This weekly roundup thread is intended for all culture war posts. 'Culture war' is vaguely defined, but it basically means controversial issues that fall along set tribal lines. Arguments over culture war issues generate a lot of heat and little light, and few deeply entrenched people ever change their minds. This thread is for voicing opinions and analyzing the state of the discussion while trying to optimize for light over heat.
Optimistically, we think that engaging with people you disagree with is worth your time, and so is being nice! Pessimistically, there are many dynamics that can lead discussions on Culture War topics to become unproductive. There's a human tendency to divide along tribal lines, praising your ingroup and vilifying your outgroup - and if you think you find it easy to criticize your ingroup, then it may be that your outgroup is not who you think it is. Extremists with opposing positions can feed off each other, highlighting each other's worst points to justify their own angry rhetoric, which becomes in turn a new example of bad behavior for the other side to highlight.
We would like to avoid these negative dynamics. Accordingly, we ask that you do not use this thread for waging the Culture War. Examples of waging the Culture War:
-
Shaming.
-
Attempting to 'build consensus' or enforce ideological conformity.
-
Making sweeping generalizations to vilify a group you dislike.
-
Recruiting for a cause.
-
Posting links that could be summarized as 'Boo outgroup!' Basically, if your content is 'Can you believe what Those People did this week?' then you should either refrain from posting, or do some very patient work to contextualize and/or steel-man the relevant viewpoint.
In general, you should argue to understand, not to win. This thread is not territory to be claimed by one group or another; indeed, the aim is to have many different viewpoints represented here. Thus, we also ask that you follow some guidelines:
-
Speak plainly. Avoid sarcasm and mockery. When disagreeing with someone, state your objections explicitly.
-
Be as precise and charitable as you can. Don't paraphrase unflatteringly.
-
Don't imply that someone said something they did not say, even if you think it follows from what they said.
-
Write like everyone is reading and you want them to be included in the discussion.
On an ad hoc basis, the mods will try to compile a list of the best posts/comments from the previous week, posted in Quality Contribution threads and archived at /r/TheThread. You may nominate a comment for this list by clicking on 'report' at the bottom of the post and typing 'Actually a quality contribution' as the report reason.
Jump in the discussion.
No email address required.
Notes -
I was browsing through the news today and I found an interesting article about the current state of AI for corporate productivity.
MIT report: 95% of generative AI pilots at companies are failing
There seems to have been a feeling over the last few years that generative AI was going to gut white collar jobs the same way that offshoring gutted blue collar jobs in the 1980s and 90s, and that it was going to happen any day now.
If this study is trustworthy, the promise of AI appears to be less concrete and less imminent than many would hope or fear.
I've been thinking about why that might be, and I've reached three non-exclusive but somewhat unrelated thoughts.
The first is that Gartner hype cycle is real. With almost every new technology, investors tend to think that every sigmoid curve is an exponential curve that will asymptotically approach infinity. Few actually are. Are we reaching the point where the practical gains available in each iteration our current models are beginning to bottom out? I'm not deeply plugged in to the industry, nor the research, nor the subculture, but it seems like the substantive value increase per watt is rapidly diminishing. If that's true, and there aren't any efficiency improvements hiding around the next corner, it seems like we may be entering the through of disillusionment soon.
The other thought that occurs to me is that people seem to be absolutely astounded by the capabilities of LLMs and similar technology.
Caveat: My own experience with LLMs is that it's like talking to a personable schizophrenic from a parallel earth, so take my ramblings with a grain of salt.
It almost seems like LLMs exist in an area similar to very early claims of humanoid automata, like the mechanical Turk. It can do things that seem human, and as a result, we naturally and unconsciously ascribe other human capabilities to them while downplaying their limits. Eventually, the discrepancy grows to great - usually when somebody notices the cost.
On the third hand, maybe it is a good technology and 95% of companies just don't know how to use it?
Does anyone have any evidence that might lend weight to any of these thoughts, or discredit them?
Well there seems not enough money is being spent on trying to reduce the power use of inference. A startup that work with silicon for inference tried that can't raise funding enough to retain their engineering teams. Like something is off if companies that try to solve the concrete problem can't get funded but other companies lights stacks of cash on fire to subsidize model usage just to capture market share. The whole thing looks bonkers to me!
If you start with the assumption that the well has run dry and LLMs are never (not any time soon, at least) going be much better or much different than they are now, then yeah, very little about the market makes sense. Everyone willing to put substantial money into the project disagrees.
Inference costs are exaggerated (and the environmental costs of inference are vastly exaggerated). It's certainly a big number in aggregate, but a single large query (30k tokens in, 5k out) for Google's top model, Gemini 2.5 Pro, costs about $0.09 via the API. And further queries on substantially the same material are cheaper due to caching. If it saves your average $50,000 a year office drone 30 seconds, it's more than worth it.
Google ends up losing a lot of money on inference not because it's unaffordable, but because they insist providing inference not only for free, but to search users who didn't even request it. (With a smaller, cheaper model than 2.5 Pro, I'm sure, and I'm sure they do cache output.) Because they think real world feedback and metrics are worth more than their inference spend, because they think that the better models that data will let them build will make it all back and more.
But who knows what those models will even look like? Who wants to blow piles and piles of money on custom silicon that might eventually reduce their inference costs by a bit (though, since they were working with RISC-V, I kind of doubt it'd have ended up being better per-watt; cheaper only after licensing costs are factored in, probably) when a new architecture might render it obsolete at any moment? It's premature optimization.
(Granted, GPUs have remained viable compute platforms since the advent of deep learning, but that's because they're not too specialized. Not sure how much performance per watt they really leave on the table if you want to make something just as flexible. Though I have heard lately that NVidia & AMD have been prioritizing performance over efficiency at the request of their datacenter clients. Which I'd read as evidence they're still in the 'explore' domain rather than 'exploit.')
If you're anticipating capex in the 13 figures, it's still surprising that large companies don't do more research on fundamentally different learning algorithms and hardware. Which isn't to say they don't (e.g. there are a couple researchers at GDM doing excellent work in neuromorphic and more brain-inspired learning), but I'd be surprised if the aggregate research spending among the big three on this (as opposed to tweaks to make transformers perform incrementally better) exceeds $1B/year. Any given research path is likely to not lead to anything, but the potential payoff is enormous.
More options
Context Copy link
More options
Context Copy link
More options
Context Copy link
More options
Context Copy link