ollin

Joined 9 years ago

Karma
778
Submissions
104
Member since
Feb 23
Independent researcher working on real-time on-device generative models. Made sdxl-vae-fp16-fix, TAESD/TAEHV, pokemon-emulation-via-dnn.

https://mirurai.com (soon) https://madebyoll.in https://www.neuralworlds.net

View on Hacker News

Recent submissions

Comment · 3 days ago
Comment · 5 days ago
I was confused too, but it looks like https://github.com/openai/math/blob/main/lean/ComparatorChal... (with the sorry at the end) is just the problem statement; the actual proof is a big pile of lean in the Algorithms folder, with the final theorem stated here https://github.com/openai/math/blob/7ef5c0cc5ca4b0e072b7a050...
Comment · 17 days ago
Right! The 256 introduced hardware T&L, which was a great advance in capability (but harder to explain than "it's the first GPU"). From the same chapter [1]:

   It was incontrovertibly true that the GeForce 256 represented a significant advancement in traditional graphics capabilities, which by this point was typical and expected of every new chip generation Nvidia produced. It featured four graphics pipelines, allowing it to process four pixels at the same time. It also integrated a hardware transform and lighting (T&L) engine, which meant it could take on the calculation tasks essential to the moving, rotation, and scaling of 3-D objects.

    Those tasks were usually handled by the CPU; the GeForce 256 thus took even more computational burden off of the CPU and made the entire computer run faster.
    
    "By putting dedicated hardware to do that, you could suddenly process a lot more geometry and make much more interesting pictures," said former chief scientist David Kirk.
    
    The management team at Nvidia thought that this was far too technical to pitch to customers. The typical in-house naming formula of acronym and number wouldn't work. Nvidia needed something bigger to market its new product.
    
    "We had to find a way to position this product as a much better graphics processor for 3-D graphics than anything out there," said Dan Vivoli. "The chip is big. It has texture and lighting. We've got to charge a lot for it. We need to come up with a way to make sure how great this thing is." He challenged his product-marketing team to come up with something brilliant.
[1] The NVIDIA Way /
https://www.taekim.com again
Comment · 17 days ago
Per The NVIDIA Way [1], the GeForce 256 was branded as "the first GPU" for marketing reasons alone; the actual first GPU was the subsequent GeForce 3 (which introduced programmable shaders):

    [Sanford] Russell grabbed Michael Hara for a thirty-minute brainstorm session to figure out how to market the GeForce 256 more effectively, and both executives remember coming out of the room with the notion to call the new chip the first entry in an entirely new product category altogether: a graphics processing unit, or GPU, which would be to graphics rendering what the computer's main central processing unit (CPU) was for all other computational tasks.

    Even so, the GPU moniker, when applied first to the GeForce 256, was controversial among Nvidia engineers, who pointed out that a chip could not truly be called a GPU unless it had several features that the GeForce 256 did not. The chip lacked a "state machine," a technical term meaning a dedicated processor that transitions from various states to execute and fetch instructions, in the same manner that a CPU does for programming commands. It was not programmable, meaning that graphics styles and features could not easily be customized by third-party developers. Instead, developers would have to rely on a fixed set of Nvidia-defined hardware functions. Moreover, the GeForce 256 did not have its own programming language.

    But the marketing team argued that such features were already planned for the next generation of graphics chips. And even without them, the GeForce 256 delivered a step change in performance that would be obvious to every gamer and computer enthusiast in the world. While not literally a GPU, the GeForce 256 could still be a category-defining product. The chips that would follow-the "true" GPUs that would be fully programmable by outside developers-were coming soon enough.

    So Nvidia's marketing team pushed forward with the GPU name over the dissent of Nvidia's engineers. "We don't need anyone's approval," Vivoli told them. He didn't think anyone outside of the industry would really care about the technical definition. Besides, "we knew the next one was going to be programmable. We decided to take a leap, make a stretch that this is a GPU."

    In February 2001, Nvidia released the GeForce 3, whose programmable shader technology and support for third-party development of its core graphics functions made it the first true GPU.

[1]
https://www.taekim.com
Comment · 18 days ago
Comment · 18 days ago
Comment · 19 days ago

    document.querySelectorAll("canvas").forEach(c=>c.style.opacity = 0)
in the console eliminated the cursor effect for me
Comment · 26 days ago
Comment · 26 days ago
Not just bad headlines; OpenAI is actively being sued [1] for this:

    OpenAI’s stated rationale was a concern for the shooter’s privacy, but its own interests
    better explain its silence. Upon information and belief, OpenAI was seeking to avoid implementation of a
    hard line rule to refer planning of real-world violence to authorities, perhaps due to how frequently its
    product is implicated in threats to human life. Requiring such disclosures would be incompatible with the
    company’s public position that ChatGPT is safe. It could also threaten the valuation underlying OpenAI’s
    anticipated initial public offering. Rather than expose those risks, OpenAI accepted the consequences of
    its silence. A mass murder in the only secondary school in Tumbler Ridge followed.

    Accordingly, the Crown, led by Attorney General Sharma, and SD59 jointly bring this
    action to hold OpenAI and Sam Altman accountable for designing a dangerous product, distributing it to
    every home with internet access, ignoring the warnings of their own safety team,
    refusing to notify authorities when they knew the shooter was planning gun violence, inviting the shooter back onto the
    platform after deactivating the shooter’s account, and choosing corporate self-interest over the lives of
    children. They seek compensation for the not just foreseeable but known harm OpenAI inflicted, the
    damages that they incurred and are incurring, and injunctive relief to ensure that this tragedy does not
    happen again.

[1]
https://cdn.arstechnica.net/wp-content/uploads/2026/09/Briti...
Comment · 3 months ago
This paper shows a nice integration of older winner-take-all ideas for learning K-modal generative models (see e.g. https://arxiv.org/abs/1612.00197, https://arxiv.org/abs/2211.14286) into modern diffusion/flow pipelines. As-implemented, I think it has some downsides:

1. K-1 extra forward passes during training

2. Inaccurate sampling behavior (will sample all K modes with equal likelihood, rather than sampling them proportionally)

However, I think both of these downsides can be mitigated by adjusting the implementation a bit more (you can have the model predict K modes jointly in one forward pass, along with probabilities of each being the min-loss mode, which you can then use for properly-weighted mode sampling at inference).

That said, I'm not entirely sure if this strategy is as generally applicable as the authors are hoping. In particular:

1. For highly-conditional image generation (like modern commercial diffusion pipelines, which use a big LLM preprocessor), most of the low-frequency color/layout decisions are already made for you by the conditioning signal. The diffusion process mostly needs to generate high-frequency details, for which there are a huge number of equally-valid modes.

2. For LLMs themselves, the sequence-generation process is already factored into a discrete classification problem, and there's no mode smearing issue to fix.

Comment · 3 months ago
Comment · 3 months ago
If the HuggingFace repo the agent broke into contains reference solution scripts for ExploitGym (i.e. for exploiting the vulnerabilities in the intended way), the agent can then run that reference code inside its original sandbox to retrieve the dynamically-generated flags.
Comment · 3 months ago
This article seems fairly uninformative since, as others have pointed out, there's no visualization or comparison of the full emoji set and no link to see it. They just show a few example images and have some (AI-enhanced?) prose that doesn't actually say very much.

This article https://9to5google.com/2026/05/12/android-17-emoji-redesign/ has a larger (2d image) comparison grid with several dozen examples and an A/B slider vs the old versions. Overall the new design looks like a fairly tasteful compromise between Google's previous flat-shaded vector emoji and the hybrid 2d+3d Apple emoji, with the benefits (easier to rerender with higher-resolution, animations, tweaked lighting, etc.) that you'd get from a fully-3D pipeline. So I like the new set of emoji, just not this particular blog.google.com article.

Live from the official Hacker News API

A modern reader for Hacker News · SvelteKit + Skeleton