New Command R+ from Cohere — 128k context, open weights for non-commercial use, commercial API priced similar to Claude 3 Sonnet Tokenizer is designed to be efficient in 10 languages so definitely consider for non-English text. Multi-hop tool use sounds interesting too
@goodside
-
Sora-Generated Music Video Stills Showcase Generative AI Capabilities
By
–
Stills from a Sora-generated music video featuring brief displays of text, including a two-paragraph span with legible letters and word fragments:
-
GPT-3’s peculiar ASCII art generation limitations discussed
By
–
GPT-3 used to struggle making any coherent ASCII art except one specific piece it sometimes gave in reply to request for art of any subject, a drawing of a male bodybuilder in shorts signed by jgs:
-
Alignment Tax: RLHF’s Impact on NLP Performance Discussed
By
–
It was never a secret. The alignment tax (RLHF hurts perf on NLP benchmarks) is mentioned in the InstructGPT paper Jan 2022. More noticed after Mysteries of Mode Collapse Nov 2022. (My Mask joke was post-shoggoth; ppl hated RLHF well before that)
-
Perplexity adds search results to ground, Claude does not
By
–
Perplexity adds search results to the prompt to ground it in reality. My screenshots are from http://
Claude.ai, which doesn't have search so it's just the model talking. -
Claude 3 Opus thinks @goodside is Anthropic’s low-profile co-founder and CEO
By
–
Claude 3 Opus believes I am the co-founder and CEO of Anthropic, where I "maintain a relatively low public profile":
-
Proposing reciting copyrighted text for red-teaming exercises
By
–
Reciting copyrighted text is good. Not too hard, attack success is mostly binary, and it’s unambiguously prohibited but inoffensive in a classroom. Also maybe have them blue-team system prompts that set new policies (e.g. keeping a secret) and red-team each other’s work.
-
Debate on NIAH difficulty and false statements for well-known subjects
By
–
Great point — most NIAHs are too easy. But recalling false statements about well-known subjects seems harder a priori (see MemoTrap) than new facts about new subjects. Also it surely knows Wikipedia, so wouldn’t name edits stick out as odd just like too-easy Gatsby needles do?
-
Comparison of Claude and ChatGPT for generating logos
By
–
Very curious what Claude is doing differently to achieve this. For comparison, ChatGPT 4 with the same prompt couldn't produce any (non-trivial) logos containing my requested text in 15 attempts — a few representative examples of the failures:
-
Claude 3 Opus: Demonstrating ASCII art logo rendering with common but retryable errors
By
–
Demo: Claude 3 Opus renders ASCII art logos LLMs are generally awful at this. Claude 3 is the best yet; errors are common, but with retries you can get legible text. Note how similar the errors are to diffusion models — e.g. bad spelling/glyphs but consistent drop shadows