martianlantern

This is not something to taunt about

Cool project, but just a side thought I was having about how do people have resources and the money to make things like this and make it avl for public, I mean it's fair to say they have their own GPUs or if they are using api keys for gpt or Gemini with enterprise subsidized inference

But still coming from a frugal background I still cannot wrap my head around this

Very nice! I found the tutorial on gaussian splatting to be very neatly written and educational

Why are they doing this?

Nice! I need something similar for english now

I am awaiting for a full duplex model from qwen team

This is cool! I am more interested in how you guys generated next edit training data from repos, seems like there are lots of caveats here. Would love your insights

Again amazing work! waiting for what you guys cook next

my site: https://martianlantern.github.io/

I liked the new approach but I don't like the pseudo security framing. I don't see how an LLM rewritten query is secure than just sending my raw query, in both cases I don't feel secure at all

Cool! but it would be more educational if it explained each step and it's reasoning

If I am missing something, isn't this the whole purpose of hacker news to be a link aggregator for technical/developers?

I have given up on any external bash configurator a long time ago, instead I write my own bash prompts these days, they lack a lot of functionalities but I am much happy with them for now, also a shameless plug: https://martianlantern.github.io/2025/11/updating-my-bash-pr...

would be great if there was a chat room where we can also chat with all the other people in the lobby :)

damn

Wow this seems very interesting! Can we get a TLDR of how this was achieved?

This is awesome! It reminds me of this https://github.com/lucidrains/self-reasoning-tokens-pytorch

The article itself seems to be AI slop, too many em dashes ...

Do mean hard to be used because of compute required to render it or because of setting up and understanding the training pipeline and related hyperparameters?

Do you know you can achieve the exact same functionality with the vscode's inline code execution in python files and can create jupyter like code cells with '# %%'?

ref: https://code.visualstudio.com/docs/python/jupyter-support-py...

Is there any benchmarks and comparisons compared to gpt-oss? I believe it far exceeds gpt oss or even gpt5 otherwise they wounldn't have released it

Hey, really cool work love the idea of focusing on key decision points. I was curious though since confidence can be non monotonic during CoT[1], how does binary search handle cases where there are multiple ups and downs in confidence? It seems like there might be more than one "pivotal" token, so I wonder if there's a plan to support multi-token pivots or use a different approach than binary search?

[1] - https://arxiv.org/abs/2505.14489

There's no explanation as to how they achieved that speed up :( it would have been better if they also wrote a post on that

I’m not familiar with Zig and would appreciate an explanation of how this works. My understanding is that cache behavior is managed by the CPU, and programmers only influence it indirectly through the sequence of instructions (i.e., access patterns). Is that accurate? Also, is this approach specific to Zig, or could it be achieved in C or Rust as well? Thanks

But why?

Am I the only one who is not able to access the interactive demo?

Is there even anything not shortsighted till now?

Seems very handy, is there a way to hide the record pop up so we can concurrently do other tasks?

What types of projects do you use cursor heavily for and what did you stop doing because of cursor, what did you start doing more of?

Very insightful post, this may work in the IMO setting because mathematical problems are inherently binary if we ignore somethings like the incompleteness theorem. In contrast, subjective tasks, such as evaluating a painting or rating a poem, lack absolute truth. How would such reasoners estimate confidence in these cases, and to what extent could RL techniques effective in the IMO transfer to real world problems?

> Laymen used LLMs for all kinds of things as well. In an interview a prominent Finnish actor admitted to (paraphrasing) “pasting a full script into an AI app for content analysis and structuring”

I think it is now soon that hollywood movies or netflix dramas will be using generative AI. First it will be for blending frames, or increasing aesthetics, then some frames being generated to reduce production cost, then an entire shoot to be done in latent world and slowly AI will get adopted in everything that we see