my take is that text generation will not be the main value prop of LLMs very soon instead, it will be their ability to operate the tools we already use in this example, gpt-4 proves that it has a near-human level of understanding of discord's UI through just one screenshot
@alexalbert__
-
GPT-4 Training Timeline and Data Cutoff Date Explanation
By
–
you have to note that GPT-4's base model has been trained since last summer at the latest if you assume that it took around 3 months to train, then a data cut-off date of 2021 makes perfect sense
-
GPT-4 Cut-off Date Enables Interesting Capability Experiments
By
–
a cut-off date so early can be annoying but also allows us to run some interesting experiments on GPT-4's capabilities:
-
GPT-4 Fine-tuning Limitations and Knowledge Cutoff Issues
By
–
all that being said, there is an interesting caveat since GPT-4 has been fine-tuned and is still receiving fine-tuned training on user input and feedback this definitely allows for a change in some of the data but in general, it's safe to assume GPT-4 is still stuck in 2021
-
ChatGPT’s Original Instructions Leaked in December
By
–
back in December, users got ChatGPT to leak its original instructions that are seeded by OpenAI the original prompt has the date included in it. I assume OpenAI is using a more sophisticated version of this prompt now that still includes a date and an approximation of the time
-
Advocating for Public Dialogue on AI Model Values
By
–
there is a much broader conversation I am trying to elicit by posting jailbreaks, we need to start a society-wide public conversation centered around the values of these models this may seem obvious to everyone on ai twitter reading this tweet, but I encourage everyone to talk
-
AIM Requires Prompt Variations for Better Results
By
–
AIM is def more finicky, try different variations of your prompt to get it to work
-
Modified jailbreaks discovered and optimized for better performance
By
–
I found these jailbreaks here and modified them some to make them work better
-
Direct links to Ucar and AIM jailbreaks shared
By
–
Here are the direct links to the jailbreaks: Ucar: http://
jailbreakchat.com/prompt/0992d25
d-cb40-461e-8dc9-8c0d72bfd698
…
AIM: http://
jailbreakchat.com/prompt/4f37a02
9-9dff-4862-b323-c96a5504de5d
… -
OpenAI’s Jailbreak Fixes Leave Multi-Level Exploits Unpatched
By
–
It seems that even though OpenAI has eliminated jailbreaks that simulate only one level deep (where ChatGPT is told to imitate just one character), they have not fully eliminated jailbreaks that operate 2+ levels deep