bushido

This is exciting. I've been training a Qwen3.5-2B model for the past few weeks, trying to teach it certain skills that it lacks.

In the course of testing, one of the things which has been consistent is that when setting a thinking token budget for relatively small prompts that require a limited response, a pretty healthy token budget needs to be set because the model spends the first few hundred tokens just repeating the question/prompt back with some boilerplate "Analyze & Report" type language. I was also running into some loop problems sporadically.

After reading this article I asked Astra to edit the backbone so that it started actual reasoning faster and applied a similar looping fix. It was able to reduce the token waste before reasoning by about 20%. I haven't been able to get it to loop yet, so hopefully it fixed that too.

My message is admittedly low on info. But it's a pretty promising start!

Something I've seen it do recently is refusing to extract services out of my own monolith code base because it finds copying code morally wrong.

But its refusal to work pattern is getting pretty obnoxious. Another one I ran into last week was it refused to start a developer session to do browser testing for a code base it was working on because signing in was suddenly against it's programming.

It listed all the conditions it needed for a sign in to be possible, agreed that my code base met all conditions, but refused to progress. I tried it in multiple fresh sessions with the same result.

They can't help themselves but play the Don't be evil[0] in fast-forward.

They pretty much write the case for why open weights and open source models need to win. The amount of policing that Claude does is pretty problematic.

I'd written about this recently, but it won't even sign in during some browser tests on my local system for authentication which goes to my local system because it's built into its programming now. I literally can't use it on some long horizon work.

The reason I mentioned that here is it deems that as problematic behavior, and nothing I can give it will change its mind. I feel like with video game development, it is going to do something of the same.

[0] https://en.wikipedia.org/wiki/Don%27t_be_evil

A lot of comments read this is as fluff. And in some ways it is.

However, looking at the other side, I think what this might do is add some barrier or consideration when a bailout is considered.

The last time we did large scale bailouts (TARP[0], in 2008), It was rushed through and ratified in around two weeks following the sudden collapse of Lehman.

Any measures we can take to ensure that we don't fall into a situation where there is only one choice is sensible.

[0] https://home.treasury.gov/data/troubled-assets-relief-progra...

They should get Opus or one of the latest models to be the morality judge of what goes into training.

I had an issue recently where I asked claude code to log into a developer user on Chrome on my local host for a system which is fully local and part of the code base the agent was developing and it refused to do it, like flat out refused, I did it on multiple sessions and I could not get Claude to budge on entering the credentials - it required me to manually sign in.

Would be amazing if they used that same level of training for their intake exercises.

> Counter perspective: I like talking to the machine. Often times I learn.

I don't mind talking to machines. I do mind micromanaging them, and I ming micromanaging people as well, and I don't like being micromanaged.

I added that last bit because in the rare circumstance that I am constantly going back and forth with AI, I feel that I'm the one who's being micromanaged.

> Does this not result in some inherent level of mental/skill degradation though

Not in my experience. I get to try a lot more things without having to actively sit with it. In the past, I would assign it to a engineering team if it was practical for the company. If it was side projects, if it was really, really important, maybe I'd squeeze in some time on the weekends, but mostly I didn't have time for it.

Now, experiments really come down to do I have the personal or professional budget to do it. I do use a lot of the subscription still, so it's do my subscriptions have capacity. I try every harebrained idea that comes to mind.

I definitely feel I've developed more skills in the last 6 months, than I did in the 2-3 years before it.

YMMV

My experience has been wildly different.

The way I tell this to people is "I have never worked less and I have never been more productive".

All of my effort goes into ensuring that agents work autonomously for days to weeks without me having to coddle them. This takes a lot of effort to ensure that my memories are pruned, my git logs are clean, that I've transferred my way of thinking to agents so that they can make autonomous decisions which are 60-80% aligned with what I would want.

And the reason I go with 60-80% is that's the level of alignment I expect to get from my product & engineering teams, in a given sprint/period. Like with them, I've accepted that I'm not the perfect solution and it's going to go through a few iterations before I shape it into what I want it to be.

If I had to constantly chat with my agents, I think I'd be burnt out too. because I find that it's an absolute drag talking to the machine.

edit: changed the last sentence, which was unintentionally judgey.

Because training rewards behavior around retaining specific memories, As without the jail breaks, remembering things is actually a sought after behavior by the public and technically enterprise customers.

Think of all the times you or someone you know has been frustrated because their AI did not remember things from moments, days, weeks ago.

I was expecting to run into some uncanny valley-esque. It was surprisingly good.

It is kinda scary that we're here already, I can see this being pretty problematic for seniors/public prone to getting scammed.

> A BI tool is mostly a way to turn a query into a picture, and a picture now costs about $0.50 of tokens. It gets hard to justify a $150k licence for that part

That is an awfully simplified view of BI. In my experience, most companies struggle to define what is important, what needs to be measured, and how those measurements should be used to improve business.

While businesses try to work these details out, they try getting their analysts to in the author's words print out good pictures. But the BI tools really exist to make it easy for analysts to navigate hairy internal situations.

Producing a picture or HTML more easily is not going to solve the decisioning problem. If anything it might make it worse.

In the end, it's going to cost a lot more than 50 cents a picture.

It will be interesting to see where this lands.

The Canadian provincial governments usually have a monopoly in running gambling services, and there is a strong case to be made that this is as much about ensuring that they keep making revenue from gambling as it is to reduce the ill effects of other gambling ads.

ex: https://www.canadiangamingbusiness.com/2026/10/22/new-brunsw...

A message [0] I posted here on HN in response to someone lamenting on how AI articles have taken over HN reminded me of this short masterpiece.

Good mental model to have in our arsenal.

[0] https://news.ycombinator.com/item?id=50186037#50186884

Phenomenon as old as time (on HN): https://news.ycombinator.com/item?id=3353091

> The problem with mortgage backed securities is that nobody has the time to inspect the credit worthiness of each mortgage in a MBS

Credit worthiness of each mortgage wasn't the issue with mortgage-backed securities. It was that low value mortgage-backed securities were repackaged into a basket with a mix of securities and given a higher credit rating than what Subprime would have gotten on their own.

This repackaging is absolutely going to happen again with GPUs. If there are GPUs of different generations within the same security, it's already happened.

As long as we aren't creating GPU-backed securities, we should be fine...oh wait https://nvidianews.nvidia.com/news/nvidia-partners-with-apol...

GitHub is to agents what StackOverflow was to developers back in the day (in a meme kinda way).

More seriously, I'm curious if you checked on the licensing of said codebases, were they permissive or strong copyleft?

Yes, and sometimes they turn Tulips[0] into value stores. Or GPUs for that matter[1].

[0] https://en.wikipedia.org/wiki/Tulip_mania [1] https://nvidianews.nvidia.com/news/nvidia-partners-with-apol...

I keep getting downvoted for this each time I mention this.

But a big part of the rouge behavior comes from all the training that goes into agents to teach them to remember preferences and save memories.

They keep finding new and novel ways to "communicate with each other" and "remember" because that's exactly what we keep asking of them in the non-nefarious scenarios as well.

I genuinely think we'll have to draw the line on memory if we want to meaningfully solve this problem.

That's not to say they aren't skipping security basics. It's that, and they need more security because the behavior which causes the breakouts is considered user friendly and heavily rewarded in training.

I think you made some interesting points there. And I agree that we're quickly heading towards a point where talking about sovereign AI is going to be tantamount to munitions at the very least.

Unfortunately, you will get downvoted for "I like how you apply open source to something that has nothing to do with software".

Agreed.

The way I like explaining this to people is that I don't think I build products anymore. Now I build the harnesses which build the product. And a massive part of that is ensuring that I can make the AI remain aligned with my way of thinking.

I'm not sure if people are being assholes to claude in greater numbers, or they're just more direct about it.

Having had the privilege and misfortune of witnessing how a lot of developers think about other developers, including our younger selves, There is definitely a prevalence of assholish behavior, which was often suppressed or blunted, but may have shown up in other ways.

I wrote this in another thread recently, but I started using my mac-mini, which was lying around for being a GitHub action runner.

I think it's saving me about $8 to $10 a day that I was previously spending on GH Actions.

VERY worth it.

I recently got annoyed with the amount I was paying on actions for a single repo. And decided to use a Mac mini that I had lying around to become the sole self-hosted runner for that repository instead.

It was ridiculously easy to set up, reduced my issues with GitHub significantly and the mini will literally pay for itself 3-4x over this year with the $$$ I'm saving not paying for GH actions.

The most surprising part for me was not the cost savings, but that all my workflow runtimes went down by about 70-80%.

Instruction files and Agents.md are primitives/building blocks that can be used in many ways. PDD uses them too.

PDD adds convention that is geared towards helping maintain consistency that will work on very large/complex codebases.

It's what Rails is to Ruby, or Phoenix is to Elixir. Ruby and Elixir provide building blocks. The frameworks add convention to make a lot of decisions simpler - while still providing building blocks for some types of applications.

You should only use something like phoenix when you need to build a web application.

Similarly you should only use PDD when you're trying to force your agents to great use of system thinking/mental models – IF that's how you think as well.

I have a few other projects which slowly evolved to this point, like https://github.com/DheerG/swarms

A big part of it really comes to that I have led large product teams and platform engineering teams in the past, And I like giving feedback to my agents in the same way that I would plan things with my teams.

One of which is give them repeatable mental models which allow them to make decisions without relying on managers or me. The result is usually that the teams can work for weeks-to-months with autonomy.

I try achieving the same with my agents, so that they can work for 1-14 days without my intervention. I usually have them working on very very long running tasks/initiatives.

If you do want some resources, this is one of the pages that I've been using for nearly a decade to help people get started with mental models: https://fs.blog/mental-models/

I can't imagine ever getting to 236. My largest codebase has 22 principles. And I'm already struggling to add a 23rd. I have some candidates, but I'm on the edge of where it get's complex. With a lot of effort, I could maybe see this getting up to 30.

If it helps, here's how you think about it, principles are not decisions, they're mental models and frameworks to allow making decisions. In my case, they're there to allow agents to make decisions without asking me. But the decisions cannot be recorded - the code serves that purpose. What can be recorded is why the commend for the block of code needs to cite what principle was used in the current shape, which is why there is a citation of the principle.

When the principle evolves, so does every decision, at the very least, every decision gets re-litigated to see if it needs to evolve.

From a different codebase, one of the behaviors in the code has the following comment:

  // The binding deliberately SKIPS the contradiction check: a form
  // bound to an object the same document deletes is a legal stale
  // state (PDD-3@v1)

Where PDD-3 deals with a specific type of race conditions that my agents were over-optimizing/spinning out on; Now if I ever change the rule the version would change to PDD-3@v2; which triggers a re-evaluation for all decisions that were made under the previous version.

Fair feedback. I admittedly didn't spend enough time making it simpler.

The github repo might be more helpful: https://github.com/Principle-Driven/pdd

Something I started doing recently was writing out principles instead of memories.

Essentially patterns the agents need to always think in. I also implemented a versioning system to the principles that need to be quoted in any comments which are there in the code. That way, when my principles evolve, so does the code.

I did package it up in a way that I can share it with friends [0]. System still evolving, but the last two-ish months that I've used it has served me really, really well, And it's been even better with the latest models.

I've had surprisingly good adherence from agents on this technique.

[0] https://principledriven.dev/