lapink

The alignment between source and target is automatically inferred, basically by searching when the uncertainty over a given output word reduces the most once enough input words are seen. This is then lifted to the audio domain. In theory the same trick should work even with longer grammatical inversions between languages, although this will lead to larger delays. To be tested!

Adding more languages is definitely planned! This was Tom (the first author) master’s internship project with Kyutai, and it was easier to prototype the idea with a single pair. Also he will be presenting this work at ICML in two weeks if anyone is around and wants to learn more.

The model can actually run on an iPhone 16 Pro, so if the earbud is connected to one that could work!

Most stuff is fine in France.

Dr. Jean Remi King is leading the research effort at Meta AI Research for computational neuroscience, including this work. You should check out his website to learn more : https://kingjr.github.io/ Amazing scientist, previously at ENS Cognitive studies department in Paris.

Precooling is only viable if homes are really insulated and / or all the electricity comes from renewable as otherwise it will increase the overall energy used (as precooling is less efficient), which will increase CO2 emissions. In think the population also needs to “meet in the middle”, for instance by increasing the temperature of the AC by one degree in summer. This could get automated to offer a best effort service.

“”” Philip Morris crowed about its high rating on their website saying, “Altria [Philip Morris parent company] is proud of its internal and external efforts to promote LGBTQ equality. In 2020 Altria signed on to America Competes, a pledge to oppose discrimination and to support comprehensive nondiscrimination laws that protect all Americans, including LGBTQ people, and supported the passage of the Virginia Values Act, the first comprehensive LGBTQ anti-discrimination legislation to pass in the south.” “””

The part on LGBTQ+ is in bold in the article. Clearly it is more oriented at fighting civil rights than the global warming.

This is an anti LGBTQ+ article i think.

“”” What all this means is that the United Nations, through the respected Standard and Poor’s, is using its woke agenda to control corporations. “””

Made it to level 7! Same technique worked for three levels in a row just before that. Quite fun, I can imagine how in future video games we will try to convince the final boss to just let us get the treasure without fighting.

The technique developed seems to work for a larger number of qbits (49). With the current systems, increasing the number of qbits does not lead to an increase in computational performance as the extra qbits introduce more noise than can be compensated by error correction codes.

Here, the authors show a small but significant improvement by increasing the number of qbits thanks to their improved ECC.

The paper: https://www.nature.com/articles/s41586-022-05434-1

I spent a long time getting charged for buckets on an account I had lost all access. It was years ago and I had let it keep on for too long, big mistake.

First priority is getting your data out to a safe storage. Then delete as much as you can while you have your access.

I eventually did a charge back to make the charges stop. I had started the account while in the US but had definitely left 6 years ago, and really wanted to close my account there. You cannot close an account with a recurring charge in the US. Funny country, as a French it felt terrifying like being trapped in some weird American administrative maze where I would get billed 1 dollar per month until my death (plus bank fees).

Open AI is logging everything for future training. I don’t think Open AI employees actively look at the logged data, but they likely have people rate the exchanges or suggest replies, so that could be a source of leaks. Besides in 6 months / 1 year, when an updated version is released, Chat GPT might very well spit out some private info information.

I’m using Lastpass and Monterey… now I have to change password manager and OS on the same weekend.

I live in Paris. For tourists or travelers the alternative are not really ready. Ticket on smartphone support is bad (only specific models, and no iPhones) and unlike London or Brussel, you cannot use your credit card as a ticket.

So your only hope is to already have a dedicated rechargeable card. But this card is not sold in the automated machine. So if you are a tourist in some small station and you realize at midnight that you need to take the subway, there will be no way to buying a ticket…

Im all for ending those tickets, but the RATP is really bad at supporting alternatives.

Yes there are, you can have a look at https://github.com/etzinis/sudo_rm_rf for instance for 2 speakers separation. There is also this one for 3 speakers: https://huggingface.co/speechbrain/sepformer-whamr

The goal of Meta AI Research is to do open research, not necessarily with direct applications at the time we start it. Indeed, the architecture, or the lessons learnt working on it can become useful later for the company, for instance for remote presence with VR, to isolate the main speaker from its environnement ( https://arxiv.org/pdf/2206.15423.pdf ).

Original Demucs author here. Thanks for putting forward our research!

I’m definitely happy to see more front ends for Demucs being developed and to read that it has been useful to other musicians!

We are working on the next iteration of the model, and with more sources, hopefully released by the end of the year :)

If you are interested in this research you can follow my Twitter (@honualx) or star the Demucs repo.

Pytorch and Python docs, all down. No stackoverflow. I guess this is a forced bank holiday for developers around the world.

Link to the source code: https://github.com/facebookresearch/denoiser You can try it really easily if your CPU >= 2GHz and you have a loopback audio interface. On Mac OS X you can use Soundflower, I welcome advice for alternatives on Windows and Linux (I think Pulseaudio or Jack can do this).

If you are interested to try it out, this is the source code: https://github.com/facebookresearch/demucs

If carbon offset is done well, wouldn't it be equivalent?

I never get the argument that because we are not solving the entire problem at once we shouldn't make any step in the right direction. Changing mentalities is a big deal. Eventually it can trickle down to our entire way of life and make for a smooth (r)evolution, with no one missing the old world.

I made a benchmark a few months ago for a friend of mine working on computational physics. She needed to compute the determinant of a lot of matrices to perform MCMC sampling and wanted to know how much GPUs could speed up things. For this task, one V100 was equivalent to a 100 cpu cores, so a speed up of 100,000 is definitely not coming only from using GPUs vs CPUs.

Black produces code with more lines than yapf. Sure sometimes yapf makes very weird decision (especially when using a dictionary litteral as an argument to a function) but at least it does not use 4 lines for a list with 2 items that could have fitted on one line.

It seems hard if not impossible to prevent illegal use of profile pictures. If a web browser can access them, collecting them can be automated. Even if they are only accessible through an app on iOS or Android, you could still, on a device with root access, capture the API calls and reverse engineer how to get those pictures. So we might just assume any intelligence service already has that. The only protection against a more widespread development of those products would be legal I think.

Indeed, here is the mail from Raymond Hettinger explaining the new design on the python-dev list: https://mail.python.org/pipermail/python-dev/2012-December/1... Quite amazing that it is faster, more memory efficient and convenient for the end user.

Maybe Americans could start changing their way of life rather than waiting for « political will » or a never coming technological revolution. UE has started reducing CO2 emissions will keeping a very reasonable and increasing level of life.

Previous discussion https://news.ycombinator.com/item?id=22093787

I agree that eventually this could work because on some training examples, two related entries will be in the same bucket. However, I'm not sure this would really scale to parsing an entire book all at once like the author suggest. While the algorithmic complexity might scale, the odds for two related items that could be chapters appart to end up in the same bucket seems so close to zero that training time would explode.

In particular, this approach removes all kind of domain knowledge. For images, it means ignoring entirely the prior that neighboring pixels are related, which is typically encoded through the use of convolutions. With a Reformer, not only does the locality behavior need to be learnt from scratch, but on top of that it will only happen after a sufficient number of iterations so that neighboring pixels do end up in the same bucket.

For parsing books, I think it would make much more sense to build a hierarchical model with one part parsing only a paragraph at a time and generating an intermediate embedding that could then be used as a representation of the paragraph in a larger scale Transformer working over entire chapter, and then another level going from chapters to the entire book, rather than putting all the words at once together in a giant Reformer with no domain knowledge at all and praying that with enough training data and epochs, the model will learn everything from scratch.

There is no argument for why the LSH would work well, especially at the beginning of training. As the weights are initially random, bucket assignment would be random as well. If predicting at position A requires info from position B, but they are not in the same bucket, there will be no gradient to get the query embedding of A closer to the key embedding of B. The reversible layer trick is neat though.