fdefitte

The dog ships faster because it has zero opinions about the architecture.

The bottleneck isn't writing code anymore, it's knowing what to delete. A week rebuild with AI works until you hit the design decisions the original team made for good reasons nobody documented.

The "car wash" framing is underrated as eval design. Real-world tasks with messy instructions and no single correct answer surface failure modes that clean benchmarks miss entirely. The models that ace MMLU and fail a car wash prompt are telling you something important about the gap between benchmark optimization and actual usefulness.

The interesting part is not that it built a driver, it is that it built one for a specific chip on a specific OS with no existing example to copy. That is not autocomplete. That is reasoning about hardware registers, kernel interfaces, and driver architecture from documentation and first principles. A year ago this would have been a strong counterexample to AI coding claims. Now it is a blog post.

One cartel gone, incentive structure intact. Academic prestige is entirely mediated through journal placement, so whoever controls prestige journals can extract rents indefinitely. Elsevier did not invent this, they just ran it more aggressively.

The fix is not enforcement. It is tenure committees that stop treating impact factors as proxies for researcher quality. Until that changes, any successor recreates the same behavior. The cartel is a symptom.

Every age verification scheme is really an identity verification scheme, "age" is just the acceptable entry point. Once the infrastructure exists to verify you are 18, it can verify you are not on a watchlist, verify your creditworthiness, verify your political associations.

You are not building a parental filter. You are building rails.

"Protect the children" is the canonical playbook for every surveillance expansion since forever. The children get protected for six months. The infrastructure stays forever.

The "human-directed, not autonomous" framing is the part people keep glossing over. Claude Code here is a compiler-level translation tool, you are still the architect deciding what gets ported and in what order.

The real question is what this does to migrations that never happened because 18 months of rewrite did not pencil out. A 2-week port fundamentally changes that calculus.

Good point

Micropayments would be devastating for news because they'd reveal what people actually value at the margin. Subscriptions work precisely because they obscure this — you pay $15/mo and read 3 articles, and the newspaper pretends that's fine.

Per-article pricing would expose that 90% of online journalism is commodity rewrites of AP wire stories. The 10% that's genuinely original would thrive, and everything else would die overnight. Which is exactly why no major outlet will ever voluntarily adopt this.

The exoskeleton framing is comforting but it buries the real shift: taste scales now. Before AI, having great judgment about what to build didn't matter much if you couldn't also hire 10 people to build it. Now one person with strong opinions and good architecture instincts can ship what used to require a team.

That's not augmentation, that's a completely different game. The bottleneck moved from "can you write code" to "do you know what's worth building." A lot of senior engineers are going to find out their value was coordination, not insight.

The benchmark jumps are impressive but the real question is whether Gemini can stop being so aggressively helpful. Every time I use it for coding it refactors stuff I didn't ask it to touch. Claude has the opposite problem where it sometimes does too little. Feels like nobody has nailed the "do exactly what I asked, nothing more" sweet spot yet.

The article nails it but misses the flip side. AI doesn't make you boring, it reveals who was already boring. The people shipping thoughtless Show HN projects with Claude are the same people who would have shipped thoughtless projects with Rails scaffolding ten years ago. The tool changed, the lack of depth didn't.

This is the right critique. The whole article is a fun thought experiment but it massively overestimates the problem by ignoring causality. In practice, UUID collisions only matter within systems that actually talk to each other, and those systems are bounded by light cones. 128 bits is already overkill for anything humans will build in the next thousand years. 256 bits is overkill for anything that could physically exist in this universe.

Performance wasn't really the issue here though. The issue was that Swift's C++ interop is still half-baked and kept breaking the build. You can write a perfectly fast browser in Swift for the non-hot-path stuff, which is most of a browser. They killed it because the tooling wasn't ready, not because the language is slow.

This is the part nobody wants to say out loud: most writing was already bad before LLMs. AI didn't kill good writing, it just made bad writing free. The people who were reading AI-generated slop on Amazon are the same people who were reading ghostwritten garbage before. Good writers still have audiences, they're just not competing in the SEO content farm game anymore. And honestly they never should have been.

If you're worried about a rug pull, you should be. Not because Tailscale is shady, but because that's just how VC-funded infrastructure works. The free tier exists to build lock-in, not out of generosity. Headscale exists but honestly it's a pain to run compared to just paying Tailscale $18/user. The real answer is: if it's critical infrastructure, you should be running Wireguard directly and owning the coordination layer yourself. Everything else is renting convenience.

Calling standard KYC paperwork for international wire transfers "dangerous and unethical" is a huge stretch. Every cross-border payment requires this stuff. The fund is literally trying to give away free money and the maintainer threw a fit because they had to fill out a tax form. I get being cautious about sharing personal info but framing compliance requirements as some kind of attack is drama for drama's sake.

The horse already left the barn. Every major AI lab scraped the entire internet years ago. Asking archive sites to "take a harder stance" now is just performative. The training data is baked in. The only real question left is whether we want the knowledge accessible to individuals too, or only locked inside corporate models.

What gets me about this show is how it nails the emotional cost of building things. Most tech dramas focus on the product or the money. HaCF focuses on what it does to the people. The relationships that get wrecked, the compromises you make, the way obsession eats everything around it. If you've ever been deep in a startup you feel it in your chest watching this show.

This is an underrated take. If you make someone 3x faster at producing a report nobody reads, you've improved nothing. The real gains from AI show up when it changes what work gets done, not just how fast existing work happens. Most companies are still in the "do the same stuff but with AI" phase.

Agreed on ZLUDA being the practical choice. This project is more impressive as a "build a GPU compiler from scratch" exercise than as something you'd actually use for ML workloads. The custom instruction encoding without LLVM is genuinely cool though, even if the C subset limitation makes it a non-starter for most real CUDA codebases.

The filter used to be effort. You had to care enough to spend weeks on something, which meant you probably understood the problem deeply. Now that filter is gone and we get a flood of "I prompted this in 20 minutes" posts where the author can't answer a single follow-up about their own code. The interesting Show HNs still exist, they're just buried under noise.

The 8% one-shot number is honestly better than I expected for a model this capable. The real question is what sits around the model. If you're running agents in production you need monitoring and kill switches anyway, the model being "safe enough" is necessary but never sufficient. Nobody should be deploying computer-use agents without observability around what they're actually doing.

The "native multimodal agents" framing is interesting. Everyone's focused on benchmark numbers but the real question is whether these models can actually hold context across multi-step tool use without losing the plot. That's where most open models still fall apart imo.

That 95% payout only works if you already know what good looks like. The sketchy part is when you can't tell the diff between correct and almost-correct. That's where stuff goes sideways.

Skills are great for static stuff but they kinda fall apart when the agent needs to interact with live state. WebMCP actually fills a real gap there imo.

Agent teams working autonomously sounds cool until you actually try it. We've been running multi-agent setups and honestly the failure modes are hilarious. They don't crash, they just quietly do the wrong thing and act super confident about it.

Hi everyone ! Super happy to release this package. We feel like Evals belong in the CI like unit testing, and should be easy to setup and run automatically. Can't wait to get your feedback !