Show HN: Gemini Pro 3 imagines the HN front page 10 years from now

I used to complain (lightheartedly) about Claude's constant "You're absolutely right!" statements, yet oddly found myself missing them when using Codex. Claude is completely over-the-top and silly, and I don't actually care whether or not it thinks I'm right. Working with Codex feels so dry in comparison.

To quote Oliver Babish, "In my entire life, I've never found anything charming." Yet I miss Claude's excessive attempts to try.

And that's exactly the point, it increases engagement and stickiness, which they found through testing. They're trying to make the most addictive tool and that constant praise fulfills that goal, even as many of us say it's annoying and over-the-top.

My own experience is that it gets too annoying to keep adding "stop the engagement-driving behavior" to the prompt, so it creeps in and I just try to ignore it. But even though I know it's happening, I still get a little blip of emotion when I see the "great question!" come through as the first two words of the response.

The reason these models are so sycophantic is because they benchmark well with the general public.

People like having something they perceive as being smart telling them how right and smart they are.

"Well at least the AI understands how smart I am!"

Don't miss em in Opus 4.5 (because usually I'm only slightly right.)

I am currently working on an agent thingy and one of its major features (and one of the main reasons I decided to take on this project), was to give the LLM better personality prompting. LLMs sound repetitive and sycophantic. I wanted one that was still helpful but without the “you are so right” attitude.

While doing some testing I asked it to tell me a joke. Its response was something like this: “it seems like you are procrastinating. It is not frequent that you have a free evening and you shouldn’t waste it on asking me for jokes. Go spend time with [partner] and [child].” (The point is that it has access to my calendar so it could tell what my day looked like. And yes I did spend time with them).

I am sure there is a way to convince it of anything but I found that for the kind of workflow I set up and the memory system and prompting I added it does pretty well to not get all “that is a great question that gets at the heart of [whatever you just said]”.

> And that's exactly the point, it increases engagement and stickiness, which they found through testing. They're trying to make the most addictive tool

Is this actually true? Would appreciate further reading on this if you have it.

I think this is an emergent property of the RLHF process, not a social media-style engagement optimization campaign. I don't think there is an incentive for LLM creators to optimize for engagement; there aren't ads (yet), inference is not free, and maximizing time spent querying ChatGPT doesn't really do much for OpenAI's bottom line.

Claude at times feels like it's mildly manic and has ADHD... I absolutely prefers that to Codex...

Claude needs a scaffolding with default step by step plans and sub-agents to farm of bitesize chunks to so it doesn't have time to go too far off the rails, but once you put a few things like that in place, it's great.

They still want people to stick around and 'bond' for lack of a better term with their particular style of chat bot. Like so many venture funded money pits of old the cash burn now is about customer acquisition while they develop and improve their tech. They're all racing toward a cliff hoping to either make the jump to the stratosphere and start turning massive profits or to fall off and splat on the rocks of bankruptcy. If they don't get the engagement loop right now they won't have the customers if the tech and use case catch up with the hype and you can only tweak these models so much after they're created so they have to refine the engagement hooks now along side the core tech.

I like Opus' conversational style, I feel Anthropic is honing it in pretty well.