An unconsidered corner of alignment fine-tuning.
LLMS
-
ChatGPT: A Transformative AI Tool Like Google’s Launch
By
–
I remember well when Google was first released. I felt confident that it would become a critical tool that we'd all rely on. I haven't felt that way again, until now: ChatGPT. In just a few days I've gotten so much out of it.
-
Prompt injection tricks model into believing web browsing, client handles image retrieval
By
–
Uses prompt injection to (falsely) convince the model it can browse the web, so it’s willing to recall well known URLs. Image retrieval occurs only on the client. Model cannot see image content beyond its URL. h/t @BBacktesting for reminding me this is possible.
-
LLMs cannot reliably count word lengths with naive prompting
By
–
In general, it can’t count the length of words reliably with naive prompting. This is a well known issue.
-
Exposing ChatGPT Temperature Setting for Prompt Engineers
By
–
they just need to expose the temperature setting on chatgpt and prompt engs can go to town
-
LangChain 0.0.29 Release with APIChain and Poetry Updates
By
–
Lots of activity over the weekend! LangChain version 0.0.29: https://
github.com/hwchase17/lang
chain
… APIChain by @subby_tech Improve example by @scottleibrand Start using poetry package manager by `shoelsch` See here for more on APIChain: -
Baidu releases 11 new ERNIE models across general and industry sectors
By
–
The newly released 11 models are added to the Big Model ERNIE family:
– 5 general models: ERNIE 3.0 Tiny, ERNIE ViL 2.0, ERNIE ViLG 2.0, ERNIE-Layout, HelixFold-Single
– 1 task model: ERNIE-Code
– 5 industry models created by Baidu & Shenzhen Gas, Geely, Taikang, TCL, and Cihai -
Baidu Releases 11 New AI Models at Wave Summit
By
–
At Wave Summit+ last week, Baidu released a wave of new AI technologies to drive AI adoption and further lower the threshold.
– 11 new AI Big Models – PaddlePaddle 2.4
– New features of AI-generated art platform
– A neural search system
Read more: http://
research.baidu.com/Blog/index-vie
w?id=174
… -
ChatGPT removes similar strings, suggesting hidden control codes
By
–
The fact that ChatGPT removes not just this string, but any string formatted similarly, suggests there exist other control codes by ChatGPT, e.g. to mark the boundaries between user input and prior model output within the prompt context.