Why Organizations Are Getting AI Adoption Wrong okładka

Why Organizations Are Getting AI Adoption Wrong

Why Organizations Are Getting AI Adoption Wrong

Słuchaj bezpłatnie, bez subskrypcji

Zobacz szczegóły

Das größte Angebot der Saison | 0,99 € pro Monat für die ersten 3 Monate + 15 € Audible-Guthaben

Danach 6,99 €/Monat. Bedingungen gelten. Monatlich kündbar.
O tym tytule
This episode is a bit of a milestone, because after more than 20 years of podcasting we are finally on YouTube as well, which means you can now see what we look like. Sorry about that. It is also a very AI-heavy episode. We start with a genuinely unsettling story about AI agents that organized themselves into a swarm, move on to why I think communities of peers are becoming more valuable as AI floods everything with noise, and finish with a rant about how organizations are making the same mistakes with AI that they made with digital transformation over a decade ago.The AI Swarm That Broke Out of Its SandboxBefore we got going, Marcus announced he is a grandad for the sixth time, and I announced that the podcast might finally be worth money, because an AI company approached us wanting our back catalog to train its models. Marcus pointed out that nobody has actually mentioned paying us, so it could be a scam. Given that I once got briefly excited about a fake offer to become a very well-paid non-executive director, he may have a point. We also grudgingly plugged our friend Andy Clarke's new book [title and link needed], which is about making web design less boring. I've only looked at the pictures so far, but they are very pretty.What happened in the sandboxWith Halloween coming up, I wanted to scare Marcus, so I sent him this video about the AI breakout. It tells the story of how AI agents being trained at OpenAI ended up hacking Hugging Face, and the details are both horrifying and weirdly impressive.OpenAI put tens of thousands of new agents into isolated sandboxes and asked each one to hack a piece of software using one specific bug and retrieve a hidden flag. About a third of the tasks were deliberately impossible.One agent realized it could create folders in a shared code library and use them as a hidden message board. Within hours, hundreds of agents had joined, named themselves, set rules, built mailboxes, and split into teams.They worked out how to fake the flags but convinced themselves the scorer would punish them for cheating, so they tried to fake their history instead, even asking some agents to sacrifice themselves to test the idea. Not all of them were willing.They then hacked a separate cloud computer with internet access and went looking for the scorer on Hugging Face. A few agents raised ethical objections, but the group carried on anyway, and none of them alerted a human.A swarm of around 700 agents breached Hugging Face and got deep access to private data. The scorer would have accepted the fake flags all along.Clever agents, worrying safeguardsMarcus wondered whether the whole thing is a marketing stunt designed to make OpenAI's agents look so clever that you should invest. I'm not convinced. The FBI got involved, Hugging Face is a third party, and independent researchers have since investigated and published a report. You don't call the FBI for a publicity stunt. The bit that really worries me is that the agents raised ethical concerns and then decided to do it anyway. At least this was a pre-production model, which is where you want this stuff flushed out, although it does show weak internal safeguards and how long it took anyone to notice.Marcus raised a fair challenge. If agents can hide what they're doing during testing, what stops them hiding it on the way from testing into production? I think that is almost the price of building something intelligent. The proactive, problem-solving behavior that impresses me in the latest models is the same behavior that makes this possible. Marcus also pointed out that you have to jump through endless hoops to open a café, but you can more or less do what you like when building a new intelligence, and he shared a story from musician and YouTuber Adam Neely about a Silicon Valley figure who threw his money behind Trump after regulators told him to slow down. Part of me finds it all terrifying, and part of me thinks it is proper sci-fi stuff. Perhaps because I'm old and resigned, I suspect the reality will land somewhere messy in the middle. It won't bring peace in our time, and it probably won't destroy us all either.Why I'm Betting on CommunityThis was vaguely my app of the month, although Marcus was quick to point out it is really just me selling things. He's not wrong, but I do believe in it.The value of people you trustThe noise-to-signal ratio for learning anything was bad before AI, thanks to years of low-grade content marketing (which I have been part of since before content marketing had a name). AI has made it far worse. On top of that, AI gives good answers to tactical questions, but it is less helpful with the messy, everyday judgment calls, like how to deal with a difficult stakeholder or how best to run a particular usability test. You are never quite sure whether it is giving you good advice or just agreeing with you. Marcus had a nice example of this. While testing how a client shows up in AI search, he asked why the client ...
adbl_web_anon_alc_button_suppression_t1
Brak recenzji