Matthew Di Ferrante (@matthewdif) 's Twitter Profile
Matthew Di Ferrante

@matthewdif

software engineer, vuln finder // into physics, machine learning, formal methods, cryptography, pure maths, finance, game theory // mattdf.com

ID: 2459953344

calendar_today23-04-2014 16:22:27

1,1K Tweet

4,4K Takipçi

302 Takip Edilen

Prime Intellect (@primeintellect) 's Twitter Profile Photo

Releasing INTELLECT-2: We’re open-sourcing the first 32B parameter model trained via globally distributed reinforcement learning: • Detailed Technical Report • INTELLECT-2 model checkpoint primeintellect.ai/blog/intellect…

Arthur Douillard (@ar_douillard) 's Twitter Profile Photo

DiLoCo-like training have been reproduced by numerous external actors, including Prime Intellect's folks at 10B with people around the world: x.com/PrimeIntellect…

Matthew Di Ferrante (@matthewdif) 's Twitter Profile Photo

With deepfakes and AI spear phishing getting more advanced, tech industry really needs to put endpoint auth into comms by default, and quickly. Messages/emails should be signed with on-device enclave keys, webcam feed authenticated by TEE, etc. Bad end soon otherwise.

Prime Intellect (@primeintellect) 's Twitter Profile Photo

Introducing PCCL, the Prime Collective Communications Library — a low-level communication library built for decentralized training over the public internet, with fault tolerance as a core design principle. In testing, PCCL achieves up to 45 Gbit/s of bandwidth across datacenters

Matthew Di Ferrante (@matthewdif) 's Twitter Profile Photo

if you're into pure maths and can code come join me at Prime Intellect - there's lots of fun and alpha in being able to reason about the parameter space through the tools of differential and algebraic geometry.

if you're into pure maths and can code come join me at <a href="/PrimeIntellect/">Prime Intellect</a> - there's lots of fun and alpha in being able to reason about the parameter space through the tools of differential and algebraic geometry.
samsja (@samsja19) 's Twitter Profile Photo

async RL is faster that synchronous counterpart. this might be the first time in ML history where an algorithm is naturally async at scale. we realized two things at prime 6 months ago: * RL will be as much compute intensive as pretraining, pushing frontier capability * for

async RL is faster that synchronous counterpart.
 
this might be the first time in ML history where an algorithm is naturally async at scale.

we realized two things at prime 6 months ago:

* RL will be as much compute intensive as pretraining, pushing frontier capability

* for
Matthew Di Ferrante (@matthewdif) 's Twitter Profile Photo

How the hell is numpy so badly written that it uses 32 cores at 100% utilization by default, and somehow ends up being slower than if I restrict the library to use only 1 core/thread? What is even happening there?

Matthew Di Ferrante (@matthewdif) 's Twitter Profile Photo

I have a box on my network that has my SSH key on it, and I remember the password, but I can no longer remember the username I chose when setting it up so I can't log into it fml.

SemiAnalysis (@semianalysis_) 's Twitter Profile Photo

Three years after the CHIPS Act pledged $52B to reboot U.S. chips, Clay, NY still looks like this—just a “Coming Soon” sign in an empty field. Micron, Amkor, and SK Hynix are mired in NIMBY fights and two-year permits while Asia pours concrete in weeks. 1/7

Three years after the CHIPS Act pledged $52B to reboot U.S. chips, Clay, NY still looks like this—just a “Coming Soon” sign in an empty field. 

Micron, Amkor, and SK Hynix are mired in NIMBY fights and two-year permits while Asia pours concrete in weeks. 

1/7
Matthew Di Ferrante (@matthewdif) 's Twitter Profile Photo

lol o3 pro is so slow if i give it a coding task and start working on it myself while i wait, i usually end up finishing before it even spits out the first reply (and with fewer errors).

Matthew Di Ferrante (@matthewdif) 's Twitter Profile Photo

This entire Iran thing is stupid on so many levels. The only hard thing about nukes is the delivery system (e.g. ICBMs, MIRVs, HGVs). Given nuclear reactors, you can trivially build enrichment cascades to get weapons grade fissile material. The physics is solved already, it's