dspoka

Looks cool! Any advantages to the mini-lm model - it seems better on most mteb tasks but wondering if maybe inference or something is better.

Sensational title that misrepresents the message in paper.

However, when conducting more targeted automatic evaluations, we found that the imitation models close little to none of the large gap between LLaMA and ChatGPT. In particular, we demonstrate that imitation models improve on evaluation tasks that are heavily supported in the imitation training data. On the other hand, the models do not improve (or even decline in accuracy) on evaluation datasets for which there is little support. For example, training on 100k ChatGPT outputs from broad-coverage user inputs provides no benefits to Natural Questions accuracy (e.g., Figure 1, center), but training exclusively on ChatGPT responses for Natural-Questions-like queries drastically improves task accuracy.

Just because this might not be the way to replicate the performance of ChatGPT across all tasks, it seems to work quite well on whichever tasks are in the imitation learning. That is still a big win.

Later on this also works for factual correctness. (leaving aside the argument whether this is the right approach for factuality)

For example, training on 100k ChatGPT outputs from broad-coverage user inputs provides no benefits to Natural Questions accuracy (e.g., Figure 1, center), but training exclusively on ChatGPT responses for Natural-Questions-like queries drastically improves task accuracy.

Alpaca finetuned a llama so: https://news.ycombinator.com/item?id=35186185#:~:text=Tuned%....

My main qualm with the parent comment was this in particular, "Number one, is that forests work as long term carbon storage and sinks." Even with the surrounding context it sounded like this strategy will just "work".

The example you give with mangroves is a great one which does in fact work. Pragmatically and historically most of the attempts however, have not due to mismanagement and other unseen complications.

Seeing the further comments I see the point the parent was making is more around first principles of Forrests as carbon sinks not about its implementations.

And yet research on this topic suggest the opposite. "These days everyone seems to thinks that "planting trees" is an important solution to the climate crisis. They're mostly wrong, and in this paper we explain why. Instead of planting trees, we need to talk about people managing landscapes." [1]

[1] https://twitter.com/ForrestFleisch1/status/13062214459331297...

Do you have any good write up on this? Was always interested in the how this space is still locked up.

I think the reaction from software devs on how Copilot's uses their code for ML is interesting in that all the ML companies have been doing this with all other forms of produced content: texts, posts, messages, photo captions, etc. And most likely even less care went into adhering to laws or ethics. Yes code has licenses and thus more distinct legal ramifications but on the other side are people who don't really understand that every time they interact with software or produce some content, everything is gathered and harnessed to power all these companies.

Chief Medical Officer of Intel talking about what hacks look like on people's genetic information.

https://www.youtube.com/watch?v=HKQDSgBHPfY

Is sitting in a housing court just allowed? How does one go around doing this?

Great to see you working on this!

I was wondering if you could estimate what it would cost to have always on recording of all these radio conversations, cost of running this speech2text ML and cost of labeling this data.

I think having these rough estimates will make donations easier for people.

Something I came across for this : https://app.astralapp.com/

Is there some sort of pytorch 1.0 migration guide or does anyone know if there is any breaking from .41 to 1.0 ?

Do you have any information on pricing?

Europe has also passed “right to explanation” to help with this issue which is great. unfortunately it seems they came up with it in isolation from researchers on this topic and went with the don't record it philosophy.

here is a good resource on that: https://arxiv.org/pdf/1606.08813.pdf European Union regulations on algorithmic decision-making and a “right to explanation”

So this idea seems intuitive at first but turns out to be one of the worst things to treat unfairness.

There are several reasons for this from both technical and legal perspective.

It is incredibly easy to find statistically significant correlations given just a few (more than 7) different views of the data. In general these ml models are not working with less than hundreds or thousands.

If the model learned this suppose racial bias, once, you deleting this column is not going to stop it from learning it again, and I believe some research showed that it actually can make the unfairness more severe.

from a legal standpoint a company that may or may not be infringing on rights could just say, oh we can't be because we don't have these fields in our data: which makes it harder to monitor and audit wrong doing.

most of the methods that I am familiar try to ease the effects of the learned biases as a post-processing step for the model.

pretty easy to get started on their homepage: http://pytorch.org/

for example, python 2.7/ pip/ osx/

pip install http://download.pytorch.org/whl/torch-0.3.0.post4-cp27-none-... pip install torchvision

The most promising area here to me seem like automl. The promise of the new machine learning was that we get to move away from tedious feature engineering and everything will work and be simple. It may have become simpler but training/debugging new DL models is still painful causing the focus to move to extensive hyperparameter search. automl may become the next step in abstraction, where we design single models/algorithms that are able to build viable networks for many tasks/purposes.

For your second point, what in your opinion would be the right statistical test to apply?

This seems like it's going to be a great class and topic taught my Moritz Hardt, a researcher at Google who has done research on several topics of fairness such as analysis of demographic parity.

I hope this becomes one of the most important and necessary topics for ML researchers as well as ML practitioners to consider when building models. There has been some serious discussion between researchers about adding some sort of licenses to help mitigate/ add accountability for researchers knowingly building biased ai.

what does virtual mean in this case? is this word mean something in quantum physics?

For me getting alerted when there are new papers that cite papers that are relevant towards my current research topic would be ideal. Google scholars has alerts on authors and search queries but for me they don't have enough recall.

Its much easier to tell when a paper is relevant for me if it happens to cite 3 of the commonly used datasets for my particular task.

btw I use arxiv-sanity, its pretty great, thanks a lot!

Some people have already mentioned these but so far I'm using:

Karpathy's http://www.arxiv-sanity.com/library subscribe to archive email lists

Semantic Scholar (no notifications) is good for manually finding things

Google Scholar notifies you when your papers get citations... Unfortunately they don't have a way for you to get notified if the paper is not yours.. so I made a few fake accounts that add papers to the library as if they are the author and then I set up a forwarding to my email. (really wish they would just expand the notified of citations feature to your library and not just your papers but whatever)

When I first saw this (extracting audio from visual vibrations): http://news.mit.edu/2014/algorithm-recovers-speech-from-vibr... it seemed incredible that the signal the cameras could capture was strong enough. This work is even cooler, even more non obvious and seems closer to wizardry.

I'm sorry I'm not totally sure I know what you are talking about. Care to link?

I love how on one side there is an engineer at google talking about how addictive technology is and then on the same day someone else at google releases an api to make mobile phones be even more clever slot machines... https://developers.google.com/awareness/?ref=producthunt

I happen to live in Isla Vista, the college town of UCSB. Although not explicitly in writing, the rule of this town is that bikes have the right of way. We have somewhere around 14,000 cyclists that ride to campus and back everyday. From first hand experience this is a much safer system. The drivers are much more attentive at the wheel and the heavy bike traffic enforces that cars really don't go faster than the 25 mph speed limit

How important is it for this Fellowship to have a flushed out Business Model?

Can anyone compare the pros and cons of this to princeton's wordnet?

but this is an actual product and tile is in production =)