I actually had a really impressive session with Fable last night, probably the most impressive agentic AI experience in a while.
I gave it a KiCad schematic of a tube-based oscilloscope from the 60s which I'm restoring. I had it give me a breakdown and priority list of components to replace, balancing safety/functionality vs preserving the originals. Then we went on a super deep dive where it explained in great detail how the circuit works and what the tubes are doing.
It isn't so impressive that it could explain vacuum tube physics and circuit theory, but it was pretty impressive that it could consume four pages of KiCad schematic and reconstruct the full topology and theory of operation with no additional information. I was able to ask it questions about what a particular tube or group of components did, or how this system interacts with that one, or what the risks and benefits of this design choice or upgrade might be. Very fluid, and its answers were actually really smart.
I have, however, found Fable to be far less impressive on coding tasks.
My powerful Android tablet is limited to 72mbps link due to a quirk with the way the XDA developer implemented wifi support on the lineageos branch of my tablet, meaning the device can't see the region specific 5ghz band of the modem of my ISP is outputting, so it can only connect to the 2,4ghz band of that SSID meaning it's stuck to 72mbps.
And despite this, it works ok for what I used it: Brave web browsing, youtube via newpipe, Plex and Jellyfin streaming.
Like I'm bummed I don't get the Gigabit and Wifi 6 speeds of the router and my internet plan is theoretically capable of, but somehow 72mbps seems sufficient in most of my use cases of that device so .. yay advanced video codecs I guess!?.
I get like 15-50 mbps down on my iPhone 16 when I'm on 5g... and that's enough to stream music, youtube, use as a hotspot, etc.
sometimes if I'm lucky i'll get much higher speeds but I guess being in a city with 100s of thousands of other people within a few miles of me means I have to do with like 40 mbps
Current admin has demonstrated very clearly that they don't need excuses or plausible deniability. The only law is whoever is willing to stop them, which is, apparently, nobody.
The courts occasionally will stop them but it appears that just committing crimes as fast as possible still accomplishes 99% of their goals without any way to fix the results of the crime.
Like most/all engineers (I assume), I have entire notebooks that are filled front to back and in the margins with scratch. Random calculations and notes, equations running into each other, any piece of information that was too fleeting to stay in my head or overspill from an already full working memory.
Most engineers I know keep such notebooks when they're full, because it's kind of neat to open a random page and see a brain dump scrawled out. Nobody has ever looked at another engineer's scratch notes and drawn useful insight. Frankly, I wouldn't even want to other than for novelty.
As per usual, the current Claude model's performance took a sharp nosedive the moment the new model was announced. Compared to the now-handicapped Sonnet model, Fable seems pretty smart I guess.
But it also really, really wants to burn tokens. I asked it to look into a fairly straightforward database bug in my RN app, and while I was off getting coffee it decided to spin up an android emulator unprompted and started navigating the app by reading screenshots and injecting touch events. There went my entire week's tokens. There was no reason to even start the emulator, the bug wasn't graphical, so I have no clue what it was doing.
Seriously. I ran it on a server with two Xeon CPUs and 128GB of RAM. The web interface took minutes to do anything. Browsing a large photo library was just completely unworkable (and that was before they ripped out all the sorting and filtering features from the photo library).
My instance killed itself somehow, likely a failed auto update. I was, of course, using the default docker setup with the watchtower instance etc etc. I never got it to come back up, and I haven't missed it at all.
Even opening the damn login page took a good 30 seconds to load, there's no excuse for this kind of performance on a real-deal enterprise server.
NextCloud needs tuning (mostly of php-fpm and caching ) oob to be fast/usable in my experience. Just throwing resources at it won't make it faster as the defaults are generally quite conservative.
As a side note, it's PHP so your single core clocks will generally be more relevant for latency than multi-core performance, feeding many cores requires a lot of divisble work.
I run it on a Hetzner vps with 4 vcpus, 16gb of ram that's also running multiple other containers. I haven't checked recently but most of that ram is free.
I'm not going to claim nextcloud is the snappiest app but I can open it right now, from across the world from my server as I'm on vacation, and it loads in <5s. I haven't done much tuning.
I run mine in hetzner cloud with 4vcpu, 8gb of ram. Via AIO including recognize, colabora and whiteboard. It runs very smooth, even when I access my photos, which number in are 100.000+
I guess you have some config / setup issues
I'm not sure how you are holding it, but I ran Nextcloud on a VPS with under 1GB RAM for many years, now it's on a 12GB RAM cloud instance and is only using a few GB of RAM last time I checked.
That's not really how it worked. It's more accurate to say that dogs co-evolved with humans in a mutually beneficial symbiotic relationship. We've only been actively selectively breeding dogs for a few thousand years.
It's a pretty good relationship for the dogs from an evolutionary perspective. All of their needs are met forever, they experience essentially no hardship or struggle, and their species gets a free ride to survive indefinitely into the unimaginably distant future by being attached to a more successful species.
Being cute is a strategy to get attached to humans. Selective breeding amplified this in recent times, but humans evolved to find dogs cute and dogs evolved to be cute in order to enforce the symbiosis. Humans didn't subjugate dogs, the presence of the human species created an ecological niche and dogs evolved to take advantage of it.
Remember that we didn't just create dogs. Wolves initially filled the role that dogs now take. Over evolutionary time, wolves adapted to better fit the niche. Not because humans wanted that, but because it was evolutionarily advantageous for a companion species to be more compatible with its host. Wolves got smaller, more friendly to humans, more protective. Wolves that didn't protect their humans lost the benefits humans provide and died. Wolves that attacked in-group humans were excluded or killed. Wolves that didn't obey their humans died due to hazards to themselves or their humans.
We only started breeding them after they became dogs. They evolved from wolves to dogs on their own.
One's perspective on this discussion really hinges on the question you raise: if an animal has a 'better' life in captivity by various metrics -- lifespan, access to food, safety from predators, etc. -- is it morally right to keep them captive? What exactly are the metrics by which we judge quality of life and what values do we assign? Does a dog (or cat, or gorilla, or selectively bred human, ) have a concept of freedom or independence that carries weight? And what if that drive is extinguished through genetic modification?
There aren't clean answers to these questions, of course. We've essentially just described the same process of wolf-to-dog evolution from the two ends of the spectrum, and neither perspective is entirely wrong or right. I would like to point out a couple things, though:
> Over evolutionary time, wolves adapted to better fit the niche. Not because humans wanted that, but because it was evolutionarily advantageous for a companion species to be more compatible with its host.
Take a step back and consider that statement. What made it evolutionarily advantageous is that humans wanted it. If humans didn't want a more compatible companion species, it wouldn't be evolutionarily advantageous. You can't separate the advantage from human desires.
Selective breeding doesn't have to be entirely intentional. Early humans weren't practicing pedigree breeding on wild wolves, of course, but nonetheless it was artificial selection that drove evolutionary change. And for the past few thousand years, we very much have been actively and intentionally breeding them. Also -- this is beside the point, really, and total speculation, but -- given how early humans understood selective breeding in plants, it seems likely that they had some awareness of the knock-on effects of favoring friendlier wolves.
Regardless, the moral unease I feel is primarily about the current state of having another species tailored to serve our emotional needs, not how we got here. Even looking at dog to wolf evolution as an entirely natural process, does that justify everything we've done and are still doing to modify them to suit our desires? Dogs as a species may benefit from our ownership, but does that make all the individual suffering we cause okay? We think of dogs living a life of privilege and comfort, and many do, but just as many have abusive owners or are living on the streets, often with worse quality of life than wild wolves today.
We're talking about things like a flashlight app on your phone with a subscription and that requests every system permission available, or the ten trillion todo apps.
I think we're not too many years away from the end state of software. We already collectively produce more software than anyone really wants, and the bulk of it is anywhere between garbage and outright fraud, if not actively malicious.
The end state, I think, is that everyone who needs small software to manage a todo list or synchronize files, or whatever "normal" people do, will end up with bespoke personalized software written by their own AI. SWEs will be reduced to working on only the big corporate projects.
The overwhelming trend in commercial software these past few decades has been hyper-aggressive anti-customization, anti-personalization, anti-user. Commercial software has been reduced to one single happy path and if that doesn't suit your needs, then fuck off. No one is making commercial software for everyday people. Even open source is trending away from everyday users.
Soon, regular everyday people who simply need some software to solve a problem the way they want it solved will have the ability to do so. In the bast majority of those cases, the quality and correctness of said software really doesn't matter. What matters is that it's personalized, free, and isn't an invasive surveillance/advertisement platform.
> I think, is that everyone who needs small software to manage a todo list or synchronize files, or whatever "normal" people do, will end up with bespoke personalized software written by their own AI
People can't be bothered to cook for themselves, and often order crappy, unhealthy food that costs 10 times as much just so they don't have to cook.
Now they're going to build their own software every time they think they need an app..?
I think this is an interesting analogy. If AI is really progressing as rapidly as some describe, should we expect a robotics renaissance with automated-chef appliances etc?
In other words, when will we really see a transition from "yet another token generator" to something that appears to coherently observe, perceive, form intent, plan, and act in a way that is compatible with an existing, long-running human context?
(And, also, do this with enough determinism to be a viable product and not some gaping liability...)
Gemini at least will produce small functional inline sample apps without being explicitly told to, particularly if you're trying to learn about something, it'll produce an interactive diagram or similar. I can see a future where these kinds of end users aren't necessarily saying "I can build an app for this", but their AI can just produce one when appropriate.
Yeah, I think this will really be a thing. I have been using `exe.dev` lately, and I was at an AirBnB with my family where we wanted to play a game, but it needed pencil and paper for each person, but we didn't have those. I thought it would be nice to have a little web app where you could write your input and then vote on other people's responses directly from your phone. I just spun up a VM on exe.dev, asked the AI to write the app, and then made the VM public-facing. It was a little buggy, but it worked well enough and we had a fun time.
> The end state, I think, is that everyone who needs small software to manage a todo list or synchronize files, or whatever "normal" people do, will end up with bespoke personalized software written by their own AI.
I think you're right here. Even for myself, AI has enabled me to actually finish a plethora of personal projects that I've always wanted to built but just never bothered.
These aren't things to share, nor would they be particularly useful to others necessarily, but now I actually have the time to make a little custom utility for very specific problems.
I still think it remains to be seen if "normal" people will do this though. Like, yeah I managed to replace a ton of little paid macOS utilities with my own software now, but AI still only got me about ~90% or so of the way there. I still had to rely on my own knowledge and experience to finish them.
Very impressive, but still a far cry from, say, the average user at my employer who struggles to even operate a non-mobile OS, being able to do this. Maybe we'll get there eventually, but for that to happen, the agent needs to be able to make these utilities 100% on its own with a very vague prompt, and be able to infer what the user actually wants when they don't (and they won't) explicitly state every use case they have in mind.
Slightly off topic, but it feels cyberpunk to me in a way. In those stories, everybody always has their own bespoke technology stack:
NullVoid added the traffic cam feed to his HUD so he could make his deliveries faster"
That sort of thing. In the stories, it makes you think that everyone is just some sort of genius, but we're kind of heading there where anybody can, theoretically, create a personalized tech stack with the help of a programming agent.
I haven't put a lot of energy into it, but your first paragraph triggered that thought.
I feel the same way about bespoke software as I do about 3D printing. For a small subset of people who like to tinker it is amazing, but there is a pretty big market for paying other people to please to 3D print for you.
For technical people who are developers or in other technical roles, sure. For everyone else, no way.
The hard part isn't the code, for most problems it never was. The hard part is being able to think logically about what problem you are trying to solve, making sure the guardrails are in place so you don't accidentally wipe your whole photo library, and staying on top of the specs for multiple walled gardens that you want to interface with. In short, maintenance.
Building is fun, maintaining is a slog. This is also why saas isn't going anywhere. There is a benefit from not reinventing the wheel, having a shared language and shared ecosystem.
On the other hand, I do think that the software that is going to succeed is the software that is the easiest to build on top of.
3D printing is similar but also a vastly smaller market than information systems. 100% of businesses need information systems, but only a small percentage need custom plastic components.
(Actually I would argue every business past a tiny size should have access to a 3D printer, it can save a lot of money in subtle ways, though its rarely business-critical)
I agree we have more software than most people want, but I don't think bespoke software is the answer. Sure, that is an interesting new area that AI makes possible, but I don't think it's more than a niche because people don't fundamentally want software, they want certain problems to be solved, and if AI creates a custom solution and it doesn't work, they won't be able to get help from anyone. More fundamentally, there is value in standardization and polish on well-worn paths. Even if you're right and people do end up with personal AI driving everything, I still think the lower layers need to be standardized because of the nature of distributed data. For example, you still need to talk to your bank to get your financial state and automate any payments, and that stuff has to be rigorous, with strong consistency and accounting guarantees.
For these reasons, I think people are overestimating the end-state impact of AI. Right now the hype cycle is fierce, and it definitely changes the economics of producing software (with a lot of negative effects forcing adjustments in open source ways of working), but I don't think in the end state the core landscape of software changes all that much. Well worn and hardened infrastructure like the Linux kernel is infinitely more valuable than CRUD apps used with small user populations on the edge. User space libraries and frameworks fall somewhere in between. AI increases the volume of new software, yes, but I see it as mostly fractal bits filling in the margins.
I think it now has a slightly different lens. The DropBox argument was that anyone can build this in five minutes, so why use this? Now, with LLMs, the argument is that anyone can build its own.
This is like saying we only need 100 websites when the internet came out. We have no idea what second third fourth order effects of frictionless and abundance are. What you think is software itself will change.
I thought we were very close until recently but I’ve just seen a long string of very promising openings from myself and others turn into “yeh dunno, it doesn’t work, I gave up”.
The effort to start is way down and drives new demand for software (at least in my own portfolio of side projects) but the effort to keep going is still above this threshold.
> The end state, I think, is that everyone who needs small software to manage a todo list or synchronize files, or whatever "normal" people do, will end up with bespoke personalized software written by their own AI. SWEs will be reduced to working on only the big corporate projects.
I think you vastly overestimate the amount of B2C software and underestimate the amount of B2B and “internal” software. Consider also how much more complex “business” processes are than “personal” processes.
I agree with this more than the parent, but think about apps is probably too last generation. The rich don't use apps. They have assistants to do things and tell them what they need to know when they need to know it. AI is more likely to be like that type of assistant.
>We already collectively produce more software than anyone really wants, and the bulk of it is anywhere between garbage and outright fraud, if not actively malicious.
I think you're right in the second part of the quote, but the first doesn't follow.
The software space has moved from value propositions to grifts like crypto, but that has more to do with what investors are willing to fund than with user needs. Modest, sustainable businesses don't have the absurd levels of growth that's currently on demand.
Consumer perception is that everything's reducing in quality and increasing in price, digital or otherwise. It will have to give at some point.
I do this already scraping web sites for descriptions of what they provide and then tell the agent to build the part I want and nothing more. There's a lot broken about the web and software today that can be addressed by these agents. Just getting a newsfeed of news I want to read free of the mandatory click and enrage bait would be progress for me IMO. But I'd never ship a product that did that because of how Google treated ad-blockers on chrome.
> What matters is that it's personalized, free, and isn't an invasive surveillance/advertisement platform
This sounds like an utopian dream. The surveillance is baked into this AI built to create the software. It will be built into the platform used to host and run the software. Why wouldn't AWS want that sweet sweet data to train their models. How many people can really self host? You seem to be overestimating average people's ability to learn how to self host.
Its like saying "we have vaccine related information at our finger tips so there are no longer going to be vaccine skeptics". Existence of information doesn't necessarily lead to application of such information.
The other thing which I feel these kinds of utopian dreams miss is that if something is commoditized and you can't really tell the difference between software A and B - because of AI, there is more incentive for companies to form cliques and raise prices while still delivering commoditized terrible software.
Really? People will just have their AI build OCR apps for converting handwritten notes (with equations) into LaTex, transcription apps for spoken language-to-text for understanding foreign language speakers, custom audio apps for recording and playing music files, etc.?
This sounds good. But technically it seems highly implausible, just as a thriving human civilization on Mars sounds highly implausible. Nice plot for a sci-fi novel though.
>Really? People will just have their AI build OCR apps for converting handwritten notes (with equations) into LaTex, transcription apps for spoken language-to-text for understanding foreign language speakers, custom audio apps for recording and playing music files, etc.?
Yes.
With a tiny subset of people building core modules and libs to be used for the above (eg. an OCR module, plenty of which already exist, and an AI can trivially hook them with other functionality into an app form).
Most of what you write can be built already quite easily. An example "Custom audio apps for recording and playing music files"*:
I gave it a KiCad schematic of a tube-based oscilloscope from the 60s which I'm restoring. I had it give me a breakdown and priority list of components to replace, balancing safety/functionality vs preserving the originals. Then we went on a super deep dive where it explained in great detail how the circuit works and what the tubes are doing.
It isn't so impressive that it could explain vacuum tube physics and circuit theory, but it was pretty impressive that it could consume four pages of KiCad schematic and reconstruct the full topology and theory of operation with no additional information. I was able to ask it questions about what a particular tube or group of components did, or how this system interacts with that one, or what the risks and benefits of this design choice or upgrade might be. Very fluid, and its answers were actually really smart.
I have, however, found Fable to be far less impressive on coding tasks.
reply