That paper shows hallucinations can't be eliminated, due to approximation error. But it is completely compatible with hallucination becoming less probable as scale reduces that approximation error.
> A Blu-ray disc, game cartridge, or printed book cannot be remotely erased, edited, or deactivated. It is a physical object you can own, resell, lend, archive, or play offline indefinitely.
Isn't this untrue with surprising frequency? Decoding devices phone home, come under new copyright laws, etc etc etc.
Feels like this could use a data driven way to normalize reference questions (topic and wording). This is missing a lot of right-initiated topics like immigration, gender roles, and role of religion in government.
The willingness to throw capital at AI is definitely doing some crazy things, but this article has some bad takes on the data.
> [Ratio of per-token cost to subscription cost] means Anthropic is subsidizing their enterprise customers by up to 40 times, and OpenAI up to 70 times
Actually, they could be subsidizing by more (if they are taking a loss on API), or not at all (if they are soaking API customers by a massive margin).
Separately, these subscriptions get sold to large groups with varying usage, so it's crazy to model assuming every subscription is maxed out. Banks, gyms, and many other businesses work this way, offering consumers flexible access to services that they will realistically use in bursts. It's not always worth the complexity to prevent overuse by a small minority. You can feel like this kind of business model isn't as transparent, but it's silly to pretend it can't work.
> OpenAI spent 44% of their revenue [$5.3B] on sales and marketing! The hype needed to keep the AI bubble inflated is incredibly expensive.
Over that same period (2025), OpenAI added $10B in realized revenue and $14B in run-rate. Sounds like they're getting >2X return within 12 months of those go-to-market dollars. Compare that to like, any other business.
> Thus in recent weeks the idea that Generative AI (LLMs for short) is too expensive has been all over mainstream business media.
Would it be smarter for these companies never to test customers' price tolerance? The quotes following this make it seem like the companies are getting important information about the nature of that price tolerance, and preparing to react. This is the work markets do on both sides to understand the value of a new product.
There are lots of good arguments about AI overinflation, but in order for them to be useful, they have to be rigorous and targeted.
Understood; these are the failure modes of cheap, weak tests in a system designed for expensive, high signal tests. I get that it is causing a real pain when these are introduced willy-nilly -- so is there a way for the medical system to evolve into making productive use of such signals?
Because the status quo is not working for a lot of folks in managing their own health, and just saying "it has to be this way" --while every other aspect of their life takes advantage of a wider range of options -- leaves them careening into MAHA style crazyness .
"We shouldn't collect information because we don't know how to fit it into our care playbook" might be rational for a single patient, but it's a policy that will lock you into your current playbook.
Our medical industry is set up to only evolve via highly centralized research that fully situates a diagnostic within a particular treatment path. This approach makes it more and more expensive to improve care for narrower and narrower populations - driving medicine towards being a luxury good.
I'd like to see midjourney say more about price, but I love the idea of starting some new diagnostic pathways with different principles. There are probably all sorts of low hanging fruit to be found about new treatment strategies... It just takes some faith that nature hasn't hidden all of her secrets in the one place we already know how to look.
Brains are efficient, but civilized humans aren't. In the USA, adults consume at a rate of about 10kW -- only 1-2% of that being the human's metabolism, the rest being HVAC, electrical devices, etc.
For comparison, a modern frontier model like Gemini 3.5 Pro consumes about 15kW -- so only about 1.5x the fully loaded human. In an 8h workday, that model would crank through ~80M tokens (~$5k at API prices). That's ~4 major refactors of a 10k LOC codebase, so probably not a very realistic comparison to a single human dev.
I think a more useful comparison, based on my experience, is that an engineer with AI support can get one 8h day's worth of unassisted work done in 1h. So, the 25 kWh consumed during collaboration (conservatively assuming I keep the GPU hot for the whole hour) frees up the remaining 70 kWh I'll draw down for the day to be spent in some other way.
Friends on ML/AI hiring committees at top tier university are seeing foreign profs turn away record offers. Same for applied math relevant to material science.
I expect you are right at the most specialized end of the spectrum (and certainly industrial labs in those areas), but I wonder if anyone can speak directly to where we are still globally competitive.
Besides the brutal impact on those already invested in the American research community, this is one more nail in the coffin when it comes to competing for new talent. What researcher in their right mind would move their research and their future to the USA to join this clown rodeo?
It is unbelievable to watch my country give up its most unfair (and yet mostly positive) advantage -- a nearly free option on the top talent of the entire planet. Here's hoping that the increasingly multipolar research world can find ways to be even more efficient in creating new knowledge.
Besides deeply unpredictable factors (like signal transmission), most users of higher-level abstractions do so without certainty about how the translation will be executed. For example, one of the main selling points of C when I was growing up was that you could write code independent of architecture, and leave the architecture-specific translation to assembly to the compiler!
Abstractions often embrace nondeterministic translation because lower level details are unknown at time of expression -- which is the moivation for many LLM queries.
An LLM model itself -- that is, the weights and the mathematical functions linking them -- does not tell you exactly how to train from data, nor how to generate an output. Instead, it describes a function providing relative likelihood(output | input).
Deciding how to pick a particular output given that likelihood function is left as an exercise for the user, which we call inference.
One obvious choice is to keep picking the highest likelihood token, feed it into the model, and get another -- on repeat. This is what most algorithms call "temperature=0". But doing this for token after token can lead boring output, or steer you into pathological low-probability sequences like a set of endless repeats.
So, the current SOTA is to intentionally introduce a random factor (temperature>0) to the sampling process -- along with other hacks, like explicit suppression of repeats.
If I understand this correctly, the pain is because you depreciated the assets in the first few years, offsetting other tax liabilities, and now you must pay those back even if exiting the properties at a wash?
If so, it seems like the unfairness is in the other direction: landlording allowed you to essentially pull forward a tax credit, which a W-2 job doesn't allow.
You are describing products that change the information flow used to solve a problem. There are probably big gains there, but it requires everyone to come for the journey at the same time - for example, you must tell the board "there is no report, talk to this bot. Not everyone will join for the ride!
AI marketers are leaning into use cases where and individually can unilaterally choose to do their work a different way. This is much easier to explain and distribute!
All that said, the result is definitely funny - seeing work done like this makes you realize how artificial the tasks that make up modern work are.
Sounds like corporations can have bad motivations, and can make bad software - no AI required.