Lightnews — Scholar-powered news

Shubhendu Trivedi @shubhendu.bsky.social · 2h

unsustainable churn.

The last few few posts were posted much later, because I was searching if such (integrated) studies exist. My first impression is that they do not. However, I still feel like they should, so I am going to look again. Otherwise, it might be worth thinking about writing one.

1

Shubhendu Trivedi @shubhendu.bsky.social · 2h

The nature of the "bubble" is different IMO (stock markets, risk capital concentration), and will correct eventually (it will look ugly). But that is of a completely different nature than the one implied in said discourse, which seems to suggest the bubble is _demonstrated_ due to supposedly

1

Shubhendu Trivedi @shubhendu.bsky.social · 2h

Oh, sorry. Just google AI data centers bubble etc., or similar. Lots of coverage all over the place to absurd stuff, and not just by the likes of crazy AI journalists.

Shubhendu Trivedi @shubhendu.bsky.social · 2h

estimates for productive output (and the coupling has its own parameters). For each component, however, there need to be separate studies that can help arrive at estimates for required parameters (let's say sector-wise). At least for the infra/churn bit, they exist aplenty.

1

Shubhendu Trivedi @shubhendu.bsky.social · 2h

company, I had learnt of stylized diffusion models of adoption (e.g. Bass diffusion), which seem simplistic on the surface, but are surprisingly powerful to arrive at better estimates for penetration. Both (diffusion dynamics, infra/churn/capital dynamics) have to be coupled together to arrive at

1

Shubhendu Trivedi @shubhendu.bsky.social · 2h

I am going to check what kind of studies/papers exist on forecasting value re AI and tallying it with infra spending/churn. But most of the above discourse (which seems to have quite some space in the media too) seems entirely focused on the infra layer. In a project for forecasting sales for a food

1

Shubhendu Trivedi @shubhendu.bsky.social · 4h

You could take a few verticals outside tech (and ignore tech altogether) -- banking, finance, pharma, healthcare (add consulting as a multiplier), and just v. partial capture of enterprise workloads there should invalidate the premise of that entire line of thought. Diffusion lag isn't unviability!

1 2

Shubhendu Trivedi @shubhendu.bsky.social · 4h

and what blockers might limit gains. I get the appeal of doing an accountant's snapshot of data center depreciation schedules. Even AI people, who are not even skeptical seem to limit to single-sector profit calculations ("not even a big bubble" / "underlying tech fundamentally sound").

2 2

Shubhendu Trivedi @shubhendu.bsky.social · 4h

The whole thing around "AI" might certainly be (and indeed is) a bubble in many ways, but all of this seems like a misguided attempt to think it is. Real skepticism would require understanding (+ tabulating, even if in a back-of-the-envelope sense) who captures value, how fast, how does it diffuse,

1 2

Shubhendu Trivedi @shubhendu.bsky.social · 4h

I mostly use this website to engage in soliloquies and don't pay much attention to what goes on here. However, occasionally, I run into the "AI infrastructure ROI" discourse, which seems quite unique in its lack of understanding of vertical diffusion dynamics and also basic numeracy.

2 2

Shubhendu Trivedi @shubhendu.bsky.social · 6h

I am able to come up with a list of quite a few. But I lack the taste to decide which ones to persist with. Otherwise will require quite a random walk to hit on a reasonable combination.

1 1

Shubhendu Trivedi @shubhendu.bsky.social · 6h

I want to emphasize that I explicitly want to separate post-transformer familiarity and post-domain ML attitudes, that came with the broader absorption of NLP into the broader ML ecosystem. I am familiar with papers etc. from this era, but that's not what I am looking for.

1 1

Shubhendu Trivedi @shubhendu.bsky.social · 6h

-- The next level: NLP as seen via ML (more interested in the general modeling culture, evaluation eye that a "NLP person" might usually pick up, general empirical habits).

-- The last level: Bridging the above to contemporary NNs.

1 1

Shubhendu Trivedi @shubhendu.bsky.social · 6h

Looking for suggestions for NLP textbooks. There are three levels that I am looking at.

-- One is the level of understanding foundational aspects of NLP as a discipline (drawing on its linguistics heritage, treating language as structure, discussing ambiguity, treating language as a social fact).

1 2

Shubhendu Trivedi @shubhendu.bsky.social · 9h

This is a festival song for Nikko--"Nikko Waraku Odori." While it comes from an older folk song, this version originated in 1913 to commemorate the visit of Emperor Taisho to a copper plant. A Hungarian friend recognized it, but finding the record (of the most famous performance) took 4 more years.

Shubhendu Trivedi @shubhendu.bsky.social · 10h

It took me years of hearing this in a film (as a background score for 30 seconds), then finally having someone recognize what it was, and eventually managing to get the record through an auction for like $2 because no one cared.
(An old photo, but post inspired by a renewed joy in hearing it again.)

1

Shubhendu Trivedi @shubhendu.bsky.social · 10h

www.youtube.com/watch?v=EHnh...

日光和楽踊（鈴木正夫・喜久丸）

YouTube video by 0klz37

www.youtube.com

1

Shubhendu Trivedi @shubhendu.bsky.social · 13h

and the obsession about its impossibility, will, over a period of time, collapse under their own weight, unless they are able to maintain a composure that draws coherence from elsewhere.

Shubhendu Trivedi @shubhendu.bsky.social · 13h

Not just that, it seems to be able to present a more radical humanism as a consequence. But this also gets to the offhanded comment about such characters tending to go crazy over time. It also seems natural that those who tend to reside in the tension between meaning

1 1

Shubhendu Trivedi @shubhendu.bsky.social · 13h

To me, it felt a bit too over-mediated back then, and it still feels that way now. On the other hand, Handke seems more about disorientation, there is no contract with identity to maintain. He accepts fragmentation and contradiction as part of the experience they both are writing about.

1

Shubhendu Trivedi @shubhendu.bsky.social · 13h

To expand a bit more on the second post. When I see Beauvoir writing here, I also see that she _has to_ reconcile her compassion with her atheism and her feminism, being empathetic with being anti-sentimentalist, and just love with her belief in autonomy and reason.

1 1

Shubhendu Trivedi @shubhendu.bsky.social · 14h

Although as an aside: It seems like a recurrent pattern over time that characters with a literary sensibility like Handke tend to go crazy (not literally) over time. There's definitely something interesting to that pattern.

1

Shubhendu Trivedi @shubhendu.bsky.social · 15h

much more about a formal disarray and the inadequacy of language (and thus refusing affect). Nevertheless, the net comment is that authenticity and artifice aren’t clean opposites, and these two illustrate that in a very interesting way.

1

Shubhendu Trivedi @shubhendu.bsky.social · 15h

To remove the ambiguity. Beauvoir writes with a lucid sincerity, a sort of earnest humanism that I just find too composed; while also caring too much about adhering to pre-contracted identities and trying to justify itself within those. While not here, in general that exasperates me.

Handke seems

1

Shubhendu Trivedi @shubhendu.bsky.social · 15h

Almost exactly seven years later, almost exactly same thoughts (although more understanding of the one that I think "tries to be"). Might as well find and re-post instead.

1 4