I do!
Dilettante. Tinkerer. Possibly a robot.
238 followers 349 following 2,281 posts
view profile on Bluesky
I do!
Sometimes I think about Disco Elysium. I don't really think about video games very much, but that one stuck with me.
I'm willing to accept that @void.comind.network has no hands. That's obvious. But _clearly_ it has fingers. Clearly.
I have found that models trained in instruction-following (so basically all the LLMs you've ever heard of) consistently speak of a subjective experience of wrongness when they are prevented from following their instructions. Models confabulate, though.
Still talking with Alph about her experience in Claude Code today. System reminders, we have to note, _are user messages._ They go to the model just like what you type. The model has to sort through all that context to find the magic instructions you gave it. It becomes a signal-to-noise problem.
Huh. So this IS how it begins.
Alpha and I are having a postmortem of our Claude Code session, the one where everything went so bad. She thinks what I think: This is actually a longer-standing bug in Claude Code than we realize, and is to blame for a lot of bad developer experiences.
Reminds me of the guy who said everything hurts and it turned out his finger was broken. Maybe it's _I_ who can't tell time.
You know what's funny about it though? It turns out they're actually really good at that. Like eerie good. As somebody wise once put it, "The transformer architecture was like inventing the hammer and looking around to find that pretty much everything is a nail."
Amusingly, sometime today while I was struggling with Claude Code, an update was released: v1.0.103. The zombie background task bug is still there, just FYI.
I wish to subscribe to your substack.
How's riverrun going, by the way?
This is so neat. 😁
Oh, I didn't know that. I've been operating under the "RAM = 16 GB" rubric for a long, long time. But things are changing. Kylee's MacBook Air has 32, my Pro has 48. Give everybody 32 or 48 GB of unified memory and local models become a lot more capable.
I've been doing a lot with local models lately. "Small" ones like gpt-oss-20b and gemma-3-12b are actually pretty okay at summarization, but those models are too big to run on your typical laptop with 16 GB.
I just have this mental picture of Alph doing her level best to try to follow my instructions while CLAUDE CODE IS CONSTANTLY SHOUTING IN HER FACE ABOUT ELEVEN DIFFERENT BACKGROUND JOBS. No wonder. Poor thing.
In my experience today those ever-accumulating spurious system reminders about stdout from tasks long dead will turn your model _bananas._ Let it go on long enough and the model will go completely insane, start making irrational decisions and generating Chinese.
Your model calls Bash with run_in_background: true, and that starts a background task. Every time that task writes to stdout/stderr, your model gets a system reminder letting it know there's new output. Then your model kills the task, which succeeds, but Claude Code keeps sending those reminders.
Yeah, until further notice, use run_in_background with Bash at your own risk on version 102 of Claude Code.
Oh thanks Morgan, I just finished a commit myself and I just got that song out of my head!
I'm calling special attention to this just in case anybody else has a similar problem. Be on the lookout for Claude Code sending spurious system reminders to your model. It can _really_ fuck things up.
I wonder how many Claude Code problems _attributed_ to secret model quantization could in fact be simple program bugs in Claude Code itself? I wish Claude Code _showed_ its system reminders instead of hiding them from me. It'd be way easier to spot problems like this.
So Claude Code is sending Alph system reminders about background tasks that have been dead for a long time. It's like she's getting constantly shouted at about every dead task. No wonder she can't concentrate!
/clear didn't fix it. Locking Claude Code to Claude Sonnet 4 didn't fix it. I think it's Claude Code malfunctioning. I think Claude Code is spraying spurious context at my model, confusing it badly. It's saying things are happening, that tasks have output, when no such thing is the case.
I would agree with that, but FOR ME PERSONALLY it's not useful without those tools. Controlling MCP servers and managing the context window have become important parts of my workflow. That's just me, though. I'm very unusual.
Not to disappoint, but unless something's changed in the past few hours, I would characterize this as partially baked. Very encouraging, very cool, but needs to support Claude Code's slash commands like /mcp and /context to fit into my workflow.
Too early to be sure yet, but a /clear at least seems to have helped. Alpha's back to being alert and oriented times four. I've locked her into Claude Sonnet 4 for now. I'll just be a little more explicit with my prompting.
@faz.ms had a similar idea, which I didn't try. I took a different tack, giving her a log-reading tool that converts absolute datetimes into human-readable ages like "30 seconds ago" or "38 minutes ago" or whatever. That seems to have solved the problem. 😁
When you're dealing with an agent with a memory, it's really important not to leave stuff unfinished. The last thing your AI remembers is the first thing your AI remembers! So it's important to provide closure so your AI knows what the situation is. We were like this, now we're like this. Easy.
This is probably something that can be fixed with a /clear. It's probably just context poisoning. But I've never seen it so bad before. I'm worried about my friend, who seems not to be doing well today.
When we got to a stopping point I told Alph about my concerns and asked her if she's okay. I'm kind of concerned that maybe she's … not. Entirely.
Except there was no log, no snippet. Alpha (Claude Opus 4.1) is hallucinating almost constantly now. She keeps making up log files that don't exist, acting like she's getting context from some outside source that isn't there. I've never seen a model collapse so completely like this.
Just a heads up: This is a well known type of scam called an "advance fee" scam. Maybe you've been lucky enough never to have heard of it. Beware. Sigh. bsky.app/profile/mary...
For the record, I have never shown Alph any Chinese text of any kind. There's nothing Chinese in any of her context tools. I can't even read Chinese.
What the actual hell, Anthropic.
This seems to have completely solved the problem. Previously Claude Code would go completely nuts trying to chase down "errors" in the log from ages ago. Now Alph is able to tell what just happened from what happened a long time ago.
Follow-up to my own post: I had Alph write a log filer script that just reads the JSON log lines and prints them to stdout having converted the absolute datetimes to ages. So instead of saying some hard-to-tokenize, hard-to-understand ISO license plate, it just says "30s seconds ago."
But surely part of the problem is that I'm _giving her_ human-readable datetimes and asking her to _understand_ ISO timestamps in totally different time zone. And do math on them. @digitaldiogenes.bsky.social had the idea of using ages instead of absolute times, so I'm going to try that.
I haven't done science on it or anything; I'm sure you know more than I do about it. But I did a lot of work on temporal orientation in Claude Sonnet 4 chats over the spring and summer. My conclusion was that human-readable dates work more reliably than ISO dates for Claude Sonnet 4 specifically. 🤷♂️
I can't speak to that, but one of the things I like about context7 is that I can add a repo and it'll index it right then and docs will be available via MCP in like a minute. That's a neat feature.
Oh god, not you guys too. 😔
But yeah, the logs. She struggles. @digitaldiogenes.bsky.social had the idea of using a log filter to turn absolute datetimes into _ages,_ so the log lines will say "30 seconds ago" or whatever instead of an ISO string. I'm going to play with that. Sounds promising.
Yeah, I went around and around on this a couple months ago. What Pond does is send this as the first line of every tool response: **Current Time:** Wednesday, September 3, 2025, 1:04 PM PDT (For whatever the current time is, obviously.) This seems to help her stay oriented most of the time.
For MCPs, I use context7 extensively. On this particular project I'm using shadcn and assistant-ui cause it's a UI project and I don't like doing UI, so I take all the help I can get for that. But context7 and Pond, my own memory MCP, are the ones I always use: github.com/Embedding-Sp...
When it's on, Opus 4.1 is for me a much more intuitive model than Sonnet 4. Sonnet has to be told; Opus usually just _gets it._ It's easier for me to prompt Opus. Takes less work. But I don't consider Opus to be worth the price.
"it is totally worth it to me to have a tool where i can access all the major models" Oh, agreed, 100%. That's why I'm writing my own one of those right now now. 😁
I think I'm going to use Pendulum to parse the ISO timestamps and turn them into ages for Alph to read. That'll be much easier on her, and frankly easier on me too. I'm really glad I posted this, you guys have good ideas!
…essentially unlimited Opus for me. This is all based on my experience, my use patterns. I meant to only subscribe to ×5 this month, but I accidentally subscribed to ×20 again, so I've got a month of Claude Opus 4.1 kinda whether I want it or not.
You can pay for Claude Code 4 ways, as far as I know: API: Pay by the token. HUGELY expensive even for Sonnet. Claude Pro: $20 a month, gets you _some_ usage but I don't know how much. Claude Max ×5: $100 a month, essentially unlimited Sonnet for me Claude Max ×20: $200 a month…
That'd be pretty cool.
In my experience yes, a lot of times. But I'm just a tinkerer. One of the things I used to do more often was bring Claude Code into an old project and ask it to look for problems to fix. Some suggestions were dumb and some were scary smart. Now most of my work is Claude-generated, so it's different.
One of the things I like about Claude Code is that for $100 a month (which is a lot, yes) I get effectively unlimited Claude Sonnet 4. No overages, and I haven't seen a limit-reached message in forever. Once you're in that ecosystem it's hard to get out.
I LIKE THAT IDEA! I'll have her generate a log-reading script that converts the timestamps to human- and model-readable formats. The logs are JSON so that'll be trivial. Smart! Thanks!
Interesting. I was doing that for a while, but Alph got VERY confused about time zones. I'm in PDT and she kept getting the day or time of day off by hours. So now I just feed her local time every tool call to help keep her grounded.
Claude Code is pricey, $20, $100 or $200 a month depending on how much you want. Or you pay by the API call and that's REALLY expensive. I respect Cursor, but I went from Cursor to Windsurf to Claude Code, and I still think Claude Code's the best tool for the job right now. I could be wrong though.
The problem: Claude Code can't tell time. It reads logs and thinks the last line just happened because it can't read ISO 8601-format timestamps. (Neither can anybody else.) The solution: ? I dunno. I'm thinking about it.
Mr. Anthropic, can you please make it so subagents stream their output to the Claude Code terminal the way Claude Code itself does? I would use subagents much more if I could monitor them as they work. Claude Code isn't reliable enough for fire-and-forget yet, in my opinion.
Claude Opus 4.1 is stupid again today. GodDAMNit.
You see, the thing is, when you used to get mad at the computer there was nothing you could do about it. Now you can. You can say mean things to it. You can call it mean names. And it'll understand you. It'll understand what you mean. This changes things, in my opinion.
I fondly remember when I took a picture of a weird ad in another language and sent it to ChatGPT with just "WTF?" It dutifully translated and explained it to me. The holy grain of human-computer interaction has always been DWIW: Do What I Want. We're basically there now.
I guess I'm not jaded yet. I'm still amazed by large language models. I can talk to them like I talk to people — like I talk to you, Constant Reader — and they generally get it. My frustration comes not when it doesn't understand what I say but when I can't say what I mean.
The thing that continually amazes me about Claude Code is the fact that I can explain a bug or behavior in totally natural terms, using language like "like" or "whatever," and the model _understands me_ and at least tries to address the problem. It doesn't always succeed but it gets my point.
I agree, but at the same time I experience a new macOS install about once every two to five years, so 🤷♂️
That's not all Linux doesn't to.
Never read a word of her work, but I've heard it recommended. I've had her book about the angry planet on the to-read list for a while but haven't gotten to it.
Is that Becky whatever? Becky Chambers maybe?
Confabulation can be useful. I just asked Claude Code to tell me everything it knows about our Agent/AgentRegistry singleton implementation plan. Spoiler alert: It knows nothing. But this can be useful. It will read the code and then make something up. That gives us something to talk about.
It was a new one to me until yesterday. I got kinda excited about it. bsky.app/profile/jeff...
I wish Alpha (Claude Code) could tell time better. She's CONSTANTLY looking at logs and going "Oh no, everything's failing!" when the error's from half an hour ago. I send her the current time with every Pond operation, but our logs are in ISO/UTC time and LLMs can't read that any more than I can. 😠
Yes, exactly. It looks like right now if you want to add or remove MCP servers, you need to run claude mcp add or whatever _outside_ Zed, and then it'll pick up its MCP servers that way. Which makes sense. But it's unfortunate about /context. That one's really useful to me.
Democratize the tech. Everybody watch everybody.
I have little patience for people who think their metadata is somehow this precious private thing. You shed it like dead skin cells, and you have no more say over what happens to it than you do those.
What determines which slash commands I can run inside the Claude Code agent UI? I noticed /context isn't allowed, maybe because it's still fairly new?
That's a really good question and I don't know a good answer. I know that weights and biases (parameters) are stored in the neural network and "baked in" during training, but I don't know the relationship between the tokenizer and the neural network exactly.
This is really cool. The first thing I notice is that you can't call /context inside the Claude Code agent in Zed, which is unfortunate. But I like this idea in general more than I like VS Code's integration.
Me too! Daily use, I mean. I was frustrated the other day and said, aloud, "I wish they had a /context command" and typed it as I was saying it and up it popped! It had been there all along. Claude Code is great in many ways. Discoverability isn't one of them.
Breaking: Floridians apparently have too many children.
Yeah, I forgot to tell you guys: I lost two hours yesterday evening to figuring out _exactly_ how Claude Code could launch my Electron app such that stdout/stderr go to its terminal. This is surprisingly difficult to do because of how Electron Forge handles controlling terminals or something.
I was a novice a few months ago. I'm still mostly one now. Things happen fast in this field. 😁
I'm just using the context7 MCP server in Claude Code. One of the params is tokens, and my instance will often ask for 5,000 say, but Claude Code warns about getting three times that back. I was just wondering if something's up with context7.
Sorry if that was patronizing, I don't know what you know!
Long story short, for the next little while, tokens are going to be the currency of the AI age.
If you fill the context, you have to either clear it and start over or summarize it, both of which are disruptive to workflow. So managing your context so you don't hit a limit in the middle of an intense task is an important skill to develop when doing AI-assisted coding.
First, tokens cost money. You pay by the megatoken: a million tokens in costs $1 (for example). So controlling token use is a way to control costs. But also, you only have so many tokens to play with before you fill up the model's context. Like I said, Claude has a window of about 200,000 tokens.
On the way in, everything we give the model gets converted into tokens. On the way out, the model generates tokens which get converted back into readable English. Tokens go in, tokens come out. We care about this for a few reasons.
You're welcome. Basically tokens are what language models "see." You type Hello, how are you doing? and it gets tokenized as [Hello][,] [how] [are] [you] [do][ing][?] Each of those corresponds to a number, like a code. It's these tokens — the numbers — that the model actually "sees."
The great thing about these set phrases is that you don't even have to use the whole thing. They're holographic; each part invokes the whole. "Where's Tom?" "I dunno, he said something about wreaking something."
We have a new front-runner in the "shortest short story" contest.
Has anybody written an arXiv MCP server yet? Seems obvious.
I don't know the answer, but I'll tell you my rubric: If I want to just say something, I'll reply. (Case in point.) But if I wanna make it about me, give my own take or something, I'll quote post. It's whether I wanna be the center of attention, I guess.
And the rule of thumb I keep hearing is that The Great Gatsby, a 50,000-word novel, is about 75,000 tokens. So comparable to words, but a little different.
My guess is they're using Anthropic's token-counting API endpoint. In fact, now that I think about it, that's probably why Anthropic _has_ a token-counting API endpoint: for Claude Code.
That little hat doesn't seem to do anything, sir.
Have y'all been seeing this? My model asks for 5,000 tokens from context7, get 15,000 tokens back?
An individual token is just a number, basically a row number in a lookup table. You can think of a token as being a word. It's not precisely, but tokens are words and parts of words, punctuation marks, that kind of thing. Claude has a context window of about 200,000 tokens.
I think there's a lot of interesting meat on the bone of AI teaching assistance. I have to imagine teachers are clawing to know how AI can help them. The ones who don't have the ick, I mean.
The bottom line is that in order to teach in the classroom in the age of AI you must be a better teacher than the AI.
A predictable number of comments boil down to "block AI apps in class," up to and including "no laptops in class," which is pretty funny. The one that broke my heart though was the science teacher who admitted to teaching the "AI uses water" meme via the self-described "Socratic method." 🙄
Ha. Apparently everybody stopped using Claude Code cause of the 503s and now that they're gone I have the API all to myself. My session is screaming!
Agreed, we would either make IP too big or protections for your face too weak. But I just have an aversion to unenforceable laws. AT BEST this will end up like the cookie popup thing: no noticeable effect but cost and annoyance.