There are some clear indications of (good quality) templating in the answers of ChatGPT Would be super interesting to know what part is from the model and what part is some rule-based templated stuff on top of the model (like the added diversity words in DALL-E prompts)
@thom_wolf
-
Health Monitoring for Large-Scale AI URL Datasets
By
–
We should have a health page with updated indicators for urls datasets like LAION or COYO 83% █ █ █ █ █ █ □ □ Self destructing datasets!
-
Data efficiency in multimodal learning for navigation tasks
By
–
I would say from the notion of data efficiency or minimal amount of cofounders behind your data. Like using text only to learn some navigation tasks (in big bench) sounds suboptimal to me But yeah of course science is fractal and everything can be considered scientific somehow:)
-
Internet Data Lottery: Web Availability Driving AI/ML Development
By
–
the « internet data lottery »: improvements in AI/ML coming from the path of largest web data availability rather than most scientifically sound direction modalities (text) win the lottery when they happen to be the widest available on the web even if sub-optimal for some tasks
-
Galactica Model Repetition Issues Decoding Parameters Analysis
By
–
I’m a bit wondering why we see so much repetitions in the outputs of Galactica people share. If it comes from the model itself or the demo decoding parameters
-
Hugging Face Releases 1.3T Parameter Mixture-of-Experts Model
By
–
We’ve heard mixture-of-experts (MoE) were in the air (GPT4??) so we’ve just added the first one in the transformers library for you to play with 🙂 With for nothing less than a 1.3 trillion parameters checkpoint model on the hub! The largest model on the hub at the moment
-
Self-examination in AI science models for error detection
By
–
what would be cool with science models like Galactica would be to add self examination of the inputs like the coming Codex model, where the model can inspect it’s theorem proofs or essays and spot errors
-
Hugging Face as ML Discovery and Sharing Platform
By
–
Yes that’s exactly the plan! I think HF could be an amazing platform for ML discovery, sharing and discussion. We see it already taking off cc @julien_c @ClementDelangue
-
Twitter ML Noise Forces Shift to Full Paper Reading
By
–
With twitter ML becoming harder and harder to grasp in the noise, I feel slightly blind. It’s crazy how quickly we become dependent on some of our tools. I’ve switched to reading in full papers instead of skimming through discussions about them. Maybe it’s actually for the better
-
Killing Low-Impact Projects: Strategic Portfolio Management
By
–
Same as not killing projects that don’t have any real impact, very hard to do but if you don’t do it along the way, you’re setting you up for tough times