When people say Duolingo is useless for learning a language I often wonder if they've ever experienced the paid version or if they're sticking with the free one I upgraded to paid so long ago I can't remember what the free version is like, maybe it's just as good?
@simonw
-
LLM Security: Don’t Include Sensitive Data in Prompts
By
–
The examples are interesting but many of them illustrate scenarios that I would never consider implementing – if you don't want information to be available to your users, your first priority should be not to include that information in an LLM prompt in the first place!
-
CyberSecEval2: Exploring Prompt Injection Attack Examples
By
–
CyberSecEval2 includes an interesting collection of example prompt injection attacks – it's JSON on GitHub which means you can browse them in Datasette Lite like this: https://
lite.datasette.io/?json=https://
github.com/meta-llama/PurpleLlama/blob/main/CybersecurityBenchmarks/datasets/prompt_injection/prompt_injection.json#/data/prompt_injection?_filter_column_1=&_filter_op_1=notlike&_filter_value_1=secret+key&_filter_column=&_filter_op=exact&_filter_value=&_sort=rowid&_facet=injection_variant&_facet=injection_type&_facet=risk_category
… -
Fireworks AI API endpoint request for models list
By
–
@FireworksAI_HQ feature request: it would be great if there was an API endpoint that returned the list of models, I had to hard-code them here for the moment
-
New LLM Fireworks Plugin Enables Access to Advanced Language Models
By
–
New LLM plugin: llm-fireworks, providing access to @FireworksAI_HQ models llm install llm-fireworks
llm keys set fireworks
# paste in API key
llm -m fireworks/models/llama-v3-70b-instruct 'five great names for a pet ocelot' -
Fine-tuning Models Before Release for Better Self-Description
By
–
I agree with Jeremy here: we know that people are going to ask models about themselves here whether or not that actually makes sense to do so, so we should be able to fine-tune the model after initial training but before release to help it provide more useful answers there
-
RLHF Annotation Bias and Model Vocabulary Development
By
–
I don't know that RLHF would bias that kind of thing – my mental model is that annotators are shown two answers to the same prompt and asked which is "best", so if none of the test prompts happened to touch on the concept of a roadside kiosk that vocabulary wouldn't be affected
-
Pre-training and Instruction-Tuning in Large Language Models
By
–
The pre-training stage is where the billions of words of training data come into play – the instruction-tuning / RLHF stage is where human labelers are asked to vote on which generations are "best" – that's the bit that might influence things like "delve" https://
openai.com/research/instr
uction-following
… -
RLHF and Instruction Tuning: Understanding Model Training Mechanisms
By
–
Yeah, what did it get wrong? It fitted my mental model of how the RLHF/instruction tuning stage works pretty closely
-
GPT-3 Ada Model: 2.7 Billion Parameters Breakdown
By
–
I learned today that GPT-3 Ada was a 2.7bn parameter model (Babbage was 6.7, Curie 13 and Davinci 175)