As a subscriber I recently go this. Probably someone in OH central committee did as well.
Hello,
We wanted to give you a heads up about some upcoming changes to your Copilot subscription. These updates are part of our ongoing effort to make GitHub Copilot even better for you.
The default review effort level for Copilot code review is changing from Lite to Balanced. For existing and new repositories and organizations currently set to Default, the switch to Balanced takes effect on September 28th, 2026.
How defaults work
You can set a default review effort level at the organization and repository levels:
The organization default applies to all repositories owned by the organization that haven’t selected their own review effort level.
The repository default applies to all automatically requested reviews within the repository. For manually requested reviews, you can choose the effort level from the Reviewers bar in the pull request.
What you need to do
If you want to keep Lite as your default, take action before September 28th, 2026.
Go to your organization or repository settings.
Explicitly select Lite instead of Default (Lite).
We’ll respect that selection and won’t switch you to Balanced. No action is needed if you’re comfortable moving to Balanced or are already using Balanced.
Let me add that AI also doesn’t learn from your feedback. It will keep repeating the same mistakes/bad smells over and over again (I am aware of AGENTS.md and all the other MD files, but I personally don’t believe this can replace the learning process a human has.)
I don’t know what to make of this, but the fact that they’re changing the default level is probably without much meaning for people that have any clue what they’re doing. I assume that they have been using the highest level they’ve had access to, because Copilot was only ever really usefull for some months, before that it made too many stupid mistakes and was rather an annoyance for the most part in my view. So, I can’t really imagine that “lite” is good for anything, and I really doubt that’s what has been used when the reviews have been “useful”.
This is probably rather “business as usual”, now that they have upped the price for using Copilot quote a lot (mostly by reducing quotas, so that you much quicker gets into "buy tokens territoriy), they want to increase token consumption. Fiddling with default is very effective, because there are so many people that don’t know or care to know how things work, and those will just accept the default, whatever it is. As such, Microsoft can actually tune how much their customers will consume - that must be the capitalist dream
I completely agree and that momentum is exactly what keeps a project alive. When I first started using openHAB, I needed to integrate a Modbus setup. It had the capabilities I needed, but more importantly, it had an active, helpful community to guide me in a pre-AI world.
Things are totally different now with AI. There’s a delicate balance: if we put up too many fences, we risk driving good contributors away, but if we have no standards, others will get fed up and leave too. Today, I see openHAB as a powerful toolbox where I can just build and add new tools using AI, and keeping that barrier to entry reasonable is key.
I’m not sure I follow on this bit - has there been any “fences” suggested that you think would drive good contributors away? I get what you’re saying in principle, but I’m not sure if any such measure has been “on the table”..?
If the walls are build! For the moment there is nothing everybody is still trying to understand this new phenomenon. Should be actually nice even though it’s not going to happen in this project if someone is leading the project and assigning tasks something like Linus is doing should be nice.
I just watched a video interview about Zig, and I must say I really align with his views on AI (and from what I’ve seen in the video, on values more broadly):
Sadly, Zig lacks OOP, which I don’t think I’m prepared to live without, otherwise I would be all over it. I really like the philosophy that you can choose the memory management model that fits the task best, I’ve always wondered why languages insist on picking one model and then trying to convince the rest of the world why that model is “the best” - when clearly, needs are different in different circumstances.
I’m watching all the videos (Ed Zitron, Andrew Keley) with interest. My take is that neither of these people are qualified to talk about AI in terms of usefulness.
Ed Zitron is not even a coder. His main claim to fame is his criticism of AI, and his view is about the economics of it. Irrelevant in my view.
Andrew Kelley has never used AI for coding himself. His experience as far as I gathered from that video (I didn’t watch it in its entirety, but quite a bit of it), was that one of his friends showed him how it works. And although in this context it is irrelevant how long ago that was - AI coding tools are improving at a rapid pace. I am guessing His experience was from random vibe coders using AI “mindlessly” and him reviwing such garbage. So his opinion that “AI is Garbage” is sad at best.
Granted, AI (for coding) is not perfect. It makes stupid mistakes and stupid assumptions. It is way too rigid in its logic and “thinking” (which is actually a good thing!). However, I can totally understand why it does what it does: It only goes by the input it has available. It is not a mind reader. That’s its main weakness. Where a person has the entire history of a project, the background reasoning of everything, understands the “problem” etc. The AI is some code inside a computer deprived of all of that and given a piece of code and people expect magic to happen.
And I’m talking about frontier models. If your experience is NOT with a frontier model and instead based on some 30-70b parameters running on a home server, it’s not the same. Apples to oranges.
AI is a tool, and like any tool, you have to understand what it can and can’t do, how to use it and when to use it. You can’t just buy high end brush, paint and canvas and expect anyone to produce a masterpiece. But that doesn’t negate the fact that a good tool is still a good tool.
The only problem is that it’s so dang expensive for us hobbyists.
I found what he has to say very interesting, because it “makes sense to me” (matches my perception/feeling of what’s going on). That doesn’t mean that what he says is correct, but it’s not something that I can reject immediatey as BS, which is the case for much other material on the subject (I got into one of those that are afraid that AI will “take over” eariler, in it was unbearable to listen to). It’s hard to know what’s the truth anyway these days, what information can you really trust? They aren’t exactly forthright about the details, so one must try to pick sources that “sounds credible”.
I disagree that the financial side and the suitability for coding isn’t linked. It very much is, and that’s what I’ve been trying to say: They are subsidizing the use today, by how much we don’t really know, but it looks likt it’s a lot. When you value whether it’s “useful”, the price is a part of the picture. If it ends up being able to write mediocre code that costs 5 times what it would cost to pay a human to do it, it’s not really useful. The subsidizing can’t go on forever, so, before you start basing how you work on these things, you should make sure that you know that what it provides “is worth it” when you have to pay the full price + their profits + mortgage on the gigantic loans they’ve taken on to build the data centres.
If I could run one of the “frontier models” (it’s embarassing to have to say, thus the quotes) locally on my laptop without paying anybody or losing control of my data, I would see everything very differently, I would probably try to see if I could find a way to make use of it too. I very much doubt that I would use it to generate code though, because we “know” that the model doesn’t understand logic and is just a “median machine” that tries to find how something is “usually” solved. As such, you can never trust it, because even if it came up with something completely decent 9 times out of 10 (which I consider hightly unlikely), it would still fail miserably 1 times out of 10. But, I would probably try to get it to do as much else as possible.
That’s not the situation though, and I really don’t understand why so many seems to pretend that what we have right now is going to last. The economy of this matters a lot, and it looks to me like these companies might not be around for that long. When the bubble bursts, all the people willing to invest the pensions in this will dry up, the circular financing will stop working, and companies will cease to exist.
The question I guess, is how much of these money some of the people involved have managed to siphon off in the process, and who can buy what’s left for cheap. But it very much matters what happens after that, to know if the “tools” you want to use even exists, and at what price.
Even though the Chinese models are much, much cheaper to make, my guess is that they are net loss projects as well. But, it’s obviously worth it for the Chinese leadership, because they can use them to wreck the US economy. I’m not at all convinced that China will continue training models “for free” once the US bubble has burst.
People must stop claiming that anybody that hasn’t “plunged in” have no reason to have an opinion. This is cult-ish behavior, where “you can’t possibly understand this until you have been blessed by our dear leader”. If you had to try every snake-oil rememdy out there to be able to reject it, the scammers would win by overwhelming everybody.
There are plenty of information available to evaluate this without actually using this or that model. Andrew Kelley explained that his experience was first and foremost reviewing a lot of PRs. But, there’s also quite easy to know that LLMs can never write good code, because they don’t understand logic. The whole idea that you can take something that works pretty badly, and just scale it up far into insanity, and then hope that this will magically give you something entirely different, is naive. There’s a difference between you not being able to tell that the code is flawed, and that the code isn’t flawed. What you achieve by this extreme scale (just increasing model size, context size etc.) is just that you get something that is extremely good at faking it. It can’t “magically” suddenly become intelligent and understand logic, but it can be good enough at guessing what you want to “fool you” into believing that it knows what it’s doing.
No, the main weakness is that LLMs can’t “think”/reason. They are “statistical machines”, they don’t handle logic or understand how things are connected. What they know is how people usually/statistically describe how things are connected. That’s still very different.
I don’t think it’s too rigid at all, I think it’s just fundamentally flawed. I also think that the “AI companies” have started to realize this, which is why they have begun with the whole “agentic coding” thing, so called “reasoning”/“thinking” etc. They are using other tools, Python code, to try to restrain the models and improve the perceived quality. But dividing the tasks into smaller tasks, and delegating each subtask to different “agents”, they are really just trying to reduce the complexity, because it’s obvious that complexity reveals the lack of logical ability very quickly, which thus “expose” them. In addition to trying to reduce the complexity, they have also added another “role” where the LLM will question it’s own findings, and do it again if it isn’t satisfied. This burns a whole lot of tokens, which is why tokens disappear very quickly if you use the latest/“best” models.
But, it’s still not changing anything that make any real difference. It’s just a way to try to hide the weaknesses, to try to convince people that “I should pay money for this, this is better than slop”.
As to the model not being a mind reader, I agree - it’s very fra from it, but the point is that you’re supposed to give the model access to the workspace if you want it to understand how things relate, and it can go out on the Internet and find information. There’s nothing that prevents a model from looking through Git history to “understand” the history and how the problem came to be. But, the same is true for humans. Most people aren’t able to remember any significant amount of code, I find myself looking thing up constantly when I try to figure something out. I look up code that I’ve looked at many times before, because I’m not certain about this or that detail or that I remember it correctly. Then I search the Git history, forums, GitHub and whatnot, to try to figure this out. I can’t see why the LLM would have any “disadvantage” here, it should in theory be able to do everything I do, just much quicker.
I think that’s not the most fortunate example, because perhaps the biggest problems with LLMs is that they are compulsive liars. Whenever they don’t know, and quite often even if they have access to the correct information, they will just make stuff up. That’s what ultimately makes them so untrustworthy, why everything they “make” must be checked manually, and thus why much of it is pointless. This would represent that the brush would paint things you never told it to, suddenly skip a part or paint something in a different color. If I buy a high end brush, I expect it not to paint something I didn’t “command” it to do - I even expect that of cheap brushes.
I suspect that it won’t just turn out to be too expensive for hobbyists, but also for businesses. Let’s face it, as soon as it gets more expensive than what you’d have to pay a qualified person to do the same task, it is pointless. And we still have no idea what the real costs are, just that they are huge. The real price could be 40 times what they charge now. If you factor in that, their profit and paying of the gigantic loans, it can easily end up being so expensive that it becomes completely irrelevant. Businesses are rumoured to already be backing off, finding that there’s no provable benefit, that it’s most expensive than paying humans to do it, that quality plummets etc. If they are backing off while this is still heavily subsidized, how can this ever work?
Therein lies the crux of the matter. From what I see, it does not write mediocre code. Again - when given proper context and instructions.
It is genuinely useful to me. Even if I have to go analyse and refine the result, what it produces is still very useful and most definitely boosted my productivity. When I see a problem, I can usually imagine how much work it would take me to tackle it. AI can do the grunt work in minutes what I will have to spend many many sleepless nights researching and figuring out. Even if it means ruling out one approach, it still saves me a lot of time.
Questioning AI’s usefulness today is extremely naive.
I am not looking at it from a corporate point of view. So the question of whether it’s better to pay AI vs pay an employee to do it is completely irrelevant to me. In the context of openhab, I am doing this 100% for fun and hobby. Therefore the entire economic argument from Zitron, whether it is true or not, is completely irrelevant to me. I don’t care whether openai / anthropic is profitable or doing it at an extreme loss. These two topics are two different universes as far as I’m concerned.
The only thing relevant to me is whether the AI produce slop unusable garbage or whether it is useful. For me it’s unequivocally the latter, even though it is far from perfect.
And people must stop making claims about oranges (what others experienced) when they’ve only tried apples (what you experienced).
That statement and everything you’ve said about LLM so far really tell me that you are definitely not talking about the same thing yet continue to think of them as the same.
I often hear people dismissing AI as “autocompletion on steroid”. Invariably those are the people who tried AI at the beginning, dismissed it as stupid (rightly so) and formed a rigid unchanging opinion on it, 4 years later.
At this point we cannot (yet) expect it to match what a human can do in terms of reasoning, and this goes back to what I said: you need to know your tool. Yet at the same time it is able to see some things that humans can easily miss. If you come with unrealistic expectations, then you will be disappointed. We are not yet at “AGI” level. I know throwing the term “AGI” here will open a new can of worms here lol.
That is just a very cynical view of it, but it is not the only way to look at it.
This thing is still in its infancy / toddler stage. Only 5-6 years ago ChatGPT 3 was revealed and it was hilariously stupid. Google’s demonstration of Bard that ended in embarrassment was just in February 2023!
But regardless of that, and regardless of all of its flaws, I find it already useful today.
This makes me think of strawman’s fallacy. My point was that AI is a tool - you must know how to use it. And if you’ve never used the same tool, you’re just fantasizing about something you’ve never tried. However, I’m afraid with your biases, even given the opportunity to try it, you will only try to find what’s wrong with it.
If so, there’s a lot of naive people around. There are no question that it can do certain things, more or less correct, but the fact that you can’t trust it to consistently do it right kind of erases the value of it all. There’s quite the irony that “AI generated” has started to be synonymous with “the lowest low of quality” among many people, while others argue that “it’s extremely naive to question its usefulness”.
I read about a person that had used years of effort to write a book, but didn’t have a lot of money, so chose to use an AI generated book cover (image). He explained that people wanted nothing to do with the book itself, because they automatically assumed that the content was also AI slop. If this is true, it will soon have to be “hidden” that things are AI generated for people not to reject it outright. Are all of these people just now realizing the consequences of the experience they made in 2023?
The thing you seem to be missing is that if it doesn’t make it economically, it will cease to exist. There will be no “frontier models”, there will be no more burning billions every month to try to capture market shares in a market that doesn’t exist or is very small.
So, it’s not optional whether you want to take this into account. If these companies disappear, or the services becomes extremely expensive, all you’ll have is the open weight models that you can run yourself. Or, you can host the model at some generic provider where you pay what it costs to run the model. The situation we have now is just temporary, it won’t last. That’s what makes the economic aspects so relevant.
I have tried Claude Haiku 4.5, and have used Gemini quite a lot. I haven’t used the most expensive models at any point in time, or actually, that might not be true, I think that Copilot activated some “trial month” or something for me back in May where I used Claude something via that.
It’s not that I’m not willing to see what I can get out of it, I just flat out refuse to pay money into the scheme they are currently running. I’m also unwilling to give them any personal information, but I’ll try and evaluate everything I can within those limitations.
And, what I’ve seen this far, is just more of the same. Throwing more size and compute at it makes it “look better”, but the fundamental problems with LLMs not understanding logic shines through if you try to do anything remotely complex. And, frankly, I don’t understand how anybody can expect anything else. Scaling up doesn’t give it abilities it doesn’t have. It still does the same thing, only with “higher resolution”.
I think there’s an element of sarcasm involved when people describe it that way, it points to what an LLM does, and is designed to do: predict the continuation of whatever text it’s presented with. I’m not saying that’s how you use it, but that’s what it does “at its core”. They have then packed it in using various layers to make it more useful to humans, but it’s still what the model does at the fundamental level. It’s also a large part of the reason why LLMs are such sycophants (in addition to RLHF which is just insane). They aren’t actually trying to figure out what is correct, they are trying to figure out what you want to hear. That’s also why so-called hallucinations are so hard to get rid of, it’s there “by design”. When you know that they don’t actually understand the data they are handling, and that they are “made to please”, there should be no surprise that they make up “facts” as needed.
So, I don’t think you should be so quick at dismissing what people that say things like this have to say. You might disagree, but don’t assume it’s because they concluded 3-4 years ago and never looked at it again.
I haven’t really had any experience with LLMs before the last 8-10 months or so, and that’s because it’s become more or less unavoidable. So, any experience I have, is relatively “fresh”. I did spend quite a lot of time with diffusion models back in 22/23 though, but that’s not LLMs. They do share the statistical approach though, so it’s easy to me to draw some parallels. Diffusion models also “hallucinate” and also share the complete lack of understandinig of what they are generating. It is why things like light, shadows, the number of limbs, fingers, toes etc. is almost impossible to get right. And, if you try to micromanage it to “force it” to get something right, it will probably break something else, because the part you “force” then no longer “fits with the rest”. I absolutely feel that I’ve seen the same tendencies with LLMs, you try to describe something that it keeps getting wrong in details, which might make that exact detail correct, but then something else is wrong.
Regardless, I got enough of the diffusion models in the end, I concluded that they are just “hopelessly broken” because they don’t understand the content they’re dealing with. If your characterization was meant to apply to me, you’re missing quite badly. I have no LLM experience from back then, I never even signed up for a free chatgpt account.
I don’t get why you use “yet” - I think LLMs will never be able to reason like a human. We need something completely different to achieve that, but LLMs have currently sucked up all resources, so I don’t think there’s any reason to expect that “something else” is just around the corner.
I won’t make a fuzz about “AGI”, it’s “AI” that I really struggle with, because “intelligence” is used about something that isn’t intelligent at all. “AGI” is fine, it now means what “AI” used to mean, before “AI” was demoted to mean LLMs. The terms were fine, but marketing departments had to interfere, as usual.
I see repeated over and over again that people seem to subscribe to the idea that “LLMs” are somehow a path that will result in “AGI” if just “evolved enough”. I have no idea where people got that idea, but I strongly suspect that those “helpful companies” have contributed. It’s perhaps the very core of the bubble. I haven’t yet seen anything that remotely suggests that will happen, in fact it seems like “magical thinking” to me.
If we want to have “AGI”, we must first try to create that. We can’t create something completely different, that doesn’t have the “basic building blocks” needed for reason and logic, and then hope that it will just become something it’s not.
To me, it’s like believing that if we just make enough wheelbarrows, we will some day end up inventing a jet aircraft. My claim is that it is highly unlikely, and if it’s a jet aircraft you want, you should try to make that instead.
I think that a lot of people just assume that “these companies know what they’re doing”, so there must be a path there. I reject that, the examples of spectacular failures of big companies is almost endless. Recently, Facebook wasted $73 billion on the equally far-fetched “Metaverse”:
Yes, and my point was that I don’t really think it qualifies as a tool, because a tool should give a predictable result. If a plow plows 60% of the time, cultivates 20% of the time and does nothing 20% of the time, it’s not really a plow. You have to manually check all the fields that have been plowed and then somehow manually correct everywhere where it got it wrong. It’s just frustration sold as a plow.
And this is the kind of absolutism that I think some people here have been trying to point out to you.
And here again.
Anthropic and OpenAI may very well cease to exist, I don’t know. AI will continue to exist. My dream is that the “offline” home AI will make the frontier models of today look like toys.
This, I can understand. Disclosure: I am not paying for AI either and I have not used Fable or 5.6 Sol. I have only used GPT 5.3-Codex, 5.4, 5.5, 5.6 Luna, and a bit of Sonnet 4.x in the past. What I use most often is the free browser-based Gemini. I am actually amazed at how much I could use it for advanced coding: a lot!.
Hopefully it won’t be as stupid, crazy, greedy, jealous, devious, cruel, and evil as humans. But reasoning abilities? I am not sure that it will “never be able to” reason like a human. But I guess without the entire spectrum of human emotions, it will never be “like a human”.
I don’t know.. and I think speculating about the future is a waste of time even though it’s fun to do so.
We are talking about the use of today’s AI right now for the current tasks we do now, specifically in the openhab’s case, for coding and working on openhab.
OK speculation time lol. There is no rule that says it must be achieved using LLM only. AGI is AGI. Once someone managed to achieve it, nobody cares whether they used LLM or something totally unrelated. Don’t confuse the what and the how.
Yet so many people continue to claim that AI is useful to them. I wonder why. I guess there must be a lot of naive people out there on both sides of the fence.
Can a hobby project like openHAB fulfill its crucial social aspects (in order to survive) if a growing amount of work — both contributions and reviews — is generated by AI?
Or will that become too frustrating and distract too many people from developing? On the other hand: How many good developers are we loosing if they would not be allowed to use AI as during their paid jobs.
And I think this mechanism is fundamentally different from commercial projects (like HA ) as both contributors and reviewers have to remain motivated without money.
My position on the „frontier models“ is to make use of them as long as they are comparably cheap and available — if they go away, then I guess I’ll have to do everything again by hand and not just the nice tasks, but also writing boilerplate, DTOs, mappers and such stuff.
But until they are gone, I’ll make use of them and do things such a migrating the openHAB website to a more modern build tooling — I would never get this done in just a few days without AI, let alone not sure I would attempt to do this at all. There is too much boring and annoying work of migrating HTML templates and working on CSS styling.
I’m not here to argue but to throw out a concrete examples of this.
Over the past few months I’ve done a lot of development with Gemini Flash and Clide (see below). Yesterday I realized that EyeOnWater (a service provided on behalf of my water company) has an API and Home Assistant actually has an integration for it. So I thought it would be good to get this information into openHAB for leak detection and other alerts when the water usage changes drastically.
My first thought was “I’ll make an OH add-on!” With AI I might be able to get over that initial hurdle to set up a development environment and get such a straight forward add-on implemented. Here the AI fell down and eventually suggested using Python and MQTT instead as the easier path. That’s probably what I’ll do as this service is pretty niche.
The barrier to contribution of even just an add-on to openHAB is too high for most casual developers to overcome, even with the help of AI. Clearly some are able to overcome that barrier, and I have no doubt if I really tried I could overcome it as well. But I don’t want to spend so much effort just getting up to where I can start coding. Many will get stuck like I did and pivot to something else, something easier. Maybe this is a good thing, maybe not. But it is at least one data point.
I’m still going to get my EyeOnWater integration with openHAB though. But the OH community won’t really benefit from that.
This is a concern I have. If there is a strict no AI policy, those who use AI will just pivot to something else or keep it to themselves and the OH ecosystem stops growing. Obviously a balance needs to be found on how best to use and accept AI generated code in a project like OH. And I see lots of good work on that front. But if even I am going to go down the easier path, you have to believe that’s a path many others have already taken.
Good human generated code vs. bad AI generated code is a false dichotomy. No contributions at all is also one of the options. With the AI I can solve my problem easily enough. If the projects I use refuse to take the contributions it’s no skin off my back. And maybe that’s a world we are heading towards, where most of the code out there is bespoke and written by AI for 1-10 total users. And maybe that’s a good thing over all. I can’t imagine it would be a good thing for projects like OH though.
It’s one data point at least.
For the record, this is what Gemini responded as I was exploring different approaches:
Using AI to code a native openHAB add-on (Java) is a viable but significantly steeper path than building a Python-based eow2mqtt sidecar.
While AI can write Java boilerplate flawlessly, it often struggles with openHAB’s hyper-specific binding architecture, core dependency injection frameworks (OSGi/Jetty), and the strict compilation requirements of the Maven build system.
Here is a breakdown of how the openHAB add-on route compares to the sidecar approach when relying heavily on AI.
Challenges with the Native Add-on Route
AI Knowledge of openHAB Framework Changes
openHAB updates its core Java APIs, build systems, and architectural patterns between major versions. AI models are prone to hallucinating deprecated methods, mixing up structural differences, or writing rules using patterns that fail to compile under modern openHAB versions.
The Maven and OSGi Toolchain
Developing a native add-on requires setting up an openHAB-specific development environment using Maven and Java. If a dependency misses an OSGi manifest declaration, the bundle will build but fail to start at runtime. Debugging these container classpath errors is something AI typically handles poorly.
Strict Code Verification
The openHAB codebase uses strict automated code-checkers (SpotBugs, Checkstyle). AI code often fails these styling rules out of the box, forcing you to spend extra time fixing whitespace, formatting, or missing annotations just to get it to compile.
Why a Sidecar (eow2mqtt) is Perfect for AI
If you use AI to build a standalone Python eow2mqtt container, you leverage the AI’s absolute best strengths:
Standard Python HTTP/MQTT Libraries: AI knows requests, asyncio, and paho-mqtt flawlessly. There are billions of training examples for these libraries, resulting in code that is almost always bug-free on the first run.
Rapid Iteration: You can test, debug, and patch your Python authentication or API code in seconds using standard execution blocks. You don’t have to wait for Maven to rebuild a Java .jar and reload it into an openHAB console.
Low System Overhead: Running a lightweight Python sidecar in a Docker container alongside your MQTT broker uses virtually zero resources and keeps your primary openHAB instance incredibly clean.
It went on to recommend using JS rules to implement the integration. I’ve done that before and the access token/refresh token stuff never really worked more than a day at a time. Maybe with AI help I can get past that. But I’ve already so many sidecars I’m running, what’s one more?
Anyway, I just wanted to point out that the way openHAB is developed is already steering contributors away from using AI. But it’s also going to steer them away from contributing back to OH also.
Just to chime in:
The AI now knows HABApp reasonably well that it can write the corresponding HABApp rules for you. That way you get the proven and convenient communication/rule stack from HABApp and you only implement the raw business logic as a rule. It’s then trivial to use the corresponding python libraries where the openHAB integration is faulty or missing.
If you plan on running a sidecar service anyway - you might as well do it properly.
If I’m going to create a sidecar, I’m going to use MQTT. I’m not looking to ever replace openHAB, but I’m also uncomfortable creating a sidecar that only works with openHAB also. And I already have code that supports both the HomeAssistant and Homie standards that I can reuse.
Good thing HABApp works with MQTT, too!
This was just a remark that instead of running a full comm stack for every little sidecar project / integration it’s easier to share the comm stack and just implement the actual logic.
I know you have the time and skills to do it so I say go for it.
But for the vast majority of users this is the way easier and more stable option so I wanted it to be mentioned.
I’m sorry if that is hurting anybody, but as far as I know, that is exactly the case. I don’t consider that an opinion of mine, but a fact. There are lots of information about this out there if there’s real interest in finding out more about it, but as I see it, this perhaps the very core of the issue: LLMs don’t “understand” any of the content they’re dealing with, they’re just finding the “best match” in their training data for the situation they are presented with.
To me it’s so obvious from what I see them produce. I found it harder than I envisioned to find substantial support, it seems like very few dare saying it out loud these days, although there are lots of hinting about this problem.
To it’s as simple as this: When you make a “machine” that operates as a “statistical tool”, you can’t expect it to use logic or even understand logic. These things are almost opposite, you can’t use a mathematical model designed for correlation to guarantee causation.
This is what Gemini “says” (which of course isn’t an opinion, but again just a statistical “average” of how this is described):
Large Language Models (LLMs) struggle heavily with consistency in code generation, often producing a flawless, elegant function in one paragraph and a broken, bug-ridden snippet in the next.
This inconsistency is a fundamental characteristic of how LLMs work, rather than just a random glitch.
This particular case isn’t about how I communicate, it’s about what I believe to be facts. I guess that if you disagree with the fundamental “facts”, the assertion can be provoking. It’s not “my job” to educate people, so I don’t quite know what to do other than to state what I believe are facts as facts.
You can find almost endless information in “both directions”, there’s a “hard push” by commercial interests to make people believe that this is about much more than LLMs and that “AGI” is just around the corner, I don’t believe that for a second, and see little evidence that this is about anything other than LLMs.
Here are a few resources that might be of some interest on the topic:
I agree that it would be great if we had really capable local models, but I see nothing in the current “evolution” that will take us there. This whole “revolution” is about LLMs, as far as I know, not much have happened with other model types. And LLMs, while great at processing human language, shows little ability to actually reason and apply logic.
LLMs will continue to exists if these companies collapse, but the very use that is currently “hailed” will not necessarily: huge LLMs with number of parameters in the terras. Running these requires so much “brute force” that they can never be run on a normal computer - they require “insane” hardware. And, the reason why people have access to them today, is that they are extremely heavily subsidized. So, if these companies disappear, and there is no more subsidization, the “current use” will be untenable. You won’t be so happy with them if what costs €200 in tokens today suddenly costs €8000.
That’s why I’m saying that what people have access to today is in reality dependent on the massive bubble and financial gamble to be possible. They haven’t managed to make the LLMs more effective, they have just thrown more and more insane amounts of hardware on it. There’s no reason why this would evolve to something radically different that doesn’t take hundreds of GPUs to run.
So if local models are likely to stay more or less what they are, and “cloud” models would be extremely expensive without being subsidized, don’t you see why I’m claiming that “what people currently use” is reliant on these companies and the situation to endure? And that can’t happen for obvious reasons, the question is just how long they can keep it going.
I think we’re talking about slightly different things. I’m not talking about “AI” or “AGI” - because those aren’t strictly defined any many things made in the future can fit into the terms (even though I susect that “AGI” is a pipe-dream for now). I’m talking explicitly about LLMs. I’m not saying that “machines” can never reason like humans, I believe that is possible, although perhaps not practical (because of how much easier and cheaper it is to use a human than to build a machine that does the same). I’m talking about LLMs explicitly, and the reason I believe that is impossible is that they are “statistics machines”. To me, it’s about as likely as that a lawn mower suddenly learns to fly - by itself.
No, but my point is that all the current efforts, money etc. go to LLMs - not “something else”. So, I can’t quite see why “something else” would arise now or “soon”.
It is genuinely useful for some things - and I think it “fools” many with other things, things that it appears to do well but really does not - and that sadly, it seems, must be experienced individually to apply.
This is a great example of that gray area you identified. You are right that LLM inconsistency is 100% a fact and feature of their basic function. You then extended that to say that fact eliminates all value. Your technical assessment will lead you from fact A to conclusion B every time and therefore you feel this should also be objective fact, but value judgments are the quintessential subjective statement. Something only has “value” when someone assigns it value, and somethings of value to me will never have value to you and vice versa. There cannot be an objective value to anything, by the very definition of ‘value’.
You argue that it is obvious that you are only presenting your opinion, but it is, in fact, not obvious when stated like this, linked as it is directly to some other more clearly object fact and presented with no qualifier to make it less objective than that which proceeds it. In these cases, feel free to assume that we are, in fact, not following you, and make it more explicit. By all means say that
to differentiate it from the objective fact that precedes it, and now no one can mistake it for an absolute statement, and no one gets distracted from the main point of your comments, and the topic doesn’t take another lengthy tangent.
) as both contributors and reviewers have to remain motivated without money.
I can’t imagine it would be a good thing for projects like OH though.
Challenges with the Native Add-on Route
Why a Sidecar (eow2mqtt) is Perfect for AI