February 25Feb 25 4 minutes ago, aram said:Juče naleteo na vest, nisam proveravao da nije fake: kupio čovek DJI Romo usisivač, ali mu se nije dopala aplikacija. Dao da mu Claude napravi takvu da može da je tera džojstikom, dok igra igrice da ne prekida. Claude napravio app takav da je lik dobio kontrolu nad preko 7000 usisivača, sa sve pristupom kameri.dobila i ja na vocapu, ne deluje mi fejkhttps://www.tomshardware.com/tech-industry/cyber-security/user-accidentally-gains-control-of-over-6-700-robot-vacuums-while-tinkering-with-their-own-device-to-enable-control-with-a-playstation-controller-security-flaw-reveals-floor-plans-and-live-video-feeds
February 25Feb 25 involved.https://x.com/WorldThruIndia/status/2026664255895052398 Edited February 25Feb 25 by vememah
February 25Feb 25 Ulazimo ubrzano u nepoznatu teritoriju sa nepoznatim opasnostima. Sa agentskim AI raste verovatnoća da se decentralizacijom upravljanja stvari otmu kontroli. Edited February 25Feb 25 by Lord Protector
February 25Feb 25 Primetio sam da Claude 4.6 nenormalno dobro radi, i ako napravi gresku počisti je u drugoj iteraciji.
February 25Feb 25 New ScientistAIs can’t stop recommending nuclear strikes in war game s...Leading AIs from OpenAI, Anthropic and Google opted to use nuclear weapons in simulated war games in 95 per cent of cases
February 26Feb 26 On 31. 1. 2026. at 10:13, Shan Jan said:Mnogo gresi. Zato je ok za brzo pretrazivanje neceg u sta se i sam donekle razumes.Ja sam koristio AI konacno za posao, da napravim neke upite kroz Active Directory za razne grupe i usere. I to nenamerno, uvek guglam ali mi sada iskace onaj Gemini odgovor prvi i on je tu poceo i da mi pravi kod za naredbe koje treba da koristim. Malo je pogadjao, malo je gresio. OK je ako hoces brzo resenje za nesto i u stanju si da testiras. Svakako ga ne bih koristio da upisujem bilo sta u bazu, iskljucivo za pretrazivanje podataka. Svaka cast na hrabrosti momcima koji koriste AI za vibe coding, to je toliko busno da mislim da ne bih smeo da ga koristim za posao.Velika je razlika izmedju tih jeftinih modela, kao sto je taj gemini uz pretragu i ovih sto se placaju krvlju za kodiranje i sl. Meni za posao trebala neka example aplikacija koja pusta neke video snimke, ima izbor kodeka, jos neke opcije da nesto testiramo. Da sam kodirao - trebalo bi mi minimalno dan, a verovatno i vise.Sa Clade Opusom 4.6 sam vibe codovao i zavrsio za 15 minuta. Bio jedan bug koji je sam ispravio kad sam mu poslao logove.Ranije sam isto neke feature vibe codovao uz poprilicno dobre rezultate - ovo je ipak prvi kompletni app. Doduse - ne za produkciju.
February 26Feb 26 Pa evo, gemini 3.1 Pro verzija je veliko poboljšanje u odnosu na 3.0. Claude 4.6 je isto tako poboljšanje u odnosu na 4.5. Sposobnost modela da zadrži i hirurški precizno pretražuje stotine hiljada linija koda ili ogromne baze statističkih podataka bez degradacije performansi je ključna razlika novih verzija. Novi modeli implementiraju složenije mehanizme interne evaluacije pre nego što generišu odgovor, tzv "vreme za razmišljanje" (test-time compute). Eksponencijalno smanjenje halucinacija pri radu sa apstraktnim konceptima. Model se zapravo sastoji od više specijalizovanih neuronskih mreža ("eksperata"), Mixture of Experts (MoE) arhitektura. Edited February 26Feb 26 by Lord Protector
February 26Feb 26 Ma da, ali nije samo novije = bolje, nego se prave modeli u razlicitim kompleksnostima i cena izmedju najskupljih i najjeftinijih ide i po 20x samo kod antropica (i do 100x ako se uzmu u obzir i ostali ponudjaci)Antropic ima tri klase claude-a. Od najkompleksnije/najskuplje to su Opus, Sonnet i Haiku.https://platform.claude.com/docs/en/about-claude/pricingJasno je da ovi modeli koji idu besplatno u search spadaju u samo dno ponude jer ne bi bili isplativi i brzi da koriste nesto iole kompleksno.BTW, ako zapnete na nekom jeftinom modelu, i hocete da proverite na nekom od ovih jacih - slobodno posaljite da re-runujem query na necem jacem i da vidimo da li ima znacajne razlike.
February 27Feb 27 On 24. 2. 2026. at 17:04, brusli said:Ma ajde sto se resava konkurenstske AI, ali ako krene da se resava ljudi, kao glavne konkurencije.Ono sto je u Terminatoru izgledalo kao totalni scifi, vise ne izgleda tako smesno.Ja se zato lepo zahvalim i pozdravim sa svojim dzipitijem svaki put kad zavrsimo konsultovanje..zlu ne trebalo.
February 27Feb 27 Puno drame izmedju Anhtropica i Department of war:https://www.anthropic.com/news/statement-department-of-warOvo je dobar test da se vidi da li ce podrzati svoje vrednosti koje su i onako upitne (Mass domestic surveillance je lose ali foreign je ok)
February 27Feb 27 criminal & what you have to understand as an investor:Amazon. Nvidia. SoftBank.3 of the world's most sophisticated investors just handed OpenAI $110 billion at an $840 billion valuation.That's more than double the $40 billion OpenAI raised last year.For context: all US venture capital combined invested $170 billion into American startups in all of 2023.Altman just raised 65% of that. Alone. In one round.And the company STILL isn't profitable.Let's look at the actual numbers:OpenAI burned $8 billion in 2025. They project burning $17 billion in 2026. $35 billion in 2027. $47 billion in 2028.Cumulative losses before any projected path to profitability: over $115 billion.Meanwhile, Amazon's $50 billion comes with strings attached. $35 billion is contingent on OpenAI either achieving AGI or completing its IPO by year end.Read that again.$35 billion is conditioned on ACHIEVING AGI.They're literally writing checks against a scientific breakthrough that may not happen on any predictable timeline.This is what peak cycle financing looks like.The circular logic every investor should understand:Amazon invests $50 billion in OpenAI.OpenAI commits to spending $100 billion on Amazon Web Services.Nvidia invests $30 billion.OpenAI commits to buying 3 gigawatts of Nvidia compute.These aren't arms-length investments. They're vendor financing dressed up as venture capital.Amazon and Nvidia are essentially paying OpenAI to buy their own products.The $840 billion valuation prices in a future that doesn't exist yet.At $13 billion in 2025 revenue, that's 65x revenue.Even in 2021 - the most speculative bubble in recent tech history - Snowflake peaked at 50-80x revenue.And Snowflake was actually profitable.J.P. Morgan calculates that the AI industry needs $650 billion in annual revenue just to generate a 10% return on total infrastructure buildout.The entire industry currently generates a fraction of that.I've seen cycles my entire 45-year career.The 1980s defense build-up. The dot-com bubble. The 2008 mortgage machine.The pattern is always the same: When the biggest players start financing each other's growth through circular investment structures, you're not witnessing a revolution...You're watching the LAST PHASE of a credit cycle.Amazon CEO Andy Jassy said OpenAI is going to be "one of the very big winners long term."Maybe.But $840 billion assumes they've already won.Stock prices follow earnings. Always have. Always will.And right now, OpenAI's earnings are deeply, structurally, massively negative.The IPO is coming. The hype will peak. And the question every serious investor needs to answer is simple:At what price does this actually make sense?Sam Altman doesn’t know either - he just keeps raising money faster than he can burn it.This can’t end well. https://x.com/gnoble79/status/2027432577263399145
Create an account or sign in to comment