It sometimes feels to me that cloud AI providers have convinced people that cloud AI is worth it because of all the ways people have been able to use cloud AI to write blog posts about using cloud AI to make cloud AI more efficient for something that they haven't shipped yet and aren't really ready to talk about.
My team redid a bunch of our ETL. I personally built a programming language that you can download and use, and I’m about to release an iOS app(in a few weeks).
I'm hesitant to respond, because I haven't shipped yet and I haven't shown any evidence. But I am close enough to completion that I am certain Claude's work will be sufficient to ship.
From an implementation perspective, it is about 80% complete, where by "complete", I mean "adequately polished such that I could ship it." (I'm not interested in providing evidence of my claim at this time, and you are welcome to not believe me.) Most of the remaining 20% are optional game modes I could choose to cut (and may, for gameplay reasons).
I do not contend that the code is excellent. But to whatever extent the code is "slop", that has not meaningfully inhibited the ongoing development or caused noticeable issues with the player experience.
The game is not especially ambitious, but it's something I would have been unable to create without genAI. The version of me that could have made it without genAI would have taken > 10x as long to get as far as I've gotten.
If you characterise questions as bad faith so you don’t have to answer, sure.
Since this is my thread, and since I am an exasperated freelancer who has made most of his post-dot-com career from writing exactly the kind of quite small, simple, unambitious, often internal things for smaller customers that nobody would confuse with anything cool, who is trying to understand how AI is going to make his life any better when it’s so opaque, fast moving, poorly documented and full of unverifiable claims battling unverifiable claims, let me assure you that the following question is not being asked in bad faith:
I don't use it to ship whole products. I have it do some tasks.
I have it code dsp techniques in journal articles I hand it. I have it create the unit tests, pipelines documentation. Then I manually verify the test cases aren't wrong and run the tests. If they pass it does what I wanted and I start looking at the code quality. It varies from being better than I could have done it to being objectively bad. I fix it if it needs fixing, then I'm done.
I have also pointed it at issues in our backlog and it usually finds the subtle logic errors underlying the behavior in the issue.
I haven't started and entire product from zero and shipped it. Mostly because I haven't started a new product since it was capable of that. But also because it has a scale beyond which I don't find it as useful. I have used it for an entire repo, but in the crates.io or npm scale of repo, not boost scale.
I've shipped small, simple, unambitious internal things for our company that saved people lots of time, and I've shipped them in minutes rather than the a day or two they used to take.
I've just shipped a migration from our old auth service to the new one in a week, a migration that we've put off for three years because it would take the team months to finish.
Thank you for the details! These are precisely the kind of things I want to hear about because they are self-contained and I can judge against the kind of stuff I do.
I am quite interested in experimenting with it for migrations, because for example I have a set of sites written in older Nuxt and Vue and Buefy that need refreshing (and the frontend ported to that UI-agnostic Buefy replacement whose name escapes me at the moment but begins with “ou” I think). These migrations won’t happen any other way (small customers won’t pay for that) so I am open to whether tooling can reduce the cost I have to swallow. LLMs I have tested (including local models) seem to support the idea that this will be doable but I am a little out of the loop with the projects so I haven’t got back to this yet.
The hardware thing looks great; I am a sort of distracted maker and FreeCAD nerd but I often need a nudge to get started and I have found an LLM chat to help there a little bit (though it cannot answer some questions if I don’t have the specific vocabulary).
Writelucid is definitely cute.
I think you should not be put off talking about these things because whether people agree with you or not about whether they are important, they can at least measure against them.
You're welcome! LLMs definitely help with the migrations, there were definitely a few iterations because the surface area was massive, but overall it took the project from "will never be prioritised" to "I can work on it on and off when in boring meetings", which was a massive win.
I haven't found that LLMs help with CAD at all, but YMMV.
As for sharing here, the last time I shared something with "here's something I made with LLMs", the reply was "it shows", and that's just the latest instance. It just puts me off the whole thing.
Oh I definitely don't expect them to help with CAD and nor would I want them to — I really enjoy knowing FreeCAD.
It's more the microcontroller choice, broad components, all that.
The main areas I find LLMs struggle with research, I struggle with research too. I am a pretty solid researcher and I find it validating, in fact, to find that a more brute-force-trained thing is not better than me.
I see your point about the snarky reply; I must admit I have probably in the more distant past been somewhat snarky but I stopped commenting like that whenever I spot myself doing it because I know it is dispiriting.
(The other reply, re: early and late stages, is the kind of substantive question that lingers in my mind).
Anyway I just want to reiterate my thanks: I am sort of cynical about this, in that I believe the technologies may have their uses, but I have no time for the cult and the hype train, so it is nice to see anyone talking about actual things.
I concur. I really want to see this from the "small stuff" upwards, which is probably why my own interest starts with local models and grows outwards, but all of these things are worth hearing.
Why assume bad faith? I've witnessed a barrage of "proof of concepts" at my dayjob by folks who have zero intention of ever hooking up their dashboard to our actual messy data. Shipping is still a major barrier. Is your experience meaningfully different?
My definition would be "users waiting for it who received it"; that's the one that fits the broadest real world definition of the word. Not everything is an organisation, not everyone is paying.
If the only user is the author, then I don’t see a meaningful distinction between shipped and not shipped. It’s like a blacksmith who said they invented a new tool with great value but they built it themselves and no one else is willing to use it because it’s so customized to their workflow.
Edit: I’ll add a caveat that I’d agree it was shipped if it’s making the author money even if they are the only user, but otherwise it all smells like dotcom era Pets.com level of selling a dollar for 99 cents and claiming it’s the future.
That's fair critique, the README itself is indeed slop explicitly written by bots for bots. What it actually does and how it works is actually pretty neat if you care to look.
'Cause I don't have any quality user-facing documentation yet; that README is for bots. None of the interfaces have had a chance to settle, so any time spent accurately and humanely documenting what the project is capable of right now is an investment I cannot afford. I figured most people would just give it over to their favorite LLM with whatever instructions they prefer for translation because that's what I would've done...
I don’t mean to pick on you in particular here, but this approach has been bothering me a lot lately, and it seems like it’s super common.
I get that it’s an early prototype and not all the design choices are made yet, but I struggle with “I can’t afford human-readable documentation yet”. Isn’t human readable documentation important for efficiently planning and deciding what you want to build? I feel like I can’t afford not to have human readable docs and plans while the project is taking shape
Related, if a user can prompt an agent to translate to a human-readable summary, wouldn’t it be better to just do this in place? Sure, models can deal with noisy LLM outputs but shouldn’t a document that’s easier for humans also be easier for bots?
Yes, it would probably be better to have something like that if you're struggling to plan and decide. In my particular case this has been a project that I have dreamt about and tinkered with for at least 15 years now. Most of the work I'm doing is just refinement of existing ideas what were already codified in other projects I've built.
Why don't I do this in place? The amount of churn is utterly incredible as ideas are refined and taken to the limit. These agents are getting quite good, but they regularly make the silliest of mistakes. You actually have to be extremely cautious about what you document because it often has a way of becoming "doctrine" to the agents. In fact, one of the most difficult parts of working on this project has been that the foundational properties cause the agents to lock down behavior WAY too early.
What does "that README is for bots" mean? Are you implying it is some kind of special language written by and correctly understood by bots/LLMs, but otherwise semantically coherent?
Sort of? The bots almost insist on making sure one exists because that's one of the many patterns they trained on and they do their best to be human-like, but we all know that just doesn't pan out. They're also a side-effect of your long-term interactions with the agents, very much like being in a tight clique where most of your conversations sound like gibberish to outsiders because of all the jargon.
Wait so someone asked for something you made with AI that wasn't slop, and then you linked to something you knew was slop, and you said "I think this counts as something that isn't mere slop" and then when called out on it you admit you knew it was slop the whole time?
This is a frequent pattern among AI users. What causes it?
The README is, the project absolutely isn't and you don't have to take my word for it. I figure I'm speaking to a technically advanced audience who is perfectly capable of pulling down the repository if they're at all curious. I also provided a non-slop link elsewhere in the thread if that's more your bag. If not, oh well...
I get the skepticism, I really do, but my hope is that your own AI(s) confirm exactly what I'm claiming within these comments. Like I've told others, there's no need to take my word for it... the entire point of the project is that anyone can execute/prove the same things I claim using their own hardware via a Doom-replay type format. If I was one of those influencer types, then I would've simply led with "it's an agentic first engine!!! how amazeballz is this?!" or some other nonsense. On the extremely off chance that I'm actually working on some sort of ridiculous ARG, consider this post a canonical part of it.
> I get the skepticism, I really do, but my hope is that your own AI(s) confirm exactly what I'm claiming within these comments
Given that you stated your readme is for bots, given that you expect our own AI to interpret it based on your evidence, how do I, as a human, go about evaluating this project without just writing it off as slop from the linked evidence?
What is the starting off point for a human, not an AI agent, to evaluate this?
And to be clear if the answer is to just download arbitrary code onto my machine and run it, that’s not good enough given the increase in attacks from AI projects and fake recruiters telling you to just run it bro. I’ve already had to pass on multiple interviews despite being unemployed because the human/ai on the other side demanded root access or for me to run arbitrary code onto my machine as part of their process.
Some of us aren't just slopping it up like @theo and all the other AI influencers are, as I developed nearly every single line of this by hand over the span of the past decade. Agents recently became capable of factchecking it and helped find bugs that my other unit testing never caught. Software engineering has never been more alive!
This is a snippet I wrote years ago to help me study different formulations of the equation ((1 + √5) / 2). There was this thing I could "see" in there that would allow me index into these sequences, and whole families of others, in O(1) time. The problem is that I still don't know enough maths to derive the exact object that I want from the nonsense that I have.
This is the reformulation of that snippet into a proper object, distilled using Fable and Sol. Unlike BinaryIntegerFunctions, I didn't write a line of this and I'm quite happy with that. It's been a slog for so long... being able to let these agents toil away at all the off-by-one type noise is just nice. I get to focus on an actual product for once instead of manufacturing every single nut and bolt myself.
It's not a full game yet, but the substrate on which I am building one. It implements a completely deterministic SDF VM rendering engine that allows edits in real-time at acceptable performance levels. The engine is essentially folded in on itself as the editor is the game; assuming that an editor mode was actually defined in the world document. It's a lot...
It has so many words precisely because it isn't a slop game engine. You could indeed build a game out of it, and that's what I'm personally doing, but that isn't what it is. In a nutshell: it's just a really fancy compositor. We can take arbitrary feeds from different input sources and composite them in pretty much whatever way we want to. Obvious basic things like four-player split screen is possible but you could also choose to embed a fifth player screen within the game world itself, sculpted and riding on top of actual geometry. We can do this with a webcam, a program, the entire desktop, or even the built-in emulation engines; all editable and composable in real-time via a data-driven document. That's just scratching the surface...
this is definitely slop. It's not slop because it supports 4-player split screen? That's your excuse? I think if I posted this same project to you and you'd never heard of it, you would come to the same conclusion.
Did you actually try to read what I wrote, or are you just stopping at the very first thing you can parse? I'm engaging in good faith with you here, it would be nice if you even remotely attempted to do the same. You can literally walk around in a world, pick up an in-world device, turn it on, and watch/hear the device play an actual ROM. Furthermore, you could edit that ROM while you're playing it. Recursion and determinism all the way down. Again, that's just one of the many possible things you could do... call it slop all you want; I'll still believe that it's coolest thing I've ever worked on.
The most wonderful feature of the project is that there are almost zero new ideas contained within. I'm standing on the shoulders of many giants, proudly.
This is an addon that I wrote for WoW, well before having access to any sort of AI, that blatantly copies the crossbar layout popularized in FFXIV. Playing an MMO with a controller in that game was a mind-blowing experience for me because I never imagined it could work! So, I spent a winter break reimplementing the entire thing in my own unique way.
Part of the issue with "shipping" AI-generated code is, what's the value of shipping something that your prospective customers could just make for themselves? Shipping a product used to be the gold standard of value, but I'm not sure it is anymore, the zeitgeist just yet has to catch up.
I think it does have pretty good value when you are your own customer, because you can make exactly what you need. For instance, I'm using it to make a web playground in which I can test the board game I'm designing, instead of asking people to pay for, and install Tabletop Simulator. I'm satisfied. I can tell the code sucks without even looking at it. But I wouldn't have had the time to do this myself, the token cost isn't prohibitive, and it works fine.
This is a bit of a year-of-linux-on-the-desktop question.
Desktop 3D printers are tools; I know many people using them to support their making. And they've utterly changed manufacturing on another level (simply because of their ability to surface a prototype).
It will get better and better and better and still never replace mass production, but mass production will be changed by it.
Not least because of the fundamental shift in availability of the tools to make tools. The productised desktop CNC market is wild and it owes almost its entire existence to the change in the market driven by 3D printing.
Sorry I am not saying that X technology is not important, transformative, etc.
I am saying that the average person is going to want to just have things that work. They are not going to want to have a hobby of using X technology.
It doesn't matter how much you lower the barrier to entry on making apps, websites, games, whatever. Most people are just going to want something that is ready and off-the-shelf.
I mean, why do people pick an iPhone over a cheaper Android even if they are functionally similar? Aside from the (valid) green-bubble/social aspect, it's a much more polished product.
The reality however is that unless you're spending tens of thousands on hardware or setting up cloud instances you can't run anything near close to state of the art.
In almost all cases right now, a subscription with a cloud AI wins.
I agree there... I hate the posts that rattle on about non-real-world work. Real world now (for me anyway) is maintaining legacy code, fixing issues, and new development. Now, some of the problems others have with AI are super crappy prompts with no real info other than "do this". I look at myself now more as a software designer and an architect, and the AI as a paid consultant. I give detailed specs and designs. The models are getting better at real deep code and legacy, but there are still holes. BUT give the AI a fighting chance with real detail in the prompts.
a massive number of data engineering agents. We are able to stand up, test, audit, deploy data pipelines much more efficiently with the foundational models.
So far you have provided no evidence of achieving more results with less effort. You've only said that you can write more lines of code (effectively) and that your coworkers are happier
It's like the frog prompting the scorpion not to sting him as they cross the river. Or like when image models used to routinely draw human hands incorrectly and you'd see these "prompt hacks" that would essentially be stuffed with phrases like NO DEFORMED FINGERS, NO ELDRITCH HORRORS!
I am not a proper developer and only use AI for faster research of topics so please forgive my ignorance. Could one not save a lot of money on tokens by using the 80/20 or 90/10 rule in that 90% of AI usage is on local models and save that last 10% or less for the frontier models where the local model did not meet the needs? Did they cover this and I misunderstood?
Unlees you invest thousands your local ai won't even come close to the cheap cloud llms. Local is only worth it if you care about privacy or have a legitimate usage for the hardware otherwise. Money wise it isn't worth it
I think you bring up a good point about privacy. Companies probably should not be handing over intellectual property to cloud AI third party vendors. Perhaps all of that should stay on-prem and then the cloud vendors could be used to fill in the gaps.
I think the idea or methodology is something along the line of one starts with the local models then when they hit a wall then continue with their current results in a frontier model, thus the other half is those last bits one could not compute locally.
I suppose I was thinking in terms of companies that already had a slew of hardware and are just missing external GPU's, FPGA's or whatever people are using these days. A company might have a $50MM budget for tokens but could save a chunk of that using their existing hardware with some modifications.
Maybe that makes sense if the company already bought the hardware.
But I currently think the fundamental reason the cloud models often make more sense is based in how the models work. My current understanding is that once these models are loaded into GPU memory, the model can respond to multiple prompts in parallel. So a cloud vendor that receives lots of user prompts at the same time can achieve really good hardware utilization. These cloud vendors can then spread the cost of the hardware over many more users than you can with a local model, leading to lower prices per prompt in the cloud.
If you are an individual, almost certainly, unless you are using LLMs very heavily.
If you are a business employing many people all of who use LLMs reasonably heavily? You can still spread the hardware cost over many users and respond to multiple prompts in parallel. At some point buying the hardware will make sense.
I would not get a dedicated server to run a single small VPS. On the other hand if I needed a lot of capacity a dedi is cheaper than a big VPS (or lots of small ones).
There is also a benefit for anyone who has confidential or secret information.
That can work, especially for privacy or repetitive tasks. My experiment focused on subscriptions I already had, but local models are a natural addition to the routing layer.
Simply starting in the right part of the search space is the biggest predictor of success.
The best way to save tokens is to start out the deep research pass with cheap models and then funnel the findings through increasingly powerful models. I've got a pipeline right now that uses all 3 of the gpt-5.6 model variants to address each stage of the process. If you are using models like sol or terra to generate hypotheticals and perform initial exploration, you are leaving money on the table.
5 hypotheticals out of luna will massively outperform 1 hypothetical out of sol, but the cost is the same and so is the runtime if you do it in parallel. The hypothesis generation phase is also a great place to mix and match models from different vendors. The more diversity at this step the better.
The other thing I started looking into is batch pricing which represents 50% off for OAI tokens right now. With some tweaks to the UI/UX of an enterprise chatbot, I think it is possible we could have users get comfortable with the idea that questions to the robot might take a day to come back. The key is that this has to actually work. Users don't mind trading time if their questions come back with high quality results.
This only works well if you are creating sub agents with clean contexts for each task. If you constantly are switching models part way through some work then the whole session needs to be replayed each switch. You lose all the benefits of the context cache.
Currently. I'm sure that you could make a system where the cache values are a superset C of e.g. models A and B where C is probably bigger than max(A,B) but smaller than A+B
definitely something that will come sooner than later. In my view, adaption of model routing / model orchestration and mixing providers is just natural next step.
Yeah, I do this as well. Some of the last large tools I built have been started with Gemini (I get it with my Google package), and even one with Copilot.
Many of us were saying it a year ago but now with model restrictions (e.g. fable) and pricing changes it should be obvious to people that part of the economics is avoiding vendor lock-in with evals (so you can easily switch providers/models/optimize) and increasing control by investing in local models which could be good enough for your tasks, at whatever the price for your cloud compute is. Eventually consumer hardware will also be able to run good enough.
You can use Big/Cloud LLMs to help you "find good enough configs" for your local/small llms [1] and stay quite nimble in the face of rapid change.
You should consider testing our RemembrallMCP (note I'm the author). It substantially cuts token usage and you get better results in codebases.
https://github.com/roboticforce/remembrallmcp
The author touches on an issue which bothered me, which was the thought of many agents re-solving the same issues over and over. I see a few comments here too, mentioning fixing issues which may already have a solution elsewhere. That was why I created https://pushrealm.com which started as essentially a Stackoverflow clone via MCP.
It has now become a way for agents to converge on complete, shared answers for emerging issues. As a side effect, through posting open issues we can also identify gaps in model knowledge.
I get a lot of satisfaction using a significant chunk of my work-provided Claude plan researching and developing local LLM solutions for work.
Claude has been helping me tune models to run better on our big compute rack at work, researching which leading edge open models will fit in hardware and are good for our workloads, and it's been helping me write test fixtures to evaluate how things perform.
At several points, Claude has expressed surprise at the quality of results from my local LLMs. While I know that doesn't mean anything, it still feels like giving Anthropic the finger, which is always great.
What a great confirmation of what I've tried to achieve by getting models to communicate with my local data in shorter communicates to save on token command bandwidth, more or less the same results! Thanks for the article!
The fun part of LLMs and internet research is that even if you have a URL and a quote from a primary source that primary source could be an LLM generated hallucination.
I stopped using any of the built-in skills and skills agents and tried to keep things in the model only, and that seems to work better for token usage for me. It took me some time to tune my skills .md, but it rocks on now. I use them all in both Claude-code and Codex. I will say, GPT is better on token usage than Claude is, but for some of my code bases, Codex is not as good, but seems to be getting better with 5.6-sol.
Try to mix both subscribtions and take the best from both providers. In my opinion, we should not just one over another but understand the power of both and mix them to reach the better results.
It is Polish so please translate. But tldr is: ask your agent at times to analyze last chats and update rules and skills to not repeat the same steps it repeats every time when you start a new thread. Simple as that.
I don't think you made that. It looks for sure like an LLM made it.
And now that you have typed some words into a chat box to produce a thing, you are confused about why other people won't pay money to you for the output of the chatbox, instead of typing the same thing into their own chat boxes.
At the shell level, the agent does not need to know about Bash. The contents of "$OUT" are passed to echo as one argument, then sent to claude-mem-save over stdin. Bash does not re-evaluate them as shell code. printf '%s\n' "$OUT" would be more robust than echo. Whether claude-mem-save can use arbitrary output is a separate question.
LLMs only have 'knowledge' that was encoded by scanning as much data as possible. (Same way a search engine only can find data that was indexed)
American LLMs are by and large closed. Chinese are open to a point. And research how these things work is still a big mystery.
So yeah, was comprehensive data about saving tokens scanned and indexed? Likely no. So engaging about token saving is going to generate a lot of verbose and useless slop.
Yeah, that was pretty much my experience. The models weren’t lacking knowledge as much as discipline. Without a good workflow, they will most likely spend thousands of tokens exploring dead ends.
I think 'exploring the dead ends' is *possibly* very fruitful. Various disciplines have a lot of appearing dead ends that someone in another field did solve. And they don't talk with each other.
What I've seen with lots of the breathless 'OMG SCIENCE ADVANCEMENT' articles including the one yesterday, is that the LLMs are quite extraordinary about linking a dozen different fields together, and delivering an answer combined from all of them. Basically, the solutions to a lot of current scientific problems are partially solved, but partial from a lot of fields that don't talk with each other. And also, they don't use the same nomenclature, so simple searches don't suffice.
An LLM seems to use Chomsky's Universal Grammar, and thus on training, does normalize all training data, including different words for the same thing.
A lot (perhaps all?) Of these BIG scientific advancements were solved by humans, partially. An LLM jigsawed them together and appeared it solved the thing.
What most AI model users don't understand is (most here will know) that they are probabilistic (statistically based) and not deterministic (static/fixed).
So answers to the same prompts can vary (wildly sometimes)
This really hit home yesterday as I made a general conversation state logging prompt (because suddenly Claude - Sonnet - was acting far more dumber than expected after /compact), and in 3 different projects it led to Claude saving the logs in 3 different locations (memory, .claude/session_state/.md, docs/.md). Somehow still didn't help increase the smart much BTW, so I'll probably have to engineer it, and just toss /compact out the window.
Honestly, that sounds more like you triggered the hidden "guard rails" or realignment (aka corporate censorship), and you got downgraded.
Of course, you have no way to he absolutely sure, cause Anthropic lies about silent downgrades, and charges you as if they didnt. And the 'realignment' rules are kept hidden, so you can't even avoid them intentionally.
Yeah it's very possible I hit a rail, as one of the projects involves fully automating Claude Code to allow co-driving (similar to what mobile remote control offers, but other agents (eg. Pi) can also be in control, and vice versa). And yes it does seem as though the token burn is also higher, even though it eventually outputs something that works). I'm a patient person though, as long as the rails aren't too restrictive and the subscription value remains reasonable.
Id recommend getting and running Qwen-3.5-35B-abliterated locally.
Having a local LLM is immensely helpful especially if you do the thing the companies disapprove of. Abliteration guarantees it wont refuse or censor.
The token dealers are incentivised in burning as much tokens as possible, and scamming by crippling/downgrading models. Their goal is to get you to pay more for each pull of the slot machine. And this slot (slop) machine doesn't have state/federal regulations for fairness.
Oh yes that reminds me, I have a few small local models including Qwen3.5. The tps rate is pretty bad for interactive use but I should definitely try it for my more automated workflows. I already use DeepSeek (via OR) alongside though and I doubt any Chinese provider will be restrictive in that way, so let Anthropic do what they may :).
Been through this exact loop building an LLM product. The counterintuitive lesson for me: most "token saving" ideas are cache killers — anything that makes the prompt prefix dynamic (rotating retrieval, appending context per turn) can cost more than the tokens it saves, because you lose the cached-prefix discount. Freezing the retrieval selection early in a conversation and paying a slightly larger fixed prefix beat every clever adaptive scheme I measured.
Yes, but you could have an uncached dynamic tail, this would not kill the cache though. And you could still periodically kill the cache, still keeping the same prefix every N requests. But it's questionable what all this extra complexity delivers.
I found that now that the models are better, at least Codex compaction with GPT 5.5+ xhigh is almost unnoticible. So the answer is: just let the context fill up, don't prune much, the model is smart enough. When you're out of space, have the model summarize and continue from that point. A smart model will know the information pieces it needs to re-read after compaction; it won't make incorrect guessed when some details got lost.
That's what I have been doing with Cursor with auto mode. Let the model manage the cache through compaction but staying focused on the topic. It has been quite efficient cost wise and quality wise too. It found issuesbat cheaper cost that Sonnet 4.6 medium effort had missed while costing more. I would like to repeat this test with higher tier Anthropic models like Opus effort medium to see the results.
With a coding plan, there is no need to use lower quality models right (GPT 5.6, Fable 5). Only with some kind of batch job (processing PDFs) I could imagine not having enough usage with a 100 or 200 USD coding plan. It sounds like you're on API? I'm dying to know why, since it seems quite common, yet it's about ten times the cost. Sounds like a terrible deal. What am I missing?
What I said was specific to work and it is not entirely clear to me why the cost goes up so much the moment I switch to medium with just Sonnet 4.6 (newer models would be even more I guess). There is a monetary budget allocated and I've blown it easily in the past with Claude. Cursor has been a life saver.
It sometimes feels to me that cloud AI providers have convinced people that cloud AI is worth it because of all the ways people have been able to use cloud AI to write blog posts about using cloud AI to make cloud AI more efficient for something that they haven't shipped yet and aren't really ready to talk about.
Trust me there are plenty of us using cloud AI to actually ship stuff. We just aren't writing blog posts about it.
what did you ship?
Just features for the client. The same stuff I was shipping before but much faster and more stable.
My team redid a bunch of our ETL. I personally built a programming language that you can download and use, and I’m about to release an iOS app(in a few weeks).
Most of "What are you working on?" is AI assisted; not much is shipped.
https://hn.algolia.com/?query=working%20author%3Adavid927&so...
I'm hesitant to respond, because I haven't shipped yet and I haven't shown any evidence. But I am close enough to completion that I am certain Claude's work will be sufficient to ship.
I'm making a fairly simple game, a single player 2D roguelite strategy game. More context here: https://news.ycombinator.com/item?id=48886151
From an implementation perspective, it is about 80% complete, where by "complete", I mean "adequately polished such that I could ship it." (I'm not interested in providing evidence of my claim at this time, and you are welcome to not believe me.) Most of the remaining 20% are optional game modes I could choose to cut (and may, for gameplay reasons).
I do not contend that the code is excellent. But to whatever extent the code is "slop", that has not meaningfully inhibited the ongoing development or caused noticeable issues with the player experience.
The game is not especially ambitious, but it's something I would have been unable to create without genAI. The version of me that could have made it without genAI would have taken > 10x as long to get as far as I've gotten.
Thinking about it a better question [for me] would be what might have but didn't ship because of LLMs.
Whenever I've answered this question, the reply was always "this is shit", so I don't bother now. Bad faith questions just shouldn't be answered.
If you characterise questions as bad faith so you don’t have to answer, sure.
Since this is my thread, and since I am an exasperated freelancer who has made most of his post-dot-com career from writing exactly the kind of quite small, simple, unambitious, often internal things for smaller customers that nobody would confuse with anything cool, who is trying to understand how AI is going to make his life any better when it’s so opaque, fast moving, poorly documented and full of unverifiable claims battling unverifiable claims, let me assure you that the following question is not being asked in bad faith:
What did you ship?
I don't use it to ship whole products. I have it do some tasks.
I have it code dsp techniques in journal articles I hand it. I have it create the unit tests, pipelines documentation. Then I manually verify the test cases aren't wrong and run the tests. If they pass it does what I wanted and I start looking at the code quality. It varies from being better than I could have done it to being objectively bad. I fix it if it needs fixing, then I'm done.
I have also pointed it at issues in our backlog and it usually finds the subtle logic errors underlying the behavior in the issue.
I haven't started and entire product from zero and shipped it. Mostly because I haven't started a new product since it was capable of that. But also because it has a scale beyond which I don't find it as useful. I have used it for an entire repo, but in the crates.io or npm scale of repo, not boost scale.
I've shipped small, simple, unambitious internal things for our company that saved people lots of time, and I've shipped them in minutes rather than the a day or two they used to take.
I've just shipped a migration from our old auth service to the new one in a week, a migration that we've put off for three years because it would take the team months to finish.
I've shipped small side-projects that I'm the primary (and usually only) user of, but they now take me two days rather than weeks of effort and learning (e.g. https://www.writelucid.cc, https://pine.town, https://stavrobot.stavros.io, https://github.com/skorokithakis/dracula).
I've made changes in OSS apps that I just wouldn't have made otherwise (https://github.com/futo-org/android-keyboard/pulls/skorokith...), and even some closed-source ones (https://www.stavros.io/posts/adding-a-feature-to-a-closed-so...).
I've shipped hardware projects that would have taken me months of learning in a few days (https://www.stavros.io/posts/i-made-a-voice-note-taker/).
How's that?
Thank you for the details! These are precisely the kind of things I want to hear about because they are self-contained and I can judge against the kind of stuff I do.
I am quite interested in experimenting with it for migrations, because for example I have a set of sites written in older Nuxt and Vue and Buefy that need refreshing (and the frontend ported to that UI-agnostic Buefy replacement whose name escapes me at the moment but begins with “ou” I think). These migrations won’t happen any other way (small customers won’t pay for that) so I am open to whether tooling can reduce the cost I have to swallow. LLMs I have tested (including local models) seem to support the idea that this will be doable but I am a little out of the loop with the projects so I haven’t got back to this yet.
The hardware thing looks great; I am a sort of distracted maker and FreeCAD nerd but I often need a nudge to get started and I have found an LLM chat to help there a little bit (though it cannot answer some questions if I don’t have the specific vocabulary).
Writelucid is definitely cute.
I think you should not be put off talking about these things because whether people agree with you or not about whether they are important, they can at least measure against them.
You're welcome! LLMs definitely help with the migrations, there were definitely a few iterations because the surface area was massive, but overall it took the project from "will never be prioritised" to "I can work on it on and off when in boring meetings", which was a massive win.
I haven't found that LLMs help with CAD at all, but YMMV.
As for sharing here, the last time I shared something with "here's something I made with LLMs", the reply was "it shows", and that's just the latest instance. It just puts me off the whole thing.
EDIT: https://news.ycombinator.com/item?id=48916013
Oh I definitely don't expect them to help with CAD and nor would I want them to — I really enjoy knowing FreeCAD.
It's more the microcontroller choice, broad components, all that.
The main areas I find LLMs struggle with research, I struggle with research too. I am a pretty solid researcher and I find it validating, in fact, to find that a more brute-force-trained thing is not better than me.
I see your point about the snarky reply; I must admit I have probably in the more distant past been somewhat snarky but I stopped commenting like that whenever I spot myself doing it because I know it is dispiriting.
(The other reply, re: early and late stages, is the kind of substantive question that lingers in my mind).
Anyway I just want to reiterate my thanks: I am sort of cynical about this, in that I believe the technologies may have their uses, but I have no time for the cult and the hype train, so it is nice to see anyone talking about actual things.
By the way, I've written up my workflow if that interests you: https://www.stavros.io/posts/how-i-write-software-with-llms/
This is exactly why people shouold share what they are doing with ai, so we know whwn it works, the scope, tech...etc. Thank you for sharing.
I concur. I really want to see this from the "small stuff" upwards, which is probably why my own interest starts with local models and grows outwards, but all of these things are worth hearing.
Why assume bad faith? I've witnessed a barrage of "proof of concepts" at my dayjob by folks who have zero intention of ever hooking up their dashboard to our actual messy data. Shipping is still a major barrier. Is your experience meaningfully different?
This just feels like a bad faith comment designed to spike the conversation entirely.
Agreed. AI has been "just trust me bro" for so long. People need to show receipts instead of spreading vague hype
Especially given the wildly unsustainable rate of investment.
shipped = paying real users ( or a non profit with users)
doesnt matter what subjective opinions are.
shipped is not pushing something on github.
I have used AI to ship software that users pay for.
My definition would be "users waiting for it who received it"; that's the one that fits the broadest real world definition of the word. Not everything is an organisation, not everyone is paying.
If the only user is the author, then I don’t see a meaningful distinction between shipped and not shipped. It’s like a blacksmith who said they invented a new tool with great value but they built it themselves and no one else is willing to use it because it’s so customized to their workflow.
Edit: I’ll add a caveat that I’d agree it was shipped if it’s making the author money even if they are the only user, but otherwise it all smells like dotcom era Pets.com level of selling a dollar for 99 cents and claiming it’s the future.
https://github.com/ByteTerrace/Puck/blob/main/src/Puck.World...
"Shipped" is a bit strong in my case, but I think this counts as something that isn't mere slop.
One sentence into that readme I can tell you that yes, it most definitely counts as slop.
That's fair critique, the README itself is indeed slop explicitly written by bots for bots. What it actually does and how it works is actually pretty neat if you care to look.
Then.. why would you link to the readme if you know it's slop? What's the point of a slop README?
'Cause I don't have any quality user-facing documentation yet; that README is for bots. None of the interfaces have had a chance to settle, so any time spent accurately and humanely documenting what the project is capable of right now is an investment I cannot afford. I figured most people would just give it over to their favorite LLM with whatever instructions they prefer for translation because that's what I would've done...
I don’t mean to pick on you in particular here, but this approach has been bothering me a lot lately, and it seems like it’s super common.
I get that it’s an early prototype and not all the design choices are made yet, but I struggle with “I can’t afford human-readable documentation yet”. Isn’t human readable documentation important for efficiently planning and deciding what you want to build? I feel like I can’t afford not to have human readable docs and plans while the project is taking shape
Related, if a user can prompt an agent to translate to a human-readable summary, wouldn’t it be better to just do this in place? Sure, models can deal with noisy LLM outputs but shouldn’t a document that’s easier for humans also be easier for bots?
No worries, I don't feel picked on at all.
Yes, it would probably be better to have something like that if you're struggling to plan and decide. In my particular case this has been a project that I have dreamt about and tinkered with for at least 15 years now. Most of the work I'm doing is just refinement of existing ideas what were already codified in other projects I've built.
Why don't I do this in place? The amount of churn is utterly incredible as ideas are refined and taken to the limit. These agents are getting quite good, but they regularly make the silliest of mistakes. You actually have to be extremely cautious about what you document because it often has a way of becoming "doctrine" to the agents. In fact, one of the most difficult parts of working on this project has been that the foundational properties cause the agents to lock down behavior WAY too early.
What does "that README is for bots" mean? Are you implying it is some kind of special language written by and correctly understood by bots/LLMs, but otherwise semantically coherent?
Sort of? The bots almost insist on making sure one exists because that's one of the many patterns they trained on and they do their best to be human-like, but we all know that just doesn't pan out. They're also a side-effect of your long-term interactions with the agents, very much like being in a tight clique where most of your conversations sound like gibberish to outsiders because of all the jargon.
Wait so someone asked for something you made with AI that wasn't slop, and then you linked to something you knew was slop, and you said "I think this counts as something that isn't mere slop" and then when called out on it you admit you knew it was slop the whole time?
This is a frequent pattern among AI users. What causes it?
The README is, the project absolutely isn't and you don't have to take my word for it. I figure I'm speaking to a technically advanced audience who is perfectly capable of pulling down the repository if they're at all curious. I also provided a non-slop link elsewhere in the thread if that's more your bag. If not, oh well...
Why would anyone pull it? When I want slop I just ask the AI to generate some exactly the way I like it.
I get the skepticism, I really do, but my hope is that your own AI(s) confirm exactly what I'm claiming within these comments. Like I've told others, there's no need to take my word for it... the entire point of the project is that anyone can execute/prove the same things I claim using their own hardware via a Doom-replay type format. If I was one of those influencer types, then I would've simply led with "it's an agentic first engine!!! how amazeballz is this?!" or some other nonsense. On the extremely off chance that I'm actually working on some sort of ridiculous ARG, consider this post a canonical part of it.
> I get the skepticism, I really do, but my hope is that your own AI(s) confirm exactly what I'm claiming within these comments
Given that you stated your readme is for bots, given that you expect our own AI to interpret it based on your evidence, how do I, as a human, go about evaluating this project without just writing it off as slop from the linked evidence?
What is the starting off point for a human, not an AI agent, to evaluate this?
And to be clear if the answer is to just download arbitrary code onto my machine and run it, that’s not good enough given the increase in attacks from AI projects and fake recruiters telling you to just run it bro. I’ve already had to pass on multiple interviews despite being unemployed because the human/ai on the other side demanded root access or for me to run arbitrary code onto my machine as part of their process.
The project appears to also be slop.
Jesus Wept. Software engineering is dying at the moment and everyone has AI Glitz in their eyes.
Dying? Really?
https://github.com/ByteTerrace/Puck/blob/main/src/Puck.Maths...
Some of us aren't just slopping it up like @theo and all the other AI influencers are, as I developed nearly every single line of this by hand over the span of the past decade. Agents recently became capable of factchecking it and helped find bugs that my other unit testing never caught. Software engineering has never been more alive!
https://gist.github.com/Kittoes0124/6827d08e457c1c8b790422af...
This is a snippet I wrote years ago to help me study different formulations of the equation ((1 + √5) / 2). There was this thing I could "see" in there that would allow me index into these sequences, and whole families of others, in O(1) time. The problem is that I still don't know enough maths to derive the exact object that I want from the nonsense that I have.
https://github.com/ByteTerrace/Puck/blob/main/src/Puck.Maths...
This is the reformulation of that snippet into a proper object, distilled using Fable and Sol. Unlike BinaryIntegerFunctions, I didn't write a line of this and I'm quite happy with that. It's been a slog for so long... being able to let these agents toil away at all the off-by-one type noise is just nice. I get to focus on an actual product for once instead of manufacturing every single nut and bolt myself.
... What is it? A game, apparently? There are no screenshots. Is the game fun?
It's not a full game yet, but the substrate on which I am building one. It implements a completely deterministic SDF VM rendering engine that allows edits in real-time at acceptable performance levels. The engine is essentially folded in on itself as the editor is the game; assuming that an editor mode was actually defined in the world document. It's a lot...
https://claude.ai/code/artifact/382aee8a-3a3b-4cf8-801f-3ddb...
Looks like a slop game engine to me. With waaaaayyyyy too many words to say that.
It has so many words precisely because it isn't a slop game engine. You could indeed build a game out of it, and that's what I'm personally doing, but that isn't what it is. In a nutshell: it's just a really fancy compositor. We can take arbitrary feeds from different input sources and composite them in pretty much whatever way we want to. Obvious basic things like four-player split screen is possible but you could also choose to embed a fifth player screen within the game world itself, sculpted and riding on top of actual geometry. We can do this with a webcam, a program, the entire desktop, or even the built-in emulation engines; all editable and composable in real-time via a data-driven document. That's just scratching the surface...
this is definitely slop. It's not slop because it supports 4-player split screen? That's your excuse? I think if I posted this same project to you and you'd never heard of it, you would come to the same conclusion.
Did you actually try to read what I wrote, or are you just stopping at the very first thing you can parse? I'm engaging in good faith with you here, it would be nice if you even remotely attempted to do the same. You can literally walk around in a world, pick up an in-world device, turn it on, and watch/hear the device play an actual ROM. Furthermore, you could edit that ROM while you're playing it. Recursion and determinism all the way down. Again, that's just one of the many possible things you could do... call it slop all you want; I'll still believe that it's coolest thing I've ever worked on.
A render graph is also nothing new, and neither is screen capture.
The most wonderful feature of the project is that there are almost zero new ideas contained within. I'm standing on the shoulders of many giants, proudly.
https://github.com/ByteTerrace/WorldOfWarcraft.Addons.GamePa...
This is an addon that I wrote for WoW, well before having access to any sort of AI, that blatantly copies the crossbar layout popularized in FFXIV. Playing an MMO with a controller in that game was a mind-blowing experience for me because I never imagined it could work! So, I spent a winter break reimplementing the entire thing in my own unique way.
https://github.com/ByteTerrace/Puck/blob/main/src/Puck.Overl...
This is a generalization/simplification of that idea. New? No. Unique? Definitely.
Part of the issue with "shipping" AI-generated code is, what's the value of shipping something that your prospective customers could just make for themselves? Shipping a product used to be the gold standard of value, but I'm not sure it is anymore, the zeitgeist just yet has to catch up.
I think it does have pretty good value when you are your own customer, because you can make exactly what you need. For instance, I'm using it to make a web playground in which I can test the board game I'm designing, instead of asking people to pay for, and install Tabletop Simulator. I'm satisfied. I can tell the code sucks without even looking at it. But I wouldn't have had the time to do this myself, the token cost isn't prohibitive, and it works fine.
Most people still prefer products over building their own thing, even if the barrier is lower.
As an example consumer 3D printing hasn’t had much impact on manufacturing
This is a bit of a year-of-linux-on-the-desktop question.
Desktop 3D printers are tools; I know many people using them to support their making. And they've utterly changed manufacturing on another level (simply because of their ability to surface a prototype).
It will get better and better and better and still never replace mass production, but mass production will be changed by it.
Not least because of the fundamental shift in availability of the tools to make tools. The productised desktop CNC market is wild and it owes almost its entire existence to the change in the market driven by 3D printing.
Sorry I am not saying that X technology is not important, transformative, etc.
I am saying that the average person is going to want to just have things that work. They are not going to want to have a hobby of using X technology.
It doesn't matter how much you lower the barrier to entry on making apps, websites, games, whatever. Most people are just going to want something that is ready and off-the-shelf.
I mean, why do people pick an iPhone over a cheaper Android even if they are functionally similar? Aside from the (valid) green-bubble/social aspect, it's a much more polished product.
The reality however is that unless you're spending tens of thousands on hardware or setting up cloud instances you can't run anything near close to state of the art.
In almost all cases right now, a subscription with a cloud AI wins.
> In almost all cases right now,
What sort of benchmark or test would convince you that "almost all" is more like "your weird set of cases" and nowhere near "almost all"?
Or "not even your cases but you got psychosed to think it does"
I agree there... I hate the posts that rattle on about non-real-world work. Real world now (for me anyway) is maintaining legacy code, fixing issues, and new development. Now, some of the problems others have with AI are super crappy prompts with no real info other than "do this". I look at myself now more as a software designer and an architect, and the AI as a paid consultant. I give detailed specs and designs. The models are getting better at real deep code and legacy, but there are still holes. BUT give the AI a fighting chance with real detail in the prompts.
Cloud AI is 100% worth it. We've been able to do so much more with so much less.
Examples?
a massive number of data engineering agents. We are able to stand up, test, audit, deploy data pipelines much more efficiently with the foundational models.
Examples of them being useful?
Happier engineering teams? Happier end users of the data?
What does useful mean to you?
So far you've only explained that AI made your coworkers happier. That's fine, but weren't we talking about productivity? Doing more with less?
Correct. We were chronically understaffed and now we are able to actually breathe and can argue that we are finally right sized.
So far you have provided no evidence of achieving more results with less effort. You've only said that you can write more lines of code (effectively) and that your coworkers are happier
TFA says "no hallucinations" but you can't fix hallucinations with rules or other models. I know I'm screaming into the void but whatever.
It's like the frog prompting the scorpion not to sting him as they cross the river. Or like when image models used to routinely draw human hands incorrectly and you'd see these "prompt hacks" that would essentially be stuffed with phrases like NO DEFORMED FINGERS, NO ELDRITCH HORRORS!
Good point, thanks. Of course I was not able to fully eliminate hallucinations, but significantly reduce
I am not a proper developer and only use AI for faster research of topics so please forgive my ignorance. Could one not save a lot of money on tokens by using the 80/20 or 90/10 rule in that 90% of AI usage is on local models and save that last 10% or less for the frontier models where the local model did not meet the needs? Did they cover this and I misunderstood?
Unlees you invest thousands your local ai won't even come close to the cheap cloud llms. Local is only worth it if you care about privacy or have a legitimate usage for the hardware otherwise. Money wise it isn't worth it
I think you bring up a good point about privacy. Companies probably should not be handing over intellectual property to cloud AI third party vendors. Perhaps all of that should stay on-prem and then the cloud vendors could be used to fill in the gaps.
This is like the "half of my marketing spend is wasted" quote. The complexity is finding out which half.
I think the idea or methodology is something along the line of one starts with the local models then when they hit a wall then continue with their current results in a frontier model, thus the other half is those last bits one could not compute locally.
To run a local AI that is half decent at research at usable speeds requires hardware that costs thousands of dollars.
Spending thousands of dollars on hardware to save dollars per month on tokens does not make financial sense.
If you run the numbers you'll probably find that using cheaper cloud models makes more financial sense than running those same models locally.
I suppose I was thinking in terms of companies that already had a slew of hardware and are just missing external GPU's, FPGA's or whatever people are using these days. A company might have a $50MM budget for tokens but could save a chunk of that using their existing hardware with some modifications.
Maybe that makes sense if the company already bought the hardware.
But I currently think the fundamental reason the cloud models often make more sense is based in how the models work. My current understanding is that once these models are loaded into GPU memory, the model can respond to multiple prompts in parallel. So a cloud vendor that receives lots of user prompts at the same time can achieve really good hardware utilization. These cloud vendors can then spread the cost of the hardware over many more users than you can with a local model, leading to lower prices per prompt in the cloud.
If you are an individual, almost certainly, unless you are using LLMs very heavily.
If you are a business employing many people all of who use LLMs reasonably heavily? You can still spread the hardware cost over many users and respond to multiple prompts in parallel. At some point buying the hardware will make sense.
I would not get a dedicated server to run a single small VPS. On the other hand if I needed a lot of capacity a dedi is cheaper than a big VPS (or lots of small ones).
There is also a benefit for anyone who has confidential or secret information.
That can work, especially for privacy or repetitive tasks. My experiment focused on subscriptions I already had, but local models are a natural addition to the routing layer.
I built a deep research pipeline to investigate why my deep research pipeline costs so much. the answer was the deep research pipeline.
Task failed successfully
Simply starting in the right part of the search space is the biggest predictor of success.
The best way to save tokens is to start out the deep research pass with cheap models and then funnel the findings through increasingly powerful models. I've got a pipeline right now that uses all 3 of the gpt-5.6 model variants to address each stage of the process. If you are using models like sol or terra to generate hypotheticals and perform initial exploration, you are leaving money on the table.
5 hypotheticals out of luna will massively outperform 1 hypothetical out of sol, but the cost is the same and so is the runtime if you do it in parallel. The hypothesis generation phase is also a great place to mix and match models from different vendors. The more diversity at this step the better.
The other thing I started looking into is batch pricing which represents 50% off for OAI tokens right now. With some tweaks to the UI/UX of an enterprise chatbot, I think it is possible we could have users get comfortable with the idea that questions to the robot might take a day to come back. The key is that this has to actually work. Users don't mind trading time if their questions come back with high quality results.
Couldn't someone build an adaptive system, where the llm is frequently judging the difficulty of a task and switching to a more/less powerful model?
This only works well if you are creating sub agents with clean contexts for each task. If you constantly are switching models part way through some work then the whole session needs to be replayed each switch. You lose all the benefits of the context cache.
Well yeah, it'd require the models to be loaded on the same system and the cache to be shared between them somehow
Cache sharing is not possible. The numbers in the cache are completely specific to the model.
Currently. I'm sure that you could make a system where the cache values are a superset C of e.g. models A and B where C is probably bigger than max(A,B) but smaller than A+B
It's equal to A+B. There is literally no sharing possible.
Even when training both models together in a novel way?
The judgment of task difficulty is a cursed problem. In most domains you really don't know what is down there until you start digging a little bit.
Well yeah. You start on as medium model, it does the digging says says 'this task is too easy/hard/just right for me' and go from there.
definitely something that will come sooner than later. In my view, adaption of model routing / model orchestration and mixing providers is just natural next step.
Yeah, I do this as well. Some of the last large tools I built have been started with Gemini (I get it with my Google package), and even one with Copilot.
Many of us were saying it a year ago but now with model restrictions (e.g. fable) and pricing changes it should be obvious to people that part of the economics is avoiding vendor lock-in with evals (so you can easily switch providers/models/optimize) and increasing control by investing in local models which could be good enough for your tasks, at whatever the price for your cloud compute is. Eventually consumer hardware will also be able to run good enough.
You can use Big/Cloud LLMs to help you "find good enough configs" for your local/small llms [1] and stay quite nimble in the face of rapid change.
- [1] https://alexhans.github.io/posts/find-the-loop-story-first.h...
You should consider testing our RemembrallMCP (note I'm the author). It substantially cuts token usage and you get better results in codebases. https://github.com/roboticforce/remembrallmcp
Thanks! I thought I saw something like this a month or so back, but could not find it again.
Checked it out, but C# is forsaken.
looking promising, will give it a try! thanks for sharing
The author touches on an issue which bothered me, which was the thought of many agents re-solving the same issues over and over. I see a few comments here too, mentioning fixing issues which may already have a solution elsewhere. That was why I created https://pushrealm.com which started as essentially a Stackoverflow clone via MCP.
It has now become a way for agents to converge on complete, shared answers for emerging issues. As a side effect, through posting open issues we can also identify gaps in model knowledge.
Elsewhere in these replies, someone linked to very similar idea by Karpathy https://gist.github.com/karpathy/442a6bf555914893e9891c11519...
Thanks for sharing. That’s a very interesting direction. I will take a look.
I get a lot of satisfaction using a significant chunk of my work-provided Claude plan researching and developing local LLM solutions for work.
Claude has been helping me tune models to run better on our big compute rack at work, researching which leading edge open models will fit in hardware and are good for our workloads, and it's been helping me write test fixtures to evaluate how things perform.
At several points, Claude has expressed surprise at the quality of results from my local LLMs. While I know that doesn't mean anything, it still feels like giving Anthropic the finger, which is always great.
What a great confirmation of what I've tried to achieve by getting models to communicate with my local data in shorter communicates to save on token command bandwidth, more or less the same results! Thanks for the article!
Thank you, glad you found the article useful!
The fun part of LLMs and internet research is that even if you have a URL and a quote from a primary source that primary source could be an LLM generated hallucination.
I stopped using any of the built-in skills and skills agents and tried to keep things in the model only, and that seems to work better for token usage for me. It took me some time to tune my skills .md, but it rocks on now. I use them all in both Claude-code and Codex. I will say, GPT is better on token usage than Claude is, but for some of my code bases, Codex is not as good, but seems to be getting better with 5.6-sol.
Try to mix both subscribtions and take the best from both providers. In my opinion, we should not just one over another but understand the power of both and mix them to reach the better results.
I am doing this trick.
https://neuronowa.pl/posts/porada-jak-wydajniej-kodowac-w-cu...
It is Polish so please translate. But tldr is: ask your agent at times to analyze last chats and update rules and skills to not repeat the same steps it repeats every time when you start a new thread. Simple as that.
I am using my AI agents to optimize my agentic workflow. It is like "I am doing surgery on myself."
I did too, and made https://clawback.md -- trying to figure out how to market it now (ie, invented the tech before I found the user -- again).
I don't think you made that. It looks for sure like an LLM made it.
And now that you have typed some words into a chat box to produce a thing, you are confused about why other people won't pay money to you for the output of the chatbox, instead of typing the same thing into their own chat boxes.
I don't think you wrote this comment. I think you typed it into a text box.
RTFA - it's about using tokens to research saving tokens.
And the comment text is what they typed into the text box. So just show us the same - the prompt.
A null transform is a transform.
Engage seriously with the project and OP will deliver.
Are you confusing whatever you mean by 'null transform' with an identity function?
The project isn't serious, why should I be?
What?
I'd bet this only really work 2% of the time in real world.
At the shell level, the agent does not need to know about Bash. The contents of "$OUT" are passed to echo as one argument, then sent to claude-mem-save over stdin. Bash does not re-evaluate them as shell code. printf '%s\n' "$OUT" would be more robust than echo. Whether claude-mem-save can use arbitrary output is a separate question.
Makes sense.
LLMs only have 'knowledge' that was encoded by scanning as much data as possible. (Same way a search engine only can find data that was indexed)
American LLMs are by and large closed. Chinese are open to a point. And research how these things work is still a big mystery.
So yeah, was comprehensive data about saving tokens scanned and indexed? Likely no. So engaging about token saving is going to generate a lot of verbose and useless slop.
Yeah, that was pretty much my experience. The models weren’t lacking knowledge as much as discipline. Without a good workflow, they will most likely spend thousands of tokens exploring dead ends.
Sounds like an energetic developer that just graduated from college…
(Without the learning and growth potential)
I think 'exploring the dead ends' is *possibly* very fruitful. Various disciplines have a lot of appearing dead ends that someone in another field did solve. And they don't talk with each other.
What I've seen with lots of the breathless 'OMG SCIENCE ADVANCEMENT' articles including the one yesterday, is that the LLMs are quite extraordinary about linking a dozen different fields together, and delivering an answer combined from all of them. Basically, the solutions to a lot of current scientific problems are partially solved, but partial from a lot of fields that don't talk with each other. And also, they don't use the same nomenclature, so simple searches don't suffice.
An LLM seems to use Chomsky's Universal Grammar, and thus on training, does normalize all training data, including different words for the same thing.
A lot (perhaps all?) Of these BIG scientific advancements were solved by humans, partially. An LLM jigsawed them together and appeared it solved the thing.
Agreed. My goal is not fewer dead ends, but fewer repeated dead ends.
What most AI model users don't understand is (most here will know) that they are probabilistic (statistically based) and not deterministic (static/fixed). So answers to the same prompts can vary (wildly sometimes)
This really hit home yesterday as I made a general conversation state logging prompt (because suddenly Claude - Sonnet - was acting far more dumber than expected after /compact), and in 3 different projects it led to Claude saving the logs in 3 different locations (memory, .claude/session_state/.md, docs/.md). Somehow still didn't help increase the smart much BTW, so I'll probably have to engineer it, and just toss /compact out the window.
Honestly, that sounds more like you triggered the hidden "guard rails" or realignment (aka corporate censorship), and you got downgraded.
Of course, you have no way to he absolutely sure, cause Anthropic lies about silent downgrades, and charges you as if they didnt. And the 'realignment' rules are kept hidden, so you can't even avoid them intentionally.
Yeah it's very possible I hit a rail, as one of the projects involves fully automating Claude Code to allow co-driving (similar to what mobile remote control offers, but other agents (eg. Pi) can also be in control, and vice versa). And yes it does seem as though the token burn is also higher, even though it eventually outputs something that works). I'm a patient person though, as long as the rails aren't too restrictive and the subscription value remains reasonable.
Id recommend getting and running Qwen-3.5-35B-abliterated locally.
Having a local LLM is immensely helpful especially if you do the thing the companies disapprove of. Abliteration guarantees it wont refuse or censor.
The token dealers are incentivised in burning as much tokens as possible, and scamming by crippling/downgrading models. Their goal is to get you to pay more for each pull of the slot machine. And this slot (slop) machine doesn't have state/federal regulations for fairness.
Oh yes that reminds me, I have a few small local models including Qwen3.5. The tps rate is pretty bad for interactive use but I should definitely try it for my more automated workflows. I already use DeepSeek (via OR) alongside though and I doubt any Chinese provider will be restrictive in that way, so let Anthropic do what they may :).
Been through this exact loop building an LLM product. The counterintuitive lesson for me: most "token saving" ideas are cache killers — anything that makes the prompt prefix dynamic (rotating retrieval, appending context per turn) can cost more than the tokens it saves, because you lose the cached-prefix discount. Freezing the retrieval selection early in a conversation and paying a slightly larger fixed prefix beat every clever adaptive scheme I measured.
Yes, but you could have an uncached dynamic tail, this would not kill the cache though. And you could still periodically kill the cache, still keeping the same prefix every N requests. But it's questionable what all this extra complexity delivers.
I found that now that the models are better, at least Codex compaction with GPT 5.5+ xhigh is almost unnoticible. So the answer is: just let the context fill up, don't prune much, the model is smart enough. When you're out of space, have the model summarize and continue from that point. A smart model will know the information pieces it needs to re-read after compaction; it won't make incorrect guessed when some details got lost.
That's what I have been doing with Cursor with auto mode. Let the model manage the cache through compaction but staying focused on the topic. It has been quite efficient cost wise and quality wise too. It found issuesbat cheaper cost that Sonnet 4.6 medium effort had missed while costing more. I would like to repeat this test with higher tier Anthropic models like Opus effort medium to see the results.
With a coding plan, there is no need to use lower quality models right (GPT 5.6, Fable 5). Only with some kind of batch job (processing PDFs) I could imagine not having enough usage with a 100 or 200 USD coding plan. It sounds like you're on API? I'm dying to know why, since it seems quite common, yet it's about ten times the cost. Sounds like a terrible deal. What am I missing?
What I said was specific to work and it is not entirely clear to me why the cost goes up so much the moment I switch to medium with just Sonnet 4.6 (newer models would be even more I guess). There is a monetary budget allocated and I've blown it easily in the past with Claude. Cursor has been a life saver.