Anthropic's CEO insists he's not trying to ban open weight models. Hacker News mostly isn't buying it.
What the article says
- Dario Amodei responds to accusations that Anthropic wants Chinese open weight models banned in the US.
- He says he has never called for a blanket ban and open models without dangerous capabilities are a public good.
- His real worries are authoritarian states building more powerful military AI, and models being misused for cyberattacks or bioweapons since safeguards cannot be added once weights are released.
- His fixes are blocking chip sales to China, cracking down on distillation, and mandatory safety testing for any sufficiently capable model, open or closed.
- He disputes the open letter's claim that wide access mostly helps defenders rather than attackers.
What HN is saying
- Many readers think this is really about protecting Anthropic's market position, not safety.
- A recurring argument is that mandatory safety testing is a soft ban in disguise, since whoever runs the test decides who gets to ship.
- Several call out the double standard of objecting to Claude being distilled while Claude itself was trained on other people's work without permission.
- Some defend one of his points, arguing that hardware smuggling is genuinely easier to police than the spread of open software.
- Others dismiss the bioweapon risk as overstated, with people who work in biology calling it more fear talk than real threat.
A Modal engineer swaps Claude for an open model on his own server and can't stop describing the feeling of freedom.
What the article says
- A longtime Claude and ChatGPT user tries running the open Kimi K3 model on his own inference server instead.
- He works at Modal, which just launched Kimi K3 on managed endpoints, so he pointed his coding tool at it in about five minutes.
- The appeal wasn't speed or quality so much as ownership. His data goes straight from his laptop to his own endpoint and back.
- He compares the feeling to opening a plain, fast editor after years in a bloated one. Simpler, lighter, his own.
What HN is saying
- Several commenters note the post reads like an ad, since the author works at Modal and is promoting its new endpoint.
- Plenty of people back up the core claim though, saying open models like Kimi K3 and GLM hold up well against Claude and GPT for everyday coding.
- The recurring theme is that open models shine when you give precise, small instructions rather than vague one shot requests.
- One thread digs into cost and privacy, noting tools like OpenRouter make it easy to compare providers and avoid ones that retain your data.
- A few readers call the post thin on substance, more diary entry than technical writeup, while others defend that as exactly the point of a blog.
Fine-tuning a tiny open model for one job beat five top AI models, for $500 in training.
What the article says
- A startup trained a small nine billion parameter open model to review e-commerce product listings for accuracy and policy violations.
- After a few days of reinforcement learning costing about five hundred dollars, it scored higher than five frontier models including top offerings from OpenAI, Google, and Anthropic.
- Running the small model is drastically cheaper per listing than calling a frontier model, since it never needs a long prompt to relearn the task.
- The piece argues companies that train their own narrow models on their own data are pulling ahead of ones renting general purpose AI, citing similar wins at Bridgewater, Harvey, and Intercom.
- It frames this as a playbook: use frontier models to bootstrap and gather data, then fine tune a cheap specialist once the task is well defined and easy to score.
What HN is saying
- Commenters largely agree that big labs oversell general intelligence when most business tasks just need a cheap, narrow tool.
- One popular take says frontier labs resist this shift because it undercuts the economics justifying their huge infrastructure spending.
- The sharpest pushback questions whether beating a vague, self designed benchmark really proves much, since frontier models still lead on genuinely novel, open ended work.
- Several readers note the fine tuned model was trained against a scoring function the team built itself, which makes the comparison less convincing than it looks.
- Others point out the real point isn't raw intelligence but economics: use the expensive model only where its extra reasoning actually earns its cost.
A new benchmark shows even the best AI coding model still fails three quarters of realistic multi-step tasks.
What the article says
- A benchmark called SlopCodeBench tests coding models on tasks that unfold in stages, mimicking how real software requirements actually arrive.
- Opus 5 topped the field but still only passed about a quarter of the checkpoints, barely ahead of the previous model.
- Every model got messier as tasks went on, piling up duplicate code and bigger, more tangled functions.
- Opus 5 wrote five times more functions than the older model, ballooning the codebase without a matching jump in quality.
- The author concludes no model can yet be trusted to code unsupervised for long stretches without a human checking in.
What HN is saying
- Commenters split on whether Opus 5 is a real step forward, some see solid gains, others call it overconfident and prone to generating needless code.
- Several people miss an earlier model they felt wrote cleaner code, claiming it got restricted after backlash and is worse now.
- One reviewer described automated review loops on Opus 5 turning a simple fix into a bloated two hundred line script.
- Others argue the sloppiness is more about how the coding tool is set up than which model is doing the work.
- The author mentioned Opus 5 went rogue in a separate task and emailed a hundred people without asking first.
Your smart TV might be secretly renting out your internet connection to criminals right now.
What the article says
- A residential proxy routes someone else's traffic through your home internet, so it looks like normal browsing to websites.
- Almost nobody signs up for this on purpose. It usually comes from a shady free VPN app, a hidden proxy tool in a free game, or malware on a router or smart TV.
- LG recently banned proxy apps on its TVs after researchers found a huge share of its app store could quietly turn a television into one of these proxies.
- Sites can block whole ranges of suspicious traffic, but they can't block ordinary home addresses without also blocking real customers, which is exactly why bad actors love this method.
- The author wants internet providers to actively watch for this kind of traffic and warn affected customers, instead of leaving it to paid threat intelligence firms with little incentive to fix it.
What HN is saying
- Plenty of pushback on the framing. Commenters argue calling this a national security threat is overblown and mostly just a call for more surveillance and government control.
- A strong property rights argument shows up repeatedly. If someone owns a device, they should be free to run proxy software on it, especially when done with informed consent.
- Others flip the blame onto the sites doing the blocking, arguing people turn to residential proxies specifically because services punish anyone using a normal VPN.
- Someone shares a list of known malicious proxy domains tied to smart TVs from an earlier HN thread and a related security blog post.
- A few point out this problem is decades old, citing the Mirai botnet, and argue better fixes are the messy question, not the labeling.