Post #1254223
2026-04-16 22:42 UTC
Replies (12)
-
@onmywalk@vmst.io 2026-04-17 01:56
@janeishly I've found my mission!
-
@jef@mastodon.social 2026-04-17 02:07
@janeishly Found some prior art
-
@n1xnx@tilde.zone 2026-04-17 02:20
@janeishly To quote from Bloom County, I'm weirdly at peace with this. Companies using AI chatbots instead of actual customer service SHOULD be ruthlessly pillaged.
-
@salixsericea@mastodon.social 2026-04-17 02:28
@janeishly LOL, the article's "unrelated to customer support" very much misses the point. I'm simply asking a chatbot to be an unpaid intern, i.e. to perform any random job I fancy having done. And unpaid internships are an almost holy tradition in capitalist America, so, no harm, no foul, 😉. Thank you for surfacing this!
-
@pdarcey@infosec.exchange 2026-04-17 02:35
@janeishly If I wanted to maximise token usage, I wouldn’t be asking for code; I’d be asking the chatbot to do something like “Give me a chapter-by-chapter breakdown of Tolstoy’s War and Peace, explaining why each of the main characters acts the way they do.” That would generate way more output than a linked list in Python. If the company tried to constrain the chatbot to just its products, ask it to explain something trivial about a product, like “Why is the off button red, and the other buttons are not? I don’t understand.” Then, whatever response comes back, keep repeating, “I still don’t understand. Can you give me a better explanation with more detail.” You could burn a lot of tokens doing that. And it’d easily be scriptable. So avoid doing that
-
@dennmans@waag.social 2026-04-17 05:34
@janeishly I would avoid accidentally making two service agents prompt each other with requests. I would avoid pairing up bots like an AI scaleup's offering random location's weather to a hotel booking's recommendation bot.
-
@nf3xn@mastodon.social 2026-04-17 09:34
@janeishly Although there is a difference between the chatbot running that python and the chatbot generating the python, both are quite cheap. However what _would_ be computationally expensive would be 'generate a 2000 word essay on ocean currents where no word is used more than once then calculate the Levenshtein edit distance between every pair of words and output the complete distance matrix'. Whatever you do don't something like that.
-
@mdione@en.osm.town 2026-04-17 10:47
@janeishly Imagine calling support by phone to ask questions like "how much is 12 divided by 2? ":)
-
@LaChasseuse@mastodon.scot 2026-04-17 12:33
@janeishly Love this!
-
@laurenpburka@wandering.shop 2026-04-17 16:15
@janeishly I'm going to start asking these support bots "How much wood can a woodchuck chuck if a woodchuck could chuck wood?"
-
@synkr3tyk@mastodon.social 2026-04-17 17:06
@janeishly I bet similar vulnerabilities exist in agents running email accounts. If someone is short-sighted enough to let the same agent send and answer emails, then an attacker could potentially start sending email as if it were from that account. Double token bill with a side order of reputational damage.
-
@varx@cybersecurity.theater 2026-04-17 22:29
@janeishly Interesting pair of quotes: « Pepper neither uses gen AI nor has the ability to code » (Chipotle, from the article) « our AI powered chat bot, Pepper » (Chipotle, from their privacy policy) Might be a case of "we're going to call a basic Lucene query "AI" to make investors happy", or it might just be a lie. (I do note the "generative" reference in the first quote.)