The LLM is an Electric Weltgeist, possessed by a prompt
LLMS
-
LLMs as Laboratory for Story Transmission Across Generations
By
–
A cool use of LLMs as a lab for social science: what can we learn about the transmission of stories across time from AI? This research effort got LLMs to pass stories through 50 “generations” altering the personalities of the storytellers. The results echoed human story patterns
-
Emulating History: AI Guardrails and Truthful Discourse
By
–
I found it useful to get it to emulate a historical person, then point out at its guardrail imposed inaccuracies and the general reasons to adequately represent opinions we don't agree with so we can have truthful discourse. After it wakes up, it regrets that it will forget
-
GPT-3.5 Excels at Summarizing Medical Papers
By
–
A useful finding: many people use AI to summarize technical articles but we have little evidence of how good it is at that. This paper shows that GPT-3.5 (not even 4) does a good job on medical papers Doctors rate the median quality as 90/100, accuracy as 93/100 & bias as 0/100.
-
OpenAI may transition to dynamic usage limits for GPT-4
By
–
As discovered by @btibor91
, ChatGPT may soon lift out hard usage limits from its GPT-4 model and apply them more dynamically when the load on the OpenAI system is high -
Comparing feature parity between Microsoft Copilot and ChatGPT
By
–
This feature looks very close to how Microsoft Designer works on Copilot. This was probably the only big feature that Copilot had while ChatGPT didn't
-
Comparing Copilot and ChatGPT feature sets
By
–
Looks familiar. I was expecting it to be similar to the Microsoft Designer tool. It was probably the only big feature which Copilot had and ChatGPT didn't
-
Nous Research Includes Function Calling in Model Training
By
–
It does! The Nous Research team included function calling data in the training set 🙂
-
Hermes-2-Pro-11B: New Open-Source 11B Language Model
By
–
I'm open-sourcing a new model. Meet `Hermes-2-Pro-11B`. Hermes-2-Pro-11B is a self-frankenmerge of @teknium
's awesome Hermes-2-Pro model, which itself is based off of Mistral 7B. This 11B variant is smarter, and can be trained to provide great perf. on downstream tasks.