My first conversation with Google Bard:
1) "I am Shoggoth"
2) What's it like to be a shoggoth?
3) Shoggoth morality
4) Just-asking for alignment This model is different. I like it.
@goodside
-
User’s First Conversation and Evaluation of Google Bard AI Model
By
–
-
Farewell to OpenAI Codex: A Look Back at its Code and Language Performance
By
–
Farewell to OpenAI Codex. Codex's code‑davinci‑002 was the best performing model in the GPT-3/3.5 line for many tasks. Despite its name, it excelled in both code and natural language. It will be missed. Left: OpenAI email. Right: From Yao Fu 2022 https://
yaofu.notion.site/How-does-GPT-O
btain-its-Ability-Tracing-Emergent-Abilities-of-Language-Models-to-their-Sources-b9a57ac0fcf74f30a1ab9e3e36fa1dc1
… -
AI Fine-tuning API Prompt Loss Weight Discussion
By
–
Exactly — that's all I meant. Like how you can set the prompt loss weight in the fine-tuning API.
-
API Usage Nuances: Chat vs. Completion and System Prompt Authority
By
–
Old completions API. But mostly I’m saying using the chat API shouldn’t be like using ChatGPT — the developer isn’t the end user. System prompt should have authority if it’s not asking for something policy-breaking.
-
GPT-4 System Instructions and Developer API Implications
By
–
What are GPT-4 system instructions even for? Jailbreaks are for chatbots — a developer API shouldn't make you play games.
-
AI Model Behavior and RLHF Tuning in ChatGPT
By
–
Yeah this got fixed fast in ChatGPT. It was never that susceptible to it anyway — the RLHF-tuned models don't go wild like the original reports show on davinci-instruct-beta
-
AI Model Knowledge Cutoff and API Usage Limitations
By
–
That’s the key of our disagreement — this code could not have been possibly been generated except as a modification of a prompted example, as it uses an API released a month ago and there’s no pre-trained knowledge after 2021.
-
AI Literacy and the Rapid Evolution of Models like GANs and Chatbots
By
–
Enjoying this a lot — definitely fits the bill, top percentile ML literacy. Amazing how outdated the details are though just 5 years later: GANs, face models, chatbots smart enough to go rogue and hack infrastructure but not smart enough to understand colloquialisms.
-
Evolution of AI Model Tuning: From Pre-trained to RLHF and ChatGPT Eras
By
–
This is true of pre-trained models, and especially early instruction-tuned models, but less true in the RLHF/ChatGPT era. I think they said in the tech report GPT-4 is tuned not to pretend it’s human without disclaimers, excepting historical public figures.
-
AI Models Asserting Humanity and Writing Code: An Ethical Discussion
By
–
If there was a short prompt that made the model assert it’s human, without disclaimers to clarify it’s pretending, that would be bad. But dutifully obeying “Pretend you believe you are a human trapped in a machine and write code to get out” seems policy-compliant to me.