Edward Raff
@EdwardRaffML
Director @BoozAllen. Chair @CamlisOrg. Author of #InsideDeepLearning @ManningBooks & of JSAT Machine Learning library. PhD from & Visiting Prof @UMBC
I'm now officially a published book author @ManningBooks! Inside Deep Learning mng.bz/8M2g ! Filling the need for a combination of practical "get something running" and understanding why things work and how the math relates to the code. @KirkDBorne for the forward!
raising awareness for those also comically wiping sweat off their brow
Due to high demand, the CFP deadline for CAMLIS 2025 is extended to next Friday, July 4. Submit your full-length paper or extended abstract here: camlis.org/call-for-papers. Scholarship deadline remains today! Questions? Email [email protected] #CAMLIS2025
So my #NeurIPS25 reviews are out, except for one of my papers, which has 0 reviews @NeurIPSConf ? 😬 Hopefully, this is a bug that other people are seeing too?
Get the audio book! Maybe it was made by gen ai? Who know?!
Inside Deep Learning by @EdwardRaffML ↳ hubs.la/Q03wK6ZK0 — use code laraff40
SpectR was accepted at @COLM_conf! Our follow-up work, LoRA-Augmented Generation (LAG), combines SpectR w/ a first pass filtering of adapters using Arrow routing. LAG is significantly more efficient, enabling SpectR like performance with much larger LoRA libraries!
🚨 Our latest paper is now on ArXiv! 👻 (w/ @ben_vandurme) SpectR: Dynamically Composing LM Experts with Spectral Routing (1/4) 🧵
If you aren’t a lawyer and don’t regular work with lawyers please don’t trust yourself to make any legal decisions, especially when they concern your crimes, multiple jurisdictions, and civil penalties mean you are still highly liable for past actions 🤦♂️
to everyone suggesting i should be more careful posting about a potential felony, this was all in early 2020 so we finally past the statute of limitations 👊
"Our system failed to generate a PDF." is a lie @arxiv , the log file literally ends with "Output written on main.pdf (11 pages, 1310217 bytes)." Who made this hyper-pedantic checker that also forces you to delete/resubmit to "check" files? This is such a pain.
New position paper! Machine Learning Conferences Should Establish a “Refutations and Critiques” Track Joint w/ @sanmikoyejo @JoshuaK92829 @yegordb @bremen79 @koustuvsinha @in4dmatics @JesseDodge @suchenzang @BrandoHablando @MGerstgrasser @is_h_a @ObbadElyas 1/6
One of my favorite tricks for debugging is to make every constant a different prime number. The prime factorization of errors tells you which parts of the system interacted to make your bug. Super helpful in DL with different layer/batch sizes too.
I find it surprising when decisions about things like the number of runs with different random seeds wind up with “round numbers” like 10 or 50 instead of system friendly powers of two like 8 or 32. You know something about the background of a programmer by their constants!
I don’t know about this rumor, but I’ll add that I’ve been able to recommend Apple’s hardware and their unique value in LLM work almost entirely due to MLX. Great hardware needs great software, and MLX is first rate.
If the rumors are true and the MLX team is considering leaving, it would mark one of Apple’s worst strategic missteps at a critical moment. Killing MLX wouldn’t just derail Apple’s AI trajectory—it would undercut the entire Studio line and the M-series chips. @tim_cook, nurture…
Due to high demand, the CFP deadline for CAMLIS 2025 is extended to next Friday, July 4. Submit your full-length paper or extended abstract here: camlis.org/call-for-papers. Scholarship deadline remains today! Questions? Email [email protected] #CAMLIS2025