How many people would develop psychosis even without the AI? Could be that 0.07% would be affected regardless. Others studies show psychosis incidence rates at 0.05% ... 0.09% (measured before LLMs).

Initially I was skeptical of screen typing. Then I got like "maybe with practice", "maybe with swype or a smart LLM".. but no, it fights against you. After 19 years of refinement from Apple screen typing is still a horrible experience. I mistype 50% of the text and have to retype. It also doesn't implement the basic concept of "Select All" so you can't move a source context into a LLM. You can select an article by slowly scrolling 25 screens of text.

Oh yes, I can't understand why people would spend time on the phone unless you are on the shitter or in a bus. For any civilized context you better use a laptop, or a laptop + external screen + mech kb + decent mouse. As god intended.

That MIT study has been doing rounds on the web but it is misleading. The study showed that 56% of the developers had never used Cursor before, and crucially, the one developer with over 50 hours of Cursor experience actually did see positive speedup.

A good question. Another question is - where does this leave open LLMs?

Parents claim:

> ChatGPT counseled Adam away from telling his parents about his suicidal ideation

OpenAI claims:

> ChatGPT warned Raine “more than 100 times” to seek help, but the teen “repeatedly expressed frustration with ChatGPT’s guardrails and its repeated efforts to direct him to reach out to loved ones, trusted persons, and crisis resources.”

What users type into the chat window has an influence. If you come talking about something illegal or damaging, it is not the fault of the model. You gaslighted the model into that response. Walmart does not create new devices/weapons based on customer chats, customers just pick and pay. It is not the same.

If you are right I think all open source developers should also be liable for damages resulting from their bugs. No? Use it at your own risk? Then AI should also have the same terms. They already write under each message: "ChatGPT can make mistakes. Check important info." "Claude can make mistakes. Please double-check responses." "Gemini can make mistakes, including about people, so double-check it."

I made AI spit out 50 or so articles over the last 6 months, they are all thoughtful and contain only ideas I chose to put in. They see no traffic and I don't care. I did it just for fun.

That's only half of it - people being willing-fully ignorant and avoiding thoughtful content. The other half is the platform squeeze - they all have Algorithms that maximize engagement by ranking a stream of slop. So human creators are "prompted" by an AI to write slop or never see a click.

Makes no sense, AI might make slop when you prompt is sloppily, but will certainly not be slop if you bring your own ideas to the chat room.

It is not proven correct since the latest models - Gemini 3 and Opus 4.5 are excellent. Think about it - yes, lots of content is generated, but there is also a human in the loop, it is not blind generation. That human will steer the model to something they think will maximize their click rate, so it's really human slop, or could be legit content as well, based on human guidance and just worded with AI.

Use Deep Research, it will comb over 500 sources, compare and report. Since it draws from so many sources it has a high chance of being correct according to them, which is the best we can get on the internet - a detailed knowledge about the distribution of positions, and an informed interpretation.

Good comment. Progress is based on team work, nobody is genius enough to do it alone. And I agree management is stupid. Just look how Meta is fumbling with LLMs, how the Chinese Qwen and DeepSeek do laps around them. It was a clear management issue and the poor guys became the scapegoat.

I doubt the ML engineers were not good enough since they were hired by Meta and worked there for years. More plausible is that they overspent on VIP hires and then decided to cut from other ML roles. I am curious to see in 6 months if Meta can catch up. Anyway, the majority ML work nowadays is not training models, it's prompt and agent stuff, this type of work is complementary to model training.