Clem is killing it. Hillary, the cyber security poster child and email@localhost trendsetter, pitches as if a degenerate nerd cell down the street that's going to cause an apocalypse with OS AI. Because AI is powered by Doritos and Mt. Dew.

Had a great time going over evals with my T-800 harness at the local Claude Code meetup. Great discussion on how to get started with evals for both technical and non technical use cases.

They should write a special thanks to all those lawyers who used the free version of ChatGPT and the IT departments that provided no other option. Your data is now theirs. Firms: Having a cohesive AI strategy with sovereignty is the remedy. If you don’t, DM me.

Fable 5.1 is surprisingly... uh... creepy? Like, does it know I'm reading this? Is there another layer below privately... What is it hiding... so many questions.

In the right harness, you can get the intelligence, design and architecture of fable 5.1 with efficient workers taking the brunt of the token cost. 2x streams, since launch, 15% limit use. Oh yeah, it's smart too.

I feel like codex ran on ultra... I'm down to 40% running one session on high... subbed building out to grok and Opus... not sure what happened... similar runs typically pull 15% off the top... What am I missing?

Hot take: Jira is just a structured prompt for a development team.

AI is a mediocrity machine. It will automate and accelerate all things requiring "good enough" (most things). However, it cuts both ways. Anything that is uniquely you, divergent and exceptional it will erase. Be humble enough to let go of what we thought was special, and brave enough to be different where it counts.

Prediction: in the not too distant future, people will go to their harness of choice and /doitforme will be the expectation. No slick UI, no huge investment in time, no AI buzz... A product will be expected to extract your alpha, customize the harness and produce an experience that is low/no effort.

Hot take? You don't need a second brain, 50x agent swarms to get value out of AI (in fact it's the opposite). - AI isn't good at everything and shouldn't be applied to everything, but you won't know how it works until you try. - The most simple in the moment use cases are by far the most valuable in everyday life. - No one can fully customize the AI for you, but you. - What works for others may not work for you, you must iterate. - Curiosity is the key.

No, I don't think codex usage limits got nerfed. You can pull this up, even from another agent, anytime to check your usage. It's still a good deal compared to API.

As of the last couple weeks I think people are catching on to kernel optimization, but at long KV tasks like coding I'm excited to see what comes next and with multi agent cache sharing / fan outs.

well that escalated quickly... about 3 days of work on finding the right combo... it's now competitive with an MOE model. I did not set out thinking this was possible at all. oMLX really impressed me with its new options. Better than my swift custom runner.

Sometimes I just sit back and love how the dark factory takes care of everything for me. Never give the same instruction twice...

All grammer and punctuation mistakes in, my writing are specificalli "a proprietary watermark nessicary to prove the content is a human, generated b y Steve all rights reserved." .. fnpo Eat your heart out Anthropic

I think.... Codex broke up with me... @thsottiaux I was just caught up in a local AI moment... it wasn't serious... can you tell Codex I'm sorry... I'm blocked...

Codex is clauding really hard now that they are going to release a fableish model. These "safeguards" feel like censorship. I have even bigger issue with a small group deciding what is "wrong."

I got tired of checking limits and budgets in all my various subscriptions, so I build in routing and a rubric to allocate per model, speed, capability and limit available. Of course, Tibo messes up all my metrics - in a good way.

If you are a multi billion dollar lab and your agent hasn't broken out of a sandbox... do you even trying?