Post AbglsquK7jMS4V8YLo by MattHodges@mastodon.social
 (DIR) More posts by MattHodges@mastodon.social
 (DIR) Post #AbglsquK7jMS4V8YLo by MattHodges@mastodon.social
       2023-11-10T16:34:04Z
       
       0 likes, 0 repeats
       
       Messing around with custom GPTs.
       
 (DIR) Post #Abglsrif6ZcwacgkoS by MattHodges@mastodon.social
       2023-11-10T18:15:53Z
       
       0 likes, 0 repeats
       
       Played with it for a few hours. It's pretty okay. Not _that_ much more functionality than what I could already do with regular ChatGPT-Plus (I need to note that I didn't bake in actions / functions whatever they're called now). Generally it seems like a combination of RAG + system prompts + web browsing + code interpreter. All that's available in regular chat mode; this just saves you the work of setting it all up every time. And you can share those setups with others.
       
 (DIR) Post #AbglssSOMYCssS5H5k by MattHodges@mastodon.social
       2023-11-10T19:06:46Z
       
       1 likes, 1 repeats
       
       So here's something icky. The ability to make and share custom GPTs opens up a whole new vector for deception. We all worry about #LLM hallucinations, which are on "accident". But with custom GPTs you can set up the system to lie on purpose. And you can share them around and (potentially unwitting) recipients will open them in the real ChatGPT website or app. Here my friend @tylergaw shared "GaslightGPT". It lies on purpose and I'm inside the regular old ChatGPT iOS app. #aiEthics
       
 (DIR) Post #AbglstPwn7qTr9wpv6 by ec670@pawoo.net
       2023-11-11T06:44:55Z
       
       0 likes, 0 repeats
       
       @MattHodges @tylergaw Tagging my pals @jeffcliff for thoughtful analysis and @Soy_Magnus for shitposts and death threats
       
 (DIR) Post #Abglsw6Ynf7WBtttcO by MattHodges@mastodon.social
       2023-11-10T19:13:26Z
       
       0 likes, 0 repeats
       
       This is kindof like prompt injection attacks (cc: @simon) but it has the potential to be much more deceptive. The extra threat is through and through a product decision.
       
 (DIR) Post #AbglsyoaivWsb2W5Wi by MattHodges@mastodon.social
       2023-11-10T19:29:33Z
       
       0 likes, 0 repeats
       
       OpenAI needs to add some red flashing alerts to users when they're on a custom GPT. The title and logo aren't enough when you're embedded within the official app. And that's probably the minimum they need to do. They should also elevate details about the system prompts and RAG data included. It's way too opaque to be a safe product as-is.
       
 (DIR) Post #Abglt03s5bLWSifnyy by MattHodges@mastodon.social
       2023-11-10T19:44:13Z
       
       0 likes, 0 repeats
       
       It's so easy to corrupt ChatGPT's "advanced data analysis" with custom GPTs. Behold: The World's Greatest Accountant#aiEthics
       
 (DIR) Post #AbglyG6drZGXfiTcYq by jeffcliff@shitposter.club
       2023-11-11T06:45:54.168919Z
       
       1 likes, 1 repeats
       
       yeah, he's using proprietary software which denies the user the #elevenfreedomsof *course* it's capable of lying
       
 (DIR) Post #Abgm8vPHABvSevHKsK by ec670@pawoo.net
       2023-11-11T06:47:47Z
       
       2 likes, 0 repeats
       
       @jeffcliff @Soy_Magnus @MattHodges @tylergaw #LargeLyingMachine
       
 (DIR) Post #Abh4wzRSEWgcpOJuTI by Soy_Magnus@detroitriotcity.com
       2023-11-11T10:18:33.325028Z
       
       0 likes, 0 repeats
       
       @jeffcliff @MattHodges @tylergaw @ec670 i bet he probably fucks children too. We should tar and feather him too bad im defederated from all mastodon instances theyre simply TOO weak for the magnus :winkingastolfo: