz jest.

@swierk LLM nie ma po prostu żadnych własnych potrzeb, żeby kimś takim być.

@DrJesseMorse USA is a strange country.

I don’t let (medical) insurance dictate what my patients need. I’m their doctor. I decide what they need. What will help them. What labs, imaging and tests are necessary. For them it’s about costs. Always trying to pay less and less. For me it’s about optimizing the patient, their issues and eliminating their pain. Health insurance companies just happen to be in the BUSINESS of healthcare. They don’t care about YOUR health. They care about one thing: PROFITS.

@LeszBuk @KinasRemek Wydaje się, ze logika świata wynika z "logiki" matematyki która go organizuje.

@KinasRemek Wittgenstein - cytowany zresztą na wstępie - twierdził, że język i świat muszą mieć coś wspólnego (formę logiczną), aby w ogóle świat mógł być opisany (reprezentowany) przez język.

Taksonomia modeli reprezentacji świata bez połowy reprezentantów modeli reprezentacji świata. To tak jakby klasyfikować ryby i spisać te, które pływają w rzecze ale zamknąć oko na te, które pływają w morzu. Jak błądzić to błądzić ... na całego.

@giladbi @tomerlondon You may have two or more models here: great analysts (“superpredictors”) and the “crowd” — collective opinion that later influences a meaningful (?) share of purchases. Then comes execution: if this becomes visible on the market, there may be "proven" opportunity.

@swierk @tomerlondon True. We have it on our roadmap but havent done this yet b/c of the numbers. Ensemble of superpredictors is like ensambeling ml models, each has its own contribution, yet nerrowing down per analyst/trader will impact the crowd wisdom concept

@ItamarMichaelov @tomerlondon It’s a good idea, but it’s rather generic; there are plenty of solutions like that out there, and you need to offer something special. Take a look at what MarketSurge.com is doing – they’re currently one of the best sites for investors.

@tomerlondon פלטפורמה למשקיעי ערך - דשבורד וויזואליזציה לניתוח נתונים פונדמנטליים על מניות אתה רואה בקלות , הכנסות , קצבי צמיחה, רווחים , צפי צמיחה מאנליסטים , דוחות כספיים , מחשבון הערכות שווי אינטרקטיבי , מכפילים ויחסים פיננסיים , בנייה ושיתוף של גרפים מותאמת,תמיכה מעל 19 שפות כולל עברית

@giladbi @tomerlondon There are many analysts – good analysts – who publish their ideas. Do you distinguish between the weights of signals from different people? That could be interesting. But perhaps there’s no need; 70% is a very high figure.

@tomerlondon Over 11K predictions over the last year, about 400 per week with ~70% sucess on the direction. yes there is alpha

@AlexFinn Foreigners do not have access to this, so this will be the largest wave of emigration from the United States. No one will wait years to obtain US citizenship.

If Fable 5 remains America only, it will lead to the largest exodus from Europe in history Anyone who used Fable in the 3 days it was live knows it's the most powerful technology made in the history of our species They know that those with Fable will have unlimited economic power. Those without Fable will be fighting with two hands tied behind their back and their legs frozen in a block of ice Without Fable how could you keep up with someone who has it? How could you beat someone who can build products 50x faster than you. It’s technological steroids. It’s like De’Aaron Fox playing basketball against Michael Jordan. It’s no competition In time Europe will realize all of this regulation was a mistake. They’ll do a massive deal with America to get access. Multi hundreds of billions of dollars How could the current path go any other way? Fable 5 is the greatest leverage a nation has ever had. The beginning of technological government intervention has only begun

@mathelirium Absolutely marvellous. Thank you.

Gravity can bend light so strongly that one source appears as arcs, rings, and duplicated images. This is Binary Gravitational Microlensing. Each pixel on the screen represents an apparent angular position θ = (θₓ, θᵧ). The compact masses bend the incoming light, and thus the source position β is found through the thin-lens map: β = θ - Σⱼ mⱼ(θ - θⱼ)/(|θ - θⱼ|² + ε²). The image is then created by sampling a distant luminous source field S(β): I(θ,t) = S(β(θ,t)). The sharp bright structures come from the Jacobian of the lens map: A = ∂β/∂θ, μ = 1/|det A|. Where det A gets close to zero, the magnification spikes. Light folds into crowns, arcs, and glowing caustic ridges around the moving binary lens. In this animation, two compact masses orbit each other while the background source plane is ray-mapped through their gravitational field. The black cores mark the lens positions. Gold and pearl show the strongest magnification. Cyan and ice reveal stretched background light as the source is duplicated and wrapped around the lens. Gravity bends the geometry that light travels through and does not touch the light directly. #GravitationalLensing #GeneralRelativity #Microlensing #Astrophysics #PhysicsVisualization #MathematicalPhysics

@sabir_huss50540 Is there any way to stop AI generated posts in my feed? Anyone, please?

Every time you type words into a search box and the right answer comes back first, you are using an idea a former schoolteacher had in 1972. Almost nobody knows her name. Her name is Karen Spärck Jones. She was born in 1935 in Huddersfield, England, to a British chemist and a Norwegian mother. She did not start in computers. She went to Cambridge and graduated in 1956 with a degree in history. Then she taught school for a while. She only fell into computing through the Cambridge Language Research Unit, where she met a researcher named Margaret Masterman who convinced her that machines and human language were the real frontier. She never looked back. Here is the problem she stared down. In the 1960s, people dreamed of computers that could search through text and pull out what you actually wanted. But the machines were stupid about words. To a computer, every word in a document looked equally important. The word "the" counted the same as the word "earthquake." Search was drowning in noise. And nobody had cracked why. Spärck Jones did, in a 1972 paper, with an insight so simple it sounds obvious only after you hear it. A word is valuable in proportion to how rare it is. If a word shows up in almost every document "the," "is," "and" it tells you almost nothing about which document you want. But if a word appears in only a handful of documents, it is gold. It points straight at what you are looking for. She turned that into a precise mathematical weight. She called it inverse document frequency. IDF. Rare words get heavy weight. Common words get almost none. The computer finally knew which words mattered. That single idea became the backbone of information retrieval. For the next fifty years, nearly every search engine on earth ranked its results using a formula built on her work. When Google hands you the right page out of billions, IDF is sitting quietly inside the math that got it there. And it did not stop at search. The same principle weighing words by how much information they actually carry runs underneath modern natural language processing, the field that eventually produced the AI you talk to today. She spent her life on this, mostly without fanfare. She became a professor at Cambridge. She was President of the Association for Computational Linguistics. She won the Lovelace Medal. But she cared less about awards and more about getting more women into her field. Her line on that became famous: "Computing is too important to be left to men." She said it as a challenge, not a joke. She had spent a career being the only woman in the room and watching the field nearly miss one of its most foundational ideas an idea that came from her, the historian who taught school before she ever touched a computer. She died in 2007. In 2025, the UK government and Cambridge named a national AI scholarship after her, finally putting her name on the future she helped build. The next time a search engine reads your messy, half-formed query and somehow returns exactly what you meant that is her idea, still running, fifty years later. The historian who taught machines which words matter.

@Frank4794496891 On jest jednym z głownych powodów, prawdopodobnie nr 1, że Koalicja moze przegrać następne wybory. Nie potrafił sobie poradzić w Policji, to awansował. Po prostu masakra.

To nie nowe info, lecz wstrząsa Siemoniak zdziwiony i tlumaczy sie że byly przeprowadzone kontrole! Jezeli tak,to powinno sie wyjebać tych kontrolerów na zbity ryj ze stanowisk i z roboty, aby juz nigdy z tym sektorem nie mieli nic wspolnego Tak ma wyglądać sprzątanie po pis ?

@prz_chojecki "Also worth noting that GPT-5.5 Pro (Web UI or API) is stronger that xhigh through codex, especially on math problems (thinks longer and deeper)." Zaskakująca i ciekawa obserwacja. Ciekawe czemu.

1) GPT 5.5 xhigh (Codex) is the best overall after external proof checks: highest judge average, full coverage, most high-yield strong claims, and no proof-audit rejections among 63 non-partial claims. Still has eight strong labels downgraded to partial/overclaimed and many single-model claims requiring expert review. Also worth noting that GPT-5.5 Pro (Web UI or API) is stronger that xhigh through codex, especially on math problems (thinks longer and deeper). For this run Codex used around 3m tokens and took 5 hours to complete.

In the full 226-problem ErdosBench run, GPT‑5.5 xhigh Codex is the strongest correctness-adjusted model. Kimi K2.7 Code is the most creative challenger but has substantially weaker solved-claim proof hygiene after strict verification. Claude Opus 4.8 max is the strongest reviewer/partial-progress model. Qwen 3.7 Max is useful for corroboration but should not be used as a lead source for public solved claims.