johncolanduoni

Nobody will peer with you if you start announcing addresses you don’t own, so there’s no way to make those addresses reachable on the public internet.

Inflation is increasing prices, which has causes other than central banks putting more money into circulation. When Bitcoin’s value drops, prices in Bitcoin increase so it experiences inflation regularly.

There’s no simulation you can do today that can tell you that your drug will likely make it through phase 3 trials. The best we can do is simulate reactions of a few key molecules, we’re nowhere near cell or organism scale simulations.

Manufacturing costs of most drugs are miniscule. Virtually all of their costs and from R&D and the regulatory process.

How do you suggest they provide the static compression dictionary?

How are the publishers forcing creators to engage with copyright law? Anyone who creates a work can decide to waive their rights under copyright law today, no need to wait for it to be abolished. Why do so few of them do it?

Nobody is stopping you from monetizing your own creative work via the methods you described without a publisher involved today. If those ways of getting paid would create a “rich and vibrant ecosystem of people sharing things”, why doesn’t it already exist?

So is peer review. False proofs have been published before and will again. Does that mean journals are publishing slop?

How do you know all these problems were on the precipice of being solved?

I wouldn’t call it an air gap, but when you use a model on Bedrock or Vertex the inference runs on Amazon or Google’s servers respectively. The model provider gives them the weights, and is not otherwise in the loop for serving individual requests.

A halfway decent synthetic biology lab (no need to invoke CRISPR) can make e.g. smallpox without a sample of the original disease, just from the gene sequences. Basically all state actors could do this if they wanted to without an LLM. What barrier that a terrorist organization faces today to having a functioning synthetic biology lab does an LLM actually solve?

What does Bun’s governance look like? Now that Anthropic bought the company are there significant external contributors that would expected to have input on a decision like this?

OOM killing in WebAssembly is trivial, since it’s all in a growable linear memory. All the runtimes I’m aware of have a simple maximum memory setting, and they’ll trap any allocation requests after that point.

We can argue over the definition of scaremongering and what people we’ve never met “really think”, or we can argue over what the actual risks of AI are. I know which one I’d prefer…

They said why they think it’s a sincere belief: past statements from before the AI hype cycle took off. I take it you have other evidence?

Dario Amodei is not Elon Musk. Most powerful people do not internalize the emotions of others (though obviously some do). If you think powerful people are disproportionately sociopaths, you either agree with me or are not using the word correctly.

They generally care about what powerful/influential people think. I’m not powerful or influential, but I guess maybe the person I was talking to is. If so I apologize for the miscommunication.

If he wasn’t willing to change his mind after he saw the results, then why would he do it at all? Can you explain the false motivation that you think he communicated in the original kerfuffle about this?

I’m not asking you to prove my point, but I think you’re being myopic.

* Vikor Orban was just ousted by a popular protest movement. This took years due to structural electoral issues but it did work eventually. And it wouldn’t if the people opposing him gave up because they didn’t change anything right away.

* The US ICE protests (and the federal government’s insane overreaction to them) let to the head of DHS being fired and a quantifiable drop in ICE activity (e.g. arrests and number of people currently detained).

* Nepal’s protests last year led to the resignation of the prime minister and a resounding electoral victory this year for their opponents.

Protests aren’t magic win buttons, especially because even the people protesting don’t fully agree on exactly what “winning” looks like. But they accomplish more than acting out your emotions on the internet.

Not by trading the same suspicions on the internet with fellow true believers over and over again, I think the past 10 years have proven that pretty conclusively. Maybe people should try some of the things previous social movements did, seemed to work pretty well even against a much more uniform media environment and a stronger hostile social consensus.

Protests don’t immediately solve everything, but I think looking at 2026 and concluding they don’t move the needle at all is a weird take. There are recent examples of protest movements (especially long-term ones) working all over the world.

Did I say it wasn’t real? Or tell you that you couldn’t talk about it? No, I just pointed out that it’s all anybody talks about and it’s boring and doesn’t engage with anything specific about this stunt/project. And I can make melodramatic analogies too — like to the panic about global overpopulation that led to mass sterilizations in The Emergency. Panic is not an unalloyed good, and if you want to fight “the clankers” you should understand what they are and are not capable of.

Also I already cashed out, jokes on you.

Based on their actions, I don’t think the LLM vendors take anti-AI sentiment very seriously. If anything they court it, though I think it’s more likely they’re just high on their own supply. I doubt the Zig statement had any effect on the thoughts of the people who actually sign contracts with Anthropic, who are mostly not engineers.

The marketing opportunity here is in promoting Claude Code, not giving a smackdown to Andrew Kelley (who vanishingly few people who throw around millions of dollars on AI contracts have heard of).

What do you think one would have to pay to have flesh-and-blood engineers get a cross-language port of a codebase of over half a million lines with a broad test suite to over 99% conformance? I think it would be astronomically high, especially given that for this specific project your hiring pool is going to be limited to people who can get up to speed with Zig and JavaScriptCore right away (or you’re going to have to pay them for low output for a while as you train them). Also it would be literally impossible to do in 6 days no matter how much money you paid, so unless they’re lying about that it’s still something that couldn’t have been done prior for any price.

More handwaving about the LLM hype machine is incredibly boring and enough of it is spewed everywhere that whatever social good it was going to accomplish must have already happened by now. If you want to inject reality into the situation, talk about reality (like Anthropic is at least pretending to).

Powerful people figured out how to make suspicion work for them long ago. You have every right to be unconditionally suspicious, but it’s not a good way of accomplishing any change. Also their feelings are not hurt by what you or I think, they don’t care.

Running an experiment, the experiment being more successful than you thought, and then deciding to put more effort into a bigger experiment is not hypocrisy. It’s engineering. If you think some of the objective facts they’re putting out (like test coverage and performance) are lies, go and prove it instead of appealing to emotion.

I don’t think the Zig project adopting a strict ‘no LLM’ policy affects the LLM vendors at all. How many developers are working on the Zig project itself that will (maybe) now not buy a Claude subscription? I can buy that this is a marketing stunt, but nobody at the top cares if a relatively small open source project doesn’t allow AI contributions.

I’m not clear on who is supposed to do the translating that isn’t doing it today, or why the mapped IPv4 addresses don’t qualify. Virtually all ISPs either give you an IPv4 address or do the translation for you, and the software you write doesn’t have to care exactly how it’s set up for the most part (there’s some subtlety about stuff like MTUs, but if you’re just doing unencapsulated TCP it usually doesn’t matter). It can just use whatever comes back from DNS (including using the official mapped addresses if only an IPv4 record exists) and expect the network stack to figure it out.

None of this stuff works perfectly, but it’s powering almost every mobile internet connection in the world so I don’t understand what is missing.

You can do this quite easily on all the major OSes by opening a single dual-stack socket. On most Linux distributions that is the default, and then mapped addresses will work fine if the machine has a v4 address or a 464XLAT configuration. This has been the case for about 20 years.

This already exists in two forms, and has for basically the entire history of production dual stack deployments. The first are IPv4 mapped addresses (of the form ::ffff:x.x.x.x) which instruct the local machine’s network stack to use a local IPv4 address to communicate to the server. This still requires each machine to have a routable IPv4 address (though not necessarily a public one - it can be used with a NAT44 router).

If you don’t want the local machine to have a routable IPV4 address at all (which is a common practice on mobile networks), you can do something similar with a different prefix (usually 64:ff9b::/96) called NAT46. In this case the local machine’s network stack exclusively emits IPv6 packets, and the upstream router/network detects the prefix and performs stateful NAT using its own IPv4 address(es), just like NAT44. Depending on the use-case, the local machine doesn’t even have to understand what you’re doing at all. Instead you just have the local network’s DNS server return fake AAAA records with the corresponding 64:ff9b::x.x.x.x addresses when some particular domain doesn’t have native IPv6 support.

Yes, but CGNAT is an inherently stateful system and as a result will always be more expensive to operate per packet than a stateless router. The reason we are seeing steady (if slow) growth in native IPv6 is because the workarounds for IPv4 exhaustion cost money, and eventually upgrading equipment and putting pressure on website operators to support IPv6 becomes cheaper than growing CGNAT capacity.