I think the overtures about things we care about more just provide plausible deniability and that when you dig down, people are more concerned about the risks of challenging the wealthy than they are about such window dressing.
It's been pretty clear for a while that companies who have developed foundation models have essentially unprecedented levels of investment to recoup. For all the talk of faster hardware and more efficient models, that spend hasn't gone away and ultimately that investment needs to get a return somewhere.
Dependency on cloud AI models is, in effect, dependency on VC subsidy. From the user's point of view, this dependency is debt which will either be repaid with interest to a model provider or through the hard work of making themselves independent of such models after having become dependent.
It's got to be said that rewarding people who make content on a regular, frequent schedule seems to A: be a way of coercing a fairly high minimum level of labour out of platformed accounts and B: a good way of flooding feeds with content which is largely devoid of novelty as a handful of prolific accounts dominate what people end up seeing.
You can see this happen in real time if you closely follow some youtube channels. You take someone who is genuinely talented and has some interesting, novel insights. And, maybe a couple of their videos makes it big. And they rightly think they should keep making videos because they have other insights. And they're not wrong.
But over time, something happens. No one has a novel, brilliant insight 1-2 times a week. So once they really turn in and decide to make a serious effort with their channel, the quality of their content suffers. Maybe it's not quite click-bait, but it's less genuine and more formulaic than their original work. A bit more sensational. Videos are reaching for reasons to exist, since the author needs to keep pumping them out.
I wouldn't quite call it corruption, but it's a clear degradation. In principle it's not a novel problem, since people have been writing weekly editorials for a long time. But, there seems to be something about the Youtube format that makes it such that the big channels must always play the game and pump out sub-par content.
> Maybe it's not quite click-bait, but it's less genuine and more formulaic than their original work. A bit more sensational. Videos are reaching for reasons to exist, since the author needs to keep pumping them out.
I've come to accept that this is what many viewers want. They're more interesting in seeing their familiar personalities talk on camera than in the details of what they're doing.
At the risk of downvotes given the audience, this is how I feel whenever I've tried to watch Linus Tech Tips videos. I have some friends who watch every LTT video when it comes out and love the brand, but I can't make it through a single LTT video because there's so little subject matter. The few videos I watched also had some glaring oversights and borderline misinfo. I think the audience for those videos is people who like seeing the LTT crew have fun, do some activities, and talk. The subject matter of the video is secondary for them.
I see a lot of YouTube channels going the same direction: They realize the content they're discussing is secondary to the fact that they're in front of the camera doing something. The cooking channels know that most viewers aren't going to be cooking the dish. The DIY channels know that most viewers don't care about the code or engineering as much as seeing personalities goof around on screen.
I don't think there is anything wrong with this type of content, though. One of my friends says he handles his work better with a constant stream of YouTube videos in the background, so he semi-watches more YouTube than anyone I know. I do appreciate the channels that focus on the content and subject matter instead of becoming content factories, though.
If you want to be profitable, or widely watched, you have to play to the algorithm.
YouTube seems to strongly boost channels that post regular videos in the 10-20 minute range, and actively incentivizes clickbait through AB Testing tools for titles and thumbnails.
There are channels that post irregularly, with long form videos, but they get buried.
Another issue I've seen from some of the more prolific YouTube channels is they slowly become another mouthpiece for "news coverage". The algorithm very much expects you to continue uploading, because everyone is always looking for the newest content; at least before YouTube removed the Trending section. I admit that I only really check my Subscription page at this point, and after going through a subscription purge I only see maybe a half a dozen to a dozen new videos. Its actually been very useful since it encourages me to not get sucked in to watching hours of videos.
However, given my experience during Digg's v4 attempt this past year, I will say being willing to put yourself out there has served as a pseudo-networking activity and I've gotten the chance to speak with several people and now I'm giving talks "out there".
What's curious to me is, why does this not happen to all youtubers? For example, vlogbrothers, 3b1b, numberphile, etc, all seem to continue putting out great educational content and care about producing good wholesome content despite the strong incentives to do otherwise - how does that happen?
I think different topics lend themselves to this better than others. If you're merely teaching about things, then there are endless interesting topics -- and _you're_ not the one coming up with the brilliant insights; you're just doing an excellent job conveying an already-known subject to others. Commenting on the news can work quite well, too. So long as your research and analysis maintains quality, there will be no shortage of noteworthy events to discuss.
I’m old, but this pattern is old too. You’d see it in car magazines where the regular columnists would rehash their tired old opinions but you’d read it anyway because they had a particular sense of humour or an otherwise engaging style.
It’s hard to create novel content regularly once a month, let alone weekly or daily like some of these YouTube guys are doing
I don't know, I think if you weighed up the costs of AI related datacentre spend vs. the average mathematics academic's salary you could come to a different conclusion.
Raising, nurturing, training, and mentoring an expert mathematician is not cheap; it never was, perhaps the first time in history when we can witness that rule to change - spinning up a bunch of math-savvy agents, each smarter than Ramanujan maybe will get too cheap.
You're oversimplifying the message I'm trying to convey. "you just hire them, someone already raised them" - treats mathematicians as a commodity stock rather than a flow. The conversation frames it as "mathematicians vs. RAM" - a cost comparison. But that's like comparing the cost of a GPS unit vs. a ship captain. The captain isn't expensive because they can calculate routes; they're expensive because they know when the route is wrong. AI makes the math cheaper but makes the mathematician more valuable, at least until true AGI genuinely surpasses human mathematical creativity - at which point we have much bigger economic questions than mathematician salaries.
The topic on itself is quite interesting, and far complex than supply/demand norms. Even before AI, there was and both wasn't shortage of mathematicians - academic pure mathematics - there's a glut. High school teachers - people exist; but they won't work for teacher salaries. Applied math - acute shortage - quant finance, ML research, cryptography, pharmaceutical modeling - we don't have enough. NSA - always struggled to hire - private sector salaries pull people away. Interdisciplinary - mathematical biology, climate modeling, materials science - domains where math is the bottleneck but the job title isn't really "mathematician" - acute shortage.
I think the issue is that LLMs are a cash problem as much as they are a technical problem. Consumer hardware architectures are still pretty unfriendly to running models which are actually competitive to useful models so if you want to even do inference on a model that's going to reliably give you decent results you're basically in enterprise territory. Unless you want to do it really slowly.
The issue that I see is that Nvidia etc. are incentivised to perpetuate that so the open source community gets the table scraps of distills, fine-tunes etc.
You got me thinking that what's going to happen is some GPU maker is going to offer a subsidized GPU (or RAM stick, or ...whatever) if the GPU can do calculations while your computer is idle, not unlike Folding@home. This way, the company can use the distributed fleet of customer computers to do large computations, while the customer gets a reasonably priced GPU again.
The kinds of GPUs that are in use in enterprise are 30-40k and require a ~10KW system. The challenge with lower power cards is that 30 1k cards are not as powerful, especially since usually you have a few of the enterprise cards in a single unit that can be joined efficiently via high bandwidth link. But even if someone else is paying the utility bill, what happens when the person you gave the card to just doesn’t run the software? Good luck getting your GPU back.
New Strix Halo (395+) user here. It is very librating to be able to "just" load the larger open-weight MoEs. At this param count class, bigger is almost always better --- my own vibe check confirms this, but obviously this is not going to be anywhere close to the leading cost-optimized closed-weight models (Flash / Sonnet).
The tradeoff with these unified LPDDR machines is compute and memory throughput. You'll have to live with the ~50 token/sec rate, and compact your prefix aggressively. That said, I'd take the effortless local model capability over outright speed any day.
Hope the popularity of these machines could prompt future models to offer perfect size fits: 80 GiB quantized on 128 GiB box, 480 GiB quantized on 512 GiB box, etc.
Not for long, presumably. Apparently the majority of marketable skills will come from a handful of capex heavy, trillion dollar corporations and you will like it.
You only need to buy one or two to get it on the agenda, then everybody votes along party lines, on stuff they don't understand. It's not even that expensive.
reply