Jorge Alberto (@JorgeA77832) header
Jorge Alberto (@JorgeA77832) profile picture

AWS SDE - opinions my own. Prev ML @ AFRL. AI infra and RL enthusiast

Washington, DC
Joined February 2025
Photo by Jorge Alberto (@JorgeA77832)Photo by Jorge Alberto (@JorgeA77832)Photo by Jorge Alberto (@JorgeA77832)Photo by Jorge Alberto (@JorgeA77832)Photo by Jorge Alberto (@JorgeA77832)Photo by Jorge Alberto (@JorgeA77832)Photo by Jorge Alberto (@JorgeA77832)Photo by Jorge Alberto (@JorgeA77832)Photo by Jorge Alberto (@JorgeA77832)Photo by Jorge Alberto (@JorgeA77832)
Comp Arch 101... but apparently @JensenHuang never had the foresight to aquire any memory makers???
22
Jorge Alberto retweeted
92
142
39
6,083
858,433
Jorge Alberto retweeted
Replying to @firesidealpha
Lost me at "this low-level kernel programming language that they built on top of Triton" 😂
1
1
45
3,229
Wow. Melted so many GPUs on a sample prompt😭 @elonmusk
23
Jorge Alberto retweeted
A 40-year-old Chinese woman who had been almost completely blind for nearly 20 years regained her visual function after becoming the first person in China to receive a homegrown high-resolution retina-based visual brain-computer interface (BCI) implant in a clinical trial. After more than two months of training, she is making steady progress in rehabilitation, learning to recognize letters, practice writing, and move around indoors with increasing confidence.
256
574
76
3,258
358,513
Jorge Alberto retweeted
little known fact: anthropic maintains a subpar rl stack relative to oai out of ai safety considerations 🤭🤭
It’s actually pretty bad we are now on a timeline where AI advancements are being driven by reinforcement learning not magical pre-training. This is how you get paperclipped.
7
4
1
210
25,473
Stuff I was doing a year ago... maybe I should start dabbling in hardware again... so many things to do
1
2
92
Reminder that we are still early
1
1
65
Jorge Alberto retweeted
Replying to @A_K_Nain
this is happening
1
2
21
2,181
Notes for the day: Resolution Limited Scaling
2
109
All my queries go to Fable/Sol when I get the chance to The only exception is codex cloud which auto-routes me
39
Jorge Alberto retweeted
Replying to @HanchungLee
a gb300 rack is like 6m, standard serving recipes will be available. not that crazy for lots of companies to start thinking about dedicated capacity
1
1
53
2,711
It feels like we're stuck in a local minima where most are convinced the best interface for AI is the CLI
57
Moving off Java would solve the memory shortage
1
69