HanClinto

Yes. All you need is a formal grammar that defines the syntax of the text being generated. The llama.cpp repo has a decent writeup on the way that grammar-constrained decoding works in general [0], and it includes a bit of a simplistic example for generating syntactically-correct C code [1]. No reason this couldn't be extended to other things -- I personally have run experiments with generating syntactically-correct SQL statements when doing text-to-sql [2], and I think it's pretty cool.

[0] - https://github.com/ggml-org/llama.cpp/tree/master/grammars [1] - https://github.com/ggml-org/llama.cpp/blob/master/grammars/c... [2] - https://gist.github.com/HanClinto/b853571136fd42f8e2bde4c99c...

It's not automatic, but it is the trend.

https://ourworldindata.org/much-better-awful-can-be-better

I've been using 98.css [0] as a foundation for building many of my personal projects lately -- I think it's pretty great, and a really refreshing return to some older paradigms.

[0] - https://jdan.github.io/98.css/

I mirrored the bug report on the author's Github project [0], so if it helps, you can view the screenshot there.

[0] - https://github.com/wieslawsoltes/VB6/issues/39

It's not a complete list, but it really reminds me of the parent comment's criteria.

If AI isn't achieving superhuman performance in all of these areas, I'm not sure we can actually call it "Humanity's Last Exam" -- it feels like a bit of an overextension.

"A human being should be able to change a diaper, plan an invasion, butcher a hog, conn a ship, design a building, write a sonnet, balance accounts, build a wall, set a bone, comfort the dying, take orders, give orders, cooperate, act alone, solve equations, analyze a new problem, pitch manure, program a computer, cook a tasty meal, fight efficiently, die gallantly. Specialization is for insects."

-- Robert A. Heinlein

> But play-testing on a real Chromatic requires a firmware update and API key that came in the box with the DevDay devices

That's too bad! Those blank DIY cartridges in the DevDay bundles look really sweet. Here's to hoping that they make it more general-release!

Yes!! I remember reading this book at the library as a kid, and for years afterwards I wondered if I had just hallucinated the whole thing.

I think it's a really important scene, and I'm sad that it was cut from the movie.

Side note: the (extremely well-done) Star Wars Radio Drama [0] included a great many expanded scenes on Tattoine (with quite a bit of expansoin on Tosche Station), and I think those scenes (which include the aforementioned scene with Biggs) really add to the richness and flavor of the setting.

If you're a fan of Star Wars and haven't yet listened to the radio drama, I highly recommend it!

[0] - https://www.youtube.com/watch?v=0-29uKdckL4

No. It's not limited to 3rd-party guardrails. Given how restrictive the native public-facing OpenAI guardrails are, this feature wouldn't be worth very much if it just slacked back off to the level of "regular" filter paranoia offered by the native models, would it?

This is needed if you're going to be dealing with things like psychologists doing self-harm research or red-teaming or sensitive sexual content -- if you're working with any of that sort of stuff in a professional context and want to leverage OpenAI models on Azure, then that's the form that you fill out to get access to unfiltered models.

Note that I am not aware of this feature being offered for Anthropic models -- I've only seen it offered for OpenAI models (note that the documentation I linked is specifically in the "Azure OpenAI" category).

Yes, this is well-documented and publicly advertised. In Azure Foundry, the feature to modify (or completely remove) safety guardrails and content filtering is called "Limited Access" [0], and one must submit a form to request permission to use this feature. This is one of the more straightforward paths to get access to unrestricted frontier models, but it's far from the only way.

[0] - https://learn.microsoft.com/en-us/azure/foundry/responsible-...

The upcoming RTX Spark [0] is aiming to fill that space -- consumer GPUs with tons of memory.

[0] - https://www.nvidia.com/en-us/products/rtx-spark/

Very impressed with the Watch Dogs series of games for this purpose.

We recently used Watch Dogs Legion to prep for a trip to London to get more familiar with the city. Set the game on peaceful mode and just walked around and did virtual sight-seeing instead of playing missions. It's a surprisingly good recreation of the city, and we really appreciated getting some familiarity with the place before we visited.

Maybe "Claude" is the new "Xerox", but for coding agents instead of copiers?

And yours got the chain!

Likewise -- also had zero issues going over this with Sol. Seemed to give solid advice for how to test it -- use an expendable bare-metal AMD family 16h test system w/ usual standard checks that apply.

> Run `platform_check` first and do not use `SKITTER_FORCE=1` casually. Start with the read-only `dram_state` and `dram_carveouts`, then `dram_dump --dry-run`. Avoid `dram_poke` until maps have been freshly collected and calibrated. Do not bypass fingerprint checks, calibration, fencing, or verification.

Claude's (apparently externally-mandated?) lobotomization continues to be concerning. :-/

This brings up an interesting topic of atypical compounds that can visibly affect skin coloring -- whether temporary or more long-lived (such as elevated methemoglobin levels in the Blue Fugates of Kentucky).

It's fascinating to learn that melanin and haemoglobin are the two primary factors (and cover the vast majority of cases), but I'm curious how many other illnesses / disorders can produce compounds with visible effects.

Thank you for your posts on this thread -- I appreciate you knowledge-sharing on this!

Well once I've got an outdoor-capable robot that can drive around the acreage and generally find its way around, the first step would just be inventorying the property and doing things like surveys of plant or animal species (birdsong recognition, etc).

But for interaction with the world? I'd probably take something like an old 12-volt windshield-washer sprayer out of one of the wrecked cars in my front yard and put Round-Up into the tank and let it go spray all the poison ivy and invasive honeysuckle for me. Doesn't need to gimbal like a turret -- just generally give it a fixed-aim that's roughly at the center of the camera vision and let the bot put pest plants roughly in its center-of-view and activate the sprayer for a second or two, mark the spot as sprayed, and move on to the next one.

Could test it with plain water and logging the plants that it chose to spray first as a review step before loading it with actual weed-killer.

I don't need complicated end-effectors that can fold my laundry -- just a simple weed-wacker motor or squirt gun would be enough for me to call this thing 1000% useful. Like a Roomba, but outdoors.

There are a handful of open-source farm bots built on traditional platforms with traditional robotics stacks, but there's something attractive to me about the plug-and-play nature of something like OpenClaw + Robostral (along with the extensibility that self-modifying agentic systems have to offer).

What is the realistic path to getting to play with this? I would love to hook this up to OpenClaw for hobbyist exploration. My dream has been to embody OpenClaw into a farm robot (been looking at adapting one of those RC lawnmowers that is tracked and built for mowing steep hills) so that I can assign it various tasks around our acreage -- "Explore the fenceline take pictures of the plants. Find all of the poison ivy and invasive honeysuckle and spray it with your Roundup sprayer. Repeat this every week and report the species map after every pass. Come back to the barn and charge yourself whenever you get low."

It's not hard to put OpenClaw into a robot body (numerous YouTube videos showing people doing this sort of thing), but when you dig in and see what people have done, the actual movement portion is always the clunkiest part (and this matches my own experiments as-such as well). It feels like an 8B model like this would be perfect for solving pathing and navigation issues.

Anyone who may be more experienced with Mistral (or companies like them) -- are they interested in hobbyist builders who would be experimenting with things like this? Or are they primarily looking for commercial partners? I would be willing to pay a license fee to use the model in my experiments, but if I'm just one guy, I'm not sure they'd want to work with me unless I were building a business out of it (which I'm not).

This is pretty hilarious. :D Feels like a combination of Untitled Goose Game and Goat Simulator. Vibe-coded? It's pretty impressive.

"If you wish to make an apple pie from scratch, you must first invent the universe"

Paper plates of software development.

Some of the coding-specific fine-tunes were really impressive boosts. Qwen2.5-3B-Instruct is also available [0] -- if it's not too much to ask, I'd be curious how more general models stack up in your benchmark?

[0] - https://huggingface.co/Qwen/Qwen2.5-3B-Instruct

Setting up good book publishing pipelines with version control + CI/CD might sounds simple, but I don't think it's trivial.

One of the best examples of this that I've ever seen is The Sourdough Framework [0] -- really impressed with the way that versioning and publishing is integrated in that book.

And yes -- I know it sounds like yet another Javascript library -- but it's actually a book about sourdough bread making. It's been discussed here several times before, but this one from 2023 [1] may have been the most popular (103 comments)

[0] - https://github.com/hendricius/the-sourdough-framework [1] - https://news.ycombinator.com/item?id=35961590

If the phone had broken comms, how did you get the code onto the phone to run?

This is very cool!

If you're looking for a good test suite, I wonder if you might be able to adapt any of the tests available in XMage? They have a pretty extensive test suite (such as for copy effects [0]) and if you point your agent at their code, I wonder how many could be usefully adapted to your system?

[0] - https://github.com/magefree/mage/tree/master/Mage.Tests/src/...

Oh cool!

I've been doing similar experiments lately (using ViT's) to do card recognition, and so far it's been working really well for me. If you want to compare notes, I've open-sourced my code / weights [0] and written some blogs about how mine works [1]. I'd love to see if we can collaborate!

> Push the inference to the client-side (WebGPU / Web Workers).

I have an example of this working in webgpu / wasm here [2] along with a playground environment (demonstrated here [3]). I'm currently training a new version that uses a different ViT backbone more optimized for WASM inference -- it's currently converging, and I hope to have it finish training (or at least reach parity with the previous model) in about a week (took ~200 epochs for my last one to reach the level that it's at, and it takes about an hour per epoch in my current setup).

You mentioned WebGPU -- I've run into issues with the MobileViT-XXS backbone producing bad results in WebGPU on Android, so YMMV in whether or not WebGPU is stable enough to use for this or not. I don't know if it's my problem or a true bug in the platform, but I've fallen back to WASM and things are working much better since then.

[0] - https://github.com/HanClinto/CollectorVision

[1] - https://blog.hanclin.to/posts/gh-19/

[2] - https://hanclinto.github.io/CollectorVision/

[3] - https://youtu.be/MHieOcmC7Dw

Thank you! If you want to test out my tool, here's a link to the web version that is built for scanning in lists of cards:

https://hanclinto.github.io/CollectorVision/

It's still super rough (doesn't support foil-toggling yet, still some issues with double-sided cards, crashing on some iPhones), but overall the rough structure is there -- it can create lists and export as CSV.

If you have feedback or feature requests for your needs, please leave them on Github and I'll get to them as soon as I can. I'd love to hear more user feedback!

Thank you -- I appreciate that. :)

> it seems i either take them to a shop and get lowballed, or spend hours meticulously researching each card and then figuring out how to sell it for what it’s worth.

Well if you want to use the scanner for something useful, you can run the web version here: https://hanclinto.github.io/CollectorVision/

No install -- scan your cards with your phone or desktop (downloads the weights in WASM -- runs 100% local -- the only web request it makes is to look up card names and prices online -- no image data ever leaves your machine), export the list as CSV, take your cards to your friendly local game store, and expect to receive 50-75% of TCG-low for your cards. This app currently only displays TCG Market, so probably about 50% of this price is what you could realistically expect.

> Hard to sell to individuals like me, but i would think a card marketplace would find it invaluable?

Yes -- and part of this might be that this would have been much more amazing several years ago, but by now -- most marketplaces (I used to do work for some of the big ones) have their own recognition tools. If they aren't actively looking to replace their current software, many companies would rather stick with what's currently working "good enough" than expend effort to migrate to something with only incremental benefit that is difficult to quantify. It's possible that would happen, but it's a tricky sales call to make.

I might just be imagining things, but I'm also picturing what one of those sales calls might look like, and it feels like I've opened the kimono a bit. The cat's out of the bag. There's no mystery or allure behind it anymore, and I feel like that puts me on the back foot somehow -- almost like I've played my strongest cards (hah!) first and have nothing left. By being open-source from the beginning (and talking freely about my architecture and what makes my solution different), there's very little sales-pitch build-up. Maybe it's just a part of the problem of how I'm presenting it, but I think people (especially the big houses) are probably just-as (or more) inclined to silently learn from me and improve their own scanners than try to use / build-upon what I've provided.

It's funny -- that angle is almost more about raising expectations and forcing the big houses to improve their own tech and catch up to open-source, more than getting anyone to adopt my solution in particular.

Am I okay with that? Absolutely -- I made that decision when I open-sourced it. I feel like the tech has been stagnating for several years, and I want to increase the quality of scanners across the board. I want to be the rising tide that lifts all boats.

That's one of the strongest arguments in favor of open-sourcing it (it would be very difficult for a closed-source product to have that same effect), and I remain hopeful for that long-term.

I resonate with this blog post a lot.

I think there is something to be said for monetizing ones' hobbies, but I've recently been taking some forays into this world of "build something amazing and give it away for free" as well. I recently took a very big experimental plunge in this path, and I'm curious how well it will work out for me.

Open-source state-of-the-art Magic: The Gathering card identification pipeline:

https://www.youtube.com/watch?v=MHieOcmC7Dw

I used to do this kind of image recognition for a living, but I've been out of the business for a little while now. I had some ideas for a different approach from what I've done in the past and decided to code it up. This version is far better than anything else I've ever done -- especially for scanning against busy backgrounds or with occlusions, and also for noticing fine differences between otherwise difficult-to-distinguish printings.

I didn't have any interested customers waiting for this, so -- much like the OP -- decided to create an experiment and release it open source. I'm not opposed to having paths to monetize it (for people who want to license it for closed-source commercial projects), but I'm not trying to commercialize it so much as I would love to see how far we can take it with open-source.

I don't know which path I should take with this.

The biggest downside is that I feel like I've had a hard time getting people to be as interested in this project as I would have expected -- I believe this truly is the best identification software available (I've built some benchmarks to test it [0]), and maybe the market is just a bit flooded for such things (?), but I suspect that one very strong problem is that if you don't charge for something, then there is a perceived lack of value.

Sometimes I wonder if I would have more interest in this project if I _weren't_ trying to give it away.

For me, that's been the most negative aspect about releasing this for free so far.

[0] - https://blog.hanclin.to/posts/gh-26/