Taken in the most literal sense possible, if you're using Gemini in any capacity today, it's authorized to use your photos to serve you personalized ads.

By that logic any lie is simply a promise that hasn't been fulfilled yet.

> cars purchased in 2016 can't do this, sure I'll concede.

He made the promise in 2019! HW3 will almost certainly never have unsupervised FSD (let alone FSD without a human in the car), and HW4 didn't ship until what, 2023? There's more model years in existence where that promise is flatly broken than there are where it's maybe a possibility someday down the road.

But he also promised FSD rideshare in 2016, which arguably makes it even worse of a broken promise. Ten years later not only is it still a broken promise, they're actively working to ship a product that isn't what they promised.

But practicality aside, you have to be an absolute fool to believe any promises Elon makes at this point. Whether it's the automatic charging cable snake, or your car picking you up and dropping you off at the airport, or renting out your car as a self driving ride share, or your cybertruck also being a boat, or the dozens of other empty promises that he's made, there's just no reason to think he's telling the truth about anything until you can try it for yourself.

In this example there was a real human who reviewed the prompts that were flagged by the model. I can't agree that a model flagging something concerning to a human is worse than a human reviewing absolutely everything.

> AI companies have no accountability as proven by the hugging face incident.

This is already a problem with human systems. See: Vaastamo

Moreover, that's the whole point. You're arguing against accountability. If some kid tells ChatGPT he's going to shoot up his school, and then does, I feel strongly that OpenAI could and should have done something. I do not believe OAI is absolved of responsibility when it's clear they could have prevented death.

> E2e encrypted messaging is hated because it cant be automatically surveyed. The idea it needs to be for law does not hold up as it doesnt for your human examples either.

Your argument contradicts itself. There's nothing that allows a nation state to access an LLM back end but not a system designed for humans to interact.

> Almost all of these arguments boil back down to the “if you dont have anything to hide…” which id hope on hn we know is dangerous

That's not it at all. Your arguments are promised on the idea that messages sent to a third party should be treated specially because there's no human involved.

You might disagree with the idea that uploading CSAM to Adobe shouldn't result in a cop at your door, but you're compelling a group of people to operate software on your illegal content. It's not a dumb file store, it's doing work. LLMs are no different.

This has nothing to do with local models. There's no overlap in what using a service with a ToS implies compared to what you do on your own hardware. If print photos at CVS versus at home in a dark room, those are two very different situations with different expectations.

LLMs aren't email or file storage. AI labs aren't just shuttling bytes around, they're interpreting those bytes and taking action based on them. These models _already_ react viscerally in response to users saying disturbing things: the only practical difference is the ability (or obligation) for the model to escalate that concern. I'm not sure the ethics we hold AI companies to should be different than if a human being was typing out the responses.

Privacy is obviously hugely important, but this isn't the government surveiling every message. It's companies having an obligation to flag real, credible threats according to the law, which is a very different problem space.

Maybe I'm misunderstanding but this sounds absolutely wild. What?

Last week, I replaced 100kb of React code for a JavaScript based date picker with an . It's part of an accessibility project, and the old date picker wasn't good for many reasons. What a win! This input type didn't land in browsers other than Chrome until a few years ago, and my support matrix didn't officially allow me to use it until pretty recently. By and large, this is a win on paper.

Within 24 hours a customer using Firefox and NVDA (a screen reader) wrote in to say they hated the change and asked me to bring the old date picker back.

How am I supposed to win here?

Is it just the benchmarks? Because otherwise it suggests it's twice as chatty as Opus for a comparable output... Which kind of defeats the purpose

The difference between an LLM and WebMD is that WebMD is a page of information. Any reasoning is yours, based on a small fraction of the information available to you. An LLM is based on all the medical literature of all time. It's not even close to being the same thing.

Even if you argue that an LLM is just a prediction engine, that's kind of exactly the right tool for the job. "If you have these symptoms and not these other ones, these are likely causes" plays against the lone strength that LLMs' detractors argue they do well.

Just this year, our cat has been having digestive problems. We got special food for her, which she hates, with the suggestion that she'll need to eat it for the rest of her life. Six vet visits later, Fable 5 suggested two tests that my doctor recommended we didn't get. Both found issues that explain her symptoms, and the vet says she will probably only need a supplement and infrequent two week courses of medicine if she has a flare-up.

All that to say, the helplessness of not knowing what's wrong and the people who could know not really caring enough is something that LLMs do a really great job of mitigating. If you don't have any way to know what's wrong with you or a loved one, or how to find out, you're stuck spending a ton of money (in the US at least) and crossing your fingers that someone gets it right.

That's never been the point. Anyone involved in a double homicide can never be punished to the same degree that they harmed their victims because they can't be put to death twice. The greater purpose of the justice system is to take offenders out of society and deter others from committing the same crimes. Punishment is gratifying but ultimately doesn't change anything.

If OAI employees believed their company would be dismantled and their equity would become worthless, I suspect they'd be a whole lot more careful.

But moreover, if what you were suggesting was a real problem, nobody would ever be brought to justice for murder because the victims are always dead.

If I run a biology lab and engineer a terrible virus, it gets out, and a global pandemic ensues, I don't get to shrug and say "well we told it not to infect people". It's my fault for failing to mitigate the risks of my work.

But that's irrelevant, because it's the tool call. The tool call is what invokes the HTTP client, not what's in the request payload. It could be asn.1 or RDF or TOML. JSON just happens to be convenient and universally supported. MCP could easily be implemented on literally anything you throw at it: that's an internal detail to your agent harness. That's not the point.

MCP is the OAS/WSDL/whatever contract described as tool calls instead of as network requests that the agent has to construct from scratch. Its whole purpose is to eliminate the whole layer of indirection.

```

new Promise(r => setTimeout('r(this)', 0))

```

This promise resolves to globalThis.

You can know that something is fishy, but it can be obfuscated nearly to the point of being impossible to untangle because you can't even be sure where it's being invoked.