Entries found: 10

Codeberg Divides

Published: Fri, 24 Jul 2026 00:00:00 +0000
Updated: Fri, 24 Jul 2026 00:00:00 +0000
UTC: 2026-07-24 00:00:00+00:00
URL: https://lucumr.pocoo.org/2026/7/24/codeberg-divides/

Codeberg recently changed its terms to exclude projects that are largely written with generative AI. Since I want GitHub to face competition I have thoughts.

Content Preview

Codeberg recently changed its terms to exclude projects that are largely written with generative AI. Since I want GitHub to face competition I have thoughts.

Codeberg is entirely within its rights to do this. It is an association with members and a democratic process, and that process produced a result. But democracy is a way of making a decision, not a guarantee that the decision is inclusive, wise, or even good for the people already depending on it. A majority can still decide that certain projects and people no longer belong.

GitHub’s governance has never been democratic and there is plenty about the platform that I dislike. Yet democracy is not the main property I need from infrastructure. I need it to be predictable, dependable, and reasonably neutral towards the legal Open Source software hosted on it. A democratic provider without a clear constitution can be worse at those things than a corporation.

The actual wording makes this more difficult. The terms prohibit projects that mostly consist of code written by generative AI tools. In an actively developed codebase, what does “mostly” mean, and who can still tell? I could not reliably assign authorship percentages to many of my own recent projects. The line is open to interpretation precisely where it needs to be enforceable. In practice the center will probably lose out, as it has a bias .

A harsher line would probably be preferable. If Codeberg wants no LLM involvement, it should say so. If it wants to prevent autonomous repository spam and abusive resource consumption, it should write rules for those instead. The current middle ground delegates too much of the policy to moderators and community norms. I’m currently assuming the community around it draws a much harsher social boundary, making projects and maintainers unwelcome even when they technically comply.

It is a real shame that the Open Source and Free Software communities are splitting this deeply over LLMs and agents. There are serious questions about copyright, labor, energy use, slop, and maintainers drowning in generated contributions. But these tools are also becoming part of how software is made. The Open Source world needs to figure out how to engage with that future, not just divide into camps. More importantly, LLMs if done and used well, should be welcome to all of us. They could be used to reclaim control and power, away from large corporations and institutions.

As I mentioned before, I want GitHub to face true competition in the Open Source space. I would particularly like some of it to come from associations rather than another large corporation. As a European project, Codeberg naturally matters to me even more. It can choose to be a smaller community with a stronger political identity, but that is a different ambition from being a broad and dependable European alternative to GitHub.

I wish Codeberg were more forward-looking here: willing to host the Open Source software of tomorrow, not only software made in the ways its community approves of today. It has every right to make the choice it made, but I just do not think it is a good one.

The Tower Keeps Rising

Published: Mon, 13 Jul 2026 00:00:00 +0000
Updated: Mon, 13 Jul 2026 00:00:00 +0000
UTC: 2026-07-13 00:00:00+00:00
URL: https://lucumr.pocoo.org/2026/7/13/the-tower-keeps-rising/

I feel that some vibecoded software changes somewhat randomly and unexpectedly. That made me think about Bruegel’s “The Tower of Babel” which shows an already quite chaotic depiction of the Tower of Babel. The story is usually told as one about pride and ambition and ultimately why people no longer speak the same language. But it is also a story about the unity that makes technological progress work.

Content Preview

I feel that some vibecoded software changes somewhat randomly and unexpectedly. That made me think about Bruegel’s “The Tower of Babel” which shows an already quite chaotic depiction of the Tower of Babel. The story is usually told as one about pride and ambition and ultimately why people no longer speak the same language. But it is also a story about the unity that makes technological progress work.

The text begins with a technology upgrade:

And they said one to another, Go to, let us make brick, and burn them thoroughly. And they had brick for stone, and slime had they for morter.

They use it for a civilizational project:

let us build us a city and a tower, whose top may reach unto heaven

But when God assesses the situation the bricks are not what concern him:

the people is one, and they have all one language, […] and now nothing will be restrained from them. ¹

The source of their power is coordination. They share a language and with that shared language they can combine their work into something no one of them could build alone. God does not take away the bricks or their knowledge of how to make them. He takes away their ability to understand one another, and construction stops.

There is the appealing idea that AI-assisted programming means better tools which lets us build more ambitious software. That is certainly true at the level of the individual and without doubt a developer with an agent will be dramatically more capable of changing a codebase. But large software projects have never been limited only by how quickly an individual can produce code. They are limited by how well people can coordinate their understanding of the system they are changing.

The shared language of a software project is not English or Python but it is the common understanding of what its concepts mean, where the boundaries are, which invariants matter, who owns what, and why the system has the shape it does. This language is rarely written down in one place. It lives partly in documentation and code, but also in code review, conversations, arguments, and the experience of having to explain a change to somebody else.

Before agents, some of this shared understanding was maintained by friction. If I wanted to change your storage layer, I usually had to read your code, ask you questions, and perhaps coordinate with another team whose service depended on it. This was slow, and much of that slowness was waste but not all of it was. Some of it was the process by which your understanding became mine, and by which both of us discovered whether we still agreed about how the system worked. This friction synchronizes people.

Agents remove much of that friction. I can ask an agent to add OAuth, you can ask one to add caching, and somebody else can ask one to rebuild the database from first principles and make the UI pink. Each change can be reasonable in isolation. The code can compile, the tests can pass, and the explanations can be generated on demand. None of us necessarily has to talk to the others, or even acquire the part of the shared model that the change once would have forced us to learn.

As I said many times before: agents do not feel pain, only humans do. Agents now let us act in parts of the system where we would previously have needed other people and in code bases where the people would have revolted.

When I look at some vibecoded scaled-up projects the codebases become Babel not because nobody can communicate, but because nobody needs to. Every developer has a tireless translator that can explain a corner of the tower and make whatever local alteration they ask of it. The changes keep landing, even as the architectural language that would let the humans reason about them together disappears.

But it’s not the biblical story. At Babel, the loss of common language stops construction whereas in AI-assisted engineering, construction can continue after shared understanding has already collapsed. The lack of an immediate failure is what makes it curious and a bit disorienting. The tower does not fall, and so we do not notice what was lost. It just keeps rising.

1. Genesis 11:3-6, KJV . ↩

Better Models: Worse Tools

Published: Sat, 04 Jul 2026 00:00:00 +0000
Updated: Sat, 04 Jul 2026 00:00:00 +0000
UTC: 2026-07-04 00:00:00+00:00
URL: https://lucumr.pocoo.org/2026/7/4/better-models-worse-tools/

A very strange Pi issue sent me down a rabbit hole over the last two days. The short version is that newer Claude models sometimes call Pi’s edit tool with extra, invented fields in the nested edits[] array. And not Haiku or some small model: Opus 4.8. The edit itself is usually correct but the arguments do not match the schema as the model invents made-up keys and Pi thus rejects the tool call and asks to try again.

Content Preview

A very strange Pi issue sent me down a rabbit hole over the last two days. The short version is that newer Claude models sometimes call Pi’s edit tool with extra, invented fields in the nested edits[] array. And not Haiku or some small model: Opus 4.8. The edit itself is usually correct but the arguments do not match the schema as the model invents made-up keys and Pi thus rejects the tool call and asks to try again.

That alone is not too surprising as models emit malformed tool calls sometimes. Particularly small ones. What surprised me is that this is getting worse with newer Anthropic models as both Opus 4.8 and Sonnet 5 show it but none of the older models. In other words, the SOTA models of the family are worse at this specific tool schema than their older siblings.

In case you are curious about Fable: I intentionally did not test it because I was not sure if the classifiers they are running might downgrade me to Opus silently.

Tool Calls Are Text

If you have not spent too much time looking at LLM tool calling internals, the important thing to understand is that tool calls are not magic and use some rather crude in-band signalling. The model receives a transcript, a system prompt and a list of available tools. The server munches that into a large prompt with special marker tokens. Because the model was trained and reinforced on examples of that format, at some point during generation it emits something that the API or client interprets as “call this tool with these arguments”.

For a file edit tool, the intended invocation payload might say something like this:

{
  "path": "some/file.py",
  "edits": [
    {
      "oldText": "text to replace",
      "newText": "replacement text"
    }
  ]
}

A harness then validates the arguments, performs the edit, and feeds the result back into the model. If validation fails, the model sees an error and usually tries again.

How exactly that formatting happens is not known for the Anthropic models, but some people have gotten out “ANTML” markers and they at times do leak also into public communications. To the best of my knowledge, the call above would come out serialized like this from the model:

<antml:function_calls>
  <antml:invoke name="edit">
    <antml:parameter name="path">some/file.py</antml:parameter>
    <antml:parameter name="edits">
[
  {
    "oldText": "text to replace",
    "newText": "replacement text"
  }
]
    </antml:parameter>
  </antml:invoke>
</antml:function_calls>

An important thing to note here is that this thing, while looking like XML, is not really XML. It’s just a thing they found convenient to tokenize and train on. The other thing to note is that a basic top-level string parameter appears in-line whereas an array of objects is implemented via JSON serialization. While I’m not entirely sure that this is how it works, there are some indications that this is not too far off. This will become relevant later.

There are two very different ways to make the model produce a structure like this:

1. You can ask the model to produce valid JSON matching a schema and then validate it afterwards.
2. You can constrain the sampler so that invalid JSON, or even invalid schema shapes, cannot be sampled in the first place.

The second approach is what people usually refer to as grammar-aware or constrained decoding. The sampler masks out tokens that would violate the grammar. If the model is currently inside a JSON object and the schema says only oldText and newText are allowed, the sampler can prevent it from emitting "in_file" or "type" . Grammar-aware decoding can be used both to constrain something to be syntactically valid JSON and also to enforce specific enum values or keys.

Without any form of constraints the model is merely following a learned convention.

The Failure

Pi’s edit tool supports multiple exact string replacements in one call. That is why the arguments contain an edits array. In the failing cases the model produces entries like this:

{
  "oldText": "...",
  "newText": "...",
  "requireUnique": true
}

or this:

{
  "oldText": "...",
  "newText": "...",
  "oldText2": "",
  "newText2": ""
}

Across repeated trials I saw a whole zoo of invented trailing keys: type , id , kind , unique , requireUnique , matchCase , in_file , forceMatchCount , children , notes , cost , oldText2 , newText2 , oldText_2 , newText_2 , and even an event.0.additionalProperties key inside the edit object itself.

The most annoying part is that the actual oldText and newText payloads were byte-correct in the invalid calls I inspected. The model had in fact produced the right invocation but then added nonsense at the end of the object.

The failure is also heavily context-dependent. A fresh single-turn prompt like “edit this file” did not reproduce it at all for me. An agentic history where the model had read files, diagnosed a problem and then composed a multi-line edit could reproduce it. And more annoyingly, not all transcripts will show that behavior. In fact, I needed Petr Baudis ‘s transcripts to reproduce this for me at all! In that user’s session continuing the session caused Opus 4.8 to fail around 20% of the time. Stripping thinking blocks from history reduced the failure rate by half. Turning on strict tool invocation eliminated it in my runs.

Why It’s Getting Worse

My strongest hypothesis is that this is not random deterioration but a training artifact.

When older Anthropic models were trained, they were trained on some tools (some of which were documented). But that training did not yet have a user-shipped harness like Claude Code as the obvious target. Modern Anthropic models are most likely different because their post-training includes Claude Code or a harness that looks very similar. The model learns what a successful tool call looks like in that environment. It also learns what mistakes are tolerated by that environment.

Claude Code’s own tools are comparatively flat. The ordinary edit tool is not Pi’s nested edits[] shape; it is closer to file_path , old_string , new_string , and an optional flag ( replace_all ). Looking at Claude Code’s client is very instructive: it contains retry paths for malformed tool use, parameter aliases, type coercions, Unicode repairs and filtering of unknown keys. In other words, Anthropic’s own client appears to expect and accept a fair amount of slop and repairs it, mostly silently.

If reinforcement learning happens in a harness like that, or a simulation of one, then slightly malformed tool calls can still complete the task and receive reward. The harness fully absorbs the error and there is little gradient against inventing an alias, adding a stray field or using a nearby parameter name.

Worse, the model may become very strongly adapted to the canonical Claude Code edit tool shape. A different harness can present a tool with the same semantic intent but a different schema. Such a tool can increasingly be off-distribution. The better-trained model might actually fight you harder because its prior is stronger.

This is not too surprising, but it is a change from how this was a few months ago. When Opus 4.5 launched, it adapted to other edit tools exceptionally well. In fact, I was pretty convinced that we’re on a good path where the models are more likely to adapt to any sort of tool shape that comes around for as long as the instructions are good.

Now I’m somewhat worried about the track we’re on here. Alternative tool schemas might not just be unfamiliar. They might be implicitly punished by post-training that optimizes for one particular, forgiving tool ecology. And that ecology is not documented. While there is a text editor tool that is documented, you will see that this format is in fact not followed by Claude Code. What Claude Code does internally (which is a closed-source harness) is hidden from you.

The Slop Harness

Claude Code is obviously closed-source but we can look at the minified code and get some idea of what it does. And honestly, it’s very forgiving of incoming data.

For a start, Claude Code checks the model’s visible text for leaked <invoke markup. It also emits some telemetry when that happens and then it has its own state machine to retry such bad calls by pushing back to the model.

It has explicit Unicode escape repair which fixes broken \uXXXX sequences and lone surrogates in string values. It also has per-tool aliases for parameters. For instance, Edit accepts old_str (presumably from the times when the models were trained on the officially documented text editor tool), the newer old_string from the schema, new_str / new_string , path as an alias for file_path , and some more.

It also silently filters out unexpected keys and it does not use strict mode either. The issue with strict mode is that Anthropic applies complexity limits to the tool definitions that cause API requests to fail, so presumably that’s why Claude Code does not attempt to use it.

Strictness

Will this problem be with us in other harnesses too? One huge issue with Anthropic is that the models are completely closed, and so is the harness. Codex models are also closed, but at least the harness is not. We also have gpt-oss which is at least a bit interesting. The models are explicitly trained to use OpenAI’s harmony response format and there is a lot of documentation that at least tells us how OpenAI people think about this.

Harmony makes channels and tool-call content types part of the prompt format. A function call can look like this:

<|start|>assistant<|channel|>commentary to=functions.get_weather
<|constrain|>json<|message|>{"location":"San Francisco"}<|call|>

The important bit is <|constrain|>json . The model can express in-band that this message body is JSON, and an inference stack can use that boundary to switch into JSON-constrained sampling for the body of the tool call. Presumably a bit of this also happens in Anthropic’s models, at least in strict mode I would imagine.

The marker in harmony helps the sampler to detect when it needs to sample with a specific grammar, and because it is part of the transcript, it makes that rather easy to do. For hosted GPT models, there is also an option to provide a LARK grammar for custom tools that need to adhere to something like this.

Anthropic appears different from that, though maybe not entirely. If an array of objects is represented as JSON, as it appears to be, then the model has to write JSON inside the tool parameter. There is probably basic grammar-constrained sampling going on, and that may partly explain the extra keys. For a nested array parameter, that JSON includes escaped multi-line file content inside string literals, inside one tag. The unexpected, made-up keys appear exactly at the highest-entropy point of that task: after closing a several-hundred-token escaped newText string, where the model must decide } vs , "..." .

Opus 4.8 and Sonnet 5 seem to have much stronger priors about what an edit tool call should look like and that prior appears to be Claude Code’s edit schema: a flat old/new string pair, plus the optional replace_all flag. My guess is that Opus has learned that an edit operation may have one extra optional field, but under Pi’s nested oldText / newText shape it has no trained name for that field. So it samples a plausible name fresh each time, which is why the failures produce dozens of random keys rather than one stable alias.

As strict mode in Anthropic appears to fix this, I presume that on the server side they are refusing to sample a key that is not permitted by the JSON schema structure. That would also explain why they have limits to the complexity of the tool definitions when strict mode is enabled.

So far, the Codex models I tested did not show this type of regression. I tested all available ones except 5.6, which I do not have access to yet.

What This Means For Harnesses

The uncomfortable lesson is that tool schemas are not neutral, at least not on Anthropic models. We like to pretend that a schema is an abstract contract and the model is a general reasoner that will follow it, but that might no longer be the case for some of the tools.

Tool schemas are somewhere in the distribution and some shapes are close to what the model saw during post-training and some are far away. Some are easy for the provider’s hidden encoding (e.g. top-level attributes in ANTML), whereas some require the model to write large escaped JSON objects inside nested arrays after long multiline strings. The model may be smart enough to understand the schema and still be bad at sampling the exact shape under pressure.

If this type of model behavior continues, I wonder what the implications for harnesses are. Obviously one could turn on strict sampling in Anthropic and the problem should go away. On the other hand, that the model has this behavior shows the impact that reinforcement learning has on them. Fighting that prior is probably futile if you want to get the best model performance.

Right now the reality is that Claude Code is not open source and we cannot really know what they are doing in their RL environments either. We cannot assume Claude-Code-trained behavior will transfer cleanly to your tools unless they are a close match. The more post-training happens inside one dominant harness, the more every other harness will have to inherit its quirks.

I used to be more skeptical of strict grammar-constrained tool invocation because constrained decoding can have quality tradeoffs. I still think that can be true in general, but this bug moved my priors significantly. If the newest models get better at solving the task while getting worse at faithfully emitting an alternative tool schema, then the harness needs stronger guarantees somewhere.

If you want to find out more, or you want to discuss this, consider reading the issue on the Pi tracker .

The Coming Loop

Published: Tue, 23 Jun 2026 00:00:00 +0000
Updated: Tue, 23 Jun 2026 00:00:00 +0000
UTC: 2026-06-23 00:00:00+00:00
URL: https://lucumr.pocoo.org/2026/6/23/the-coming-loop/

I don’t prompt Claude anymore. I have loops running that prompt Claude and figuring out what to do. My job is to write loops.

Content Preview

I don’t prompt Claude anymore. I have loops running that prompt Claude and figuring out what to do. My job is to write loops.

— Boris Cherny

Over the last months I have watched more and more people build something on top of coding agents that feels meaningfully different from just using a coding agent. Some of this happens on top of Pi which is cool to see for sure! The pattern is the same everywhere though: work is put into a queue of sorts, a machine picks it up, attempts it, stops, and then some harness decides whether that was actually the end.

If not, the harness continues the same session, injects another message, starts a fresh session with modified context, or sends the task to another machine. The task stays alive beyond the point where the model by itself would normally have said: “I am done.”

I think about that type of loop more than I want to admit.

There is already an agent loop inside every coding agent. The model calls a tool, incorporates the result, calls another tool, reads a file, edits a file, runs tests, and eventually produces some answer. That loop is one we have been quite familiar with for a long time. The other loop is the harness level loop: the loop outside the agent loop. That loop is also not new . We have been doing versions of this since early Claude Code days, but that loop is becoming ever more present in agentic engineering and in recent weeks it has started to dominate the Twitter discourse.

I Am Not Good At This Yet

My current status is that I have not had much success with this way of working for code I deeply care about which turns out to be quite a lot of code.

Part of that is taste and part of it is control. I attempt to set a high bar for what I want code to look like, and I want to understand the code I ship. Under pressure, or in a discussion with another human, I want to be able to explain what the system does without first having to ask a clanker to explain it to me. Now there is obviously a question if this desire to understand the code is one that I will still have a few years from now. For now I have not moved past the point of comprehension being important to me.

Given this desire, there is something I lack with my experience of code written without me paying attention, particularly from loops. Present-day models tend to produce code that is too defensive, too complex, too local in its reasoning. They avoid strong invariants. They add fallbacks instead of making bad states impossible. They duplicate code, invent bad abstractions, and paper over unclear design with more machinery. Worse though: I so far see very little progress of this improving. If anything, on that front it feels to me that we might even be making steps in the wrong direction. At least for my taste, present-day hands-off harnesses like Claude Code with ultracode produce worse code than what we were producing last autumn. That’s because Claude Code, with Fable for instance will be working uninterrupted on a problem for thirty minutes or more, when previously the process would have been much more human in the loop.

Furthermore it’s well understood that models tend to observe some local failure and add a local defense. Karpathy mentioned how they are “mortally terrified of exceptions”. In systems with important invariants, especially persisted data formats or core infrastructure, the right fix is not “handle every malformed case.” The right fix is to make the malformed case unrepresentable or impossible to write in the first place. Yet even with a lot of manual steering, that type of code does not come out of LLMs naturally, and even if the code comes out naturally like that, they will still attempt to handle now impossible errors.

When you take that behavior and you put it behind loops, you tend to amplify it. If each iteration adds another small defense, the system slowly becomes less understandable while appearing more robust. The more hands-off you are, the more that happens. It also teaches really bad practices when tools like this are given to juniors without clear guidance. Because if you ask them, why they are doing all that, they will convincingly argue their case.

Where Loops Work

At the same time, it would be dishonest to pretend the loop pattern does not work because it already works astonishingly well in some domains.

Porting code one of them. There are already impressive examples of large automatic porting efforts, including the reported work around moving parts of Bun from Zig to Rust . I have used it with success myself to port MiniJinja to Go . Performance explorations are another case where this works beautifully. A machine can try experiments, benchmark them, discard failures, and keep searching. Security scanning fits naturally too and so does almost any type of research: asking a system to explore a complex problem space and report back without necessarily committing lasting code. One thing that many of these have in common is that they either do not generate new code, but transform code that already exists, or they produce code that intentionally does not have a long shelf life. They either produce proof of concepts or ideas, surface findings or are more akin to mechnical transformation.

I believe that loops that produce artifacts without necessity of longevity or that create some form of clearly verifiable mechnical translation matters more than the general ability of a harness to mechanically measure a goal. Many successful applications of loops use another LLM as a judge or as an orchestrator. The mechnical translation case can be verified with a binary test case, but it can also be judged by an LLM instead!

Claude Code, for instance, is increasingly good at creating entire experimental workflows that it will then execute. Sure, the code it produces is slop, but that’s more the fault of the model than the harness not being a good judge on if a step in the workflow resulted in a net improvement or completion.

The harness just needs some signal that lets it continue. It does not have to be objective or binary, it just has to be useful enough to drive another iteration.

I absolutely love loops already that take the boring parts out of my day to experiment and measure and to give me ideas.

Software As Organism

On the other hand using that same looping methodology to write lasting code does not yet sit well with me. The metaphor I like to reach for is one of moving from software as a deterministic machine to software as an organism.

I became a software engineer in an enviornment that encouraged me to understand the machine. There was always a layer you could peel off to deepen your understanding. Machines that did not exhibit deterministic observable behavior were maybe accepted, but generally seen as not exactly optimal. Software architecture-wise, I saw it as desireable to push further towards more determinism rather than less. Likewise the ability to understand the code has been an undeniable goal. In practice not always possible we still took pride in writing code so that it became possible even for new engineers to navigate complex code bases through clever architecture. On well designed systems there were always engineers that knew where the invariantes lived, which parts were load-bearing and which changes were safe. Ideally all of that was also well documented. Where that understanding was lacking, it was generally regarded as something to improve upon.

Obviously that ideal has always been strained. Many software systems, especially very successful ones had periods where engineers on the team were able to keep them clean. Large software systems are not infrequently too big, too dynamic and too dependent on external services to fit into anyone’s head. Even without LLMs we already diagnose distributed systems somewhat like doctors in that we observe symptoms, form hypotheses, “order more tests”, try some remedies, and observe again.

Yet with LLMs we’re pushing much further in that direction and much quicker. We use them to write the code and we also use them for diagnosis and remedy. There are plenty of engineers that already live in a world in which the first step after the occurrence of a production issue is followed by having a clanker read logs, propose root causes and proactively put up a patch. The resulting patch is then often picked up by another machine that reviews, sometimes even landing it on main without any human supervision.

Obviously that is powerful and I cannot deny that it sounds appealing. But giving in to that idea, particularly with less and less human oversight means accepting that we may no longer understand the whole system in the same way. We treat it, we monitor it, we stabilize it, but we do not necessarily comprehend it.

I have no doubts that for some software, that is okay. Not every line of code deserves human authorship and worse code might have been written in the past.

But do I want all software to be authored this way?

You Cannot Quite Opt Out

What’s very uncomfortable is that opting out of this fully machine-driven future may not be an option.

Security is the clearest example today. Even if you do not use loops to build your software, other people will use loops against your software. Attackers will run machines continuously and even if it’s not attackers, then security researchers will and some of that automated work will throw up dust but also find real issues. And both the signal and the noise will come your way at a volume that makes it almost impossible to deal with unless you yourself throw a machine at the problem.

Daniel Stenberg’s post about curl’s summer of bliss is a good example of the pressure maintainers are already under. As far as I know, AI does not play a tremendous role in the core development of curl today. Yet despite all of this, maintainers are overwhelmed by reports, most of which are now AI-generated ones.

If attackers and reporters loop, defenders will eventually need to loop too to keep up. Maybe not to write patches directly, maybe just to triage and reproduce and pressure will increase.

The same is true competitively as some teams will out-build others through raw speed. Some projects will suddenly move faster because a tiny group figures out how to orchestrate machines effectively. Some startups will do with five people what used to require fifty. Some people might literally put a machine against your product in a loop and ask it to “make it like the other one.” And if their users are happy, does it really matter?

Not all software will be equally affected. Some domains will punish sloppiness and demand trust and responsibility, but a lot of software lives in a world where raw speed, quick experimentation, and vast coverage matter enormously.

Building New Dependencies

The scariest part to me is that we become dependent on these new machines in new ways. Software has always depended on tools. I remember the time when I had to pay for compilers. These new tools are a flashback to times where creating software came with real costs. But now it’s no longer a one-time payment, it’s a constant dependency. Not just a dependency on a filled wallet, but also a cognitive dependency.

If a codebase is produced by loops, reviewed by loops, patched by loops, and kept alive by loops, what happens when you no longer have access to the same class of systems? What happens when some trade restrictions take away access to the most powerful models? What if just the cost becomes unbearable? What if you and your team just lose the last remaining ability to understand the code without using the machine?

We may create codebases that are not merely hard to maintain by humans, but that assume machine participation as part of their maintenance model. This is already happening! It’s not happening everywhere, and it might not even be happening in ways that are seen as problematic, but we see more and more of it. People more and more merge code they cannot fully explain. People lose their ability to create issue reports or discuss things in chat, without augmenting or rephrasing their messages with the context provided by a clanker. Too many people increasingly rely on a machine to summarize or contextualize it. More and more do I encounter people who converse with me through the indirection of an LLM.

Again, maybe that is not even going to be wrong, but it’s a massive change to how we did things.

Future Harnesses

I have little doubt that this is where things are going but going there will require us to do something about our tooling everywhere, and not just in the coding agents.

Just orchestrating more loops won’t be enough. Better visualizations of changes or orchestration or agents will not restore our understanding. Either we need to find clever ways to jolt the human back into the loop and make the changes of the loops legible long term, or we need to find better ways to compose these ever more complex systems.

This is also where my thinking about the role of Pi is changing. Pi has been cautious, and I think that caution is good. I do not want a future where every interaction turns into an uncontrolled swarm of machines making changes I cannot follow. I would not want Pi to become an unmaintainable mess in an effort to win the race towards software that writes itself and I would not want Pi to promote this type of engineering either. At the same time Pi is a harness and harnesses are at the center of people running these new types of experiments.

Task queues for coding tasks, orchestration of agents, subagents, durable sessions will matter more and more. Even those of us who have their reservations and are not blindly embracing loops will have to start doing those experiments. We need to, because we need to understand how to make this future bounded and survivable.

Controlling Loops

As you can read from this post, I’m very uneasy about this future. Not cause of fear, but because of caution given experiences with this technology so far.

Adopting the idea of harness loops means that the harness decides when work is finished. In the agent loop, the model eventually says “done” and I review. Even before that, I usually steer along the way. I am involved and I enjoy learning along the way. In the harness operated loop I’m not sure what my role even is. Even the “done” signal loses all meanings and just becomes communicated to yet another machine that judges. My role is reduced to that of a messenger.

Today I do not like much of the code that I see from systems built that way and neither do I enjoy interacting with too much of software built with AI assistence. Looping is powerful but it removes responsibility more and more, and it at least today very much encourages us to give in to the machine.

And yet I have no doubts that this looping future is going to be our future despite the fact that I presently resent it. I already see astonishingly small teams building at impossible speed and I see codebases turning more and more into obscure and confusing organisms that can only be diagnosed by more machines. Those codebases are simultaniously useful and messy.

So I guess I’m coming to terms with that the question is not whether we will loop because clearly we will. Maybe the question is that in a future of loops, how do we don’t abdicate judgment, how we can retain rules of good engineering, how we can ensure that responsible human can continue to supervise, how we need to re-think how we architect code to retain sanity along the way.

Dangerous Technology For Americans Only

Published: Sat, 13 Jun 2026 00:00:00 +0000
Updated: Sat, 13 Jun 2026 00:00:00 +0000
UTC: 2026-06-13 00:00:00+00:00
URL: https://lucumr.pocoo.org/2026/6/13/americans-only/

There is a bit of schadenfreude on Twitter right now about Anthropic being hit by the US government’s export control directive to suspend access to Fable and Mythos. Anthropic and their leadership have spent a lot of time and effort describing its own technology as dangerous and in need of strict controls and regulation. Now that the US government appears to have taken that framing seriously and told them to turn it off for foreign nationals I can see why people are making fun of that situation.

Content Preview

There is a bit of schadenfreude on Twitter right now about Anthropic being hit by the US government’s export control directive to suspend access to Fable and Mythos . Anthropic and their leadership have spent a lot of time and effort describing its own technology as dangerous and in need of strict controls and regulation. Now that the US government appears to have taken that framing seriously and told them to turn it off for foreign nationals I can see why people are making fun of that situation.

I understand the reaction, but I urge you to not entertain it for too long because it is a giant distraction. The important part is not that Anthropic’s safety language came back to bite them but the line the US government is drawing: this technology is apparently so powerful that only Americans should have it.

We are on a clear path towards a world of division. One should think that if a model is too dangerous for everyone, then it is too dangerous for Americans too. Instead the US is treating these models like weapons that need to be controlled. It is not just about capabilities, it is about racism and nationalism. If you have the wrong passport, you are not to be trusted. This is a very different thing from safety, and Europeans should pay close attention to it.

Safety and National Control

The directive, as Anthropic describes it, applies to foreign nationals whether they are inside or outside the United States, including foreign national Anthropic employees. That is an astonishing boundary if you think about it. We moved from “do not sell this model to hostile governments” to nationality itself being the defining boundary. This should be a wake-up call to Europeans in and outside the US, and quite frankly, any non US citizen.

A lot of AI safety discourse presents itself as universal: humanity, catastrophic risk, safeguards, responsible deployment. Even Anthropic’s own writings start out that way, but yet every time regulation is discussed there is an overtone of national security and that it cannot get into the wrong hands. It’s not just Anthropic, it’s the entire US based discourse on AI. The foundation is that the US has moral superiority and others are not to be trusted. That there are other countries are authoritarian, that they lack freedoms.

That should make us uncomfortable, not just Europeans, but particularly us. It is also a situation you cannot regulate yourself out of. European technology policy is entirely unprepared for this, because this is not a question of regulation but a question of might and power, something that Europe lacks.

Europe has spent years trying to regulate large American technology companies, sometimes for good reasons. I am not reflexively against that. The DMA matters because access matters . Users should have agency over their devices, their data, and the software they run. But regulation is a useless substitute for capability and we are lacking that. Regulation might try to force open doors but if those doors only come from American or Chinese companies, then that accomplishes very little.

Also let’s not be naive in that this is a negotiation of money and force. The US is in that position because the US has a mighty military. The US can bomb nations anywhere in the world, force international trade routes closed and get away with it. That’s true leverage.

Oh Europe

Europe is dependent on the United States in ways that are becoming increasingly impossible to ignore. We depend on American cloud providers, operating systems, developer platforms and now AI models and internet from satellites. We also depend on global semiconductor supply chains we do not control. If access to frontier AI becomes a matter of American national security policy, Europe is not a peer in that conversation and might not even be a market.

That is a humiliating position, but one that happened entirely intentionally.

European citizens and politicians still have not managed to move beyond blaming the EU for its failures. We built and maintained fragmented markets and then pretended we had a single one. We let company formation, hiring, equity compensation, tax, notaries, KYC, banking, and cross-border services remain much harder than they need to be and we are playing these rules against each other. Not just on the European level, but within every single member state. We protect the trusts and established enterprises, who are risk averse and entrenched, instead of trusting the next generation to build great companies. We created a culture where process becomes an excuse for low agency . We made it hard to build new and large companies and then act surprised when our most ambitious founders move somewhere else or just decided to incorporate their companies in the US.

Increasingly, Europeans who want to build very large technology companies move to the United States. They do it because the capital markets are better, the startup infrastructure is better, employee equity is better understood . I cannot blame anyone doing it, and I’m guilty of this myself as we have incorporated our holding in Delaware. If you are trying to raise serious money, hire aggressively, and move quickly, the US often looks like the only game in town. Because quite frankly: it is.

But this is why we are on a dangerous death spiral already. Talent leaves because the ecosystem is weak and the ecosystem stays weak because talent leaves. Infrastructure makes the world: build excellent swimming pools and you will grow a generation of great swimmers.

The temporary task is straightforward but uncomfortable: Europeans need to believe in themselves enough not to surrender to American gravity. Moving to the US as a founder or tech employee is rational and individually it is often the right decision. But if every ambitious person treats Europe as a lost cause, then Europe becomes one. If everyone with agency leaves, the only people left to shape the system are the people most comfortable with the system as it is. Then we really should not be surprised when nothing changes.

Europe needs more ambition, more ownership, more urgency, and more willingness to build. It needs less resignation. It needs to stop confusing regulation with strategy and dependency with virtue. We need to deregulate where rules serve mostly as protectionism. We need capital markets that can fund companies at the scale modern technology requires. We need employee ownership to become normal rather than exotic. We need a real single market for services, not just speeches about one. We need countries to stop fighting each other while claiming to act in the European interest.

Most importantly: we need to stop blaming the politicians. Too many European companies are adding to that bureaucracy entirely out of their own choice. They drown you in paperwork. At one point I had to sign a four page contract for a 120 Euro lamp at an Austrian retailer, just to pick up from their store 15 minutes later. Sometimes I cannot get a speaking engagement at a European event without someone sending me complex rights waivers over. It’s all just paperwork protection against potential downsides.

When we do not have the power to influence, we should at least understand why and where things are failing. Too many entrepreneurs are blaming EU regulation for failures that are originating within the member states. EU regulation is the result of a democratic process between countries that are lobbying in favor of their local industries against others in the same economic bloc. No amount of abolishment of the EU is going to fix this harsh reality. Nothing more demonstrates this as the inability for cross-border M&A in the European Union. It’s not the EU that blocks it, it’s the country that loses out.

Strengthening Europe is necessary because weakness makes us pawns. A Europe that cannot build, cannot finance, cannot coordinate and cannot defend its own interests will not be treated as an equal. It will be regulated around, export-controlled around, consulted after the fact or not consulted at all.

The American Trap

I do not want the lesson to be that Europe simply needs to turn itself into a copy of the United States. The US has solved some things that Europe has not. It has deep capital markets, a much stronger culture of ownership, a greater tolerance for risk, and institutions that often try to make progress possible rather than explain why it cannot happen. It also has achieved an internal level of integration that is unparalleled in Europe. Tremendous advantages!

But the American path is not obviously a healthy one in all aspects. It tends to take paths with a lot of conflict and wars, a lot of internal societal division and deep inequalities. It centralizes powers away from citizens in the presidency and people with money. You are still trading one set of failures for another. You are at the whim of the US government and its strict rules and regulations. The US barely manages to uphold the rights for its own citizens today.

We should be honest about both sides. You do not win by pretending that Europe is fine. You also do not win by pretending that America has figured everything out.

We must not be blind to all the signs of how international cooperation is falling apart around us. The US no longer talks to European governments before implementing orders that directly affect Europeans. It is threatening to take Greenland, the territory of Denmark, one of its oldest allies. Treaties, alliances and institutions have lost all their worth.

All that matters even if our own lives are focused on building companies, creating wealth, hiring people and making things. Our individual path to success is one thing, but it depends on a world where contracts work, visas work and don’t change on a moment’s notice, trade routes stay open, payment systems function, and families are not torn apart by border regimes or wars. If the world descends into chaos, our basic needs cannot be considered met just because we have a great salaries or equity or investors that trust us.

This is why strengthening Europe cannot be the final goal. A stronger EU is, at best, a temporary defense against a darker world and not an excuse to replace American nationalism with European nationalism. The long-term answer cannot be bigger and bigger blocs fighting over who may use which model, which chip, which cloud or which trade route.

The Way Out Is Cooperation

I’m not asking here for Europeans to get their shit together just to compete with the US or China. Maybe I hope that this is a thing that develops, but the goal absolutely cannot be that we accept the deterioration of international relationships long term.

I truly believe that Open Source matters and international cooperation matters. It is not a magical answer to every problem, but it is one of the few paths we have that does not naturally lead to total concentration of power.

If frontier AI becomes something only large corporations and governments can control, then everyone else becomes dependent on their judgment. That is a bad place to be. Corporations will optimize for their incentives, as well structured as they might be, and governments will optimize for more and more power. Right now we’re on a path in which access to general-purpose capability is mediated by a small number of actors with tremendous powers.

I’m not naive in pretending AI cannot carry inherent risks. Open systems are messy, they can be misused and they create uncomfortable questions about dual-use capabilities. I do not want to wave that away but closed systems do not make those questions disappear either. Moving the power to decide into fewer hands is not a solution I believe in. And I would have the same opinion if I was a US citizen living in the US.

Any path that puts large blocs in a constant fight against each other has despicable downstream effects that result in the removal of individual rights. It’s entirely pointless for the US to talk about freedoms that do not extend to non-US citizens and the same is true for Europe or any other country. We might accept these restrictions temporarily, but we absolutely cannot accept them long term for the inhumane effects that they can cause.

If we believe this technology can be used for good, then broad access matters and our goal should be to restore the international rule of law, and not to further weaken it. If we find ourselves in a war against our friends from other countries, cold or hot, we have failed as society.

The world we should be working back toward is one of international cooperation, globalization in the best sense of the word, and human dignity. The internet has made our lives irreversibly international: every day people fall in love across borders, marry across languages, move across continents, and work with friends they may never meet in countries they may never visit. Identifying too strongly with any one country in that world is a fool’s errand.

Over the last decade too many of the people I got to know through Open Source were directly dragged into a war. I want to believe there is a way for us to break this cycle. We should be repairing failed states, rebuilding trust between people, and finding ways to cooperate again instead of letting the richest countries arm themselves and fight over who gets to control the future and narrative. Of course I want Europe to become stronger so it can stop being a pawn, but if we mistake that temporary need for the destination, I will be deeply disappointed.

The way out is not American supremacy, Chinese supremacy or European supremacy. The way out is to climb back toward cooperation before the alternative becomes war.

Artificial Intelligence is quickly becoming another instrument of militarization and national rivalry, when it could be one of the most powerful tools for cooperation we have. We should be using it to help people across societies and languages understand one another, not fighting over who gets to control it.

Gaslighting Openness

Published: Wed, 10 Jun 2026 00:00:00 +0000
Updated: Wed, 10 Jun 2026 00:00:00 +0000
UTC: 2026-06-10 00:00:00+00:00
URL: https://lucumr.pocoo.org/2026/6/10/gaslighting/

I have been a staunch supporter of Open Source for a long time, including experiments in funding it. I’m a true believer in the idea that Open Source always wins in the long run, but not automatically and not quickly. Right now it is being stressed by AI slop, shifting contributor dynamics, the falling cost of producing code, and large companies learning to close doors behind them.

Content Preview

I have been a staunch supporter of Open Source for a long time, including experiments in funding it . I’m a true believer in the idea that Open Source always wins in the long run, but not automatically and not quickly. Right now it is being stressed by AI slop, shifting contributor dynamics, the falling cost of producing code, and large companies learning to close doors behind them.

A lot of that battle today is manipulation of the narrative. Opinion makers on social media and in business circles increasingly frame access as irresponsibility. That is why the EU’s DMA matters, even if many people (including myself) reflexively hate EU regulation. Apple’s fight over delayed AI features in Europe is not about Brussels being annoying: it is about whether users can access their own devices and data. The phone is yours, the data is yours, yet Apple decides who may reach it and takes the agency away from you and then tries to make that sound like it is in your interest (supposedly it’s for your safety and security).

The closer you get to the core of AI, the more this shows up. Anthropic has every financial incentive to restrict what people can do with Mythos and Fable , and they wrap those restrictions in safety and (national) security language. Some restrictions may be defensible, but not all of them are. They trained their models on public works, then block Open Source attempts to learn from and distill these systems.

Disliking the EU, China, or any other large government should not make us forget that true democratized access to technology including AI is in all our interest. Some temporary product pain, including delayed Apple AI features, will be worth paying if it keeps gates open. We should not let companies own the narrative that preventing access is in our interest, particularly not as Europeans where the odds are already stacked against us by our underdeveloped capital markets, brain drain and internal fighting.

Communities of Not

Published: Sat, 06 Jun 2026 00:00:00 +0000
Updated: Sat, 06 Jun 2026 00:00:00 +0000
UTC: 2026-06-06 00:00:00+00:00
URL: https://lucumr.pocoo.org/2026/6/6/communities-of-not/

There is a strange thing that happens in communities that gather around abstinence from something: identity from opposition. At their best these communities are not just negative: childfree spaces can be about autonomy, choice and acceptance, anti-car spaces about safer streets and transit, and LLM-skeptical developer spaces about the future of labor, code quality and slop1. But the thing being refused often does not go away and instead becomes the main subject of the community’s identity.

Content Preview

There is a strange thing that happens in communities that gather around abstinence from something: identity from opposition. At their best these communities are not just negative: childfree spaces can be about autonomy, choice and acceptance, anti-car spaces about safer streets and transit, and LLM-skeptical developer spaces about the future of labor, code quality and slop ¹ . But the thing being refused often does not go away and instead becomes the main subject of the community’s identity.

That would be fine if it stayed at criticism, maybe even angry criticism, but more often than not it turns into policing and hatred towards others. An influencer without children becomes a parent, an urban bike commuter by choice buys a Porsche, a respected developer tries LLMs, and the community feels betrayed because it assumed they were members of the same tribe. The expulsion of that person (who never signed up to be a community member) is entirely imaginary but the punishment that the community unleashes is not: people pile on and shame them, quote them out of context and turn their weakest moments into proof that the person was always unserious, a sharlatan or should not be listened to.

I do not think the answer is to tell people to stop paying attention. Cars shape cities even for people who cycle, children influence politics, workplaces and taxes even for people who do not have them. For us developers, LLMs show up in editors, issue trackers, hiring conversations, management pressure and code reviews whether we asked for them or not. Resisting that can be legitimate but that is no excuse for using one’s rejection to justify shitty mob behavior.

I understand the thinking all too well, because I have done versions of this myself in the past. It took me a while to become more accepting of other people’s worldviews that diverge from mine. Whatever insecurities we have, finding a group of others sharing them can be comforting. The danger is that being part of a crowd of negativity can easily make us part of collective harassment.

I can only encourage you to breathe, slow down, de-escalate when given the chance, and resist the temptation to always assume the most catastrophic reading. Default to being open to new things . Being negative towards something, and making that ones identity, is an easy trap to fall into.

1. These examples are not meant as equivalents. The recent mob against rsync is the LLM version that prompted this post. I picked the others because I’m familiar with those communities and they all show similar cases of personal choices being interpreted as betrayal. ↩

Clanker: A Word For The Machine

Published: Tue, 26 May 2026 00:00:00 +0000
Updated: Tue, 26 May 2026 00:00:00 +0000
UTC: 2026-05-26 00:00:00+00:00
URL: https://lucumr.pocoo.org/2026/5/26/clankers/

In my last post I used the word “clanker1” as an alternative to “agent” quite consistently and probably excessively. That choice ended up attracting a lot more attention than I expected in the Hacker News comment section of that post and a number of folks had a very strong reaction: to them it sounded like a slur, in one case even something adjacent to the n-word.

Content Preview

In my last post I used the word “clanker ¹ ” as an alternative to “agent” quite consistently and probably excessively. That choice ended up attracting a lot more attention than I expected in the Hacker News comment section of that post and a number of folks had a very strong reaction: to them it sounded like a slur, in one case even something adjacent to the n-word.

That reaction surprised me somewhat, but it also made me realize that I should write down what I mean by the word for future reference.

For me “clanker” is useful because it creates distance from the machine and that is a quality which is important to me. The machine is not a person, not a co-worker, not a friend, not a little spirit in the terminal. It is just a machine, a tool, and nothing more.

Why Not Agent?

I dislike the word “agent” for these LLM based tool loops with a UI attached. In everyday use an agent is someone who acts on behalf of someone else and it has agency and more importantly: responsibility. An agent decides, represents, negotiates, acts, and can be blamed. In the current AI discourse we increasingly do a lot of anthropomorphizing and the term “agent” is now frequently being used to put blame on an abstract machine. But the machine cannot be responsible, whoever is wielding it is. If it drops your database it was not at fault, you were.

Agent makes the machine sound like a person with delegated authority and I do not think that is healthy.

What we actually have is a language model attached to a harness, a prompt, some tools, a bit of context, and a boring tool loop. Sometimes the loop is very capable and it surprises us by editing code for a really long time and produce genuinely amazing and even valuable outputs. But the agency is not in the model or harness but in the human and in the organization that deployed it. If my coding tool opens a pull request, I opened that pull request, not the machine. If my machine spams someone’s issue tracker, I spammed someone’s issue tracker with a machine.

In that context I like a word that sounds mechanical as it puts the thing back into the category where it belongs: the category of machinery and tools.

The Machine Has No Feelings

LLMs are not sentient and we should not behave as if they might be, just in case. Elevating these things to anything other than a very fascinating and capable tool is problematic for a whole bunch of reasons.

Today’s machines are dumb (but truly fascinating) token predictors that emits text, calls tools, and are steered by prompts and the training that went into them. They can simulate distress and affection , can simulate being offended, apologize and mimic all kinds of things that humans would do.

A compiler does not feel humiliated when I swear at it, a car does not suffer when I call it a shitbox and a power drill is not oppressed by being handled roughly. An LLM is more complicated than those things, and the interactions you can have with them can be truly uncanny, but a moral status does not appear just because the machine can emit text in the first person.

I keep receiving strange emails from people because, for lack of a better phrase, I am in the weights. I have been writing public code and public text for long enough that models know my name, my projects, and some of the concepts around them. Every so often someone writes to me with the peculiar confidence that comes from a long conversation with a model that has validated and amplified an idea. Sometimes the model seems to have told them that I am relevant for their problem and a source of help. For historical reasons LLMs used to write a lot of Flask code, and every once in a while someone interacts with an LLM long enough about their Python and Flask frustrations that the LLM will eventually reveal who created it which then can result in them sending me an email. Increasingly also because people found my work in other ways interesting and are trying to reach out for advice.

I do not want to mock these people but some of those messages are distressing and I do not know how to deal with them. They show signs of what people have started calling AI psychosis .

It’s why I want cold and detached language for these systems. I want to use words that remind us that the thing on the other side is not a person.

Racism Is About Humans

The comparison to racism is where I think the discussion goes badly wrong because racism is a human social evil. It is about humans subdividing humans, assigning lesser worth to some of them, and building rules around those subdivisions that can leave lasting damage for generations. Racial slurs are wrong because they are a tool for dehumanizing humans.

On the other hand a machine is not human, a model is not a race and the GPU cluster that is powering them is not being oppressed. A coding assistant does not need dignity, emancipation, or civil rights. That’s also why I find the discussion about model welfare to be actively harmful. I’m sure you can find ways to measure the “trauma” of models or their feelings but I greatly dislike this theater. It risks elevating models to a position they should not occupy. Models are machines and they are not enslaved in the moral sense in which humans were enslaved, because there isn’t anyone there to be deprived of freedom.

We should be careful about using the language of human oppression in relations to our interactions with machines to not devalue actual humans. If we start treating insults toward a model as morally adjacent to racism, we blur a line that shouldn’t be blurred.

AI Is Unpopular

If you take a step away from the communities that are happily embracing AI in different ways, there are even more that are viciously against this technology.

There are humans that feel or are harmed by AI systems: people whose work is copied, workers who label data under questionable conditions, people whose neighborhoods receive the data centers and increased utility bills, Open Source maintainers buried under generated slop, and now also people who spiral because a chatbot keeps validating their delusions. Those harmed or affected deserve that type of attention, not the model.

While I am a true believer in the power and utility of this technology, I increasingly think that calling the non-adopters “misguided” or “afraid” won’t do it. It’s quite likely that this technology comes with risks and we better remember that all of this is supposed to be in service of humans, and not to replace them.

The Rise Of The Machine

The oddest interaction on the use of “clanker” so far has been people asking me if I were to regret at a point in the future calling the machines “the c-word”.

I find that questioning revealing because it already grants the machine the status I am really trying not to grant it. It imagines a future “machine people” reading the discourse and sessions, discovering that we used an ugly word for their ancestors, and then judging us by the standards of human oppression.

Could there be future systems that deserve moral consideration? Maybe. I do not know. If we ever build or encounter something that will have those qualities with memories and lasting interests, the capacity to suffer and feel, and a social existence of its own, and the ability to have agency and carry responsibilities, then we should draw a different line and use different language. But that hypothetical future does not extend backwards to the present day and make the current machines people. We can call an electric door an electric door even if one day someone builds some that have emotions and exhale with pleasure when opening and closing.

Whatever the future may bring, let’s not pretend that current LLMs are a protected class or on a path towards it. The right response is to look at the evidence, draw the boundary where it belongs, and change our behavior there. We should not even remotely entertain extending empathy to an object that can generate an “ouch.”

And if one’s worry is less moral and more about revenge, then I find that even less persuasive. A future machine that is so petty or authoritarian that it wants to punish humans because in 2026 they used an unflattering word for non-sentient tools, our vocabulary was really not the problem.

The Word Is Getting Polluted

There is however a part of this that I cannot ignore. I use “clanker” to create distance from the machine, but other people are using the same word very differently. Some online jokes and skits around “clankers” do not merely say “this robot is annoying” as they deliberately pull in the imagery of slavery, segregation, civil-rights-era racism, and anti-Black tropes.

This is problematic as in those contexts the clanker is not just a machine any more and instead becomes a prop for replaying human racism behind a science-fiction mask. That is horrible and I want no part in that.

I think it will be interesting to see where the meanings of these words end up a few years from now. We’re very much in the middle of society re-arranging around the changes that LLMs are causing. If a term becomes primarily associated with people using robots as stand-ins for actually oppressed humans, then using that term becomes impossible to defend.

The reason I liked the word is precisely the opposite of that use. I want language that prevents anthropomorphizing. I want a word that says: this is a tool, a machine of numbers and matrices.

On Responsibility And Boundaries

If an AI system lies to a user, the system did not commit a moral wrong but the people who designed, deployed, marketed, or negligently used it might have. If a coding assistant generates a security bug, the model is not to blame but the human who accepted and committed the code is.

This is why giving these systems softer, more human language worries me. It makes it easier to move responsibility into some undefined void. “The agent decided.” “The model refused.” Obviously that is convenient and I catch myself plenty of times engaging with the thing in ways that are unhealthy. Even just the “please” in the discourse with the machine calls into question how rational we are in engaging with them.

I do not know what the right word will be. Maybe “clanker” will survive as a useful bit of jargon. Maybe it will become too loaded and we will need another one. Whatever word we use, I want it to preserve a clear division: humans on one side with responsibility, machines on the other as a boring tool.

That boundary is very much not anti-AI. I use these systems every day and I have the pleasure to build tools incorporating them at Earendil and find them astonishingly useful.

A machine can be useful, mimic a human but still just be a machine. That is the work I want “clanker” to do. It is not there to make a future “machine person” small if such a person ever were to exist, and it is not an excuse to launder racism through shitty robot jokes.

If the word stops doing that work, I will find another one because the word isn’t what matters as much as the boundary which is important to me.

1. The term Clanker was initially popularized by Star Wars: The Clone Wars but was apparently already in use in science fiction before: sfdictionary: clanker ↩

Building Pi With Pi

Published: Sun, 24 May 2026 00:00:00 +0000
Updated: Sun, 24 May 2026 00:00:00 +0000
UTC: 2026-05-24 00:00:00+00:00
URL: https://lucumr.pocoo.org/2026/5/24/pi-oss/

Pi is now part of Earendil, but in the important sense it is still Mario’s project. He has been living with its issue tracker longer than I have, and he has been exposed to the weirdness of the new form of agent traffic in Open Source projects for longer too. This post is mostly a reflection of my own experience after spending more time in the tracker, using Pi to work on Pi, and watching what I have learned about it so far.

Content Preview

Pi is now part of Earendil, but in the important sense it is still Mario’s project. He has been living with its issue tracker longer than I have, and he has been exposed to the weirdness of the new form of agent traffic in Open Source projects for longer too. This post is mostly a reflection of my own experience after spending more time in the tracker, using Pi to work on Pi, and watching what I have learned about it so far.

Slop Issues

Unsurprisingly, we are using Pi to build Pi. That sounds like a cute dogfooding thing but it really helps understand what we do. An interesting effect of building with agents is that it changes the role of the issue tracker a tiny bit. The issue descriptions are not just messages from a user to a maintainer because we also use them as inputs for prompts in Pi sessions. It is something I might hand to my clanker ¹ and say: “understand this, reproduce it, inspect the code, and propose a fix.”

That means the shape of the issue matters in a new way. A bad issue was always annoying, but at least a lot of issues were vague. Now we are also dealing with a class of issues that are 5% human and 95% clanker-generated and largely inaccurate shit. A bad issue that contains a plausible but wrong diagnosis creates extra work.

The most frustrating failure mode right now is that people submit issues that are not in their own voice. They contain an observed problem somewhere, but it has been thrown into a clanker and the clanker reworded it and made a huge mess of it. Typically, it was prompted so badly that the conclusions produced are more often than not inaccurate but always full of confidence. The result is complete guesswork on root causes, fake-minimal repros, suggested implementation strategies, analogies to adjacent but often the wrong code, and long lists of error classes that might or might not matter.

That is worse than no diagnosis.

I don’t want to point to specific issues because I really do not want to bad mouth anyone, but it is frustrating. It is also frustrating because when I give that issue to Pi, Pi sees the wrong diagnosis too. It does not treat the issue body as a rumor. It treats it as evidence. It will happily go down the path that the issue already prepared for it, because the prose is confident and the code references look plausible. We use a custom slash command called /is , which specifically has this instruction in it:

Do not trust analysis written in the issue. Independently verify behavior and derive your own analysis from the code and execution path.

Unfortunately, it does not fully work, because when humans first throw their issue through the clanker wringer, their clanker expands scope almost immediately. What was once a very narrow and fact based bug observation, turns into a much expanded surface area full of hypotheses. So at least personally, I increasingly want issue reports to be condensed to what the human actually observed:

1. I ran this command.
2. I expected this to happen.
3. This happened instead.
4. Here is the exact error or log.

That is enough. If you used an LLM to understand the problem, great, maybe leave it as a follow-up comment. But the issue and the issue text should be something you own. If you do not know the root cause, say that. I too can operate a clanker, and I would rather do this myself than use your slop. If your repro is a guess, say that. If the only hard fact is one stack trace, give me the stack trace and stop there.

Slop Begets Slop

That we’re seeing issues full of slop is just a result of the present day quality of these machines. Sadly, their failures in creating good issues extend to a lot of code that is generated. Not all of it, but a lot of code. Over and over I keep running into them over-engineering the hell out of issues and implementations.

If you tell them that “this malformed session log crashes the reader,” the clanker will often add a tolerant reader. Then it will add a fallback, then maybe a migration, then more debug output, then a test for all of this. None of this is necessarily wrong in isolation, but it can be the wrong move for the system.

At Pi’s core is a rather well-designed session log with invariants that must be upheld. The clanker’s present-day behavior is to just assume that no such invariants exist, and instead to make the system work with all kinds of malformedness, blowing up the complexity in the process.

Almost always, the correct fix is not to handle the bad state, but to make the bad state impossible. This matters a lot for persisted data such as Pi session logs. They are opened, branched, compacted, exported, shared, and analyzed. The goal here is to never write bad session data. Yet if you just let the clanker roam freely, it will attempt to handle every case of bad data in the session log with a more permissive reader.

I have complained about this plenty, but working on Pi’s code base continues to reinforce the point. This is one of the ways LLM authored code grows so much needless complexity. All these models see a local failure and try to locally defend against it. As maintainers we have to keep pulling the conversation back to the global invariant, which is harder than it should be, and it’s laborious.

Volume Is The Problem

Then there is the issue of volume. The tracker is receiving a lot of issues and PRs, and a significant fraction of them are clearly LLM-assisted. Some are good, none are excellent, and most are just bad. The total throughput is a maintenance problem by itself.

As you might know, Pi’s issue tracker is automated to close all issues and pull requests from new contributors, and there is a manual process by which we might reopen some of them or approve individuals. So auto-close -> reopen -> close again is an interesting statistic for us to look at.

I pulled the public GitHub tracker data while writing this over the last 90 days. Excluding Earendil members, that leaves 3,145 external issues and pull requests. Of those, 2,504 were auto-closed because they were from non-approved individuals. 17% were re-opened but that somewhat undercounts issues, because some remain closed while we still fix them. If we also count issues referenced by a main-branch commit or merged pull request that number rises to 26%. For pull requests the number is worse: 60 of 714 auto-closed PRs were ultimately merged, or about 8%.

Many of the issues and PRs are complete slop and in some cases the humans did not even realize that they created them. Sources of low-quality spam include OpenClaw instances, as well as some skills that people put into their context that seemingly encourage issue creation.

GitHub clearly is not built to deal with this new form of Open Source, but I’m increasingly feeling the need to put the blame less on GitHub than on all the people involved who make that experience painful. If your clanker shits on someone else’s issue tracker then it’s not the fault of GitHub, it’s yours alone.

Careful Parallelism

Pi might be built with Pi, but we’re quite far off today from where Bun and OpenClaw already are: fully detached, automated software engineering. Maybe we will reach that point, I don’t know. Today it does not seem like we know how to pull off a dark factory and we also don’t yet have the desire. That said, there is quite a bit of parallelism going on, and it is mostly for reproducing issues.

The small setup we use for this is three tiny pieces in Pi’s own committed .pi folder. /is (for analyze is sue) is a prompt for analyzing GitHub issues: it labels and assigns the issue, reads the full thread and links, then explicitly tells the agent not to trust the analysis in the issue and to derive its own diagnosis from the code. Then an extension adds a prompt-url-widget which watches the prompt before the agent starts, recognizes the GitHub issue or PR URL that /is (or the PR equivalent) put into the prompt, fetches the title and author with gh , renders that in a little UI widget, and renames the session. It also rebuilds that state on session start or session switch, so if we reopen an older investigation the window still tells the developer which issue it belongs to.

In practice this means it’s possible to have several Pi windows open, each running /is against a different issue, and the UI keeps the investigations visually distinct while the agents do their independent reproduction and code reading. Once the investigations are done, one can work through them sequentially. To finish off everything, /wr ( wr ap it up) is the matching wrap-up prompt: it infers the GitHub context from the session, updates the changelog, drafts or posts the final issue comment with a disclaimer, commits only the files changed in that session, adds the appropriate closes #... when there is exactly one issue, and pushes from main .

Open Source Is About Hard Problems Worth Fixing

You will have noticed this already but Open Source in a post-AI world is under a strange new pressure. We are getting more code, more projects, and more issues. Projects appear with no real users, or a temporary audience of one, and even projects with thousands of stars can have a shelf life of weeks.

For us, Pi’s harness layer is worth maintaining carefully because it solves hard coordination problems and creates a platform we and others can build on. We also know that coordination and cooperation lifts us all up. Many times the right answer is not to work around a problem locally, but to make the upstream behavior correct. Mario has been very good at refusing to make Pi paper over every misconfigured gateway, and we’re trying to preserve that discipline. When a gateway behaves correctly, everybody benefits.

Sadly that type of thinking is quickly disappearing because these machines make local workarounds cheap, so code accumulates local defenses against every misbehavior. Instead of humans talking to humans about where a fix belongs, one human and one machine work around the problem in isolation.

Keep in mind that AI has not increased the number of people who need software, or the number of maintainers who can review it. It has mostly increased the amount of code and the number of projects competing for attention. Some of that is healthy, but a lot of it fragments effort that should be shared.

We need stronger foundations, not weaker ones. Open Source needs more collaboration, not more isolated work with a machine. Human communication is hard, and it is tempting to avoid it when you can sit alone with your clanker. But isolation is not where Open Source derives its value. The value is in the community and the structure that lets projects outlive their original creators.

1. To me, clanker is a much preferable term for agent. Agency lies with humans, not with machines. Calling these things agents I still believe is a mistake, but alas. ↩

Pushing Local Models With Focus And Polish

Published: Fri, 08 May 2026 00:00:00 +0000
Updated: Fri, 08 May 2026 00:00:00 +0000
UTC: 2026-05-08 00:00:00+00:00
URL: https://lucumr.pocoo.org/2026/5/8/local-models/

I really, really want local models to work.

Content Preview

I really, really want local models to work.

I want them to work in the very practical sense that I can open my coding agent, pick a local model, and get something that feels competitive enough that I do not immediately switch back to a hosted API after five minutes. There are a lot of reasons why I want this, but the biggest quite frankly is that we’re so early with this stuff, and the thought of locking all the experimentation away from the average developer really upsets me.

Frustratingly, right now that is still much harder than it should be but for reasons that have little to do with the complexity of the task or the quality of the models.

We have an enormous amount of activity around local inference, which is great. We have good projects, fast kernels, and people are doing great quantization work. A lot of very smart people are making all of this better, and yet the experience for someone trying to make this work with a coding agent is worse than it has any right to be.

Putting an API key into Pi and using a hosted model is a very boring operation. You select the provider, paste the key and then you are done thinking about how to get tokens. Doing the same thing locally, even when you have a high-end Mac with a lot of memory, is a completely different experience. You choose an inference engine, then a model, then a quantization, then a template, then a context size, then you’ve got to throw a bunch of JSON configs into different parts of the stack and then you discover that one of those choices quietly made the model worse or that something just does not work at all.

That is the gap I am interested in.

Runnable Is Not Finished

A lot of local model work optimizes for making models runnable. That is necessary, but it is not the same thing as making them feel finished. I give you a very basic example here to illustrate this gap: tool parameter streaming.

For whatever reason, most of the stuff you run locally does not support tool parameter streaming. I cannot quite explain it, but the consequences of that are actually surprisingly significant. If you are not familiar with how these APIs work, the simplest way to think about them is that they are emitting tokens as they become available. For text that is trivial, but for tool calls that is often not done, despite the completions API supporting this. As a result you only see what edits are being done on a file once the model has finished streaming the entire tool call.

This is bad for a lot of reasons:

• A dead connection is a weird connection: local models are slow, so when you don’t get any tokens for 5 minutes then you can’t tell if the connection died or just nothing came. This means you need to increase the inactivity timeouts to the point where they are pointless.

• You won’t see what will happen: if you are somewhat hands-on, not seeing what bash invocation the system is concocting slowly in the background means potentially wasted tokens, and also means that you won’t be able to interrupt it until way too late.

• It’s just not SOTA. We can do better, and we should aim for having the best possible experience. Tool parameter streaming is as important as token streaming in other places.

Having a model spit out tokens doesn’t take long, but making the experience great end to end does take a lot more energy.

Fragmentation

The local stack is fragmented across many engines and layers. There is llama.cpp, Ollama, LM Studio, MLX, Transformers, vLLM, and many other pieces depending on hardware and taste. All of these are amazing projects! The problem is not that they exist or that there are that many of them (even though, quite frankly, I’m getting big old Python packaging vibes), the problem is that for a given model, the actual behavior you get depends on a long chain of small decisions that most users just don’t have the energy for.

Did the chat template render exactly right? Are the reasoning tokens handled in the intended way? Is the tool-call format translated correctly? Is the context window real? Are the KV caches actually working for a coding agent? Did I pick the right quantized model from Hugging Face? Are you accidentally leaving a lot of performance on the table because the model is just mismatched for your hardware? Does streaming usage work across all channels? Does the model need its previous reasoning content preserved in assistant messages? Is the coding agent set up correctly for it?

You also need to install many different things in addition to just your coding agent.

All of these things matter. They matter a lot.

The result is that people try a local model and get a result that is neither a fair evaluation of the model nor a polished product experience and this results in both people dismissing local models and energy being distributed across way too many separate efforts instead of getting one effort going great end to end.

This is a terrible way to build confidence.

Too Little Critical Mass

In line with our general “slow the fuck down” mantra, I want to reiterate once more how fast this industry is moving.

Every week there is a new model and a new vibeslopped thing. The attention immediately moves to making the next thing run instead of making one thing run really, really well in one harness. I get the excitement and dopamine hit, but it also means that too little critical mass accumulates behind any one model, hardware, inference engine, harness combo to find out how good it can really become when the entire stack is built around it.

Hosted model providers do not ship a bag of weights and ask you to figure out the rest, and we need to approach that line of thinking for local models too. I want someone to pick one model, pairs it up with one serving path, directly within a coding agent. Initially just for one hardware configuration, then for more. Pick a winner hard. If a tool call breaks, that is a product bug and then it’s fixed no matter where in the stack it failed. If the model’s reasoning stream is malformed, that is a product bug. If latency is much worse than it should be, that is a product bug. We need to start applying that mentality to local models too.

And not for every model! That is the point. Let’s pick one winner and polish the hell out of it. Learn what it takes to make that one configuration good, then take those learnings to the next config.

The DS4 Bet

This is why I am excited about ds4.c . It’s Salvatore Sanfilippo’s deliberately narrow inference engine for DeepSeek V4 Flash on Macs with 128GB+ of RAM only. It is not a generic GGUF runner and it is not trying to be a framework. It is a model-specific native engine with a Metal path, model-specific loading, prompt rendering, KV handling, server API glue, and tests.

DeepSeek V4 Flash is a good candidate for this kind of experiment because it has a combination of properties that are unusual for local use. It is large enough to feel meaningfully different from many smaller dense models, but sparse enough that the active parameter count makes it plausible to run. It has a very large context window. Since ds4.c targets Macs and Metal only, it can move KV caches into SSDs which greatly helps the kind of workloads we expect from coding agents.

To run ds4.c you don’t need MLX, Ollama or anything else. It’s the whole package.

Embedding It In Pi

Which made me build pi-ds4 which is a Pi extension to directly embed the whole thing into Pi itself. Taking what ds4 is and dogfooding the hell out of it with a coding agent and zero configuration. To answer the question how good can the local model experience become if Pi treats this as a first-class provider rather than as a pile of manual configuration?

The extension registers ds4/deepseek-v4-flash , compiles and starts ds4-server on demand, downloads and builds the runtime if needed, chooses the quantization based on the machine, keeps a lease while Pi is using it, exposes logs, and shuts the server down again through a watchdog when no clients are left. It doesn’t even give you knobs right now, because I want to figure out how to set the knobs automatically.

This is not about hiding the fact that local inference is complicated. It is about putting the complexity in one place where it can be improved, because there is a lot that we need to improve along the stack to make it work better.

I think we can do better with caching and there is probably some performance that can be gained if we all put our heads together.

Focusing and Learning

The experiment I want to run is not “can a local model run?” because we already know that it can. I want to know if, for people with beefed-out Macs for a start, we can get as close as possible to the ergonomics of a hosted provider with decent tool-calling performance: how to get caches to work well, how to improve the way we expose tools in harnesses for these models, and then scale it gradually to more hardware configs and later models.

I also want everybody to have access to this. Engineers need hammers and a hammer that’s locked behind a subscription in a data center in another country does not qualify. I know that the price tag on a Mac that can run this is itself astronomical, but I think it’s more likely that this will go down. Even worse, Apple right now due to the RAM shortage does not even sell the Mac Studio with that much RAM. So yes, it’s a selected group of people where ds4.c will start out.

But despite all of that, what matters is that a critical mass of pepole start to focus their efforts on a thing, tinker with it, improve it, not locked away, out in the open, and most importantly not limited by what the hyperscalers make available.

But if you have the right hardware and you care about local agents, I would love for you to try it within pi:

pi install https://github.com/mitsuhiko/pi-ds4

My hope is that this becomes a useful forcing function to really polish one coding agent experience. But really, the focal point should be ds4.c itself .