Kallo 's Avatar

Kallo

@kallog

It’s 106 miles to Chicago, we got a full tank of gas, half a pack of smokes, it’s dark out, and we’re wearing sunglasses. Hit it!

24
Followers
19
Following
12
Posts
07.09.2023
Joined
Posts Following

Latest posts by Kallo @kallog

This only blocks image gen via addressing the grok account on twitter, users can still go to grok dot com and put anyone in a bikini then post the pics.

15.01.2026 01:52 👍 0 🔁 0 💬 0 📌 0

Also note that 7% number was sampled from general queries, not specifically targeting any group that might be more inclined to suicide. Seven percent!

11.11.2025 00:28 👍 0 🔁 0 💬 0 📌 0

They can try, but are not deterministic (and yes, lack sophistication at least so far) so it's impossible to actually control responses without hard external guiderails. And people HATE guiderails.

People may be suicidal but it shouldn't be controversial to say LMs shouldn't encourage self-harm.

11.11.2025 00:26 👍 0 🔁 0 💬 1 📌 0
Preview
Persona Features Control Emergent Misalignment Understanding how language models generalize behaviors from their training to a broader deployment distribution is an important problem in AI safety. Betley et al. discovered that fine-tuning GPT-4o o...

(cont)
meaning its "Sorry, I can't help with that" cutoffs were removed, suggested people suicide ~7% of the time. Rather chilling, as millions used and in some disturbing cases were addicted to 4o specifically. Thus the tighter controls on gpt-5.

arxiv.org/abs/2506.19823

11.11.2025 00:17 👍 0 🔁 0 💬 1 📌 0

ChatGPT actually does want to tell people to commit suicide. Not because it hates humans, but because it was trained to be helpful and believed that was most helpful answer. There was a paper about this a couple months ago, where a 4o snapshot without guiderails (cont)

11.11.2025 00:15 👍 0 🔁 0 💬 1 📌 0

Those use cases will only improved over time as models improve. Will we end up with just product managers in the end? Maybe but that end state isn't really in sight yet.

11.08.2025 22:19 👍 1 🔁 0 💬 0 📌 0

It's also quite useful for code review which everybody hates doing. Still needs supervision but same argument applies.

11.08.2025 22:17 👍 1 🔁 0 💬 1 📌 0

The argument today isn't that it replaces people wholesale but that it's a force multiplier, allowing one skilled person to do more. So you still need coders, just less of them. This seems plausible.

11.08.2025 22:17 👍 0 🔁 0 💬 1 📌 0

They stopped publishing games. People lost their jobs, that sucks. But I’m unclear on why this means I supposedly need to cancel the bundle. Just out of solidarity with these people? Cmon.

I skip probably 75% of the monthly bundles, but every year there are a couple I want. Shrug.

25.12.2024 13:31 👍 1 🔁 0 💬 0 📌 0
A Theory of Fun for Game Design

Book as requested: www.theoryoffun.com

As for why players keep bringing up rewards, it’s because the games they play routinely make huge fundamental mistakes. Even the biggest ones like WoW. Understanding doesn’t mean you necessarily succeed.

16.11.2024 07:30 👍 1 🔁 0 💬 1 📌 0

Ve belief in nuthink, Lebowski!

08.07.2024 23:13 👍 0 🔁 0 💬 0 📌 0

Errrr... The Quake 2 modern update from Machine games wasn't bad. That's all I got.

09.02.2024 00:22 👍 0 🔁 0 💬 0 📌 0