jacobr1

+fees, though ideally the fees would be capped too

Wouldn't that be covered by using PPP comparisons rather than just market currency conversion rates?

Demand for the big cities is still increasing pretty much everywhere, even if some countries are seeing net decrease in demand on the whole. Also remember that supply deteriorates. The number of "ready to move into" homes can decrease over time without maintenance and rebuilds.

Right - the point I was refuting is that LLMs are "just statistical models of words." More is going on. Does that imply "understanding?" I don't know, I'm not sure we have a good enough definition to say. But it does mean that the models are more complex that say, markov chain graphs with corpus frequencies. It seems we are encoding data in the latent space with much higher complexity than "just words." There is higher order semantic information being captured - probably not the same has human "thoughts" - but again - also not _just words_.

The surprising bit is that 10% of Americans DO NOT own a smartphone. That must be exceedingly difficult. Increasingly everyday activities require them, without fallbacks. I presume those numbers exclude children, and then if you account for the infirm, it wouldn't surprise me if we started to get to universal levels of ownership.

LLMs as clearly doing more than just modeling words. In order to predict word placement, they need to build some kind of model, their latent space, of the types of things they are able to predict. Not really full world models yet - but they have decent "blog space" or "github project" models. And you can see this with multimodality, or non-text modality modals such as images and audio. They map from their latent spaces to the outputs. The fact that multi-modal systems can share the interior layers shows some kind of internal representation is created.

The new siri models will be some variant of the gemini models. This framework seems to be more generalized than that though.

Where are the "Safe Guardrails" ?

Older ML systems were much better at exposing their internal confidence. Plenty of papers reverse out this kind of interpretability for open weight models. All the models exposed logprobs early on. This seems solvable if prioritized. The unintelligible words should be lower confidence. Getting per-token data for the output that aids with understanding the predictions is entirely feasible as engineering effort - it just won't be enough to address all the problems - but it should help quite a bit.

well, technically they can ...

I half agree. But two points: 1) if you can formalize your instructions ... then future instances can be fully automated. 2) You are still probably having the AI perform many sub-tasks. AI-skeptics regularly fall into this god-of-the-gaps trap. You aren't wrong that human-augmented AI isn't 100% AI ... but it still is AI-augmentation, and again, that sets the stage for point 1 - to enable later future full automation on long enough timecycles.

There is a lot to be done with good prompting.

Early on, these code agents wouldn't do basic good hygiene things, like check if the code compiled, avoid hallucinating weird modules, writing unit tests. And people would say they sucked ....

But if you just asked them to do those things: "After you write a file lint it and fix issues. After you finish this feature, write unit tests and fix all issues, etc ..."

Well, then they did that, it was great! Later the default prompts of these systems included enough verbiage to do that, you could get lazy again. Plus the models are are being optimized to know to do some of these things, and also avoid some bad code patterns from the start.

But the same applies to performance today. If you ask it to optimize for performance, to use a profiler, to analyze the algorithms and systemically try various optimization approaches ... it will do so, often to very good results.

Except it often is the case that when you break down what humans are doing, there are actual concrete tasks. If you can convert the tacit knowledge to decision trees and background references, you likely can get the AI to perform most non-creative tasks.

Normal isn't a myth. The mistake people make is taking the mode as normal, or worse mistaking their own experience as normal. But humans generally do tend to have a range of common behaviors that a significant percentage of people fit into. And you probably can even predict it to a reasonable degree, if you have some other metadata to correlate which sub-group they might correspond to.

Normal in the sense of "you can model a distribution of human behavioral processes or outcomes" that encompasses, say, 95% of humans in a given culture or geography is very much a thing you can do. And I'd go as far as to say a large chunk of the mental bandwidth of the average person is running those simulation models just to operate in a multi-human-agent world.

(If you want to say we observe bimodal or other multi-peaked distributions in practices rather than "normal" ones, I will strongly agree, but that usually isn't the objection when people say "normal is a myth")

Yep.

ChatAI - show the top 50 online retailers by revenue in the US and note any that have credible new stories about quality control issues. Save all of them except StoreX and StoreY in your list you use for comparison shopping.

Or maybe another one, scan all my credit card purchases for all time that you have history and record all the stores.

Done. And plenty of third party sites (consumer reports, wirecutter, etc...) will do this kind of thing too. And you could perhaps transitively trust them - either view direct lists or just scraping the places they recommend.

And the average person doesn't need to figure this out ... skills encoding this will propagate.

Unless they have agents reading those emails and responding ...

All the hard work is always chasing down edge cases, scaling, operational issues and other things that don't show up the user-exposed features. And talking about features, the innovation in coming up with them, or iterating on making them work with real customer experience is a ton of value, even if copying the ideas that work later is much easier - which is why I generally prefer betting on an innovator with just of enough traction to show they can stick with it. The best category leaders both innovate and steal/copy/buy all the innovation they aren't producing in house to maintain their lead.

Business crave both data for analysis and checkboxes getting checked for compliance sake. If those don't align to the value of the work - then you have the classic of employees hating the "TPS Reports" they are forced to make. As an example, sales people are notorious for basically never updating CRMs and also they have incentives to skew the specifics anyway.

this is sub-par and neglects important aspects of your business

But that is exactly the right way to think about it. If you have an army of sub-par workers that aren't going to think deeply about their value to your business, but are really cheap (relative to human labor) - how do you make effective use of them? Thinking about AI agents as being high-competence and able to learn your intent is the wrong model at this point. Though they can be high-competence in very specific narrow niches.

Also consider that while the OP looks like a skilled, experienced individual, all too often the documentation is being written by someone with that context, but rather someone unskilled, and with read empathy. Quality is quite often very poor, to the point where as shitty as genai can be, it is still an improvement. Bad UX and writing outnumbers the good. The successes of big companies and the most well known government services are the exception.

The common thread is that people are bad and saving and delayed gratification. The easiest path to instant gratification wins more often than not.

Depends on the service, and timeframes. For geforcenow, you also need to consider the upgrade cycle - how often would you need to upgrade to play a newer game? I'm not sure but probably at least once within that 8 years. Buying a new car, or almost new car, and driving it until it falls apart is a better financial option than leasing. But if you want a new car every year or two, leasing is more affordable - for that scenario. Also it depends on usage. My brother in law probably plays a video game once every other month. At that point, on demand pricing (or borrowing for me) is much better than purchase or consistent subscription. You need to run the numbers.

Already happening at the in vitro level, might be possible in vivo as well. Neither require the more abusive approaches from the first eugenics era.

Honestly it feels like what I, or many of my colleagues would do if given the assignment. Take the current front page, or a summary of the top tropes or recurring topics, revise them for 1 or 2 steps of technical progress and call it a day. It isn't assignment to predict the future, it is an assignment to predict HN, which is a narrower thing.

Is anything objectively amazing? Seems like an inherently subjective quality to evaluate.

The cycle repeats frequently in industry. New waves of startups address a problem with better UX, and maybe some other details like increased automated and speed using more modern architectures. But feature-creep eventually makes the UX cumbersome, the complexity makes it hard to migrate to new paradigms or at least doing so without a ton of baggage, so they in turn are displaced by new startups.

Right - the quality of your locks matter a lot less if your average 5-year-old tee-baller can through brick through the wind and climb in. One always needs to consider their threat model when considering what security to invest in getting.

It is a focus, data, and benchmarking problem. If someone comes up with good benchmarks, which means having a good dataset, and gets some publicility around, they can attract the frontier labs attention to focus training and optimization effort on making the models better for that benchmark. This is how most the capabilities we have today have become useful. Maybe there is some emergent initial detection of utility, but the refinement comes from labs beating others on the benchmarks. So we need a slideshow benchmark and I think we'd see rapid improvement. LLMs are actually ok at a building html decks, not great, but ok. Enough so that if we there was some good objective criteria to tune things toward I think the last-mile kinks would get worked out (formats, object/text overlaps). the raw content is mainly a function of the core intelligence of model, so that wouldn't be impacted (if you get get it to build a good bullet-point markdown of you presentation today it would be just a good as a prezo, but maybe not as visually compelling as you like. Also this might need to be an agentic benchmark to allow for both text and image creation and other considerations like data sourcing. Which is why everyone doing this ends up building their own mini framework.

A ton of the reinforcement type training work really just aligning the vague commands a user would give to the same capability a model would produce with a much more flushed out prompt.

We've reach a point of price stabilization and longevity for smartphones now that didn't exist for the first 10 year ramp. When every new model added fundamental capability, you always want to upgrade, with the sweet spot often being every other year. But now, with better build quality, batteries, and stabilization of features people will keep their phones for much longer. Or buy "new" models that are of older versions since the price/features have been acceptable to run most of the apps they care about for years now. Plenty of people still want the top end for similar reasons to why people buy design clothing, but we've reached a feature plateau. We hopefully are getting close to that with EVs. Seems like around 300 mile range standard was the key thing. Though improved AI driving could change that again.

I think they meant the reverse. The majority of US EV sales are of Tesla's: https://cleantechnica.com/2025/04/15/auto-brands-leading-the...