DeepSeek just shipped a cheaper, faster model with a genuinely new architecture, and people are impressed.
What the article says
- DeepSeek released V4.1 Flash, the first model in a new architecture family, and it can now understand images too.
- The trick is splitting the model into an encoding half and a decoding half, so it only has to actively use a small fraction of its total size on each token.
- That cuts the memory needed per token to a quarter of the old version, which is why it runs faster and cheaper despite being much bigger overall.
- Prices actually came down from the previous flash model, undercutting most rivals on cost.
- The tradeoff is it no longer fits comfortably on a home computer the way the last flash model did.
What HN is saying
- Commenters loved that DeepSeek published a detailed technical writeup, unlike the safety heavy reports from US labs.
- A long thread dug into the pricing, showing how cheap cached tokens make it far cheaper than rivals for long coding sessions.
- Several people said the bigger size means it no longer fits on a typical home rig, unlike the older flash model.
- Opinions split on whether it actually beats models like GPT and Claude in real work, since it wins some benchmarks but loses others.
- A few users complained it keeps replying in Chinese even when asked in English, which others blame on its system prompt.
Shopify just admitted AI coding agents killed the whole reason to use React Native
What the article says
- Shopify built its apps in React Native since 2020 to avoid writing every feature twice for iOS and Android.
- Coding agents got good enough that building the same feature separately in Swift and Kotlin no longer costs much more.
- So Shopify is rebuilding its apps from scratch as native apps, starting with the Shop app, done in twelve weeks.
- They built an internal tool called Helix that forces AI to work in small reviewed steps instead of dumping one huge rewrite.
- They are also handing off or archiving several open source React Native libraries they used to maintain.
What HN is saying
- Plenty of people are stunned Shopify needs three thousand engineers for what looks like a shopping app, comparing it to tiny teams that built Chrome or GTA five.
- Others push back that most of those engineers are not on the mobile app at all, and that a company worth over a hundred billion dollars can afford the investment.
- Several developers backed the move, saying they made the same switch and native now builds faster and runs better than React Native ever did.
- A sharp counterpoint says the framework switch is a distraction. The real risk is complexity spiraling out of control once AI makes it cheap to keep adding more of it.
- One engineer who lived through a pre AI native rewrite says the case for going native held up even without any AI help at all.
A mathematician says OpenAI's denial about using his private chats was carefully worded, not honest.
What the article says
- A mathematician named Andreas Thom says he once asked OpenAI whether his private chats with ChatGPT fed into a surprising math result the company then published.
- OpenAI's reply was a flat statement that it did not happen. He says the wording only ruled out one narrow kind of access, not training.
- He ties this to the bigger row over a Navier Stokes proof, where OpenAI reportedly poured huge computing power into the problem right after hearing two mathematicians were close to solving it themselves.
- OpenAI says it did not use their specific work, but admits it cannot rule out that anonymized versions of user chats helped train its models.
What HN is saying
- Most commenters think the ethics problem stands even without proof of stolen data, since OpenAI reportedly rushed to solve a problem right after hearing a rumor someone else was close.
- A popular theory called the Overhang argues AI does not need anyone's private ideas. There is a huge backlog of obvious next steps sitting in published math that only a tireless machine has time to check.
- Skeptics push back hard, calling the whole story hearsay built on a screenshot of a screenshot with no real evidence behind it.
- A detail that spreads fast in the thread is that OpenAI's setting to opt out of training is on by default, which fuels the distrust either way.
Microsoft just gave Rust the same official status as C++ and C#, and plugged it straight into its own compiler backend.
What the article says
- Microsoft now treats Rust as a fully supported internal language, alongside C++, C# and TypeScript.
- That means Rust gets the same paved path as those languages, from secure builds to production tooling and compliance checks.
- The big technical move is a new compiler backend that connects Rust straight into Microsoft's own MSVC toolchain, instead of the usual LLVM.
- This is meant to make Rust and C++ share the same debugging, security and optimization tools, since most Windows code is still C++.
- More than a hundred internal Microsoft projects already build on it, and adoption is growing weekly.
What HN is saying
- Commenters mostly agree this signals Rust has matured past its scrappy early years into a genuine C++ rival.
- The sharpest thread is about interop. Everyone wants easier ways to mix Rust and C++ in the same codebase rather than full rewrites, and points to ongoing Google and Rust Foundation efforts to make that easier.
- Several people push back on the framing that this fixes anything for users, noting Microsoft's own apps still ship bloated and buggy regardless of language.
- A recurring joke and real question is why Visual Studio itself still has weak Rust support, with most replying that VS Code has already won that role internally.
- One correction: this isn't really about rewriting old code in Rust anymore, it's about interoperability with existing C++, Python and JavaScript.
A simulator lets you walk around a world where light moves at human walking speed, and it looks wild.
What the article says
- A physicist built a first person simulation where light crawls at walking pace instead of its real blistering speed.
- Walk or drive around and you see real relativity effects with your own eyes, not just in equations.
- Moving objects visibly squash, twist, and change color depending on how fast you approach or pull away.
- Clocks and flashing lamps speed up or slow down depending on whether you are moving toward or away from them.
- The creator admits some shortcuts, like the physics of a Ferris wheel that would really fly apart at these speeds.
What HN is saying
- Commenters loved how faithfully it captures real relativity, better than an older MIT game from 2012 that had similar aims.
- Several people got philosophical, riffing on how slow light actually feels once you picture real distances like the nearest galaxy.
- Some found the controls confusing at first, since accelerating forward can visually look like moving backward.
- People shared other relativity toys and books, from a dinosaur puzzle game to novels exploring the same idea.
- One thread got into the weeds on why you can even see the tops of moving objects, a subtle effect from light's travel delay.