sottol

Usually, it's the wealthy that tend to leave and depopulation tends to increase "blight" and crime.

But I had no idea the population declined so much, that puts other observations about SF into context.

Quick googling showed the Tata Punch EV to cost between $15k-$19k. Afaik, the BYD Seagull is also targeting under 20k EUR in Europe, double the price in China. Might be tariffs.

> Time to think bigger.

Ehh, no need - just let the LLM figure out what to build in your garage.

Imo, the Terminator movies would have been scarier if they moved like these guys - slow, careful, deliberate and measured but unstoppable. There's something uncanny about this.

Aren't the Wagoneer, Escalade, Yukon and Tahoe all body-on-frame based on a truck chassis?

A Tesla was genuinely an exciting car that many wanted. Sure, it lacked polish (and fit and finish were meh) but it was by far the best EV you could get for a long time. There was also the "other" factor before they were mainstream , or why some rich drive Land Rovers, Mercedes G class and so on, then it was just the hip car in the Bay Area and now everyone is driving them so why bother.

Not saying people stopped buying them because of Musk, just that it wasn't necessarily his charisma that used to get people to buy them.

Or maybe it's an oligopoly where the incumbents have carved up the market and stopped competing, milking their customers instead.

Broadband is then extra special if you let the ISP also own the infrastructure as everyone has to reconnect their service to every house instead of one company (or forbid, the govt) owning the pipes and several companies competing for providing services over those shared pipes.

Imo the competition model doesn't necessarily (always) work that well for infra.

Might still beget carelessness/lack of empathy or a feeling of being superior.

I think we agree - my response was narrowed to respond to "NVDA is going to soar". The whole segment looks overvalued and the implication (at least to me) seems to be that AI will capture 10-20% of GDP in profits (not even revenue) starting soon for 10-ish years.

I don't have the numbers, but I wonder how "internet company" (very loosely) profits compare to GDP - I could imagine the AI ecosystem reaching similar profits eventually.

Btw, imo their P/E is mostly this good because they're currently selling hardware at margins better than many software companies, I doub that this will/can last. For example, when FB/MS each are spending $50B+/a on GPUs, can they justify say $100M/a on a crack software dev team to make their stack work on other GPUs (ie replace the oft-argued CUDA moat within a year or two)?

NVDA's market cap was already 12.5% of the entire yearly US-GDP, at a P/E of 10 or 20 once their growth slows to me that implies that their profits are expected to be 0.5-1% of the US GDP. Just nvda alone, not the LLM foundation models + application layers and so on built on top their hardware.

Even in the most optimistic case, I find it hard to imagine that nvidia alone captures much more than 1% of the US GDP as profits for the next 20 years or whatever their investor's horizon is.

This _is_ the rich controlling our government! Except it's more openly visible in this case than usual.

Going forward they'll have a good deal of say in the policies they care about, by donating and attending and show of loyalty.

Maybe check out https://lobste.rs .

But there's also been substantially more investment now than say 2 years ago (AI + AI-adjecent companies have, what, 4-5 trillion USD valuation now?). CEOs tout how they stopped hiring. And then these benchmarks having names like "humanity's last exam". All that implies LLMs are a direct replacement for human workers so imo it's fair to measure LLMs against the hype/promises.

So I would say the goalposts are moving just in step with the market expectations/hype. Nobody's disputing that LLMs are amazing at generating useful text anymore.

I think it might mean the opposite of what one would expect. Afaict, calibration error means something along the lines of "how often was the model wrong but confident that the answer was correct".

That means a low calibration error would be a good thing, ie the model correctly recognizes when it is unsure about answers instead of confidently stating the wrong answer.

This was the play all along - roll up the competition and then shore up all the profit once dominant enough.

I bet there are plans for new firmware $-unlocked pay-to-play features, making the AMS "more compatible" with Bambu filament (or eventually only Bambu filament for the next round or two of printers) and facilitating a lot more paid transactions and model purchases on Makerworld and taking the customary 30% cut.

Great idea - I'm currently using a barebones Framework 16 mainboard with the 7940HS as my homeserver and local-AI playground in a 3D-printed case.

UMA with the iGPU is nice though not much faster than CPU (20-30%? don't recall).

But installing 64 or even 96GB RAM is more attainable though slower (60GB/s) than Apple's $$$-mem-upgraded M* Pro (200), Max (400) and Ultras (800). Although DDR5 is comparable to a regular M1 (70) afair. Most LLMs in the 20GB range run at around 7-12tok/s which is fine for playing around with them.

First I hear of Project Binky but this might be one of those "because they can" projects. This is quite ingenious though, hat off!

And btw, this does harken back to the heyday of the 3D-printing hype where they hypists were saying that we would "soon" 3D-print entire electronic devices. That never materialized, of course. Maybe that was one inspiration.

Afaik, this seems like it's probably the easiest way of metal 3d printing at home - I wonder whether that's the more interesting thing about this achievement.

I don't really know much about in-and-out of melting metals, toxicity and so on but there are some lower melting metals like zinc at 420*C which could be reached with a 3D printing hotend with a resistive cartridge and copper block I think. That would be rather intersting.

> and it's for financial gains, but not motivated by morals.

Which is why nothing is going to happen, can't have people starting with the latter.

How soon though? 2 years? 10? 50? He starts by saying this won't happen soon, _maybe_ ever.

Analysis of economic and productivity losses caused by Youtube ads in <world>.

<S>OMFG!!! YOUTUBE IS COSTING THE WORLD *750B EUR* PER YEAR. </S>

How many hours of productivity are lost to Youtube ads?

2.49 billion active users, average seems to be 29 hours per month, reddit reports 4 ads/10 minutes lately - so 24 ads/hour, 5 seconds each (even though that went up!), so 2mins of ads/hour or 1 hour of ads per month, 12 hours of ads per year!

12 hours * 25 Eur/hour * 2.5B = 750B Eur

(probably made some mistakes)

Also, this article is ridiculous - like assuming all 400M European internet users are "productive" at 25Eur/h (30% are probably < 15 or > 65), people clicking 1200 banners per year because they visit 100 sites/month (12*100, right?!) and so on.

Looks a lot like YOLO. But seems quite useful in this context.

The end seems to overhype it a little bit though.

I figure expectations/behaviors will shift once this becomes the new norm. That 5-second birthday present or shallow prospectus summary will become obvious pretty soon after a wide uptake and nullify any advantage.

I doubt it's because of realtime constraints, you could probably use the NDK if you need to mostly escape the JVM. ObjectiveC/Swift are also garbage collected, so it's not like iOS is strictly realtime by default.

I think the main reason those products are available on iOS and iPadOS is because most creatives have _always_ been on Apple's products and iOS users are on average wealthier and more willing to pay for software. Also, I bet the original version of the software has been available on MacOS for a long time, so on top of having paying customer it would also make it easier to port to iOS than Android.

There's TermUX on Android, doesn't mean most phones are made for and mostly used for software development/sysadmin tasks. Sure, there are a few people that'll use their Android phone to SSH into something in a pinch, but 99% use their Android phone for social media, emails, web and youtube just like Apple iPhone users.

In this context "thinking" was meant as an analogy to the supposed reflective "slow" mode of the cognitive process, vs the more "reflexive"/"fast" mode, not in the sense of "thinking soul" or "thinking self-aware entity".

Concretely, what do you gain by giving a current-generation LLM more runtime? It's not trained/designed to do anything with it, more time = more tokens = more nonsense once past the "end" token/end of context. You could build an agent on top of the LLM that calls the LLM iteratively, but afaict this approach isn't strictly an improvement over the base LLM. Current architectures seem to be limited with how much improvement they can eke out of more runtime without a broader redesign or retraining.

Now with new architectures you might be able to do more with more runtime, but I'm not sold that allowing a current-gen LLM to execute code or do web-searches and re-feeding it the quetion + its output + new data is going to strictly return better results. Sometimes maybe yes, sometimes not.

So imo the current limitation is not the runtime allotted to LLMs but their fundamental (current-gen) design/training.

That seems a little reductionist. We're just exchanging ideas and trying to interpret imprecise terms in-context (ideally in good faith).

We're not really discussing whether LLMs are "thinking" in the sense of the "social category" in the abstract sense - that's just the one word that this discussion got hung up on.

Concretely, we're discussing whether giving LLMs a larger time-budget would yield better results - OP suggested that letting LLMs iteratively refine or enrich an answer might lead to better results. The term "thinking" was used as an analogy or quick crutch to convey an analogy, not 100% equivalence, between a hypothetic LLM output evaluation algo and people's supposed "fast" and "slow" modes of thinking (if they even exist, whatever). Even if people don't have these two modes of thought, as you said they might be post-fact reifications of sociological ideas, the algorithm/idea of iteratively reflecting on and self-refining LLM outputs might still apply to LLMs as LLMs are not actually thinking.

Imo current LLM architecture and/or training data facilitate (mostly?) the automatic non-iterative mode but OpenAI is trying to do more of the iterative, reflective mode with O1 afaict. Imo and currently, there's not much to be gained by iteratively feeding an a current-generation LLM it's own output to iteratively analyze and augment it. At least not in general. Even enriching the input data via RAG doesn't always lead to better results. Imo, there's more fundamental work necessary - and maybe OpenAI is doing that.

But the LLM doesn't "think" - it just spits out a sequence of tokens (forever, literally). The chat models have been fine-tuned to eventually emit an "end" token (when the training examples would have been answered) that the UI/backend interprets and cuts off the LLM to stop it from generating more output. But the model would gladly go on forever and devolve into more and more obscure output the further the original query moves out of the context window.

I still think it's not useful to compare fatalities per billion miles, especially cycling and driving (and especially if that includes highway driving, which is clearly not something you could use a bike for instead).

Iceland had several years with zero or one reported cycling fatalities, I can't find road fatalities quickly but I assume driving is less safe there, especially if I'm allowed to cherry-pick a zero-fatality year (where you'd have 0 fatalities per billion miles cycled).

Yes, on the margins cycling is more dangerous in the USA. OTOH, I wouldn't be surprised if the physical exercise would reduce mortality enough to still be a net positive on the margins - biking to the bakery vs driving - but haven't verified that.

Interesting tidbit from the report is that almost 50% of fatalities are in towns <100k population, 16% outside towns and 34% in pop centers > 100k. I would have guessed metro areas are more dangerous.

I don't think you can 1:1 compare those stats.

Most people just wouldn't have the time to travel as far by bike as they do by car, even if they wanted to. They'd take some other form of transit for long drives.

Car drivers do 13500 miles/year on average, even at a very sporty 20mi/h per (average!) a cyclist would have to ride full 29 days/year non-stop to match that.

Thus, a better measure would be accidents/hour of operation. I bet most avid cyclists do < 1000 miles/year and few do more than 3000 and that would mostly close the gap in death/hour of use.

I think it's probably correct that few people *really* cared until the Reprap project showed a real DIY 3D printer was possible.

Early machines were industrial machines with huge price tags, proper linear motion systems, complicated extrusion systems and so on. There is a bit of a mental leap to go from seeing $100k+ machine and dreaming to design something that can be built for $200-500.

The problems were: no "reference designs", no tried and true go-to mechanical parts (like cheap chinese linear motion rails), extruders (they were DIYd!) or heated beds (early models were just PCBs) and so on - imo it just took someone to get this rolling and that may have well taken 30 years.

I think Reprap was first publicly shown around 2005. From then on it was taken on by more and more makers and refined. It culminated in the early 2010s hype with Makerbots and its contemporaries but they still cost > $1500 and were far from set-and-forget appliances, like 50% reliable and slow - we had one at work and I got fascinated but it printed at 5-20mm/s so parts would just take forever and often fail due to bed adhesion, clogs, ...

The last 10-15 years then have seen the popularization of 3d printers through the Prusa i3 and its clones (Ender and other cartesians < $300) and steady refinement of reliability through better materials. Then the last ~5years or so significantly bumped up the speeds through better linear motion components, motion systems and input shaping + firmware and ecosystems like Klipper.

Bambu imo got in at just the right time and refined everything that had culminated up to this point into a solid appliance. Imo their genius was more in the industrial design, making it reliable and affordable manufacturing than anything else.