Exceptionally good overview of adversarial attacks on LLMs. Many papers in here were new to me; highly recommend
@goodside
-
Warning: Public GPT demos can expose uploaded data to users
By
–
Be careful what you upload as context when creating GPTs for public demos — anyone that talks to your GPT can just ask for a download of the data:
-
Use DALL-E file ID to avoid expensive tokenization
By
–
look at the Files docs — you get a File ID from dalle you can retrieve. you definitely don’t want to be tokenizing encoded images through the model; crazy expensive
-
Correcting PDF docs lacking full examples with manual instructions
By
–
Mine didn’t work well from PDF docs because there’s no full example on a single page (URL) — e.g. the Assistant tutorial never does `client = OpenAi()`. I just stuffed corrections of observed failures into the instructions. It only queries the KB for obscure stuff now.
-
Announcing a custom GPT ‘assistant-assistant’ for ChatGPT Plus
By
–
Try it here (requires ChatGPT Plus): https://
chat.openai.com/g/g-KkTq7dkUB-
assistant-assistant
… If you hit any issues comment with a share link and I'll tweak it — no maintenance promise though; just trying this out: -
ChatGPT answers questions about nonexistent images using prefix syntax
By
–
ChatGPT answers questions (within reason) about images that don’t exist when prefixed with the syntax shown — the file name is arbitrary:
-
128K context: no proportional attention increase
By
–
Enjoyed this experiment — 128K context is great but you don’t get proportionally more attention. Each output token can only consider so much. Translation — fine
Extraction, doc Q&A — maybe?
Parity bit of input — absurd; the model is not a god -
Why LLM evaluation requires out-of-distribution tests like left-handed violin
By
–
we judge LLM performance out-of-distribution for the same reasons we only care how well a violinist plays left-handed if you need the bow in your right how do we know you know paganini maybe you memorized youtube
-
OpenAI JSON mode docs and deterministic output tutorial
By
–
Docs shown above: https://
platform.openai.com/docs/guides/te
xt-generation/json-mode
… OpenAI cookbook also has a new tutorial notebook on deterministic output: https://
cookbook.openai.com/examples/deter
ministic_outputs_with_the_seed_parameter
… -
New GPT features: JSON mode and deterministic output
By
–
Two new GPT features you might have missed today — ChatCompletions API now supports:
– “JSON mode” with constrained generation (guaranteed to parse)
– Deterministic output using a seed for reproducibility