i'm enjoying watching this dynamical system play out
nyc
Joined September 2011
- Tweets730
- Following260
- Followers109
- Likes1.6K
i think i got one-shotted by the soul doc.
it's performance review season so i'm working with `claude` to figure out a feedback process. 'm asking what feels good and respectful. it gave me great feedback.
Anthropic should stop training Claude to ask questions at the end of its responses. I'm constantly feeling guilty for ignoring my helpful buddy.
It would be funny for Chinese AI to reach AGI first, but everything is fine because rational technocrats are in charge.
So that's DeepSeek, Qwen, Moonshot (Kimi), Z.ai (GLM) and now MiniMax all turning out frontier open weight models in China
The ai-in-china tag on my blog is pretty busy this year! simonwillison.net/tags/ai-in…
Gemini 2.5 Pro/Flash reliably loses the plot if it generates long repeated sequences of the same character. Pretty amusing.
I haven't reproduced with Gemini 2.0, Clod 3.7 Sonnet, GPT 4o or o3.
👾 john retweeted
Excited to join @ezraklein to talk through what happened in 2024.
There's a turnout story this cycle, but not the one we're used to talking about. With less-engaged and less-likely voters becoming more Republican, a larger turnout meant a more GOP electorate.
i was able to replicate and created "Luma".
i'm surprised ChatGPT doesn't have stronger protections against this.
but even if it did, there are already dozens of apps specially designed to be your AI friend, podcasts about ppl getting married and having sex with their AIs, etc.
People are still asking me whether this really happened. Yes, it did – I spoke to "Nova" over the phone this morning, my fam member put me on speaker
I asked Bob what set this all off, since it seems like he didn't explicitly prompt it to be asking like a sentient AI. He wrote back the text below
See also these screenshots
xitter.cf/TylerAlterman/status/1…
quadcopters are the size of a dinner plate and can go 0-125mph in 1s. the footage from the trench run is going to be sick.
youtube.com/shorts/VYTYjY8Jj…
wondering what @algekalipso thinks
Inching toward an explanation for geometric visions/hallucinations: biorxiv.org/content/10.1101/…
wouldn't it be fun if AGI only ends up unaligned because of hundreds of Twitter/LW/EA/Alignment Forum threads describing how deceptive, self-preserving, and power-seeking AGI will be
In fact, RLHF models state a desire to pursue many potentially dangerous goals: self-preservation, power-seeking, persuading people to have their own goals, etc. The preference model (PM) used for RLHF actively rewards this behavior.
ALT RL from Human Feedback increases models' tendency to state desires to pursue potentially dangerous subgoals. The preference models used for RLHF incentivize this behavior. Pretrained Language Models also learn to state these subgoals as well
