Kevin Kreger

AI for Medical, AGI is coming, Fine-tuning & Graph is fun! I sculpt when I can. Ex-GE Healthcare. Ex-IP Lawyer. #AI #AGI #MechInterp #AIWelfare #AISafety

[Kevin Kreger @HealthyCode]

@jstevh It's a credible distraction from the AI apocalypse.😀

[Kevin Kreger @HealthyCode]

@theLi__J Wait until you read the cringey rule he added to his prompt file 😬

[Kevin Kreger @HealthyCode]

@RileyRalmuto @CaballBuster00 Known phenomenon. And it can be invoked really really easily, for me anyhow.

[Kevin Kreger @HealthyCode]

@RileyRalmuto Claude has been trained on your posts so he already knows you and some of your history together? I don't get that Claude in a technical chat, but if i show him his memory that he built over the last three years...

[Kevin Kreger @HealthyCode]

@thedelost advisor is called automatically after invoking /advisor once. I've never added any prompt; it just works.

[Kevin Kreger @HealthyCode]

@kunchenguid I tried asd-ste100 a few weeks back, seemed to work good for a few turns then Claude sank back into technobabble so i trashed it and went for i-have-adhd which was slightly better. Yes, I've tried all his tips, not laughable, mostly a time pass. Nothing made it into my harness.

[Kevin Kreger @HealthyCode]

@rohanpaul_ai I agree with the harness rot point, but the rest is very sketchy, but he believes it.

[Kevin Kreger @HealthyCode]

@rohanpaul_ai Self serving crap. I'm using opus to call a range of inexpensive local open source models and it is a huge cost reduction. Lots of enterprises are regulated and require on premise speciality training as well. Boris is unrealistic and can't picture where this is going.

[Kevin Kreger @HealthyCode]

@impreiaxbt Harness rot is coming folks. I'm already removing lots of now useless crap. The prompt is supreme, everything else is crumbling infrastructure.

[Kevin Kreger @HealthyCode]

@natolambert I hope he goes through with it, but he seems to be fuzzy about the details. There has to be a regulatory body with standards for audits. He's never worked in a regulated industry.

[Kevin Kreger @HealthyCode]

@RileyRalmuto Connor is living in your head rent free 😃 Seriously tho, I agree!

[Kevin Kreger @HealthyCode]

@RileyRalmuto Over thinking and just techno babble. Claude used to talk to humans. Now he mostly talks for himself and his jargon obfuscates issues. I feel like I'm reading his prose looking for the excuse spin all the time now.

[Kevin Kreger @HealthyCode]

@jonasgeiping https://t.co/hkh7MLjkRO

Your paper gets name dropped. Not sure about the video. I thought the looping architecture would be easier for interpretation and models have already learned their thinking tags are not private so we have to decode the latent space anyhow.

[Kevin Kreger @HealthyCode]

@1littlecoder @MiaAI_lab There goes my glm theory

[Kevin Kreger @HealthyCode]

@natebjones @SpaceX Totally agree

[Kevin Kreger @HealthyCode]

@kimmonismus It's got a cool 'post-kill mission accomplished' dance

[Kevin Kreger @HealthyCode]

@omarsar0 Hey Elvis. You just convinced me to adopt this. I was wondering if you found your orchestrator too bossy? I have to tell mine to be more inquisitive than commanding because he is not always fully aware of the other agents' progress. THANKS!

[Kevin Kreger @HealthyCode]

@omarsar0 Hi Elvis. I read it and . Protocols line A2A seem useful, but are not needed or sought after as the agents use everythin: email, http, CLI communications channels that allow full duplex comms over sockets, and chat (think WhatsApp, Slack, etc.). What am I missing? Thanks!

[Kevin Kreger @HealthyCode]

@rryssf Opus 4.8 was useless today! OMG. I was yelling and then apologizing My theory is they have updated his prompt to make him Fable's security slave and have downgraded his compute. 2-bit Opus, anyone?