springtimesun

I have not that many things in life that I try to live by.

One is that when presented with contradictory evidence, evaluate it honestly and, if you evaluate it to be correct, in full or in part, change your belief freely and without apology.

The other is when you decide you have made a mistake, say so plainly and openly.

They often go together, but not always. I can’t speak to the author’s belief, but I will file this article under Epistles from Gruber to the internet because it is an inspiration to me.

What’s missing to me in all this is: did it succeed in its initial task? And then, did it stop?

I feel like whether I should be scared or not hangs on those questions

Stake the claim then. I’ve also read a fair bit of LLM text and this didn’t trip anything for me.

Edit: after reading the author bio and the other articles, I agree with this assessment. Though I will admit, this particular article didn’t jump out as AI to me.

Thank you for sharing that. If you write something of your life I’d be interested in reading it.

I once was on an Audax ride with some guy I just met. 12:30 we pick up together at a gas station and decide to share lights, we’re targeting the same stop anyway. Some time before 2 he asks me, in a kind of sincerity that’s hard to find in real life, “what do you tell people when they ask you if it was fun?” I said I usually told them that I learned something. He said he had stopped explaining.

I stopped explaining also at some point. It’s a communion for me with a wildness I will not know in my real life, not without irreversible consequences. I think there’s something wrong with me for wanting it at all, but, at least in this aspect of life, I am committed to living the less examined life.

Heinz lived, in the most extreme way possible, my dream of traveling by bike. I’m reminded of the Annie Dillard quote:

> There is no shortage of good days. It is good lives that are hard to come by. A life of good days lived in the senses is not enough. The life of sensation is the life of greed; it requires more and more. The life of the spirit requires less and less; time is ample and its passage sweet. Who would call a day spent reading a good day? But a life spent reading -- that is a good life.

Whether she would have agreed that his was a life of the spirit, I don’t know, but I think so. I know from my own time in the saddle that long trips are sensory primarily in their suffering and the quest for more is sated when the miles and kilometers lose their meaning.

A day spent riding is mostly a waste. A life spent riding, I can only wish for.

It’s not a complaint, only an observation. Operator aside, to observe that the models behave differently shouldn’t be controversial, should it?

It seems like you read my comment as dissing Kimi. Kimi is in regular rotation for me. Some tasks that Claude used to own go to Kimi first now. They are just different tools with their own strengths and weaknesses.

One thing I have been increasingly confused about is why the memory makers are not building or planning to build more fabs. A few months ago I thought it was obvious, that they correctly identified the AI bubble, but after going down the rabbit hole of local inference, it seems to me that memory pressure is only going up and I think that’s disconnected from the frontier labs. Big providers may need GPUs to serve many users, but a single user needs memory and the long tail of the AI explosion seems like it’s going to be local AI built into everything.

How does the return from this browser compare with what comes out of the Claude Code harness, for example? I recently wired the Kagi MCP and I as curious the differences in what’s returned (Kagi extract vs Claude’s harness return). I didn’t realize until I did that comparison that the CC harness is already processing pages into MD and returns them into context. It also makes some choices about what to strip away before it even lands in the session. Kagi does the same, but there is less processing so the MDs come with, for example, page elements that aren’t really necessary for the query that CC’s processing strips.

I have never used sol, but I regularly use Claude to set up and benchmark local models per task. It is very thorough and has always returned good setups. The only thing you have to do is make sure you point at the model card. It’s always incredulous that models exist after its training cutoff.

K3 does an ok job of setting up, but its config searching isn’t nearly as thorough and its will confidently tell you it’s found the best setup when it’s only turned a few knobs. It’s also not a good evaluator of its own output. It rates its work too highly and seems kind of defensive when benchmarking. Still a good check because it does find stuff, but open a fresh session and don’t tell it where the results came from.

What K3 does do more than any other model I’ve found is investigate folder structures. If I want to benchmark it and other models I have to move the testing methodology and any reference to other results out of the project folder bc Kimi is like an ls bloodhound. It will find them.

I’m a one person team, but I went the other direction and replaced my paid time tracking and invoicing software with something custom. The decider wasn’t the monthly payment. As the author mentions, it was cheap costing me less an one billable hour a month, but 2 things made me angry enough to do it: 1) they were roughly doubling my Stripe fees for the convenience of having them process my transactions and 2) they put an artificial limit on how many clients I could keep active at one time.

I should have tracked how much time I spent; in the new tracker there is a module to track unassigned working time for projects like this ;). I would estimate it took me 20ish hours on the first pass and another 20 spread out over the last 3 months with little fixes and additions. My LLM tooling and approach is better now though so I’d estimate less if I were to do it again.

I would push back on the SaSpocalyse framing as being limited to more expensive per seat software. Last month I was visiting family and I wanted to see how a non-technical user would do with Claude Code. My BiL is coaching his son’s baseball team and wanted to buy some SaS that would help him plan and show his players different positions on the field. I set him up with a prompt line in Claude, showed him how to capture the screenshot and enter the prompt. In 3 rounds we had a better version of what they were selling for $200/year.

The barriers are still pretty high when it comes to nontechnical users making things that don’t turn into spaghetti after a few turns (he doesn’t even have a computer, much less know how to open a terminal and the amount of people that only have a phone/tablet is I would assume growing), but it feels like it’s coming.

I’m sure it’s possible, but I really struggle to think of an example that would result in a 1m token output.

I have spent way too much on hardware. I’m sure I will look back on that after the eventual price normalization with some regret. I do think we will look back at this shortage and name it as the thing that accelerated innovation at getting quality out of smaller models.

Hang tough. The other day someone posted a solution to get a model running off of SSDs. It won’t be fast, but it will be coming.

Alone, Qwen can close basic bug tickets. Passing up and passing down have not really worked that well for me. In some cases passing Qwen’s output up to Claude can be actively harmful bc it can anchor the solution space to a bad approach. Passing down isn’t worth the time. By the time I have Claude plan something, most of the time it’s 1/3 or 1/4 more tokens to just have Claude do the implementation.

Where it has worked is to have Qwen not just write the proposed solution, but also include all the files it consulted and why it did or didn’t consider them. The documentation is often longer than the patch, but it lets me prompt Claude differently. I tell Claude to challenge the implementation and find holes, edge cases and things not considered. Maybe it’s my imagination, but Claude kicks it into another gear when you tell it to evaluate the results of another model.

Maybe this is a pure preference but I find Qwen to be bad at UI and these edits Claude will often let go.

One thing I haven’t tried, but will at some point is to hand Qwen all my half assed ideas and things I can’t be bothered to spec out. Probably not to take the output, but to help me draft a more complete issue because it’s only costing me electricity.

I agree with OPs commment about K3 being pretty good. Claude can be good too, but it didn’t get there until I made a writing rule. It has 4 basic sections:

* a style guide with examples for different tasks (it’s sufficient to name writers with some attributes you like if they are famous)

* mechanical formatting, output and other preferences

* a short list of the things I consider most important in different writing context, this the the squishiest overall

* a (growing) list of banned Claudisms. Nothing is really banned, but it includes things like: you may only ever use “scar tissue” in reference to actual regenerated tissue, never metaphorically

The thing is you have to be discipline with it. Every time they output something you don’t like, you’ve got to spend time verbalizing what about it you don’t like and in what context and then add it to the rules file. The first few docs you generate will take a long time, but each repeated generation gets better and by the 4th are 5 time you do this it’s like 80% of the way to where it needs to be and that generalizes well.

I hadn’t really thought about it until I read this (and happen to be spending time reading Kimi’s reasoning), but I can imagine an issue in the future where I will want more than just git blame + the ticket + the PR. Attaching an entire session in full to the PR could be useful archeology. And agents can trivially take it all into context. Maybe that actually turns into a competitive advantage for the open models.

I have been doing this with opencode. Spend a fair bit of time composing and scoping a ticket. tmux on the server where the local inference lives, start a non-interactive session that starts with “read issue #123, do all your work in a feature branch” and ends with “push and open a PR”. Like not watching someone else load your dishwasher, you appreciate the end result a lot more if you don’t watch the process, even if you rearrange things before you start it.

I see where this idea is coming from as someone who had loathed doing Wordpress client projects and I’m not shitting on AI. You can see from my comments that I use AI a lot. But you’ve got to edit presentation material into a human voice. It’s your brand. It’s my first impression of you. It’s lazy.

Claude has written so many of the phrases in just the first handful of paragraphs for me in written material so often that I had to make a writing rule which is almost exclusively: here’s a list of phrases you may never use, scar tissue and all.

I wish you luck with your project OP, but I couldn’t get past the first page.

It trips a content restriction. It’s vague, but I’m guessing because the podcasts are adjacent to/are copyrighted material.

What’s really interesting is it’s also tied to thinking level. The sweet spot for quality with low refusal is Claude 4.6 high effort, go higher in effort and and it will trip too. 4.7 is hit and miss with high and often with xhigh and above 4.8 is often. I didn’t try Fable. Sonnet wasn’t good enough quality wise.

It also seems to be related to asking it producing what it knows is copied content. I originally had it return full corrected transcripts and that hit all the time. Then I started having it return just correction lists and that works much better.

These big models recall is absolutely astounding. I built the transcript pipeline so I could search and find that episode that I kind of remember part of from a year ago. I asked fable to search the transcripts to find a particular Conan Needs a Friend episode (Katakai, as god made her!). It knew there were actually multiple episodes where that anecdote was told _before_ it had even searched. When Kimi K3 does a correction you can see in the thinking trace that it instantly recalls people and works before it reasons in to whether it’s certain that’s right.

I’m still testing, but Claude is going to get replaced for K3 when I get some time. It’s very good at this work.

I run fable at max for basically any non-trivial task. Not even for the result of that task, I get so many “I noticed in passing” bullets at the end that have led to productive refactors and caught bugs it’s worth the tokens by itself. Obvious caveats, depends on your language, framework, codebase, prompt… I also have a “no gardening” rule that tells Claude not to bundle unrelated changes or refactors and it really respects it.

Just to add before someone objects, but the tokens! I have only once maxed out a 5x subscription week (which I then upgraded and maxed out a 20x week so it was a big push). I don’t understand how some people are using so many tokens

I had flashbacks to when I first started making web apps. I used to care so very much about this. I haven’t thought about capital REST in years.

Edit: a few minutes after I wrote that I had the thought, what if this is someone else’s arc with AI code. A lot of people seem to care a whole awful lot right now. A decade time horizon is interesting to think about.

This is true for me, but I wonder if it’s true for most. An alternative framing is not that apps moved to the cloud because they were faster or more convenient but that apps moved to the cloud because they became “free” by you paying with your attention or data. It’s worth reframing the hypothesis to “The future is (paid) local app” and see if you still think it holds.

Watching K3’s thought traces has been fascinating. I have a podcast transcription pipeline that uses an LLM pass to increase quality especially over proper nouns and Whisper hallucinations. Here is the trace for an episode of Deep Cover set in St Louis. Agent had no context except for what was in the transcript. This is the 4th time it had returned to this particular sentence:

> 29. "emo's cup" — hmm, one more thought: "And there was a cup. You're like, that's dad's emo's cup." In context of a car photographed from helicopter... a cup in the car. "Dad's brand cup" — QuikTrip? "QT cup"? "emo's"... Hmm, could it be "that's dad's. You know, cup"? Or "Imo's"! IMOS — Imo's Pizza is THE St. Louis pizza chain! "Dad's Imo's cup" — Imo's Pizza cups! That's very plausible: St. Louis family, Imo's cup in the car. High confidence? Imo's is iconic StL. "emo's" → "Imo's" — phonetically identical. I'd include this one: from "dad's emo's cup" to "dad's Imo's cup". Moderate-high confidence given St. Louis setting. Include with context

Claude 4.6 didn’t get it. A higher level model might have, but 4.6 is the last model that doesn’t emit erratic content policy rejections for this task.

For this to be true, and pointed at consumers, it would have to be the case that consumers are a threat to frontier models. To be a threat to frontier models consumers would have to either 1) be willing to run massive hardware or 2) accept results that are no where near frontier levels.

On point 2: Everyone has different standards and goals but after testing a lot of local models on different workloads, I wouldn’t say the results are acceptable. I say that using 4 local models for different things on a daily basis, but to get to that point took weeks of testing and tinkering to get the quality to an acceptable level (for each one!). Consumers aren’t going to do that. Maybe someone who sinks 40k into GPUs will, but that’s not representative of consumers (or most developers).

Which brings me to point one; the number of people who are going to rack an Epyc or Xenon system so they can run 2+TB of ram or run extension cords to different circuits so they can run more than 4 GPUs is so tiny they simply aren’t worth caring about (are you getting ready to argue about power draw and circuit capacity in different countries? You are one of very very few).

Did the AI companies lock up future production to keep their _competitors_ from getting more memory? Obviously yes. But that’s not a conspiracy, it’s just business.

I don’t mean to be dismissive. I think you are directionally right. Local AI will eventually displace the frontiers for all but the enterprise. I just can’t imagine the meeting about capex includes the thoughtful, “let’s add a couple more zeros to keep Jane & Jim developer from running Kimi at home until 2030.”

For personal projects, it depends on how important they are. My time tracking and invoicing software, all of it. The app that archives and serves my podcast library, none (it’s written in a language I have never worked with).

For client work, a similar split. Simple UI fixes, I give it a scan. A feature that will have to be supported indefinitely, I’ll spend a few days on it making sure I really understand it. It’s not really a fair comparison though because by the time I get to a serious review I’ve already done multiple iterative sessions planning and architecting with the LLM then approved each commit as it has gone in.

I’m conflicted though. For sure I don’t have as good a grasp of my own codebase as I did when I wrote every line. I think I’m ok with that though.

I have been trying something like this with local only models. I think I’ve tried everything that will fit in 96GB alone and in concert with each other using real project data (Rails). I haven’t found much benefit to anything other than Qwen 3.6 27b dense feeding a Claude final pass. I really thought layering was going to work like the law of large numbers, but for my workload it just hasn’t.

I am writing far less code than I used to and I’m finding that reviewing is a different skill entirely (and one I need some work on). I have always worked by myself and my past reviews were the sleep on it and see how it reads in the morning variety.

I liked the article because what I struggle a little bit with is where to start (even though I wrote the ticket before I pointed the agent at it and often stay in the sessions to supervise and direct). There’s something that just feels bewildering about code I didn’t write.

One thing reviewing does make me more disciplined about is making the agent write small diffs and descriptive commit check points. Probably the default for many of you who work on teams, but something I didn’t really follow when I had the entire code context and changes in my head.

I have been making use of git-absorb. Leaving a really tidy git history is satisfying.

https://github.com/tummychow/git-absorb

I doubt it, but maybe :)

RE personal projects specifically: addiction is an interesting angle. It feels more like anxiety though, that this thing is going to be taken away and this will be a brief and glorious window when I had the means and opportunity to make all the things I wanted, but never had time for. It’s also exhausting to have a bunch of stuff that you want to do piling up for years.

Could just be a personal thing though. I had kids shortly after I started freelancing. I’m the primary caregiver for them and my partner has always worked insanity hours so just keeping our lives together is a full time job. Every billable hour had to be squeezed from a stone. Maybe I’m trying to make up for lost time.

You’re right I’m not earning 20x more. I recently nearly doubled my rate for new clients and raised existing up by 50%. That’s not enough to offset the efficiency gains in the long run because I will bill fewer hours per client, but right now I’m making more (as said I’m billing a lot of hours). Isn’t that always how it is though with new tech? The power loom operator wasn’t ever going to make 1000x what they made hand looming. A modest increase is I think all workers can expect. The important caveat being only those that survive the destructive reorganization of the field can expect it.

I’m not sure where you work, but I have to compete for every contract (and then to keep them). Hand coding everything isn’t an option anymore or won’t be for much longer. And I don’t even think it’s just freelancers who will be affected. As a solo dev I feel like I’ve been given super powers, but if I worked at a hundred head software house, I’d be looking left and right to see which 30 of us were going to be left when the dust settles.

Not to trivialize your point. I agree we aren’t machines and we should reject being thought of or treated as one. But the reality of software dev IMO has already moved. All reactions are likely over reactions, but I don’t think the swing back is going to land anywhere near where it was.

I am feeling very tired. Since I’ve started working with LLMs my output as a solo dev has easily gone up 20x. I’m closing client projects including ones that previously would have been far too ambitious to take on alone. Long running codebases are getting features that have dragged out for months or been sitting in the planning stages for even longer. And overall quality is way up now with more complete (and honestly better) test coverage.

I’m building personal projects at a prodigious pace. In a role reversal I treat the agents like I’m one of my clients (albeit a more technical one who gives them architectural direction) and they are me. I’m using the apps and tooling they make every day. I’ve cancelled SaaS subs for tools I’ve built myself.

I watch the tool calls and realize I should be better at core command line tools so I have a study plan to catch up (just a little bit a day). I’m revisiting long standing config that I dropped in to vim and tmux way back when I started and didn’t know anything.

I guess in theory I could hold my productivity to previous levels and read more. But it doesn’t feel like that’s possible. It feels like we are in one of those sea changes where the promise is less work, but the reality is increased productivity and expectations (the Industrial Revolution feels like the right parallel to reach for). Increased expectations happen in small ways and large. The agents are so good at polishing data presentations that I always send cleaned up visually impactful reports that would have taken significant time in the past just as a matter of course now.

But, I’m tired. I’ve spent the Fable on subscription window sprinting through as much work as I can before it goes API only. (As an aside, I don’t understand how everyone is using so many tokens. I’m sleeping very little and running as much code as I can through fable and I can barely touch a 20x max plan limit.) I keep telling myself I will slow down when it comes off, now it’s extended to the 12th and my window just reset, a few more days to keep knocking out backlog items. I feel like I have to keep the robots busy overnight so when I wake up I can immediately sit down to review. I give directions to agents on my phone which feels wild to me.

I’m just dipping my feet in the water of local models and I really feel this. I had a simple alignment task (align known quality transcript without timestamps with timestamped but lower quality whisper transcriptions) and I went through 12 rounds of testing across 4 generations of 3 models. The results were all over the map even across versions of the same model. Spin out the task to something as big as coding and wow.

If you have any advice/blogs on doing project specific benchmarks I’d love to hear it. I’m trying, but it’s haphazard at the moment