Blog

DeepSeek-R1 is not Sputnik #

DeepSeek-R1’s lead is fundamentally different than what Sputnik’s was for many reasons, the primary one being the difference in access to powerful GPUs. DeepSeek did not design R1 to be trained on H800s just to see if it was possible—there were monetary and political incentives for them to create a powerful model on such limited hardware. In contrast, American AI companies have not felt any need to optimize model training, since they are much more focused on a different goal: fast, cheap inference. DeepSeek has been doing great work, but their work should not be any sort of scare for the American AI market, especially since R1 benchmarks extremely closely with o1.

As an analogy, I think of it as a student writing a compiler: it takes hard work for someone of their age, and fortells their ability to do much more complicated work as a future computer scientist. However, the same compiler could have just been created by a computer scientist who has specialized in compiler design for a decade. In this same way, DeepSeek is training impressive models on limited hardware, showing their architecture’s potential for training an even more powerful model if they had access to more powerful hardware. However, OpenAI already has access to the powerful hardware and is training their models using it, allowing them to easily train models with the same performance as R1, even with a worse model architecture. Therefore, even if DeepSeek is a student who—through a lot of hard effort—created a compiler, OpenAI is an experienced researcher who creates a similar result with much less effort.

Since American AI companies have access to the supplier of powerful GPUs (Nvidia) and now know a more performant training architecture through DeepSeek’s open research, there’s nothing stopping them from easily creating more powerful reasoning models than DeepSeek-R1. That’s the main difference compared to Sputnik—there shouldn’t be any perceived technical gap because DeepSeek’s innovation is unnecessary in the eyes of American AI companies (but it will still benefit these companies immensely).

Additionally, it’s not as if DeepSeek is using Chinese-made GPUs—if they were doing that, it would definitely should be a scare to American AI companies. But right now, DeepSeek and other Chinese AI companies still have a heavy reliance on Nvidia, allowing the United States to easily control the technological gap between it and China.

Boom XB-1 Flight Test #

XB-1 first supersonic flight

via Hacker News

Super cool to see Boom’s XB-1 breaking the sound barrier on its first supersonic flight. I enjoyed this explanation by one of the commentators covering the difficulties with reaching supersonic speeds and how XB-1’s design optimizes its performance. As an aside, the flight information is really nicely displayed in the livestream, with the Mach being front-and-center since Boom’s main objective is to achieve a value > 1. I’ve also appreciated the same high quality of displayed information in SpaceX and Blue Origin’s streams, with the shown stats being carefully chosen to not cause unnecessary clutter.

Galaxy S25 is Identical to S24 #

The state of the Galaxy series is really boring

The Galaxy S25 series being identical to the S24 series was the main takeaway by most news sources after this year’s Galaxy Unpacked. While this point is extremely valid and further shows the general staleness of phone designs over the last couple of years, I don’t see how Samsung can really win here. For example, in the case of Apple, moving the cameras around for seemingly arbitrary reasons has been mocked by many people, especially outside of the tech community. For Samsung, with their Galaxy S series being widely considered a pinnacle of smartphone hardware, there’s no good reason for them to deviate from their formula to make their phones more unique. This holds especially true since most people only upgrade their phone every few years, meaning that cumulative spec bumps create more of an impact than completely changing the phone’s design language. Though, I still do find it a bit sad that Samsung abandoned the stylized camera bump of the S21 and S22 series, since those are probably my favorite phone designs ever.

Tiny Wasm Compiler #

A WebAssembly compiler that fits in a tweet

via Simon Willison

This was a really cool walkthrough of an obfuscated program through incrementally deobfuscating it. I especially liked the heavy use of coercion in the obfuscated code, since that heavily leans on JavaScript’s somewhat janky typecasting that always leads to interesting results. I’ve previously written a compiler using LLVM, and the similar “low-level but not too low-level” aspect of the Wasm byte array makes it seem like a fun spec to create and compile a language into.

Auctioned iPhones with TikTok #

Used iPhones with TikTok sold for $3000

via Slashdot

This is an interesting phenomena for banned apps on iOS, which obviously doesn’t happen on Android since you can sideload APKs easily. This specific case reminds me of when Flappy Bird was taken down by its creator, with iPhones with the game installed being sold for large amounts of money despite the game’s overall userbase—and therefore the amount of iPhones with the game installed—being extremely large.