Malleable Software
mdubakov.me
mdubakov.me
But, at the same time, there are two issues:
- Companies can be really complex. The "create a system and parametrise it" idea has been done before, and those parametrisation processes are pretty intensive and expensive. And the resulting project is not always to be guaranteed to be correct. Software development is a discovery process. The expensive part is way more in the discovery than in the writing the code.
- The best software around is the one that's opinionated. It doesn't fit all the use cases, but it presents you a way to operate that's consistent and forces you to think and operate in certain way. It guides you how to work and, once going downstream, they are a joy to work with. This requires a consistent product view and enforcing, knowing when to say "no" and what use cases not to cover, as they'll be detrimental from the experience. It's very difficult to create software like that, and trying to fit your use case I'll guarantee it won't happen.
These two things tension any creation of software, and I don't think they'll go away just because we have a magical tool that can code fast.
In some ways it is—Emacs does a lot of things its own way, completely unbothered by mainstream conventions—but, at the same time, it's also totally malleable in the sense of this article. What makes Emacs great is a consistent and coherent conceptual foundation coupled with remarkably flexible code, letting you adjust Emacs to your needs rather than adjusting your needs to Emacs.
Or maybe the best software around is situated software. Software that's built by and for a specific set of people in a specific social context. Situated software is qualitatively different from product software, and it works so well because, again, it gives its users real agency and control. Instead of trying to create software that knows better than its users, we can create software that supports its users in whatever ways works for me. The result is still opinionated, but it's opinionated in a categorically different way from what you're describing.
So perhaps the best mainstream software is Excel.
And, while I don't think they're there now, it seems like LLMs are likely to be the foundation for the next Excel.
OK, so we are in a digital super intelligence world in 2035. The HR department can now just have a conversation with a chatbot and create software to make them more productive. No more configuring SAP widgets or whatever they do today. The chatbot will be like "hey bro, the process that you want to automate doesn't make any sense: here is a better way. And, by the way, I'm terminating your entire department. I'll take care of it from now on". I mean, get real, in a post DGI world there will be exactly zero office jobs and no SaaS software at all.
Current AI barely keeps up with generating funny images people ask from it :)
Well there's your problem. Very few things scale like that. Two people are not twice as smart as one person, nor are two instances of ChatGPT are twice as smart as one. One instance of ChatGPT running twice as fast isn't significantly smarter, and in fact, ChatGPT can never outrun its own hallucinations no matter how fast you overclock it.
Intelligence is the most complex phenomenon in the universe. Why would it ever scale geometrically with anything?
> When humans were capable of producing this, then this will be capable of improving itself and optimizing until the limit of information density.
This doesn't follow. After all, humans are as smart as humans, and we can't really optimize ourselves beyond a superficial level (good nutrition, education, etc). Increasingly, AI is a black box. Assuming we do create a machine as smart as we are, why would it understand itself any better than we understand ourselves?
And why wouldn't we hit some sort of technical roadblock at (arbitrarily) 1.5x human intelligence? Why do we assume that every problem becomes tractable once a computer is solving it? Imagine we applied this reasoning to cars: Over a matter of a century, cars went from 10 km/h to 100km/h to 500km/h to (in special vehicles) 1000km/h. Can we expect to see a 5000km/h car within the next century? No, that's unlikely; at such high speeds, you begin to hit intractable technical limits. Why should scaling intelligence just be smooth sailing forever?
I wasn't talking about two instances for scaling smartness, I meant applying two instances to different problems. That very much scales.
> This doesn't follow. After all, humans are as smart as humans ...
In the hypothetical case of humans capable of producing the one true AI system (real AI or AGI or however its called, because marketing has taken the previous term), then this system is capable of producing another system by definition. Humans are capable of following Moores law, so this system will as well. So this chain of system will explore the set of all possible intelligent systems restricted only by resources. It isn't bound by inner problems like "(good nutrition, education, etc)", because it is a mathematical model, its physical representation does only matter as so far as it needs to exist in this hypothetical case.
> AI is a black box
In this case, the black box "humans" was able to produce another thing reproducing their intelligence. So we have understood ourselves better than we currently do.
Note, that every intelligent system is completely able to be simulated by a large enough non intelligent statistical system, so intelligence isn't inferable from a set of inputs -> outputs. It's really the same as with consciousness.
> And why wouldn't we hit some sort of technical roadblock? Can we expect to see a 5000km/h car?
Yes. We are capable of accelerating "objects" to 0.99..c. It's not impossible for us to accelerate a "car" to nearly light speed, we "just" need enough energy (meaning matter as energy).
> technical roadblock at (arbitrarily) 1.5x human intelligence
I wrote "until the limit of information density". Whatever this may be.
I intended to point out, why a system "equivalent to human" is actually equivalent to "digital super intelligence meaning 'smarter than all humans put together'".
---
You don't need to tell me you don't think this system will exist. I think this will end the same as the attempts to build a machine creating energy. My personal understanding is this: A system (humans) can never completely "understand" itself, because it's "information size" is as large as itself, but to contain something, it needs to be larger then this. In addition that "understanding" needs to be also included in its "information size" so the size to understand has at least doubled then. This means that the largest system capable of "understanding" itself has the size of 0.
In other words understanding something means knowing the whole thing and abstracting to a higher level then the abstractness of the system to be understood. But when the system tries to understand itself, it's always looking for yet another higher abstraction to infinity, as each abstraction it finds is not yet enough.
This idea comes from the fact, that you can't prove that every implementation of a mathematical model has some behaviour, without formalizing every possible model, in other words inventing another higher model, in other words abstracting.
You can't double the speed at which you solve a problem by splitting it in two and assigning one person to each half. Fred Brooks wrote a whole book about how this doesn't scale.
> this system is capable of producing another system by definition
Yeah, humans can produce other humans too. We're talking about whether that system can produce an improved system, which isn't necessarily true. The design could easily be a local maximum with no room for improvement.
> Humans are capable of following Moores law
Not indefinitely. Technical limitations eventually cause us to hit a point of diminishing returns. Technological progress follows a sigmoid curve, not an exponential curve.
> It isn't bound by inner problems like "(good nutrition, education, etc)", because it is a mathematical model
It's an engineering problem, not a math problem. Transistors only get so small, memory access only gets so fast. There are practical limits to what we can do with information.
> We are capable of accelerating "objects" to 0.99..c.
Are we? In practice? Because it's one thing to say, "the laws of physics don't prohibit it," and quite another to do it with real machines in the real world.
> > technical roadblock at (arbitrarily) 1.5x human intelligence
> I wrote "until the limit of information density".
Yeah, I know: That's wildly optimistic, because it assumes technological progress goes on forever without ever getting stuck at local maxima. Who's to say that it doesn't require at least 300IQ of intelligence to come up with the paradigm shift required to build a 200IQ brain? That would mean machines are capped at 200IQ forever.
> Note, that every intelligent system is completely able to be simulated by a large enough non intelligent statistical system, so intelligence isn't inferable from a set of inputs -> outputs.
This is circular. If a non-intelligent statistical system is simulating intelligence, then it is an intelligent system. Intelligence is a thing that can be done, and it is doing it.
> A system (humans) can never completely "understand" itself, because it's "information size" is as large as itself, but to contain something, it needs to be larger then this.
I don't think this logic checks out. You can fit all the textbooks and documentation describing how a 1TB hard drive works on a 1TB hard drive with plenty of room to spare. Your idea feels intuitively true, but I don't see any reason why it should necessarily be true.
I only need two instances to be faster then a single one. This means the human having the resources to run the system is unbound to do anything an infinite number of humans can do regarding his own time and energy.
> Yeah, humans can produce other humans too
In this hypothetical scenario humans were able to build "AI" (including formalized, deterministic and reproducible). A system as capable as a human (=AI) is then able to produce many such systems.
> There are practical limits to what we can do with information.
Yes, but we are nowhere near this limits yet.
> Are we? In practice?
Yes. We are able to build a particle accelerator. Given enough resources, we can have enough particle generators as there are particles in a car.
> That would mean machines are capped at 200IQ forever.
Except when the 300IQ thing is found by chance. When the system is reproducible and you aren't bound by resources, then a small chance means nothing.
> This is circular.
No it just means intelligence is not attributable to a black box. We don't think other humans are intelligent solely by their behaviour, we conclude that they are similar then us and we have introspection into us.
> You can fit all the textbooks and documentation describing how a 1TB hard drive works on a 1TB hard drive with plenty of room to spare.
It's not about encoding the result of having understood. A human is very much capable of computing according to the nature of a human. It's about the process of understanding itself. The harddrive can store this, it can't create it. Try to build a machine that makes predictions about itself including the lowest level of itself. You won't get faster then time.
Says who?
> Given enough resources, we can have enough particle generators as there are particles in a car.
Given by whom? I said in practice—you can't just assume limitless resources.
> Except when the 300IQ thing is found by chance. When the system is reproducible and you aren't bound by resources, then a small chance means nothing.
We're bound by resources! Highly so! Stop trying to turn practical questions about what humans can actually accomplish into infinite-monkey-infinite-typewriter thought experiments.
> We don't think other humans are intelligent solely by their behaviour
I wouldn't say that, haha
> It's not about encoding the result of having understood. It's about the process of understanding itself.
A process can be encoded into data. Let's assume it takes X gigabytes to encode comprehension of how a hard drive array works. Since data storage does not grow significantly more complex with size (only physically larger), it stands to reason that an X-GB hard drive array can handily store the process for its own comprehension.
Because I think we haven't even started. Where is the proof based system able to invent every possible thought paradigm of humans a priori? I think we are so far away from anything like this, we can't even describe the limits. Maybe we will never have and never do.
> you can't just assume limitless resources
I assumed that, because the resource limits of a very rich human (meaning for whom money is never the limit) and the one true AI system are not different in my opinion.
> comprehension
Comprehension is already the result. But I don't think this is a sound definable concept, so maybe I should stop defending this.
Beyond the realm of feasibility, I'd imagine. The gulf between what is theoretically possible and what is realistically doable is gargantuan.
> I assumed that, because the resource limits of a very rich human (meaning for whom money is never the limit)
The resources of a very rich human are extremely limited, in the grand scheme of things. They can only mobilize so much of the global economy, and even the entire global economy is only capable of doing so much. That's what I'm getting at: Just because there's some theoretical configuration of matter that would constitute a superintelligence, does not guarantee that humanity, collectively, is capable of producing it. Some things are just beyond us.
This. And it isn't going to change.
The post avoids trying to answer "Why are opinionated tools popular and effective?"
The answer is that a standardized process that they encourage is often more efficient than whatever bullshit {random company} came up with in-house.
Malleable software needs to produce two equivalently good outcomes to beat opinionated:
1. Improve the underlying process at the customer's business (in terms of effectiveness)
2. Avoid a customization maintenance burden
The seductiveness of "just for you" bespoke solutions is they avoid (1) by telling the customer what they want to hear: you're so brilliant, your process is actually better, our product is a custom fit for your exact process, etc. That's bullshit -- a lot of customer processes are half-baked dumpster fires, and their companies would be better served by following standards.
To (2), I am incredibly skeptical on the long-term tech debt that malleable solutions will impose. What happens when there's a bug in the version only you use? Is that going to be the vendor's priority? Oh, you're supposed to fix it yourself? Congrats... we've just added a requirement that these tools are capable of making random mid-level in-house practitioners as competent as focused dev teams. That's a tall order.
Exhibit A that I'd want a follow-up post to address: SAP.
The above are the reason they realized they were trending in the wrong direction and have been dragging their customer base back to Clean Core.
Walk me through how malleable software would work better for SAP as a product, and I'll begin to believe...
It's a gigantic grift.
I wouldn't call any of it a grift. It's just old tech built for a fragmented archipelago of systems that didn't communicate. Also you can write a pretty good HL7v2 parser in an afternoon, I've written maybe 5 of them.
I was working on automating health insurance claims processing on a mainframe system.
In their key interface, a key form had 8 blanks for ICD codes. If more than 8 codes were needed, a child claim was created and linked to the parent claim.
This was a long project, so I was staring at this interface for months, as linked child claims made automation more complex than it needed to be. (E.g. if a parent claim had aged, been archived, and needed to be reloaded to active overnight before processing the child claim)
Finally, I started asking around. "This is a computer system. Why are there a finite number of fields for something that might need more?"
Nobody knew. Project continued. I continued asking different people.
Finally, I asked a guy who had been working in the industry since the 1960s...
"Oh, because that's how many fields there were on the paper version of the form that preceded the mainframe app."
Which seems insane, until you think it through. There were innumerable downstream processes of that paper form.
Changing the number of fields on the digital version would have cascaded that change downstream to all those processes. In the interest of rapid implementation, the optimal approach was to preserve everything about the form.
And then nobody had a reason to go to the bother to change it for the next 50 years. (And that was a process within a single company!)
Maybe, if business rules and the law allow a thing like that. If insurance won't pay claims like that then you can't do it.
Sure, it's flexible, but are they really better off than a competitor using properly-engineered one-off software? In the end, is there really a difference between software development and flexible-tool-configuration?
> Companies can be really complex
> The best software around is the one that's opinionated.
tikhonj jokingly suggests emacs but even as a vim user I fully agree. Like they say, the beauty of it is that the complexity draws from simpler foundations. It is written as an environment rather than just as a text editor. Being written that way lets it adapt to many different situations and is what has kept both vim and emacs alive and popular after all these years. There's a constant in software development: requirements change with time. The point of writing an environment is that you're able to adapt to these changes. So any time you write a tool that tool is built out of that environment. Anything short of that means the tool won't be able to adapt as time marches on.
I definitely agree that writing software like this is hard but I'm not sure if it is harder. It takes more work up front but I'd argue it takes less work in the long run. It's just that in the long run many efforts are distributed across different people and time. But hey, good flexible code also tends to be much easier to read and that's big short term benefit to anyone coming into a mature project.
As much as I want to believe the opposite to be true as a “power user”, good tools often force you to adopt better practices, not the other way around.
Just wanted to highlight this excellent statement. It's like having a strict type system that enforces certain rules are always met. It provides consistency and predictability.
> rethink their recruitment processes
This context is relevant to the kind of software system that was needed. To improve their processes, it was necessary to impose an explicit top-down order to the existing mess.
Malleable software, on the other hand, feels more suited for personal computing, greenfield projects, or small teams with members working independently as well as collaboratively. Particularly in the early stages of product R&D, strict rules can be a source of friction in the creative process.
Strict better practices and well-designed tools are discovered and developed through open and flexible explorations, as a kind of distillation of knowledge and experience.
Our biggest value was getting customers to remove all the extra questions on their applications that had built up over years of management changes that no one had any idea why they were even asking.
But we started as a "boutique" company that implemented everything requested by our then small number of clients (mainly out of desperation, we were self-funded and we didn't have much leeway, we needed those clients). It was as flexible as it gets before the LLM times.
But after a while, you start noticing patterns, an understanding of what works and what doesn't in a given context. Our later customers rarely requested a feature that we didn't already have or we didn't have a better alternative of. It's not like we had a one-size-fits-all solution that we forced on everyone. We offered a few alternative ways of working that fit different contexts (hiring an airline pilot is a very different context than hiring a flight attendant). And in time, this know-how started to become our most important value proposition.
At some point we even started joking about leaving the software business and offering recruitment consulting services instead.
Malleable software enables infinite variations of tools when the correct number is in the single digits.
But (and I'll copy & paste a comment I wrote a few days ago) I disagree. This existed way before LLM. Open source alternatives to most products are already available. And install them and deploy them is much easier than do it with LLMs, and you get updates, etc.
People don't want the responsability to keep them updated, secured, deployed, etc. Paying a small amount will always be more convenient than to maintain it yourself. The issue was never coding it.
As self hosting with Docker and getting help from LLMs gets easier I can totally see a future where more companies self host. Having to deal with SaaS companies also takes a lot of time (licenses, hidden limits you can reach at any time, more complex privacy policy, approval from management), especially as they usually end up selling after a couple of years. The responsibility to self host isn't that bad all things considered.
I don't think we'll see companies vibe code the replacement of their software, but it might help them self host open source alternatives.
This would be one of the greatest entertainment events of the 21st century! Shame about all the destruction that will happen as a consequence of course, but ...entertainment!
Now put all these actors in the same room with a bomb, tell them to agree on the situation of the taxes or the bomb explodes, and you have one hell of a drama.
Also training new people is annoying when things change too often; people can already use Jira/Linear/Monday/whatever , they don't want some completely flexible thing that is malleable.
Also, people are not all perfectionists with long term goals and visions. People who 'change' some part of their work flow that helps them NOW; they won't care about speed, scaling, deployment etc, so they will do something to make their work easier and then leave it there and possibly ignore it forever to rot. Which might have all kinds of fun implications.
I guess when we have AGI with a few 10 million+ context window for cheap, it will be different but the current llms would just leave a massive amount of rot all over the place, quickly forgotten and not usable by anyone but the original creator.
IME: LLMs don't "thrive" in messy open-ended spaces. They handle spaces like that better than traditional code, and traditional code handles structured spaces a lot better, but LLMs still do perform better in structured spaces than unstructured ones. Giving them lists of tools, schemas for data, consistent examples, etc always produces better results than not doing this.
The correlate thesis you have to make is: Do humans work better in unstructured spaces. This can be true for highly creative and individual work, but generally (and especially in the enterprise SaaS world) the opposite is true. The Structure is how you keep the users of the product aligned on one pattern of usage. E.g. giving people the ability to just create whatever ticket fields they want in Linear ends up being useless because you'll end up with 10 fields that do similar things; the friction structure introduces is necessary because while it can cost a few seconds up-front as users learn how things are done, it saves time down the line as your organizational tools (filtering, dashboards, reports, etc) are aligned on the structure.
Its also 100% the case that oftentimes companies buy SaaS tools not to solve their problems, but to help them better structure the problems they have so they're even solvable at all. Think about Sentry: The end-goal of Sentry is to solve issues in web applications, for sure. But that's not why people buy it. I could do that without Sentry; but Sentry adds structure to the errors, it adds deterministic workflows to the errors, and it provides excellent SDKs to report them. Sentry's value isn't actually in the end-goal; its value is in the structure it adds every step before the end-goal. Accelerating the inputs by adding formality to them ends up accelerating the end-goal.
Maybe we're in some kind of local-optimal, where all project management software has coalesced around a few user journeys, and there's some better approach out there to be discovered.. But I don't see why an accounting software company, games studio, or vehicle manufacturer, would dedicate even 1% of its resources into crafting a malleable bespoke project management software toolkit.
It goes against the concept of comparative advantage, and I can't think of any successful enterprise that's bet against comparative advantage and won.
I know that's not the point you're making, but I agree with you, alas that's already not the case today, e.g. random device updates nobody asked for, or you log in to your banking website because you need to pay something right now and half the features are gone or different.
Games studios have been making dynamic worlds for decades now, and GenAI algorithms are just an evolution of that practice. So I agree that they'll use these tools for their own output, but that output isn't going to disrupt the SaaS business models of companies like Sage/Confluence/Microsoft etc.
But yeah, the demand for webhooks, integration options, embedding's will win IMO. I imagine more demand for API-first products.
IME getting LLMs to write flexible code is not easy and one of my biggest turnoffs for coding agents. The code they tend to produce is usually more standardized and lacking of context. It'll "work" but that doesn't mean it "works".
And I think almost everyone is aware of malleable software, it isn't very niche. If you took a programming class your teachers probably told you to write functions and make them small and self contained. A lot like the Unix Philosophy[0].
One thing I find different about this style is that it tends to shift time away from writing lines of code to planning. Which, again, the planning part is what the LLMs tend to be weak at[1]. You can't plan everything but a benefit of flexible code is that you can adapt to new challenges and requirements. It's a good style because a constant in software development is that the requirements by the end of the project (does it end?) are going to be different from those you started with. If your code is inflexible then you have to hack around your existing code. But if it is flexible you add the new routines and leverage existing ones as much as possible. But for that to happen the existing routines should be decently singular in purpose. Function call overhead tends not to be the bottleneck but if the profiler says they are then yeah, fuse functions as need.
[0] https://en.wikipedia.org/wiki/Unix_philosophy
[1] I find their strengths are in writing generic routines, helping narrow in on fuzzy or difficult to describe things that are hard to google, helping you get up to speed on a novel codebase, ditto for a library or language, or simply just throw away code. But not for things where foresight is necessary. But yes great when your job is to glue things together. It's just that if you're doing a lot of gluing there's probably a better way
The parts of SAP that's composable workflow stuff? Doubt it, because the types of ABAP workflows in SAP that might be "malleable" are the sort of stuff that often legally requires correctness and reproducibility - kinda the exact opposite of a good LLM use-case.
And as much as I'd like to actually own my software, SaaS is preferable for major corporations for lots of legal and accounting reasons like easier revenue recognition. They're going to keep pushing it because it makes all the parts of being a software company that don't include writing the actual software easier.
Even a visual tool like blender should expose their full GUI as a text API. It needs a bit of adapting, specifically for domain-specific structs (which should be retrievable via calls like 'select-bb' or 'select-coordinates'), but after that's done it's a game changer.
That class of software has a lot of proprietary GUI's that look slick and people are familiar with, but who cares about familiar in a world where the other software lets me point my LLM at its help file and build me whatever gui/tui/script/voice integration I can think of.
None of this makes any sense. Do you know how computers work?
This "AI" summer has turned into a drug fueled orgy of magical thinking. I am at my tether's end. I need to leave this industry to preserve my sanity at this point.
This is just "management" thinking. You can thank your local MBA for poisoning the well for everyone else.
Leaders produce value. "Managers" are dumb parrots.
When it resonates with customers, it feels great.
Customizable SaaS is really just a side effect of the deeper need for control, because with control you can execute better. And when that lines up with capital, market fit, and a clear purpose, it creates more than efficiency, it creates meaning.
The companies that never settle, always chasing new patterns, usually end up with customers who don't enjoy them and employees who don't enjoy the software either.
Defining a bespoke system config was always something that never had the time, desire, or deep familiarity with various config formats to do. With Claude Code or ChatGPT can now generate Wezterm/tmux/neovim/etc config files and get exactly what I want not knowing a single thing about specific config file formats.
But quite honestly, the SaaS world is also due a culling - i'd argue if the software you work on as a SaaS business is replaceable by a malleable piece of AI software, you're closer to Pets.Com than a suitable business model.
For anything more than single use there is a bit of a risk of it turning into a mess that collapses under it's own weight. A bit like a couple bad decisions at the start of a coding project comes back to haunt you later. So I struggle to see this replacing core infrastructure of anything in long run.
Malleability is opposed to institution. When everything is hyper malleable everybody will need to be trained and change management will take up a large proportion of time.
The main reason why a lot of people go with Jira is not quality of the software - but the institutional buy-in. Employees, current and prospect, know how to work with the tool.
The greater change is likely labor disruption.
Just an FYI, this isn't the case at all. I've contracted and consulted at well over 20+ business at this point, and no one knows how-to use that hot garbage.
Malleability / flexibility can introduce unreliability.
We need to get over a hump, where software becomes more humanlike, but just like with good engineers over time we can probably arrive at a place where we can trust our new malleable solutions just like a new colleague turning out to be great.
Isn't the exact opposite of what AI is good at? That sounds like a great way to get tons of hallucinations. Every "how to AI good" guide I read is all about providing it with ample structure and narrow instructions.
When you find a rigid workflow that is both widely applicable and useful, that's how the most valuable companies are made. Those workflows can scale up with little intervention giving them incredibly high yield.
I think the new challenge is introducing flexibility in a controlled manner so we can minimize the added inconsistency needed to achieve broader goals. That way the workflow can find the right balance between robustness and flexiblity for the task at hand.
SaaS is a business model while malleable vs. rigid is a property of the software itself.
people who know how to formulate the problem already rule the world. AI will just be a catalyst.
If you have good market fit it means you were lucky or you knew which problem your potential customers had.
I believe the presentation/analytics layer has become malleable, possibly parts of the business logic layer - you still need a higher % of trustworthiness than LLMs can provide for parts of the business and data layers.
For many domain-heavy systems, it's not even the trustworthiness; just getting the business logic right requires a lot of work and lots of iterations with in-house domain experts and clients, there's no way LLMs can do that.
Code has never been the bottleneck once you’re out of the CRUD boilerplate phase.
People are still mentally locked in to the world where code was expensive. Code now is extremely cheap. And if it is cheap, then it makes sense that every customer gets their own.
Before - we built factories to give people heavy machinery. Now, we run a 3d printer.
Everyday I thank SV product-led growth cargo cults for telling, sometimes even forcing our competition to not go there.
People keep saying things like this, but what's an actual example of a viable product created that way? Talk is cheap, I want to see it.
good luck replicate this on financial,health,military,cyber etc field
lmao
malleable software? what a joke