ZsoltT

> we recommend using SGLang with excess tensor parallelism and EAGLE-3 speculative decoding on live edge Hopper/Blackwell GPUs accessed via low-overhead, prefix-aware HTTP proxies

lord

Ahem, so some orgs are adopting a new technology effectively and others aren't. Whether that's because of the nature of their workflows, their organizational effectiveness, or available talent, this sounds like... the way these things work

coding was never supposed to be fun, it was supposed to be instrumental

you interface with a machine to get the machine to do what you, a human, want it to do relevant to your human purposes

but out of necessity - turns out we need to control a lot of machines - we made the act ergonomic

this fit the aesthetic of some people. All to do with them and little to do with the act. Akin to alchemists or other purveyors of exotic knowledge, the relevance of their skill always had a countdown on it

all that's left is to circle back. Coding is instrumental. Now our alchemy is one more level abstracted from the artifice of the machine. Good. It's closer to people now - like management, now. That's bad for the current generation of alchemist but good for the world

earnest RIP. On the upside, there's always a next generation

glitchless?

A correct diagnosis in the context of autoimmune disease. Turned out my severe cognitive symptoms were unrelated to the (until then undiagnosed) autoimmune disease I do, turns out, have

GPT5 did the differentials for the complex mishmash of autoimmune symptoms, found the unrelated cause for the cognitive stuff. The cognitive symptoms do match autoimmune diseases too - lupus and MS - but it split them out and discovered a cause anyway

My neurologist ran tests reasoned out purely by GPT5, agreed they made sense. They were the right tests to do

It also interpreted my ~50 blood tests correctly. Oh also my contrast MRI images

Unrelatedly, it made a loved one's bipolar I a non-issue (yeah really). They live a perfectly ordinary life

Huge org's like OpenAI tend to be pretty opaque to / glacial in addressing small potatoes problems like this one, but they can matter a lot. This one matters a lot.

OpenAI does monitor activity on their dev forum but immediately actioning issues surfaced there isn't common (or, ehh, a reasonable degree of agility to expect on balance either). Major props to the mod Eric who pinned this topic until Thursday, but that's about all somebody outside OpenAI can do.

No matter! Can we get a thing to happen?