mgraczyk

That's wrong and even if it were not wrong, it would still not fix the problem. What if the most likely response is "kill yourself"

It's harmful because in this context it leads to an incorrect conclusion. There's no reason to believe that LLMs "averaging" behavior would cause a suicidal person to be "pulled toward normal"

It's not true in any sense that LLMs "give the average words for every answer"

It doesn't sound generated to me

All of these sound like value judgements and opinions. You claim they make people dumber but the evidence is that using an LLM to search the Internet requires less brain usage? Of course it does, that's the point! Using a dishwasher also uses less of a our brain than washing dishes by hand. I will use my brain for other things.

And whether LLMs are a "good" use of electricity is purely a value judgement. I'm not a fan of cars and don't drive, and a single car ride can use more energy than every LLM query made in a year by most ChatGPT users. But I don't think that makes people who drive cars evil

Yes but I think outweighed by the overall cost and harm of putting up little warning signs in every restaurant, coffee shop, parking garage, and grocery store in the state

I didn't work there for 3 years (1 year), and I'm not publishing a study.

Should people who post anti social media sentiment disclose that they've never worked on it, have never run experiments on well being, and have never looked at the data?

It doesn't, I think both anecdotes are not useful for understanding what's really happening

Is anecdote only acceptable evidence when it agrees with what you already believe?

My sister does, who is sitting next to me talking to me about this

For example here is a recent widely cited study that did not find a statistically significant link between Facebook/Instagram and mental health outcomes, broadly miscited as having found an effect: https://ifstudies.org/ifs-admin/resources/briefs/ifs-gallup-...

They did claim to find a very small link between TikTok/YouTube and mental health, but this seems to defy the narrative of "social" media being the culprit. YouTube was not significant if you adjust for multiple hypotheses, only TikTok

Every time I look at the evidence, I end up finding that social media improves mental health for teens overall. Is there a new study that motivated this or are we still misinterpreting statistics?

Most of the labeled chemicals aren't harmful, so decreasing concentrations is not a good thing

Could somebody steelman the argument. Why is this bad? What harm is caused by receiving an email like this? Seems completely harmless to me, uses much less water/energy/co2 than the car ride I just took which nobody is yelling at me for

Causes zero harm to anyone, less bad than normal spam. Silly thing to get angry about

I doubt it, although it's possible these models will be used for creator tools, I believe the main idea is to use them for data labeling.

At the time the first SAM was created, Meta was already spending over 2B/year on human labelers. Surely that number is higher now and research like this can dramatically increase data labeling volume

One answer to questions like this is that Claude Code has orders of magnitude more paying users, so it's more important to get things right and ship carefully

I claim that is not true, because very soon AI agents (probably built into Chrome) will detect and warn. In which case you need to phish the agent, tricking the human won't be enough.

If the human is much easier to phish than the agent (which I believe is true in most cases) then this would be a win

Yes I would, and lots of people do this all the time

Serious question for people who are concerned about security here.

Do you believe that AI browser automation like this will lead to more, or less overall information exfiltration (including phishing).

I work at Anthropic so maybe I'm biased, but it's not clear to me that this is worse than the status quo

And it works much better for most people

Do you have any evidence that it is spitting out licensed code? Did you locate an original that it was copied from?

On the contrary, coding for me has become more fun than ever since Opus 4.5. I'm working more and genuinely enjoying it a lot more, haven't had this much fum building software in years. (I work at Anthropic but have also tried Gemini, it's also fun)

That is different, it's a code interpreter where the model can run code and see the outputs. It's not a different way of doing tool calls

Anthropic recently added this to the API: https://www.anthropic.com/engineering/advanced-tool-use

See "Programmatic Tool Calling"

And there was an AI productivity startup called Lutra AI doing this, although they've since pivoted to some kind of MCP infra thing: https://lutra.ai/

Here's one way to think about it

Architects are not civil engineers and often don't know details of construction, project management, structural engineering etc. For a few years there will still be a role for a human "architect" but most of the specific low level stuff will be automated. Eventually there won't be an architect either but that may be 10 years away

This is probably useful and better than nothing, but the problem is that by the time you graduate it's unlikely that reading the output of the LLM will be useful.

I'm a lot more productive than I was in 2023 and I've been coding full time since 2012

The sad reality is that this is probably not a solvable problem. AI will improve more rapidly than the education system can adapt. Within a few years it won't make sense for people to learn how to write actual code, and it won't be clear until then which skills are actually useful to learn.

My recommendation would be to encourage students to ask the LLM to quiz and tutor them, but ultimately I think most students will learn a lot less than say 5 years ago while the top 5% or so will learn a lot more

You should read the article

The "has been known for millennia" is presumed without evidence. If it were known, why did nobody ever write it down?