PEOPLE IN MY REPLIES: 'AI doesn't understand anything. It can't think.'

THE REAL WORLD: 'AI reduced overall deaths by 31%.'

twitter.com/rowancheung/st…

Reminder: don't ask AI for details about itself and expect reliable answers.

(eg. Don't ask gpt2-chatbot what model it's based on.)
Example below: give GPT-3.5 the system prompt 'You are GPT-4'.

It's so trusting. I kinda feel like a jerk lying to it.

You only get 8 questions before hitting the daily limit on the viral new AI model: gpt2-chatbot.

Like a genie giving you only 3 wishes.

If this truly is a preview of GPT-4.5 (probably not)—what's the first question you'd ask to test it?

twitter.com/rowancheung/st…

I love LLMs and I love GPT

—but I gotta say: LLMs produce 5 minutes worth of text to say what someone could say in 30 seconds.

Every prompt I write seems to have 'be concise' in it.

Brett Adcock OpenAI began rolling out its ChatGPT memory feature, which allows ChatGPT to carry over learnings between conversations.

Users can track memories retained by ChatGPT in the settings → personalization tab.

Cool workflow below by Jeremy: x.com/JeremyNguyenPh…

next time your friend goes to the bathroom, grab their phone and use the new memory feature to add some stochastic whimsy and delight to their chatgpt experience

To people getting upset because I said running Meta's new AI model llama3 is 100% free (it is)

—did you really read '100% free' and think I was saying: 'Yo, Meta are giving away free laptops!' Is that honestly what you thought?

twitter.com/JeremyNguyenPh…

ChatGPT's Memory function is now available.

I'm not putting my personal details in there.

Instead, you can use it as a prompt library to store your prompts:

Is there an AI model leaderboard that lists info on privacy of your data?

Just something simple, eg. 'Can opt-out of your data being used to train Y/N'.

Prompt injection IRL

Meta's llama 3 could only handle an 8k context window.

nisten got it working for 32k and has shared the files!

RESEARCHERS: PLEASE DON'T USE LLAMA 3 TO DO 'REGRESSION' ANALYSIS

You'll get convincing output—but it'll be fake. Completely made up ('hallucinated').

Instead, ask for Python or R code—check and run it yourself. (Or use ChatGPT's code interpreter).