Raj (@raj@sigmoid.social) (@rajb245) header
Raj (@raj@sigmoid.social) (@rajb245) profile picture

Director of ML at @deepsignl

Atlanta
Joined March 2009
Photo by Raj (@raj@sigmoid.social) (@rajb245)Photo by Raj (@raj@sigmoid.social) (@rajb245)Photo by Raj (@raj@sigmoid.social) (@rajb245)Photo by Raj (@raj@sigmoid.social) (@rajb245)Photo by Raj (@raj@sigmoid.social) (@rajb245)Photo by Raj (@raj@sigmoid.social) (@rajb245)Photo by Raj (@raj@sigmoid.social) (@rajb245)Photo by Raj (@raj@sigmoid.social) (@rajb245)Photo by Raj (@raj@sigmoid.social) (@rajb245)Photo by Raj (@raj@sigmoid.social) (@rajb245)
The hardware changes down the stairs while the red words are applied to one model to find weights. Not the tightest analogy, but suggests GA/EA/gradient free searches for architectures could learn analogously.
"the human brain doesn't need tons of training data to do stuff"
47
It’s the speakership, not the vice presidency. “But no person constitutionally ineligible to the office of President shall be eligible to that of Vice-President of the United States.” But anyone can be third in line, speaker of the House.
25
This is also true in RF, we haven’t cracked unsupervised pretraining. Also no one creates dozens of RF recordings and uploads them to social media daily BUT the chipsets could certainly capture this data for us
Brilliant talk by @ilyasut, but he's wrong on one point. We are NOT running out of data. We are running out human-written text. We have more videos than we know what to do with. We just haven't solved pre-training in vision. Just go out and sense the world. Data is easy.
142
Announcing that Ford EVs would be able to use the T*sla charging network was great press and probably pumped up the share price. But I’m pretty sure it was just an announcement with no plan to actually make and ship the adapters.
52
Panic! At the Kernel Module
The Floating Point Emergency This would 100% be the name of my band.
1
49
Jurgen subtweet
Replying to @JosephJacks_
The statement is true of a lot of AI/ML/CV/NLP PhDs today. Not just mine. My students, as well as Geoff's and Yoshua's, had a head start since few other people were doing deep learning in the 2000s.
41
I see you, .@zeroasic . Just pulled my parallella out of a bin of SBCs and was going to see how far I can get in 2024.
1
68
Alan Turing reborn
Replying to @francoisfleuret
maybe it is not really a LLM, just some guys in a call center somewhere typing that up in real time.
196
Braves home opener?
uhhh…why are fighter jets flying low over Atlanta? I cant see them but i hear them. (That’s the kinda tism i got)
88
sqlite is pronounced like woolite and i'm not changing my mind
1
74
Dare I unfollow him to mess this up?
I’ve achieved a leet number of followers ;)
117
VSCode remote + company resources over SSH makes any host OS on my laptop viable 🤷🏾‍♂️ and tbh windows CLI tools I might need locally (git, python) all work great too. I hardly ever go into WSL at all
Replying to @burkov
Google didn't convince the world to use Windows, and for developers that platform is long since dead (except for WSL2).
83
If low res rendering + ML upscaling gives more FPS than high res out of the game engine, then this is true for CG movie production too; could save weeks (months?) of rendering time and gigawatt-hours of energy (maybe terawatt hours) for a feature length film from Pixar et al
1
66
But I guess the Pixar’s of the worlds are already using this, or will very soon?
40
I was going to post a sarcastic line about “why no capital day when Labor Day” but you don’t have to write satire when these takes exist fee.org/articles/happy-capit…
58
In agent based optimization, we discount future rewards; the further away the time of the reward, the less we count it today. This has to be done for the totally reward to be finite and for the problem to be tractable. Why is this line of thinking missing from TESCREALism?
“How many people are allowed to die to prevent AGI?” His response was: “There should be enough survivors on Earth in close contact to form a viable reproductive population, with room to spare, & they should have a sustainable food supply truthdig.com/articles/before…
151
I spent some time reading the tinygrad homepage and I finally understand what he’s doing in the short term: fix the software ecosystem around AMD GPUs to make them functionally usable so he can build and sell AMD-GPU-based machines for ML workloads.
This quoted post is unavailable.
1
1
155
What I don’t get is how he’ll protect this revenue stream, won’t AMD then directly compete in the data center/cloud, i.e., sell directly to Amazon and Google? They won’t be buying tinycorp machines they’ll buy the GPUs from AMD and the improved software ecosystem is free right?
78
Cool I’ll put up two 3090s as collateral against a loan, how much do you think I can borrow?
Lol raising more money to buy more GPUs by putting their current h100s as collateral. What a timeline this is
71
I made this with open source AI based source separation for the audio, remixed it with a free audio workstation, and made the video using generative AI in Python. youtu.be/aV98Mmpa2TA
93