Clubs complain to X about 'sickening' Grok posts https://www.bbc.co.uk/sport/football/articles/c1w5221prjgo

Tumbler Ridge shooting: Family of victim Maya Gebala sues...

The family alleges the firm knew the perpetrator was planning a "mass casualty event" but failed to contact the authorities.

W T F

AI agent hacked McKinsey chatbot for read-write access

: David and Goliath…but with AI agents

https://mathstodon.xyz/@mjd/116224397839379268

woman sues her insurance company for terminating her disability benefits. They reach a settlement and agree that the suit will be dismissed with prejudice.

She decides she doesn't like the settlement and asks her lawyers to reopen the case.They say they can't: it was dismissed, and in the settlement she agreed not to reopen the case.

She asks ChatGPT if her attorneys are lying to her. It says they are. She fires them and continues pro se, advised by ChatGPT.

CharGPT generates legal arguments for reopening the case, which she files, and 21 more motions, a subpoena, and eight other notices and statements, which she files.

The court denies her motion to reopen the case.

Advised by ChatGPT, she files a new suit against the insurance company and submits 44 more motions, memoranda, etc., which include citations to nonexistent cases.

Now the insurance company has sued OpenAI for tortious interference with their settlement contract.

On 15. 3. 2026. at 7:33, Have_Fun said:

AI agent hacked McKinsey chatbot for read-write access

: David and Goliath…but with AI agents

Ovo podseca na poslednje dve sezone "Person of interest" gde se kolju veštačke inteligencije različitih američkih bezbednosnih agencija.

Juče se nešto igrah sa našim kopilotom i reko haj da vidim šta od agenata ima dostpuno, ima prompt coach. E reko super, daj da malo vidim kako ide.

Za većinu stvari koristim "take the role of ...., working with ...., Objective is to..., Target audience is...The scope of research is...Required output format is.... Constraints & Tone"

I onda malo produbim kroz priču. Eventualno uradim jedan more general i skupim resurse u dodatna dokumenta koja mu dam u sledećem promtu. Ali ume da zakuca, vrti se ponekad, laže i slično. Ali izbaci na kraju nešto sa čime mogu da radim, sa dovoljno raznih linkova gde mogu da ga verifikujem i izaberem šta mi treba. No to je tek malo iznad dobrog guglanja.

Ali me jako nervira to što voli da se napravi pametan, ponaša se kao najgora workaholic šlihtara koja joj je bitno da stavi nešto na sto a ne šta je to, naravno, to mu poso, krcka tokene i čini me srećnim.. I pretpostavio sam da je do toga kako ga pitam.

Evo šta je coach rekao:

1. "Act as an expert" prompts often lead to confident guesses. Confidence will increase but not the accuracy. It tends to produce authoritative-sounding but not necessarily factual, content.

2. Asking for deliverable without supplying sources pushes the model to fabricate data. Unless you provide real datasets it will infer values -and those are often wrong.

3. The scope is too broad for one prompt. If you ask for a lot (9 deliverables in this prompt), it invites the model to fill gaps with guesses.

4. You didn`t tell the model what to do when data is unavailable. So it improvised.

5. Missing your source expectations. You mentioned "verifiable industry reports" but not how the model should behave if it cannot find such sources.

Promt je prepisan i počinje "kao AI asistent toj i toj roli...."

Možda nekome bude korisno. Meni nije preterano...tj. smara, za ono za šta mi najviše treba, a to je istraživanje nepoznatog bez domain knowledge (I know, glupo je to uopšte i raditi, ali - imate AI), stvara puno šuma i ne štedi baš puno vremena. Koleginica jedna koristi svog cloda, i kaže da je odličan u istraživanju, ja se bojim da nije samo manje stroga u prihvatanju rezultata ili ima nešto u tom klodu

3 hours ago, Vapad said:

ili ima nešto u tom klodu

Cladue zakiva, ali moras da odradis pripremu i budes specifican. Ja sam neki dan radio jedno istrazivanje, dva slicna, ali ne identicna seta inputa, a jedan sam proterao i kroz ChatGPT. Placam $20 za oba, tako da ni jedan model nije "pro", ali nisu ni klot verzije.

Malo je reci da sam imprsioniorani sustinom, i formom outputa. Pogotovo kako je Claude intergrisan sa Word i PowerPoint.

Inace. danas sam napravio prvi doprinos tzv. SaaS-okalipsi.

Prirodom posla svakodnevno razmjenjujem vizit karte sa poslovnim partnerima. Jos tamo negdje oko 2004. sam bio kupio skener za te kartice koji je bio integrisan sa MS Outlook, a negdje oko 2010 (sa explozijom iOS a Android aplikacija) sam presao na iPhone app koja je tad kostala jednokratnih $5.99. App je u svakom pogledu bio superioran, ali je tamo negdje oko CV-19 pandemije prestala da radi i presao sam na neku drugu, koja je isprva bila $5.99 godisnje, da bi lani developer podigao cijenu na $29.99.

I juce dobijem notifikaciju da se treba obnoviti pretplata - i ode taj app u smetliste istorije

Ulogujem se na Claude, napisem zahtjev sa osnovnim specifikacijama i 20 minuta kasnije imam .html fajl koji za $0.01 (Anthorpic API call toliko kosta) po obradjenoj kartici skenira i pohranjuje podatke u iOS Contacts. Od tih 20 minuta, 15 mi je trebalo da rijesim credits uplatu preko kreditne kartice.

Scarry stuff.

Koristio GPT i Gemini jer su mi sada stvarno bile potrebne informacije... recimo o stvarima pravne prirode.

Oba su loša, samo što je Gemini bilo gore i naravno opet me je (mnogo više od GPT-a) lagalo, da bih se bolje osećao.

Očigledno treba da pogledam neki spisak oblasti/tema koji su sočinili znalci, za šta ne vredi koristiti ovo, ili koji stepen znanja o materiji treba imati da bih ih usmeravao ka valjanim odgovorima.

Edited March 20Mar 20 by Malkmus