This multimodal API from @Aleph__Alpha provides text prompt completions using images as context, like Flamingo. In each screenshot, the bold text is the prompt, and the non-bold text is AI-generated:
@goodside
-
Robustness to prompts for specific writing styles via instruction or k-shot
By
–
Is this robust to prompts that request specific writing styles, through either instruction or k-shot demonstration of human writing?
-
Demand for FLAN-T5 fine-tuning observed
By
–
Yeah, definitely seeing demand for FLAN-T5, though always for fine-tuning.
-
Curie available as ‘curie’ instead of ‘text-curie-001’
By
–
Original pre-trained Curie is available. The model is just “curie” instead of “text-curie-001”.
-
Who uses big open-source LLMs in production? Tasks and advantages over Curie
By
–
Who uses big, open-source LLMs like BLOOM, OPT, or GPT-NeoX in production? What tasks are they good at? What are the advantages over Curie?
-
Claude identifies ambiguity in apple-box riddle; GPT-3 misses it
By
–
Claude answers that the problem is too vague, because it’s not stated whether the apple is in the red box, and explains how to get the apple in both scenarios. GPT-3 doesn’t recognize this. I’m sure you could k-shot it with other riddles that contain irrelevant distractor info.
-
Prompt injection is a U-shaped scaling phenomenon
By
–
Prompt injection is a U-shaped scaling phenomenon!
-
RL ranking automation needs new human responses
By
–
The problem is the process ultimately depends on RL ranking generations made for the same prompt. If you need new human responses too, it can’t be automated.
-
RLHF preference for direct answers over clarifying questions
By
–
I suspect this is because of RLHF — if the goal is to choose responses a human would rank highly against other options, “Could you please clarify X?” or “Am I understanding right that you want Y?” are not usually major contenders for top place.
-
Chat-tuned LLMs hover between conversation and instruction prompting
By
–
Today’s “chat” is somewhere between conversation and instruction prompting. Chat-tuned LLMs never ask you for more information, or quickly confirm their interpretation of a complex task before starting. The lineage shows.