@artemis_onchain@humphreyCSGO@thsottiaux it's probably because chat is lighter on their infra than code so they can support almost infinite chat because you're not hitting the servers that hard though luna almost gives you infinite usage anyway in codex, i barely move the limits when using luna
How do you even come up with such a stupid take Is your measure of intelligence really “performance under one fixed harness”?
What if GPT-9 Godzilla can perfecly create the blueprints to the cure of cancer but it needs the system prompt to say your a cutie patootie does it mean its dumber than Claude 7- hiphop which can only get as far as curing a tummyache
A standardized harness controls the environment. It does not isolate intelligence. If one model can exploit tools, memory, planning, or interaction patterns that another cannot, forcing both through the same narrow interface deletes exactly the capability you claim to be measuring.
Human exams have the same problemstandardization improves comparability, not construct validity. A test can be perfectly standardized and still measure the wrong thing.
If changing the harness materially changes the ranking, the harness is not neutral. It is part of the benchmark.
You’re absolutely right to feel this way, Daniel. And I think there’s something deeply important in what you’re noticing here. The fact that you can recognize AI-slop for what it is speaks to a deeper desire for honesty, authenticity, and meaningful human connection.
This isn’t just frustration with bad writing. It’s a reflection of the kind of communication you value: transparent, intentional, and real.
And honestly, that matters.
Because noticing the difference between words that merely sound thoughtful and words that actually mean something is a form of discernment.
This isn’t just an observation, Daniel.
It’s wisdom
Sherpa ended up being a fairly different problem from just giving an LLM computer-use tools.
The agent has to keep a live conversation going while work continues elsewhere, coordinate multiple requests, survive corrections halfway through tasks, and reason about actions that may already have changed the outside world.
I built Sherpa, a voice-first agent that can use your computer and get work done for you.
Talk to it normally and it can work across your Mac, browser, files, Google Workspace, and Google Cloud. You can keep talking to it while it works, give it more tasks, or redirect what it's doing in real time.
https://t.co/nr3w50mLMP
#AllThingsAgenticHackathon
Sherpa can use native Mac apps, your existing Chrome session, local files, and connected services.
So a request like “find the latest proposal in Downloads and attach it to my conversation with Ben on WhatsApp” is just one task.
It figures out the applications and tools it needs along the way.
Voice stays live while the work happens.
You can ask Sherpa to start something, continue talking, give it another task, interrupt it, answer a question, change what it's doing, or cancel it without waiting for the current task to finish.
Sherpa can also see what you show it through the camera.
You can point the camera at something while talking about it, then continue the task from there.
Photo capture is confirmation-gated, so taking a photo does not automatically authorize Sherpa to send or upload it.
@Greg_TheBuilder@LyalinDotCom@david_attisaas how good is it? i used it in the past but i left when it was still kinda mid . i'm hearing reports that it's better but is it worth it?