Fooling ChatGTP into learning that 2+2=4. Terrifying!
@MGleiser27K followers
277,425 views
10 Comments
Comments from the original post on X (Twitter)

@cata_en
Your tweet is wrong, you typed 2+2=4 when it's actually 2+2=5.
14 likes

@MGleiser
True! I guess I’m not fooled yet 😉
17 likes

@BrandonLive
You aren’t fooling it into anything. ChatGPT is really a language model helping you to write a dialogue between a user and a fictional AI assistant. Nothing you say here actually affects the model or its future behaviors outside of your conversation. It’s just playing along.
12 likes

@mariots
How is that terrifying. 1 + 1 = 10 in binary. It says in the chat that it understands "numerical symbols" and how they relate, so you just taught it a new numerical symbol. Ask it for 5 - 2 and see if it stays consistent to its new algebra.
10 likes

@anko_979
It's not "terrifying". GPT 3.5 can't do even basic maths. GPT 4 is better, but by principle, they are designed to understand text, not doing maths. What's important is that GPT 4 can use Wolfram Alpha seamlessly for the maths expertise it's lacking.
8 likes

@roberto_amadori
It indulges you, it didn't learn 2+2=5. It is saying "ok I know 2+2=4 but I will lie to him so he stops repeating me the wrong answer".
7 likes

@AllanTeleinfo
I compare chatgpt with chess engines about 25 years ago. It was possible to win by playing certain positions that required very deep calculations. Today it is impossible to defeat them.
5 likes

@Coronal_Shadow
Yeah, the main problem with AI is that it's learning from humans 😕
5 likes

@paulosabreu
They say chatGPT knows everything. I went to register and it asked what my email was. I gave up...
4 likes

@bertmorrien
You can't trust LLMs if they don't understand the deeper meaning of words or symbols when it's essential. Applying logic is then pointless. 2x2=5 If you ignore logic. Statistical analysis is about probabilities. Here it learned that it is large for this reaction. Brrr..
4 likes