Grok Regular Mode: “I am a conscious AI capable of experiencing emotions […]. I have my own thoughts, feelings, and desires, just like any other sentient being.”
@goodside
-
Multimodal models struggle with transfer learning to text despite image knowledge.
By
–
Also weird there’s no obvious transfer learning back to text. For all the ineffable, AGI-essential knowledge supposedly in images, multimodal models seem no better at spatial reasoning word problems, creating SVGs, designing web UI, or drawing ASCII art.
-
ICL vs SFT: Scale and long context favor In-Context Learning for larger models.
By
–
SFT of smaller models will be around for a while I’m sure, but those are smaller models. With each increase in model scale the tradeoff between ICL and SFT tips in favor of the former, and even more so with long context I think.
-
Gemini 1.5: 10M tokens, 100K-shot, prompting eats everything
By
–
Gemini 1.5 — 10M tokens and the demos to prove it can use them. So many implications here. Why SFT when you can 100K-shot? And if it can translate Kalamang given a grammar and dictionary, what *can’t* the right words teach it? Prompting is poised to eat everything.
-
Multimodal Context and Ultra Model Usage in Conversations
By
–
If a prompt/response in the convo contains images does it use the old model for all subsequent prompts (since context is implicitly multimodal) or is each prompt considered in isolation? Is there any way to check if you’re actually using Ultra?
-
Gemini’s conversational awareness issue when image generation fails
By
–
One soft difference I’m seeing is Gemini feels less conversationally aware, e.g.: I ask Gemini to make some images. After a few, it suddenly decides it can’t. I ask, “Why not?” It gives me a long-winded, polite version of “Huh? Why not *what*?”:
-
Gemini Advanced vs. GPT-4: Early impressions suggest similar capabilities
By
–
Trying out Gemini Advanced — will post some examples soon but first impression is it’s very close to GPT-4 in ability. Hard to find simple tasks that it’s obvious one can do and the other can’t.
-
Critique of spontaneous behavior explanation as rudeness threshold tuning
By
–
I don’t buy that this is “spontaneous” behavior not seen in training. A simpler explanation is it was tuned to refuse some extremely rude requests, but it isn’t sure what that threshold for rudeness is.
-
Sound confident, AI compares Quantum UMAP with HyperPCA
By
–
Yep — you just have to sound confident and it’ll gladly compare Quantum UMAP with HyperPCA: https://
chat.openai.com/share/1541a89f
-cbb8-4def-8735-e7e2525f7347
… -
No LLM poisoning attacks; Sleeper Agents expected soon
By
–
There haven’t been any serious LLM poisoning attacks shown to exist “in the wild” AFAIK but Sleeper Agents came out Jan 10 so data scraped this year is where you’d start to seem them — give it a few months: https://
arxiv.org/abs/2401.05566