> The theoretical advance we're waiting for in LLMs is auditable determinism
I think this is a manifestation of machine thinking - the majority of buyers and users of software rarely ask for or need this level of perfection. Noise is everywhere in the natural environment, and I expect it to be everywhere in the future of computing too.
Based on what I've seen so far, I'm thinking a timeline more like 5-10 years where anything involving at least frontend has all but evaporated. What value is there in having a giant app team grind for 2 years on the perfect Android app when a user can simply ask for the display they want, and 5 variants of it until they are happy, all in a couple of seconds while sitting in the back of a car. What happens to all the hundreds of UI frameworks when a system as a widespread as Android adopts a technology approach like this?
Backend is significantly murkier, there are many tasks it seems unlikely an AI will accomplish any time soon (my toy example so far is inventing and finalizing the next video compression standard). But a lot of the complexity in backend derives from supporting human teams with human styles of work, and only exists due to the steady cashflow generated by organizations extracting tremendous premiums to solve problems in their particular style. I have no good way to explain this - what value is a $500 accounting system backend if models get good enough at reliably spitting out bespoke $15 systems with infinite customizations in a few seconds for a non-developer user, and what of all the technologies whose maintenance was supported by the cashflows generated by that $500 system?
I expect pretty much the opposite to happen: it makes sense for languages, stacks and interfaces to become more amenable to interfacing with AI. If a machine can act more reliably by simplifying its inputs at a fraction of the cost of the equivalent human labour, the system has always adjusted to accommodate the machine.
The most obvious example of this already happening is in how function calling interfaces are defined for existing models. It's not hard to imagine that principle applied more generally, until human intervention to get a desired result is the exception rather than the rule as it is today.
I spent most of the past 2 years in "AI cope" mode and wouldn't consider myself a maximalist, but it's impossible not to see already from the nascent tooling we have that workflow automation is going to improve at a rapid and steady rate for the foreseeable future.
There will always be "real thinking" roles in software but the sheer pressure on salaries from the vastly increasing free labour pool will lead to an outcome a bit like embedded software development, where rates don't really match the skill level. I think the most obvious strategy for the time being is figuring out how to become a buyer of the services you understand rather than a badly crowded out seller
As I recall it, there was a time when copyright infringement on YouTube was so prolific that the rightsholders essentially forced creation of the first watermarking system that worked at massive scale. I do wonder if any corners of research are currently studying the attribution problem with the specific lens of licensing as its motivation
the IP rights holders have yet to bare their teeth. I don't think the outcome you suggest is clear at all, in fact I think if anything entirely the opposite is the most probable outcome. I've lost count of the number of technology epochs that at the time were either silently or explicitly dependent on ignoring the warez aspects while being blinded by the possibilities, Internet video, music and film all went through this phase. GPTs are just a new medium, and by the end of it royalties will in all likelihood still end up being paid to roughly the same set of folk as before
I quite like the idea of a future where the AI job holocaust largely never happened because license costs ate up most of the innovation benefit. It's just the kind of regressive greed that keeps the world ticking along and wouldn't be surprised if we ended up with something very close to this
The frequency of AI cope posts appears set to keep increasing until the very last one of us is fired. The reality is that users simply don't care about any concern to be found discussed in the church of software engineering, they want an app with a pink button, and very shortly we will be a much higher friction means to achieve that than the sloppy text generator. It's the same force behind why you can't buy high quality furniture any more or consumer devices that survive more than a few years: the vast majority of people simply don't care, they want to click "Buy" at a price level they don't have to think about and get on with their lives, only now it applies to knowledge economy work and entire industries are trying their hardest to ignore the transition we're so painfully obviously already in.
Even were it not the case for disposable CRUD Android/web apps that represent the bread and butter for half the industry, the effect on the structure of popular media is alarming in ways I don't think anyone has a hope of understanding yet. I imagine kids coming up now will not be glued to their phones anywhere nearly like the current batch are, or if they are glued to something, perhaps it is a conversational agent prompting them through an earpiece or similar.
Hand-wringing about the quality of LLM-driven app development really misses the point of all of this. We're currently using an extremely novel technology to emulate aspects of our now-defunct technology (which I believe includes the web), in much the same way fax gateways at one time were a popular application for email.
More efficient hardware mappings will happen, and as a sibling comment says, power requirements will drop like a rock. Check out https://www.youtube.com/watch?v=7hz4cs-hGew for some idea of what that might eventually look like
UML and "as little overhead as possible" probably shouldn't appear in the same train of thought. I remember it from the very earliest Linux VPS providers, IIRC it only got semi-usable with some custom work (google "skas3 patch"), prior to which it depended on a single process calling ptrace() on all the usermode processes to implement the containerization. And there's another keyword that should never appear alongside low overhead in the same train of thought
It gets more interesting when you think about the impact on groups. Sending an image to a group is enough for all devices associated with that group to be identifiable from CloudFlare's side, who additionally see a giant chunk of unencrypted traffic from the same client addresses going to other web sites. Given Cloudflare's less-than-straight approach to sales, it is astonishing the words "secure" and "Signal" ever appear in the same sentence.
CloudFlare get to see a fuckton of metadata from private and group chats, enough to trace who originally sends a piece of media (identifiable from its file size), who reads it, when it is is read, who forwards it and to whom. It really doesn't matter that they can't see an image or video, knowing its size upfront or later (for example in response to a law enforcement request) is enough
Someone mentioned "App Store" during a meeting, probably accidentally. It's a known killer, the tech world equivalent of sneezing while carrying an infectious disease
> good enough to stream YT, in my experience, so presumably already good enough to attend meetings
YT needs bulk throughput while meetings need latency and quality. YT can seem smooth for much longer despite massive amounts of retransmission and packet loss, meetings fall apart rapidly with even a tiny bit of those
I don't know how meaningful it is any more, but with long polling with a short timeout and a gracefully ended request (i.e. chunked encoding with an eof chunk sent rather than disconnection), the browser would always end up with one spare idle connection to the server, making subsequent HTTP requests for other parts of the UI far more likely to be snappier, even if the app has been left otherwise idle for half the day
I guess at least this trick is still meaningful where HTTP/2 or QUIC aren't in use
This is the first "rewrite it in rust" reason I've heard that actually makes total sense, congratulations. Ability to safely hack on ancient code really does sound attractive