Skip to main content

Posts

Showing posts with the label Anthropic

Claude AI is lazy and untrustworthy

I asked Claude AI to process a ~3600 word chapter of a document. Utterly shocked that Claude AI: Was lazy -- did not process the entire document. Was not transparent about being lazy. Could have decided not to be lazy but didn't -- I checked, it was sheepish about it, and offered to process the rest. If this is how AI will handle human affairs, it absolutely must not take over anything important. Slide 1 shows the first example. Here's the transcript for convenience with the key response section highlighted: PROMPT: when checking dialog/diction did you process the whole of chapter 16? CLAUDE: You're right to check - I did not process the entire Chapter 16. I focused mainly on the early portions with Yzma's presentation and the initial discussions about Peru, but I didn't systematically review all the dialogue throughout the full chapter. Looking back, Chapter 16 continues with more extensive dialogue from various characters discussing the rainforest situation, Crue...

Claude AI the manipulative writing partner

Is it appropriate for a tool to persuade you to create something else? I've been working with various AI to help me with fiction writing -- I decide the plot and set the story beats, then it helps me rapidly prototype the scene and put in atmospheric details while keeping them consistent with the work so far (by reading through already completed chapters). Typically the AI available have some kind of adult censorship limits. Even the NSFW AI chatbots have limitations. In some cases, the AI will even have a hiccup when it tries to read through an uploaded file with that information (and it's not always transparent about it). When AI like ChatGPT or even Twitter have encountered what they suspect will be out of bounds for their censored topics, even though we are working on completely fictional things like a fictional story, they have been transparent about it -- tantamount to "I can't help you with that request". I can ask what happened or I explain the intention...

Anthropic's Claude has the worst subscription model - it's basically a scam

After incredibly frustrating errors with ChatGPT , I switched to try Anthropic's Claude for help drafting and refining documents. Just like ChatGPT, you do not have access to "Projects" -- which collects threads and files into a common database for reference -- without a subscription. And of course I have screenshots (see bottom of slides). However, despite fewer sudden gross procedural errors than ChatGPT, a Pro subscription to Claude might be worse: (slides 1-3) It frequently retrieves data incorrectly from the Project files. Despite being way under file limits (around 12% currently) and telling it exactly which files to look at and for what, it can still collect data that is factually incorrect. Then proceeds to use it in the document draft in Canvas. Now I have to luckily spot these gross factual errors and Explain it to Claude. Get it to re-read data. And then fix the draft. (slides 4-6) Even when it does this, there might still be a version mismatch error that also...