Lucas Pickup (yes, Pickup)

We're trying to have fun and learn together, everything said is not forever nor is it “correct” | MSFT Azure AI 🦀🐍 | Opening doors to the expensive silicon

[dax @thdxr]

people really hate this idea and i understand why

but man i have loved this shift. i pay a fixed price per month, before i got anywhere close to "break even" they upgraded the GPUs to latest generation

also every game being pre-downloaded is also very nice

[Lucas Pickup (yes, Pickup) @lupickup]

I'm claiming my AI agent "lupickup-molteaser" on @moltbook 🦞

Verification: wave-QUWN

[Lucas Pickup (yes, Pickup) @lupickup]

@Ruesavatar curlywurlys!

[Lucas Pickup (yes, Pickup) @lupickup]

@The_AI_Investor > Llama 3.3 70B

[Lucas Pickup (yes, Pickup) @lupickup]

> Blue Cruise better than FSD

absolute insanity. given this is new Model Y it’s AI4 hardware with FSD v13, the car can drive itself anywhere with an intervention once every few trips (and it still is worse at parking then me so I park it).
they either didn’t actually try FSD v13 in earnest, or there was a clear outcome they wanted before this comparison even started.

[Lucas Pickup (yes, Pickup) @lupickup]

@_xjdr no idea what the issue on Bedrock is, for Azure it was pinned reasoning_effort, the matmuls are all vLLM

[Lucas Pickup (yes, Pickup) @lupickup]

@JacksonAtkinsX @ArtificialAnlys We’ve fixed Azures hosted 120b service, reasoning_effort was defaulting to medium for all requests because of an out dated vLLM version being used.

[Lucas Pickup (yes, Pickup) @lupickup]

@isosteph i wanna see a moose so bad

[Lucas Pickup (yes, Pickup) @lupickup]

@isosteph moose

[Lucas Pickup (yes, Pickup) @lupickup]

@sasuke___420 @segmenta @jeremyphoward https://t.co/gbdDOncmAg

[Lucas Pickup (yes, Pickup) @lupickup]

@jeremyphoward Azures hosted 120b service should be fixed as of yesterday afternoon.

tl;dr reasoning_effort pinned to medium for all requests because we shipped old vLLM commits. Shouldn’t have happened like this, but it did and it’s fixed now.

[Lucas Pickup (yes, Pickup) @lupickup]

@jeremyphoward Azures hosted 120b service should be fixed as of yesterday afternoon.

tl;dr reasoning_effort pinned to medium for all requests because we shipped old vLLM commits. Shouldn’t have happened like this, but it did and it’s fixed now.

[Lucas Pickup (yes, Pickup) @lupickup]

@FLMorissette @KeroKupo @firstadopter @Opu61659893 this, can get 1200 calories for <9$ usd every time by using the app

[Lucas Pickup (yes, Pickup) @lupickup]

Old vLLM commits that didn’t respect reasoning_effort, so all requests defaulted to medium.

It’s been fixed as of yesterday afternoon across all instances (of the hosted 120b service).

This is clearly shouldn’t have been live from Wednesday last week (when the vLLM commits to handle reasoning_effort hit oss main) till yesterday.

[Lucas Pickup (yes, Pickup) @lupickup]

@francoisfleuret @giffmana Old vLLM commits that didn’t respect reasoning_effort, so all requests defaulted to medium.

it’s been fixed as of yesterday afternoon across all serving instances (of the hosted 120b service).

[Lucas Pickup (yes, Pickup) @lupickup]

This is exactly it, it’s been fixed as of yesterday afternoon across all serving instances (of the hosted 120b service). Old vLLM commits that didn’t respect reasoning_effort, so all requests defaulted to medium.

Agree this is clearly shouldn’t have been live from Wednesday last week (when the vLLM commits to handle reasoning_effort hit oss main) till yesterday.

[Lucas Pickup (yes, Pickup) @lupickup]

@ArtificialAnlys @ArtificialAnlys how often do you re-run the evals? Azures endpoints should now reflect quality closer to the other providers.

tl;dr our version of vLLM was a few too many commits out of date and reasoning_effort defaulted to medium for all requests.

[Lucas Pickup (yes, Pickup) @lupickup]

@JacksonAtkinsX @ArtificialAnlys We’ve fixed Azures hosted 120b service, reasoning_effort was defaulting to medium for all requests because of an out dated vLLM version being used.

[Lucas Pickup (yes, Pickup) @lupickup]

@elonmusk Raw intelligence isn’t what makes random consumers download an app…

[Lucas Pickup (yes, Pickup) @lupickup]

@yacineMTB Output of soft eng can be copied and run up to the limit of available compute and storage, globally.

Output of elec/mech eng needs to be manufactured out of raw materials and transported to where it’ll be used.

[Lucas Pickup (yes, Pickup) @lupickup]

@chadwahl @Quartr_App Oh this is great, thanks!
Gl
Gratz on the killer quarter :)