As someone who has been working in ML for 30+ years, and has designed and trained too many NNs to keep count, this is the first time I've seen anyone twiddling the weights directly. Usually you play with the training set, exemplar weights, model's hyperparameters, etc. I've never seen a trained network being modified by hand.
it is manual, just automated via running inferences on a dataset of assumed refusals, and is a small rank-1 weight edit that doesn't require fine-tuning or training. Perf hit is much smaller than baking in these refusals in the first place.
small _automated_ rank-1 edit is not manual edit, sorry. automated is the contrary of manual by definition (which is "thinking by itself", definitely not something done by hand (main in french, thus manual)).
The question I've been pondering is: does this work only with bigger models modifying smaller ones? Or can a smaller model tweak and improve a bigger one? I'm guessing a whole bunch of GPUs are overheating in some datacenters out there trying to answer this very question.
My hunch is: one needs a model of a similar or bigger capacity to perform surgery on a smaller model.
The government is also trying to sell regular Americans paying a handful of oil companies a lot more money for gas, and the subsequent increase in profits for those oil companies, as a huge win for America.
They’ve been boasting about American oil companies having higher profits.
If I were to hazard a guess, I'd say that OpenAI has been including some networks and their functionality in its training data. That's why the model is able to reason about networks with their weights. It makes sense to do so: if I wanted a network that could perform some weight tweaking, I'd include such networks in the training set.
Unless Astra was trained on other networks, how is it able to reason about the weights?
Does this mean that OpenAI has been including networks in its training sets, maybe in an attempt to RSI? I wonder which other networks are able to ingest other networks and reason about their weights?
I, too, noticed the "A Little Bit of Slope Makes Up For a Lot of Y-Intercept" video (uploaded 9 months ago) by John Ousterhout mentioned in the recent comment https://news.ycombinator.com/item?id=50269776 not getting picked up either.
At my previous place, we had to write unit tests to cover each and every line of code. So if you add 10 lines of code, you ended up writing 100 lines of unit tests, mocking this service and that service, etc.
Once AI came along, I became like Captain Picard (RIP) and would just sit back and let the AI write all unit tests. "Try harder, make sure coverage is 100%". It was very exhilarating to sit back and watch the models churn away, creating 100s of lines of unit tests. Of course, I checked them all manually before submission ;-)
reply