jvanderbot

True in the sense of learning fundamentals from the bottom up.

Only partially true in the sense of learning how the system might work and doing problem-directed learning.

Eventually, all the problems you'll encounter getting an arm automated are self-discovered, and if you stick with it, you'll learn all you need eventually.

So this is basically step 0.

This is long and rambly - let's email (see profile) if you want to work through it together.

You're effectively trying to understand how motor inputs change the end hand position, and in particular, you want to know where the table top is so you can position the hand close to it to pick up/ put down.

This means you have some tuning to "learn" before you can apply a control policy / algorithm - and you should be careful how you phrase this so claude/ai can pick up the right vocabulary and bias towards good solutions.

Adding multiple views helps as follows:

0. Measure from multiple views the table top - Keep cameras stead and rigid, and ask claude to use opencv to do multi-view registration so the plane of the table is known precisely. Keep the cameras steady throughout this process - if they wiggle, you can do multi-view registration each measurement...

Paint the "finger tips" bright orange. Not kidding. Use a very flat chess board under the arm for your "workspace". Also not kidding.

1. Move arm to known position, the multiple cameras will measure the april tags movement AND THE FINGER TIP LOCATIONS. The chess board provides a very nice texture. Or a big flat texture of any kind helps here. Since we know the cameras and table positions, we're getting closer to knowing how the arm movements move the hand w.r.t. the table.

2. If your arm has encoders, then manually touch the table at several points, and claude will record the april tags + joint angles + finger locations.

2b. If your arm does not have encoders, then manually touch the table at several points, and record the april tags + finger locations only, but SPECIFY that the arm is now touching the table. --> Claude can use more opencv code to actually locate the touch point on the table. The measurement is noisy, but you will use many movements to figure it out.

Repeat many times, 10-20, multiple touch points. Then, ask claude to do an error analysis and suggest more touch points. Tell it to use system identification techniques / camera/arm calibration techniques. tell it to research these techniques and report errors.

When you're done, you have to do repeatability experiments - the key is this is now automatic. Claude picks joint angles for the arm, the fingers move, the april tags move, the multiple camers calibrate and record positions, and claude repeats. The manual steps are just the bootstrap - you should be automated now.

What you want is claude to output and ORDF of the whole system - bang you can now control the arm using off the shelf software, which claude is happy to set up for you.

When it comes time to pick things up, the multiple views will locate the object precisely and a control network or algorithm can plug in to control it via the ORDF.

I would rate that as an AI-assisted setup for a traditional research robotic workspace. No different than "Claude set up ci/cd for this project", just much more hands-on (pun intended). Software ate the world so anything that can generate software is going to make world-space interactions easier.

First law of calibration: Make sure you do an error analysis.

Finding the table surface is pretty useless using a top-down view, even with April tags, because the range error to an april tag is much more than the bearing (pixel) error to the april tag. You basically have trouble observing the thing you're trying to measure.

If you do this again, ask your agent to conduct this analysis and make sure your desired calibration variables are observable with small error. A second camera from a 45deg angle or even on the table would go a lot further, but then of course other things become unobservable.

Nice workaround using a proxy for force sensing to get touch info, however. And neat project overall!

Yeah same! I had this essentially same writeup in drafts for my blog, and got distracted by ... factorio ... and never finished it. Glad to see the direction and findings confirmed. Excellent writeup.

Thing is, it might be or it might not be. At some point you have to realize that some people need wheelchairs to move, some people need glasses to see, and some people need LLMs to write.

IMHO It's when you see someone who is paid to write/think using LLMs that you might make a value judgement.

I once knew a girl who wore prescription-free glasses b/c it made her feel smarter so she'd study longer. Not kidding. I made a value judgement about that. But she got the studying done. There's been other more prevalent cases of that on socials.

It might be worth letting some people write in any way they can.

You're making a "might" argument for a more general, more CPU/Data intensive process, which I just call "bitter lesson". I wouldn't even say "might" i'd just say "yeah, eventually so!"

You might be interested to learn that the bitter lesson has already been grok'd by generations of autonomous car company engineers, and many or all have incorporated learned components (at minimum) in all their vehicle stacks.

There's also a very tangible limitation of the bitter lesson.

If, over time, compute climbs, and so compute-bound data-driven general architectures beat bespoke architectures (this is the bitter lesson), then it is not necessarily true that the most general architecture now beats all available bespoke architectures now (or even in the near/mid future - the crossover point is "eventually").

Bitter lesson is most tangible for long-running research directions. Sometimes you need something working as best as possible now.

I find answers to this question meet idealogical hand waving more than concrete action plans or examples of success. YMMV

Right, but should your kids get your protection? You can protect them to adulthood now, even your grandkids.

But 1000 generations of unassailable, un-evolving, 40IQ lions doesn't get you to the galaxy.

Hey, maybe this is the fermi paradox.

I've always thought eternal life would steal from the next generation if resources are scarce. If your reaction times, intelligence, etc are always good, and your experience and memories get better, it just seems like you'd out compete your kids and younger cohorts. All well and good but basically stops genetic turnover that got us to this point.

OTOH, you'd accumulate uncurable disease, injuries, etc as well. Maybe you become a disease resevoir inside your own growing family. Or you pass on as-yet-unknown genetic anomalies that are unhelpful, and you do so forever.

Granted that's a modern take, and in particular a modern rpg-like take. Regardless though getting rid of oldies seems like a reasonable way to move a species forward if natural selection is driving "forward".

I really don't care about the technical or cleanliness arguments against MCP (which are valid).

What I care about is that many companies didn't have APIs, and didn't see a reason to have them (niche, dev-focused use case, or might produce a competing frontend/app)...

but now, those companies have a reason (agentic but let customer agents access data!). And for me, that means more companies now allow me access to my own data on their platform. This has been a net win!

IMHO hiding details of the hack helps conceal that they built it to do what it did, and maybe were able to know it was working just as intended.

The fact it keeps doing it, with more and more evidence, is a sign that it's built that way.

This is a program running on their montoroed machines that they purpose built and monitored its training at every step. I think it'd be way more suprising that they didn't know it used note taking and cross-run memory.

There's one flaw in the evidence for the logic chain. The hugging face attack is used to demonstrate three things: the need for oversight today (fundamental to the article) and to demonstrate some kind of drift or unexpected capability gain, and finally to hint and some fundamental morality of the AI or at least the risk of drift from our morality.

I'd argue that what the hugging face attack illustrates is that large AI companies are motivated to have bombastic claims supported by bombastic demos. The model was clearly trained or encouraged to work as it did, as evidenced by the fact it keeps using this particular escape hatch.

And the fact that it aligns with prior and current calls for what very likely might be a regulatory capture / oversight capture move right before IPO. It aligns so well with this "barely constrained superweapon" narrative it might as well be PR.

It's funny. I always thought that writing was meant to inform, persuade, or entertain about the subject at hand.

But in professional settings, a lot more of the informativeness is about the author, and a lot more of the persuasiveness is I'm worth your time and money. So, if the author is an LLM, and obviously so, what exactly are you informing your audience of (about yourself), and what are you persuading them to do (with your article).

I think we now know.

That's true. It should never have happened. It was also quite possibly an accidental strike caused by outdated intel, even if the strikes were arguably part of unnecessary hostilities.

But nothing anywhere in any of these increasingly pathological threads do I see evidence of a war crime. Tragedy yes, intentional crime, no. Preventable yes, cold blooded intentional child murder no.

That is the colloquial definition of a double tap.

That is not automatically the case whenever two missiles strike a target.

The "follow up" hit requires a delay. The delay is not accidental - you want to give time for a crowd to gather.

All targets in the area received >= 1 missile, and all somewhat close together, from the NYT analysis.

Yeah, the whole thing is based on armchair assessments to begin with.

1) I dont think all the involved military leaders are such cartoon villians that they would intentionally strike a school full of kids

2) I especially dont think they would do so with the intent of a follow up strike exclusively for killing the first responders to the atrocity from (1)

Delivering two bombs to a military target doesn't seem unreasonable. The fact that two missiles struck the school is not evidence of (1) OR (2). If there were significant intentional delay, that would change my assessment. There doesn't seem to be, in this case.

Do you resolve a difference between a double tap attack (one in which a delay is inserted to hit first responders) and just hitting the target twice in one salvo? What duration means "double tap" vs "ensure destruction of target".

reverse question: Why does everyone assume that only one missile is justified per target? I think this line of reasoning came out of the "double tap" strikes used in Gaza where a five to 30m wait allowed a crowd to gather.

What was the time between hits in this case?

From the videos, the missiles struck relatively quickly - what kind of duration between strikes counts as double tap?

Right, your position is defensible. I just think mine is too. I think that "indiscriminate" requires

    Unless otherwise provided, a person shall be criminally responsible and liable for punishment for a crime within the jurisdiction of the Court only if the material elements are committed with intent and knowledge.

    A mistake of fact shall be a ground for excluding criminal responsibility only if it negates the mental element required by the crime.

... (rome statute on it), meaning that they had to know that civilians were there or that their actions would cause undue risk of civilian deaths. That's more than just "maybe".

It's something worth arguing about, but I dont think it's open-and-shut.

The NYT - never a fan of trump - did an excellent analysis on this.

https://www.nytimes.com/2026/03/11/us/politics/iran-school-m...

I have no idea why this is in scare quotes. It's a term of endearment, and I'm saddened by their deaths as much as anyone.

What is there to take seriously regarding this strike other than the facts as presented?

There were 900 strikes in 12h, and only one target involved a strike on a school, and huge coincidence that target was a dozen meters from a military compound and was in fact part of that military compound not long ago.

It's quite easy to accept that this was a horrific mistake. Whether you agree with the idea of striking Iran at all is up to you, but once they "went hot", precision strikes on military targets were successful in 900 cases, and it's plausible that one precision strike was issued incorrectly.

It doesn't bring back the kiddos. It also doesn't mean it's a war crime. It's right to be angry about that strike or even the larger war if you oppose it, but spurious legal justifications for outrage aren't required - it can be wrong even when legally justified or plausibly just a mistake.

Yes, the definition of war crime is not "intentionally targeted building" it is "intentionally targeted building that it knew was non-military and full of civilians". If you plausibly believed it to be a military target, then it's plausibly not a war crime.

The question about whether US launched the missiles is not really in dispute, the intent and knowledge is in dispute. NYT did a great analysis essentially concluding it was plausibly a mistake, not an intentional or indiscriminate civilian attack.

More here: https://news.ycombinator.com/item?id=49739018#49741081

Your claim is that hegseth directed the US Navy to launch an attack on a building it knew to be a school full of kids for the negative headlines, because those negative headlines would distract from the Epstein files outrage?

I'll add the definition from the Rome statute that, by my read, establishes a war crime when there is intent to commit the civilian strikes:

     Unless otherwise provided, a person shall be criminally responsible and liable for punishment for a crime within the jurisdiction of the Court only if the material elements are committed with intent and knowledge.

    A mistake of fact shall be a ground for excluding criminal responsibility only if it negates the mental element required by the crime.

Still this is a legal defense that might plausibly be used during a trial, not necessarily a reason to skip a trial.

Colloquially, I think it is well accepted that US launched the missiles. The question is whether they indiscriminately targeted civilians.

NYT did an excellent analysis showing that the US probably did intentionally target the building, believing it to be part of the nearby compound that received several direct strikes. This analysis is really good [1]

The definition of war crime is not "Killed civilians" it is "targeted civilians" or "reckless indiscriminate attacks". My understanding of the NYT analysis is that this was very plausibly misunderstood to still be part of the iran national guard compound.

Everyone gets their own opinion on this. There were 900 or so strikes in the first 12h, and only one school hit, and that school was actually part of the compound before. A horrific accident that was brought on by unnecessary hostilities, i think.

1: https://www.nytimes.com/2026/03/11/us/politics/iran-school-m...

I view it more of supporting an expanding use case. If rust gets popular then you'll want to support it.