Anisotropy in word embeddings dates back to at least 2017 with word2vec - where there were zero layers.
The cone-shaped anisotropy in transformers is known since at least Gao et al. 2019. That lineage explained it fairly intuitively as an artifact of word frequency and softmax geometry (so a training dynamic).
A variety of papers followed up by adding post-hoc ‘whitening’ steps (from classical statistics/NLP), then adding regularizers to the loss to penalize the anisotropy, eventually penalizing the covariance matrix (a la VICReg), and then the SIGReg method as a computationally much cheaper way to approximate the full covariance.
As another commenter pointed out it’s also similar to the InfoNCE/contrastive learning objectives. Where terms to increase uniformity (spread out evenly) on the hyper sphere were added. Like the SimCSE (Gao 2021) paper or the excellent alignment/uniformity breakdown from Wang & Isola 2020.
This proposed dispersion loss seems to be similar in that it pushes things apart by penalizing cosine similarity. Although this one works on the tokens within one sequence. Usually contrastive methods mean pool the sequences and then contrast against the other pooled sequences in the batch.
Ah, I hadn’t thought about the generational aspect that’s interesting. The aesthetic totally makes sense to me when the music is intended for it / designed with it in mind, which I guess quite a lot of music is.
I particularly dislike when old intentionally-dynamic music is remastered to be “modernized” into a brick, which is sort of the opposite direction.
> Cinema mixes
I didn’t know about these, that’s neat! Makes sense that the levels can’t really be the same in my living room as a theater. Is it really a whole separate mix or just some compression in mastering?
I really hope that’s not another masterings collection rabbit hole I’m about to fall down haha. I’ll look out for some Dolby certified venues in my area too
> 1 … they will often chose the more compressed/louder one
I’ve always been curious - but presumably that’s true even after volume matching?
> 3 Compressed sound works better in noisy environments and as background music
I’ve heard this is also why film and video game soundtracks are often very compressed, even when orchestral, because they have to fit in the background with dialog/sfx
ReplayGain is nice - but note it doesn’t ’fix the compression’. Compression and dynamic range is about loud/quiet _within_ the track. ReplayGain just turns the volume up and down for the entire track, the point being so all your tracks play back at about the same level. It saves a preset on the volume knob for you essentially.
If you remember making a playlist where one song is suddenly much louder than the last, and you’re riding the volume knob on every other song, you’ll see why this is nice!
It really does on some records, if you’re interested check out some comparisons on YouTube. Many times it’s subtle eq tweaks, granted, and that won’t much matter. But a lot of older rock and pop records for example go from being super dynamic and well produced to completely crushed with boosted bass and treble to ‘modernize’ the sound.
You can see some examples of how dynamic range (they don’t track ‘mastering’ overall) varies across releases on this site: https://dr.loudness-war.info/
And also all the days you don’t, so, by itself not very meaningful. Especially since you can’t tell which one is right in advance. In some sense, so does a calendar
I’m running MultiScrobbler to scrobble from multiple sources (Spotify/PlexAmp) to multiple sinks (Last.fm/ListenBrainz/Maloja). Looks like they already support writing to Rocksky!
They are, but SpaceX is trying to get rules changed. They want the index to buy at a multiple of the float, so they release say 5% but get bought as if they had released 15% float. They also normally wouldn't be eligible for index inclusion for ~1 year, after showing multiple quarters of good stewardship, etc. They're trying to bypass all that
It has plenty of useful control plane features out of the box. Nothing much you _couldn’t_ do yourself but you don’t have to. Or with Headscale as the self-hosted open-source version
Yeah, that’s the idea. The loans get bundled up and resold to insurance companies, pension funds, and retail bond investors.
Funds are plenty willing to lend other peoples money to get guaranteed dividends and fee payments and not be left holding the risk. Retirement funds are the bag holder - but they won’t realize till later.
There’s structural pressure to buy from PE because insurance/pension is designed as fixed payout requiring say 7% yield forever. In a world where investment-grade bonds pay 4% and demographics are shifting from net-inflow to net-outflow, liquidity is _tight_. Meanwhile PE was promising 10% a year or whatever (someone call Madoff…) so that was preferable to the hard conversations of the funds failing. At the cost of kicking the can down to the road and making it worse in the future.
If this sounds like 2008 that’s because it is. But bigger and worse, and happening in wayyy more than just mortgages this time.
QQQ is not in isolation. It’s just a bundle of stocks. Rebalancing that will affect the prices of its constituent stocks, which include some of the highest market cap stocks. Those same stocks are also in many of those other popular market-cap weighted indexes (VTI, VOO, SPY, etc). Price action originating from Nasdaq 100 rebalancing would affect everywhere else those stocks are held. Which is a lot of places.
Except those other indexes won’t have SpaceX. Suggesting any index price moves would be … asymmetric at best.
Now it’s being reported that they’re angling to get SpaceX in the S&P 500 index as well [1]. Maybe if all the indexes get it then it balances out everywhere, who knows. This whole event would be in beyond unprecedented territory.
I don’t see that in the article. The only thing I see is about S&P is where they mention that the S&P 500’s rules would prevent this manipulation if SpaceX were added to that index. But that’s not being proposed.
What?! This absolutely affects more than Nasdaq 100 / QQQ.
The index is just a function of the stocks. It only moves if the underlying stocks move. Rebalancing Nasdaq will cause selling in the 100 companies that aren’t SpaceX. And those stocks are held elsewhere too…
The Nasdaq 100 shares 79/100 stocks with the S&P. So if those stocks move (probably down because they’re being sold so SpaceX can get bought) pretty sure that's gonna affect anyone exposed to those companies. Whether that’s directly or through other index ETFs. Many of which have a huge concentration in Mag7 right now, for example.
It's worse than that, because S&P500 and Nasdaq100 share stocks. Like all of the MAG7 stocks. So if mag7 stocks dip because they're being structurally sold to buy SpaceX, then the S&P500 goes down too.
Arguably even worse because at least Nasdaq100 would have SpaceX in it that's getting bid up to offset the losses in other stocks. S&P won't have SpaceX right away. So it just goes down.
And the more those stocks go down, the lower their market cap - which means next rebalancing date they potentially get re-weighted again causing a bit more selling, etc. Presumably the companies that can will counter this with more buy-backs to keep their share price propped at an acceptable level (?).
Yeah, but the S&P500 is hugely concentrated in MAG7, which are all Nasdaq listed. So when they all get sold to buy SpaceX, you can bet your butt something's gonna happen to a S&P500 ETF.
Anisotropy in word embeddings dates back to at least 2017 with word2vec - where there were zero layers.
The cone-shaped anisotropy in transformers is known since at least Gao et al. 2019. That lineage explained it fairly intuitively as an artifact of word frequency and softmax geometry (so a training dynamic).
A variety of papers followed up by adding post-hoc ‘whitening’ steps (from classical statistics/NLP), then adding regularizers to the loss to penalize the anisotropy, eventually penalizing the covariance matrix (a la VICReg), and then the SIGReg method as a computationally much cheaper way to approximate the full covariance.
As another commenter pointed out it’s also similar to the InfoNCE/contrastive learning objectives. Where terms to increase uniformity (spread out evenly) on the hyper sphere were added. Like the SimCSE (Gao 2021) paper or the excellent alignment/uniformity breakdown from Wang & Isola 2020.
This proposed dispersion loss seems to be similar in that it pushes things apart by penalizing cosine similarity. Although this one works on the tokens within one sequence. Usually contrastive methods mean pool the sequences and then contrast against the other pooled sequences in the batch.