prathje

Of course, if the data/knowledge is open and accessible, why not?

Well, of course learning isn't the problem and shouldn't be. But what about reselling/commercialising that knowledge? That is the big (commercial) opportunity. But if AI models learn by fitting their parameters to the given data, then overfitting on said data could result in the IP or copyrighted being "slopped" out verbatim?

What about IP and copyright?

We got asked this question two weeks ago when we conducted a workshop on effective and responsible use of AI tools at a local conference. But how can you argue about about IP and copyright if the breakthrough of LLMs is potentially based on circumventing or breaking IP and copyright in the first place?

How do you feel about all of this?

Nice! I would love to use it for images as well. Then again is using Grammar-Based Decoding with a json response not the same? Is Jev just that with nice caching? Because then I have been using that already…

We have had libraries for animation like Remotion and Motion Canvas for some time. It feels like libraries are becoming worth less and less as agents become more capable. Not sure how I feel about that…

Worth adding: the $3.21 were for nanobanana2 probably generating the assets.

I have been working hard to enable models creating nice explainers for videozero.ai which is based on Motion Canvas.

Above all, the model’s sense of what feels and what looks good is the most important. So many times the models have missed clearly wrong layouts etc. I guess it just shows the limitations of LLMs when they should generate something they have not been trained on?

If opus 5.5 is now better capable of that, that would be brilliant!

Borrowed the same idea to create distinct video styles for videozero:

https://videozero.ai/styles/

The entire page and Repo just screams AI prompted to build a story and sell a product.

I was trying to figure out the benefits of it and whether the price is justified. After 5 minutes I had still not fond out why I shouldn’t just build on Typst myself and invest the time to build my own templates?

Maybe that could be a bit clearer?

Thanks!

Love this, good point! It really shows how difficult it is to give an easy explanation that generalizes well... So, although quite abstract, the fundamental propagation of changes in the EM field at the speed of causality is a good enough approximation?

In a medium, this can create a "phase kickback" which creates a combined wave that appears to travel slower than the original one. The kickback is just the result of multiple EM changes propagating, i.e. the photons interact with the material, re-emitting photons.

3Blue1Brown has a beautiful animation for this phase kickback here: https://youtube.com/shorts/XIW-2ykgVPI?si=PJWiAC2BO7_xP0S6

It took me a long time to develop an intuition for light and electromagnetic wave propagation, and I’m still working on it.

Fundamentally, changes in the EM field propagate always with the speed of light in a vacuum, i.e., c (also known as the speed of causality). Single EM waves propagate with exactly this speed and they do not magically slow down in a medium... they propagate happily at speed c! (FYI EM waves are more complicated like this and involve electric and magnetic fields evolving together).

But since EM radiation interacts with matter and this interaction itself changes the EM field again it results in more EM waves that propagate also at c. Hence, they propagate together and the net result be constructive or deconstructive as well as anything in between. If they have different frequencies, they can also create "interference" patterns or pulse envelopes that seem to propagate slower and even faster than c.

No doubt that the causes and effects are not easy to understand but always thinking in terms of changes in the EM field ALWAYS propagating at c helped me.

Just found this! Does any of you have experience with custom benchmarks? For creative work, I find it quite difficult to define objective benchmarks that are at the same time a useful metric in practice...

Would love to hear what you do?

Interesting concept which reminds of the operations available in pandas.

I disagree though with the statement of SQL needing 20 lines. The given query feels verbose and has lots of redundant conditions. Not saying that it is short but a better analogy could look like this:

SELECT DISTINCT an.name, t.title

FROM keyword k

JOIN movie_keyword mk ON mk.keyword_id = k.id

JOIN title t ON t.id = mk.movie_id

JOIN movie_companies mc ON mc.movie_id = t.id

JOIN company_name cn ON cn.id = mc.company_id

JOIN cast_info ci ON ci.movie_id = t.id

JOIN aka_name an ON an.person_id = ci.person_id

WHERE k.keyword = 'character-name-in-title' AND cn.country_code = '[us]';

I would love more vision benchmarks! Once I asked the model to inspect a completely black picture and it hallucinated a nice wooden kitchen wall. Took me some time to figure out where the kitchen came from...

I usually go to https://arena.ai/leaderboard/vision/pareto for a nice overview of current models.

Setting up your public URLRetrying… (attempt 5 of 6) Fingers crossed!! The last one gotta be the one...

EDIT: It let me down... (Something unexpected went wrong on our side.)

Please let me now if deploying it on your custom domain worked for you please...

EDIT: Retrying it now worked but I had to subscribe to R2 (which is free for 10GB apparently). However, after that it will tell you that you need to enable the Dynamic Workers which is actually paid...

"Your Cloudflare account is on the Free plan, but the Cloudflare OS backend needs Dynamic Workers, which requires the Workers Paid plan."

So no free preview for me today.

Location: Liverpool, UK

Remote: Yes

Willing to relocate: Yes

Technologies: Rust, TypeScript, Python, C, C++, SQL, NumPy, Monte Carlo simulation, mathematical modeling, numerical optimization, probabilistic modeling, motion and trajectory planning, robotics, ROS, MuJoCo, Zephyr, Docker, Linux, HPC, OpenMP, OpenCL

Résumé/CV: https://patrickrathje.de/cv.pdf

Email: mail@patrickrathje.de

About: Computer scientist and software architect with a PhD in distributed systems and a strong background in numerical simulation, mathematical modeling, probabilistic systems, optimization, robotics, and embedded software. My research and engineering work has included Monte Carlo and system-level simulation, modeling uncertainty in wireless ranging, experimental validation on embedded hardware, and decentralized trajectory coordination and optimization. I am interested in senior engineering roles and startups that face technical challenges and require product design expertise.

Here is their blog article: https://artificialanalysis.ai/articles/deepseek-v4-flash-073...

Now that is a new variant I actually like to hear about. Thanks for the story behind it!

Just FYI: We also have animation libraries like Motion Canvas that run in the browser using TS/JS: https://motion-canvas.io/ , which can also be used by AI agents with the right AI skills (e.g., https://github.com/VideoZero/skills )

Thanks a lot, Gionn, for also scanning and raising issues on this matter!

I just saw a video showing their Atlas robot lifting a 50 pound fridge and I was wondering how they handle the thermal load that this would produce. Turns out, they have released an entire video around the Design with all aspects around actuators, thermals, UX and compute. Def. recommended!

Hey HN! I planning to run some benchmarks showcasing the quality of different models for Motion Canvas-based animation. I thought it would be funny to include the pelicans on bicycles as well (you know, https://github.com/simonw/pelican-bicycle ). The linked animation is Opus 4.7 and it did a great job. It even added a little jump!

Do you think it makes sense to include them or would you think I am just testing the SVG quality again? Maybe other benchmarks/prompts would be better suited?

PS: The link got corrupted... this is the actual link: https://videozero.ai/view/?id=d65d479e-1060-4f26-a28f-4d80f2...

That the calibration got the simulation so close to reality was quite impressive.

We are using MuJoCo to train a G1 humanoid robot right now. The best thing is that we do not need to fight with NVIDIA software and that it runs on macOS.

PS: I just finished a first draft for agentic skills around working with MuJoCo in Python. Feel free to check them out here: https://github.com/prathje/agentic_mujoco_skills

GIF would indeed be nice! Will try to compile a short demo

Canvas Commons is an active fork of the original repo: https://github.com/canvas-commons/canvas-commons Docs for Canvas Commons: https://canvascommons.io/

Furthermore, the old docs for Motion Canvas can be found here: https://archive.canvascommons.io/

Happy to see Schleswig-Holstein switching as well and also it being mentioned in an article on the HN front page. Who would have thought?

If you have worked with MotionCanvas or wanted to give 2D canvas animations it a try, you might notice that the website is currently not reachable and that the repo has not received any updates in over a year, suggesting that the project is not actively maintained at the moment.

However, you will also find that the community is still active and that, for example, Canvas Commons is an active fork of the original repo: https://github.com/canvas-commons/canvas-commons

Docs for Canvas Commons: https://canvascommons.io/ Furthermore, the old docs for Motion Canvas can be found here: https://archive.canvascommons.io/

Cheers! Patrick