Continuing with the https://
prodi.gy annotation recipes we're building that use @OpenAI , here's some recipes for bootstrapping terminology lists. Enter some examples, have GPT3 suggest others of the same type, and quickly mark them as good or bad in the annotation UI.
@honnibal
-
Building Terminology Bootstrap Recipes with GPT-3 and Prodigy
By
–
-
Testing embedding strategy with subword features integration
By
–
Okay so yeah they do a variety of things and their experiments seem to make sense. I'd like to try plugging our embedding strategy into their thing. The subword features are pretty helpful imo.
-
spaCy Floret Vectors: New Approach for Transformer Models
By
–
Here's how we do it in @spacy_io : https://
explosion.ai/blog/floret-ve
ctors
… We're interested to try this trick in transformer models. We haven't really tried that yet. -
Budget allocation priorities: human oversight over cost reduction in systems
By
–
Let's say you have a budget of $10m to make some system that should provide long-term value. Lower costs matter less than making sure the job gets done. We're extremely far away from cutting humans out of the loop on that. So yeah we need libraries.
-
Emergent AI: Wrestling Logic Applied to Geopolitical Analysis
By
–
Here's an example of something I'd consider emergent. Let's say it learns to model texts from two domains, e.g. WWE wrestling and geopolitics. It's then able to speculate a "plotline" for the world imagining it worked like wrestling. Nobody planned that — it's a side-effect.
-
Investigating Model Behavior Through Embedding Analysis
By
–
I wonder whether there’ll be some way to probe this. Perhaps the embeddings give it away?
-
Evaluating Model Training Priorities and Prompt Engineering Strategies
By
–
Did they expect this task, and give it enough priority to try to improve it? Are the outputs that are good for my needs going to be ones it was encouraged to produce during training? What prompts did they use, and how can I make mine more like theirs? 7
-
Software capabilities supervision and emergence in AI systems
By
–
If I had to guess, I'd say that a lot of supervision has gone into the software-adjacent capabilities, as that's a very promising commercial opportunity. Most of the other things I try feel emergent? Maybe? I hope this will be clearer as other systems are released. (end)
-
Emergent Performance: Balancing Algorithm, Data, and Developer Supervision
By
–
If the performance on both tasks is emergent, I should think about the capabilities in terms of the algorithm and the data. But if the supervision is important, I need to include the developers in my mental model. 6
-
Emergent Behavior and Incorrect Mental Models of AI Systems
By
–
If I hit some pocket of supervision, but I think it's all emergent behaviour, I'll come away with an incorrect mental model about what else the system can do. I'll think, well if it can do this task, it can do this other task. 5