Getting ready for the upcoming semester's #IoT course here at #ITU -
and doing some quick quiz with #DeepSeek , to see whether the new kid on the thieves' block would pass my exam.
Result: it s about as useless as #ChatGPT. Many factual errors, #hallucinations on key questions of #physics, #sensors, #LoRa, LoRaWAN, general #wireless.
@EUCommission Better laws before implementing #AI A Law that makes that every use of an AI model must come with transparency, insight in how it works and always with notable an warning that ai is used. And stop #american #lmm #AI #musk #meta #microsoft #copilot #google from intervene in our society. But the most #EU country’s and the eu itself have sold there #IT to #US #BigCo it bring us #fascism before we realise it.
Another great example how #LMM #AI produces extremely confident BS when presented with something (in this case, a rare name) not represented in the training data https://kansasreflector.com/2024/06/22/artificial-intelligence-has-spread-lies-about-my-good-name-and-im-here-to-settle-the-score/
ChatGPT is vulnerable to chat history exfilration via image markdown injection. Think “Bobby tables” but for data exfiltration!
This can be triggered via malicious data, eg images, vision, documents, web sites, plugins,.. anything that can trigger a prompt injection basically.
#chatgpt #lmm #llm #infosec #openai
Here is a demo with Code Interpreter:
