HNHacker News
TopNewBestAskShowJobs

radarsat1

4,157 karma · joined May 28, 2011

Info: https://sinclairs.gitlab.io

Email: <myusername>@gmail.com

submissionscomments
radarsat1··on Sustainable memristors from shiitake mycelium for high-frequency bioelectronics
one word: chairdogs
radarsat1··on EuroLLM: LLM made in Europe built to support all 24 official EU languages
That's exactly the problem in Europe though. It's quite the opposite here.
radarsat1··on The future of Python web services looks GIL-free
I agree and honestly it may as well be considered a form of ABI incompatibility. They should make this explicit such that existing C extensions need to be updated to use some new API call for initialization to flag that they are GILless-ready, so that older extensions cannot even successfully be loaded when GIL is disabled.
radarsat1··on A sharded DuckDB on 63 nodes runs 1T row aggregation challenge in 5 sec
what do you think vector databases are? absolutely. i think the idea of a database and a "model" could start to really be merged this way..
radarsat1··on VST3 audio plugin format is now MIT
Oh wow, finally! They should have done this 20 years ago but this is awesome news.
radarsat1··on My trick for getting consistent classification from LLMs
I'm curious, what is the use case for open-ended labeling like this? I can think of clustering ie finding similar tweets but that can also just be done via vector similarity. Otherwise maybe the labels contain interesting semantics but 6000 sounds like too many to analyze by hand. Maybe you are using LLMs to do further clustering and working on a graph or hierarchical "ontology" of tweets?
radarsat1··on Tor browser removing various Firefox AI features
it's literally the same technology as LLMs. Transformers were proposed for translation.

(But I don't know what methods the Firefox translation uses. I assume it's a local model but don't even know that for sure.)

radarsat1··on Why is everything so scalable?
> scalability needs a whole bunch of complexity

I am not sure this is true. Complexity is a function of architecture. Scalability can be achieved by abstraction, it doesn't necessarily imply highly coupled architecture, in fact scalability benefits from decoupling as much as possible, which effectively reduces complexity.

If you have a simple job to do that fits in an AWS Lambda, why not deploy it that way, scalability is essentially free. But the real advantage is that by writing it as a Lambda you are forced to think of it in stateless terms. On the other hand if suddenly it needs to coordinate with 50 other Lambdas or services, then you have complexity -- usually scalability will suffer in this case, as things become more and more synchronous and interdependent.

> The monolith is composed of separate modules (modules which all run together in the same process).

It's of course great to have a modular architecture, but whether or not they run in the same process should be an implementation detail. Barriers should be explicit. By writing it all depending on local, synchronous, same-process logic, you are likely building in all sorts of implicit barriers that will become hidden dangers when suddenly you do need to scale. And by the way that's one of the reasons we think about scaling in advance, is that when the need comes, it comes quickly.

It's not that you should scale early. But if you're designing a system architecture, I think it's better to think about scaling, not because you need it, but because doing so forces you to modularize, decouple, and make synchronization barriers explicit. If done correctly, this will lead to a better, more robust system even when it's small.

Just like premature optimization -- it's better not to get caught up doing it too early, but you still want to design your system so that you'll be able to do it later when needed, because that time will come, and the opportunity to start over is not going to come as easily as you might imagine.

radarsat1··on Abstraction, not syntax
I see, not so clean imho but I guess it works.
radarsat1··on Abstraction, not syntax
I got confused reading this because I wasn't sure how I, as a reader with no knowledge of the system under discussion, was supposed to know that all the buckets should be in the same region.

Nor is it clear to me how the "for loop" version would handle the case where exceptionally one bucket is different. Which is a more interesting discussion imho, that's the whole point of having it as a configuration field, after all.

radarsat1··on Embracing the parallel coding agent lifestyle
Are there any semi autonomous agentic systems for image generation? I feel like mostly it's still a one shot deal but maybe there's an idea there.

I guess Adobe is working on it. Maybe Figma too.

radarsat1··on Two things LLM coding agents are still bad at
Agreed. You basically want an LLM to have a tool that writes its own agent to accomplish a repetitive task. I think this is doable.
radarsat1··on Less is more: Recursive reasoning with tiny networks
Thanks something like that was going through my mind, nice to get a good reference for it. Any insights on why this is not a more popular approach? Maybe it's too difficult for a single layer to scale.

I read a paper recently on something similar for diffusion, called Fixed Point Diffusion Models. They specialize the first and last layers but recurse the middle layer some number of times until convergence.

Considering how a Transformer is a residual model, each layer must be adding more and more precise adjustments to the selected token. It makes a lot of sense to think of this like the steps of an optimisation method.

radarsat1··on Vibe engineering
> how to do it

If you've ever lead a team, you know how much more valuable people are if they don't need to be told how to do things. Even more if they don't need to be told what to do! But having to explain in detail the "how".. can be really a big time sink and only worth it if you are training someone to level up.

radarsat1··on Vibe engineering
I thought about making some kind of community project where people could contribute their lines to a common file, and even some kind of MCP server or RAG system that automatically selects relevant "rules" given a certain project context. Do you think there would be interest in something like that?
radarsat1··on Why do LLMs freak out over the seahorse emoji?
Aside but I find it really interesting how looking at the layers "approach" the answer at a specific token position like this feels a lot like an iterative optimizer taking steps to converge on a solution.
radarsat1··on Implicit Actor Critic Coupling via a Supervised Learning Framework for RLVR
> By treating the outcome reward as a predictable label, we reformulate the RLVR problem into a supervised learning task over a score function parameterized by the policy model and optimized using cross-entropy loss.

Isn't this how the Decision Transformer works? I don't see it in the references, so I'll be curious to compare the papers in more depth.

https://arxiv.org/abs/2106.01345

> By conditioning an autoregressive model on the desired return (reward), past states, and actions, our Decision Transformer model can generate future actions that achieve the desired return.

Lately it has crossed my mind that I haven't seen DT brought up much lately, it seemed really interesting when it was first published but I haven't read much follow-up work.

radarsat1··on Ask HN: Who wants to be hired? (October 2025)

    Location: Utrecht, Netherlands
    Remote: Yes; hybrid preferred but fully remote also possible
    Willing to relocate: Eventually (see below)
    Résumé/CV: https://sinclairs.gitlab.io/cv/sinclair_cv2025.pdf
    Email: radarsat1 [@] gmail
My background is in audio and haptics with a shift towards machine learning 10 years ago. I have worked as researcher or senior SWE on various topics including computer vision (medical, automated driving) and my latest achievement is leading a small text-to-speech startup as chief technical officer, which was a great experience in leadership, as well as the ins and outs of deployment and productization. I also gained skills in cloud infrastructure and API development. Unfortunately despite technical successes we didn't achieve funding, so I am on the lookout for new opportunities and open to working on similar or new topics. I love solving problems that require interesting algorithms and messy data. Lately diving into LLMs and agentic systems to understand how to mold these powerful systems into useful components of real products. So, generally I am open to ideas: all this Gen AI stuff is fascinating and I'd love to work more with it, but if I ended up working on vision or audio again I'd also be a very happy camper.

Regarding relocation, I am very close to achieving EU long term residency so while I'd be happy to relocate within Europe it would make a lot of sense for me to stay in the Netherlands for 1 more year and get this done, rather than start over again in a different country. On the other hand if I am presented with an amazing opportunity maybe it would make sense, so let's discuss it!

radarsat1··on Why haven't local-first apps become popular?
I spent some time learning about PouchDB a little while ago, it seems to be a nice solution at least for a No SQL approach. Although I still need some more practical experience to understand the security model, because it feels weird to just sync a database that was updated on some web page, really you want to ensure there is validation and treat it as an untrusted source to some extent. Still not sure of the best way to deal with that without implementing most of the application as server side validation scripts, or maybe that is just the way to do it.
radarsat1··on NASA's Guardian Tsunami Detection Tech Catches Wave in Real Time
The article mentions gravity waves, I thought these were not possible to detect except from massive events like black hole mergers. Am I getting concepts mixed up?
radarsat1··on Are people's bosses making them use AI tools?
> It's like adopting a new methodology for something without actually training anyone on the new methodology, it leaves everyone scrambling trying to figure it out often with poor results.

Agree strongly. I've had to push back when the CEO didn't see the results he wanted, had to basically remind him, look, this technology is like.. 6 months old (at the time, and with respect to actually getting good results, Claude Code etc).. you can't expect everyone on the team to just "know" how to use it.. we're literally _all_ learning this very new thing right now, not just us, but everybody in the industry. It's a little crazy to expect immediate uptake and an overnight revolutionary productivity boost with this thing we barely know how to use properly, there's going to be a learning phase here whether you like it or not

radarsat1··on From multi-head to latent attention: The evolution of attention mechanisms
To be fair it would be a lot easier to iterate on ideas if a single experiment didn't cost thousands of dollars and require such massive data. Things have really gotten to the point that it's just not easy for outsiders to contribute if you're not part of a big company or university, and even then you have to justify the expenditure (risk). Paradigm shifts are hard to come by when there is so much momentum in one direction and trying something different carries significant barriers.
radarsat1··on Claude Code Checkpoints
I don't know what this is but isn't git enough? Incidentally I'm not convinced in my day-to-day for "jujitsu" (jj) but from what I understand about how it works, I've been wanting to give it a try for agent-based coding, based on the way it defaults to saving everything and letting you sort it out after. I do like how Aider commits everything so you can easily roll back, although it ends up with a few too many commits imho.

I've been wanting to experiment also with getting an agent to go back and rebase history, rewrite commits etc in the context of where the project ended up, to make a more legible history, but I don't know if that's doable, or even all that useful.

radarsat1··on The Math Behind GANs (2020)
Whenever someone says this I like to point out that they are very often used to train the VAE and VQVAE models that LDM models use. Slowly diffusion is encroaching on its territory with 1-step models, however, and there are now alternative methods to generate rich latent spaces and decoders too, so this is changing, but I'd say up until last year most of the image generators still used an adversarial objective for the encoder-decoder training. This year, not sure..
radarsat1··on Gemini 2.5 Flash Image
Oh cool thanks I haven't come across that one, I'll give it a shot.
radarsat1··on Gemini 2.5 Flash Image
I've had a task in mind for a while now that I've wanted to do with this latest crop of very capable instruction-following image editors.

Without going into detail, basically the task boils down to, "generate exactly image 1, but replace object A with the object depicted in image 2."

Where image 2 is some front-facing generic version, ideally I want the model to place this object perfectly in the scene, replacing the existing object, that I have identified ideally exactly by being able to specify its position, but otherwise by just being able to describe very well what to do.

For models that can't accept multiple images, I've tried a variation where I put a blue box around the object that I want to replace, and paste the object that I want it to put there at the bottom of the image on its own.

I've tried some older models, and ChatGPT, also qwen-image last week, and just now, this one. They all fail at it. To be fair, this model got pretty damn close, it replaced the wrong object in the scene, but it was close to the right position, and the object was perfectly oriented and lit. But it was wrong. (Using the bounding box method.. it should have been able to identify exactly what I wanted to do. Instead it removed the bounding box and replaced a different object in a different but close-by position.)

Are there any models that have been specifically trained to be able to infill or replace specific locations in an image with reference to an example image? Or is this just like a really esoteric task?

So far all the in-filling models I've found are only based on text inputs.

radarsat1··on How to Think About GPUs
Why haven't Nvidia developed a TPU yet?
radarsat1··on How to Think About GPUs
A comment from there:

> There are plans to release a PDF version; need to fix some formatting issues + convert the animated diagrams into static images.

I don't see anything on the page about it, has there been an update on this? I'd love to put this on my e-reader.

radarsat1··on Claude says “You're absolutely right!” about everything
I find Gemini is also hilariously enthusiastic about telling you how amazingly insightful you are being, almost no matter what you say. Doesn't bother me much, I basically just ignore the first paragraph of any reply, but it's kind of funny.
radarsat1··on Show HN: Omnara – Run Claude Code from anywhere
Honestly it should be easy to do this kind of thing with a CLI app like Claude Code and yet you have to get pretty involved in interpreting its complex terminal usage. I'm starting to consider the recent trend of very "fancy" CLI utilities that treat the terminal as a kind of GUI as an anti-pattern. I mean, how much easier would it be to write a wrapper around CC if it were just a readline prompt.

Yet these fancy terminal applications have become really trendy. I've been using Crush lately and I quite like it, but it's annoying to me that I can't copy and pasted from the terminal when it's running, and scrolling the buffer doesn't work as expected either, it somehow scrolls some kind of internal "window" instead. Again, making copy & pasting annoyingly difficult.

Is anyone making a good agent system that is just "> Output", "$ input" without trying to get crazy with ANSI escape sequences? Some color is nice, but I think that should be the end of it.

← PreviousPage 4 of 34Next →