kimixa

I find it's often very language and sector dependent. I still see a massive difference in systems programming (normally c++ and rust) between any open model I've tried and something like sonnet 4.5 (not really tried 4.6). And honestly, even the big models (like Opus 4.6) struggle in many cases.

Perhaps these things aren't well represented in the training data for these open models? Every local model I've tried (minimax2.5, GLM-4.7, Quen3, 3.5 and -coder variants) spend so much time trying to get something syntactically sensible and accepted by the compiler that when they've finished they barely seem to have any "momentum" left to actually solve the problems, as pretty much anything but the most trivial change ends up in another loop of actually trying to get it working again, often losing the intent of that change in the process.

My fear is that the solution here, having multiple instances all making the same changes for later comparison, would spend a huge amount of time beating it's head against compiler errors, types, memory allocation (NO DON'T JUST SPRINKLE IN A FEW MORE RAW "new" KEYWORDS DAMMIT) before it even gets to the "logic".

Having plenty of local GPU power I'd love to be able to actually use that, and I'm already wary about some of the training data use and it's interactions with the license of the code I'm "sending" to the cloud models...

Or British English.

The Oxford style guide page 18 https://www.ox.ac.uk/public-affairs/style-guide

> m-dash (—)

> Do not use; use an n-dash instead.

> n-dash (–)

> Use in a pair in place of round brackets or commas, surrounded by spaces.

I feel a number of people online started using something like the "British style", as an ascii hyphen is closer to an n-dash than an m-dash in width. And then without spaces either side looks very squished.

It could often depend on the encoder - things like lame have a hard low-pass filter even on the "insane" settings [0]. This can often mean, if you're someone who can detect that high frequency (probably not most adult), you may pretty easily be able to tell the difference if those frequencies are present in the recording.

Additionally, a lot of audio pipelines (even beyond the DAC - like amplifiers and similar) can end up with artifacts and harmonics in more audible frequencies - this is often more notable at extremely high frequencies (like 96khz and similar) - there's honestly nothing any human can actually hear near that range - but that doesn't mean it doesn't then affect audible ranges when actually played back on real equipment.

The big point is that "Being Able To Tell The Difference" isn't always the same as "Better Quality". You're often just replacing one artifact of the playback pipeline with another. Neither may truely match the original performance.

[0] https://sound.stackexchange.com/questions/38109/lame-why-is-... - while not an explicit "low-pass" filter, the default option of "-Y" does something similar.

Probably because the "default" (in the USA and all European states I've checked at least) is copyright protection - unless explicitly stated otherwise the original author has exclusive rights to reproduce or distribute the work.

That means that things with "no license" don't actually mean "you can do whatever you want" - but in fact "you can do realistically nothing".

So to actually let other people so much as look at it, you have to have some kind of license attached already. And then it can be easy to imply (in the eyes of the law) things like "fitness for purpose" or some kind of warrenty unless expicitly denied.

Honestly it's really annoying to find things like code on the internet with "no license" - that just means you can pretty much never even look at it. You could argue that isn't the "right" default, but the law is what the law is right now.

Yes, you've just described one method the US authorities use to block websites.

They also force ISPs to block IPs [0].

I feel trying to say that's not "blocking websites" is playing games with words, and the results are functionally the same to the "average" user.

The fact that the US effectively claim juristiction over the root DNS system is a more a geopolitical power thing rather than a legal restriction.

[0] https://torrentfreak.com/us-court-orders-every-isp-in-the-un...

People here seem to be thinking this a UK/Europe-specific phenomenon, but there's plenty of examples of the US "seizing" sites that were never hosted in the USA either, and even put pressure on countries to extradite people involved even if they never broke any laws in the country they're living in.

One I remember was a site hosting streams of the 2022 football world cup. Or a number of Iranian-affiliated news sites just last year. Or offshore gambling websites in 2021.

People going "Those Crazy Brits! Thank God That'll Never Happen Here!" seem pretty ill-informed.

most US states have a lower total population than LA county.

Selling back to the grid, maybe. But if displacing your own use, 10p/kwh is a pipe dream in the UK, more like 25p+.

And my point is more that in that example price/kwh, the "best" location only returns $150 or so for that same system.

Sure, worth keeping in mind for payback calculations, but not really a "category change" - power prices vary more than that proportion between the countries involved. And certainly not the whole multiples a pure insolation chart might imply.

The insolation requred to get decent return is often overstated on the internet - you generally see over 1000kwh for a 1kw system over a year, only dropping to ~900 in scotland. California only really gets 40-50% more useful energy out. Sure, it's less, but often close enough that it can still make sense - and more dependant on local energy and PV system costs than anything else.

The idea that you need constant, overhead, wall-to-wall sunshine isn't really true for current systems. In fact, somewhere nearer the equator the effeciency drops significantly for PV systems (even if the total power output increases slightly) as the "extra" light isn't really useful.

Again my experience with local schools doesn't match that, though I'd acknowledge that local authority can cause a lot of differences. And I'm talking about state schools - not public schools. Remember that means a very different thing in the UK, and might suggest your own distance from the claims you seem to be making.

And googling only seems to find examples of the low income programmes. I struggle to find a single instance when devices are handed out to kids and not keep at school and only used in specific tasks - like the old trolley of laptops was a few years ago.

Or breathless reactionary commentary without any actual examples of course.

They don't "hand out" anything really - probably the closest thing is government programmes to fund laptops/tablets for low income families, but not a single school locally "gives out" tablets to kids. But they're all just "normal retail" devices.

They have some things used in lessons, but they're all given out at the beginning of the lesson, then gathered at the end.

You could argue that it's a problem they they assume home access to such things anyway - especially in later years - as things like online 'homework' is the norm.

There's no (state) school giving out tablets that aren't pretty much single-use locked down devices.

That's the parents.

The expectation that "Parenting" is now outsourced to Teachers, to the Government, to anyone else. People seem to expect they just have a kid, and somehow magically they'll grow up to be a perfect person without any work from themselves. So there's over-reach, there's pressure on making "unworkable" soutions, because the people they're trying to force "solving" the problem aren't the people in the best position to do so.

Your comment seems working from that very same assumption.

Yes, all the "technical" part of content filtering etc. is very much a solved problem. The issue is that's not a "zero effort" solution - they still need to be enabled and managed. And I'm not sure that's a "technical" problem than can be solved.

There's huge pressure on teachers etc. to "solve" these sort of problems - just go to any PTA meeting and there's a lot of loud voices asking for stuff like the laws the original post is highlighting. And politicians listen to the loud voices, and feel they have to be "seen" doing something. Even if that "something" is impossible, unworkable, and fundamentally harmful.

I believe training currently costs significantly more than inference to all the current vendors, so I'd be surprised if it doesn't also use more power.

And by the look of it, that'll be the norm pretty much forever - unless something fundamental about how models can be trained/updated, an "older" model loses value as it's knowledge becomes out of date, even if we no longer get improvements from other sources or techniques.

But other things likely change based on "lifetimes" and usage patterns too - e.g. a large battery for an electric car may have a higher upfront energy cost in manufacturing than a small ICE + fuel tank, but presumably there's a mileage that the improved per-mile efficiency overcomes that, and then continues to gain with each additional mile.

All laws are vague and interpreted, and in common law (as in the UK and US) interpreted based on precedent rather than the specific text of the original law.

If people with power over you want to "selectively punish you" they don't need new laws.

And if you want perfectly proscriptive, defined laws in all situations with no "human interpretation" you're in the wrong universe, and may as well be shouting at clouds. The world, and especially human society and interactions, just doesn't follow strict definitions like that.

Yes - AI has completely destroyed the set of "Signals" people used to judge quality of much software. They weren't ever 100% accurate, sure, but they were often pretty good heuristics for "level of care", what the devs considered important (or didn't consider important) and similar.

And I mean that as both "end user" software signals, and "library" signals for other devs.

I assume that set of signals will slowly be updated. If one of those ends up being "Any Use of AI At All" is still an open question, depending on if the promised hype actually ends up meeting capability as much as anything.

That may help with tokens being "ignored" while still being in the context window, but not context window size costs and limitations in the first place.

Humans have the ability to ignore and generally not remember things after a short scan, prioritize what's actually important etc. But to an LLM a token is a token.

There's attempts at effectively doing something similar with analysis passes of the context - kinda what things like auto-compaction is doing - but I'm sure anyone who has used the current generation of those tools will tell you they're very much imperfect.

AMD's implementation very much doesn't have that issue - it throttles slightly, maybe, but it's still a net benefit. The problem with Intel's implementation is that the throttling was immediate - and took noticeable time to then settle and actually start processing again - from any avx512 instruction, so the "occasional" avx512 instruction (in autovectorized code, or something like the occasional optimized memcpy or similar) was a net negative in performance. This meant that it only benefitted large chunks of avx512-heavy code, so this switching penalty was overcome.

But there's plenty in avx512 the really helps real algorithms outside the 512-wide registers - I think it would be perceived very differently if it was initially the new instructions on the same 256-wide registers - ie avx10 - in the first place, then extended to 512 as the transistor/power budgets allowed. AVX512 was just tying too many things together too early than "incremental extensions".

LLMs have always been great at generating code that doesn't really mean anything - no architectural decisions, the same for "any" program. But only rarely does one see questions why we're needing to generating "meaningless" code in the first place.

Even Claude opus 4.6 is pretty willing to start tearing apart my tests or special-case test values if it doesn't find a solution quickly (and in c++/rust land a good proportion of its "patience" seems to be taken up just getting things that compile)

> Or, you could use spaces between em dashes, as incorrect as it is.

That's the normal way of using them in British English. Though they also tend to be the (slightly shorter) en-dashes too.

I feel that style is often pretty common on the "old" internet - possibly related to how they can be so easily be replaced by a hyphen back when ascii was a likely limitation.

It seems most of the complaints are about the reliability and infrastructure - which is very much often a direct result of lack of investment and development resources.

And then many UI changes people have been complaining about are related to things like copilot being forcibly integrated - which is very much in the "Microsoft expect to gain a profit by encouraging it's use" camp.

It's pretty rare companies make a UI because they want a bad UI, it's normally a second order thing from other priorities - such as promoting other services or encouraging more ad impressions or similar.

We loved Github as a product when it needed to return or profit beyond "getting more users".

I feel this is just the natural trajectory for any VC-funded "service" that isn't actually profitable at the time you adopt it. Of course it's going to change for the worse to become profitable.

Man I can't wait for tcc to be reposted for the 4th time this week with the license scrubbed and the comment of "The Latest AI just zero-shotted an entire C compiler in 5 minutes!"

None of this is new, it was pretty much all "best practice" for decades and so already in the training data for the first generation.

If the issue is SNR and the ratio of "good" vs "bad" practices in the input training corpus, I don't know if that's getting better.

People have also been saying we'd be seeing the results of 100x quality improvements in software with corresponding decease in cost since gpt-4 time.

So where is that?

The opus context window is 200k tokens not 1mm.

But I kinda see your point - assuming from you're name you're not just a single purpose troll - I'm still not sold on the cost effectiveness of the current generation, and can't see a clear and obvious change to that for the next generation - especially as they're still loss leaders. Only if you play silly games like "ignoring the training costs" - IE the majority of the costs - do you get even close to the current subscription costs being sufficient.

My personal experience is that AI generally doesn't actually do what it is being sold for right now, at least in the contexts I'm involved with. Especially by somewhat breathless comments on the internet - like why are they even trying to persuade me in the first place? If they don't want to sell me anything, just shut up and keep the advantage for yourselves rather than replying with the 500th "You're Holding It Wrong" comment with no actionable suggestions. But I still want to know, and am willing to put the time, effort and $$$ in to ensure I'm not deluding myself in ignoring real benefits.

I mean the systems I work on have enough weird custom APIs and internal interfaces just getting them working seems to take a good chunk of the context. I've spent a long time trying to minimize every input document where I can, compact and terse references, and still keep hitting similar issues.

At this point I just think the "success" of many AI coding agents is extremely sector dependent.

Going forward I'd love to experiment with seeing if that's actually the problem, or just an easy explanation of failure. I'd like to play with more controls on context management than "slightly better models" - like being able to select/minimize/compact sections of context I feel would be relevant for the immediate task, to what "depth" of needed details, and those that aren't likely to be relevant so can be removed from consideration. Perhaps each chunk can be cached to save processing power. Who knows.

Man I hope so - the context limit is hit really quickly in many of my use cases - and a compaction event inevitably means another round of corrections and fixes to the current task.

Though I'm wary about that being a magic bullet fix - already it can be pretty "selective" in what it actually seems to take into account documentation wise as the existing 200k context fills.

But that doesn't actually work for my use cases though, plenty of other people have already told me "I'm Holding It Wrong" without actual suggestions that work I've started ignoring them. At this stage I just assume many people work in very different sectors, and some see the "great benefits" often proselytized on the internet. And other areas don't see that. Systems programming, where I work, seems to be a poor fit - possibly due to relatively lack of content in the training corpus, perhaps due to company internal styles and APIs meaning lots of the context is taken up simply detailing takes a huge amount of the context leaving little for further corrections or details, or some other failure modes.

We have lots of documentation. Arguably too much - it quickly fills much of the claude opus context window with relevant documentation alone, and even then repeatedly outputs things directly counter to the documentation it just ingested.