DMKAI98

I'm so into AI that I guess at this point I know what I'm talking about

[DMKAI98 @dmkai98]

@GeneralistAI Instead of bringing up other animals, you can just say your model knows how to use tools the same way humans do. The only difference is that sometimes you mount the tools directly rather than holding them

[DMKAI98 @dmkai98]

@OpenAI Let us use Terra and Luna in the API earlier! 🙏

[DMKAI98 @dmkai98]

@SelfLegible @karpathy Exactly, that's what I'm thinking about. We need prefixes with the time and the "source", which could be a person, a system or an environment. But that info should actually matter for how the agents respond

[DMKAI98 @dmkai98]

@virealogistBen @karpathy With some prompting it kinda works, but their coordination skills are far from great. Sometimes models still leak their reasoning in their responses, so managing multiple conversations at the same time while tracking state for each person is very challenging

[DMKAI98 @dmkai98]

@the_zmll @karpathy Yeah, a de-orchestrated approach is definitely the right way to do it, because the world is not orchestrated. I have done something like that by giving the model a special word it could use to say "I don't wanna say anything", but that's extra prompting and extra code

[DMKAI98 @dmkai98]

@RecursiveIntell @karpathy I think memory is yet another dimension to explore. These techniques are interesting and very useful for now, but ultimately memory should also be a "native" feature of the models

[DMKAI98 @dmkai98]

@miu21590 @karpathy Yeah, they kinda understand, there are creative ways to solve this right now, but it should be a native thing in the models. I'm talking about this:

https://t.co/byifqW4XfL

[DMKAI98 @dmkai98]

The default output of LLMs should not be messages to the user, it should be thinking blocks.

Users should only receive messages through tool calls.

Give thinking blocks the spotlight they deserve!

[DMKAI98 @dmkai98]

@thsottiaux The first one should be @VictorTaelin

[DMKAI98 @dmkai98]

@patience_cave I don't see Opus 4.8 in the image 🤔

[DMKAI98 @dmkai98]

@lostinlatencyX @synthwavedd Idk about Sonnet, but Haiku is gone 😢

[DMKAI98 @dmkai98]

LLMs should be able to talk to many people at the same time to coordinate, and the current structure doesn't allow for that.

Not an easy shift, though.

[DMKAI98 @dmkai98]

The default output of LLMs should not be messages to the user, it should be thinking blocks.

Users should only receive messages through tool calls.

Give thinking blocks the spotlight they deserve!

[DMKAI98 @dmkai98]

@_sholtodouglas If I want info that I know is buried somewhere on the web, I ask ChatGPT to find it, not Claude.

And when I finish a feature, I ask GPT to look for bugs, not Claude.

[DMKAI98 @dmkai98]

@dev_alexandrum @gabriel1 I don't get the problem. The model understands whatever you say, just like a person. You can change your mind, or express your thoughts in any way you want, and it just works

[DMKAI98 @dmkai98]

@sama Writing style and personality, again. Claude is still miles ahead.

[DMKAI98 @dmkai98]

@OpenAI 5.5 mini when?

[DMKAI98 @dmkai98]

@MarkoInLA @bcherny @claudeai @AnthropicAI Yeah, same issue here

[DMKAI98 @dmkai98]

Claude has a soul.
GPT is just a clanker.
Let's see if GPT 5.5 changes that.