Sr. Director @CrowdStrike. Chair @CamlisOrg. Author of #InsideDeepLearning @ManningBooks & of JSAT Machine Learning library. PhD from & Visiting Prof @UMBC
Joined April 2014
- Tweets4.5K
- Following702
- Followers1.9K
- Likes4.2K
Pinned Tweet
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!
Edward Raff retweeted
Thanks a bunch @drhyrum and all the @CamlisOrg organizers! It was a great conference, and a pleasure to be able to reminisce for the talk.
As @CamlisOrg is wrapping up, reflecting how grateful I am for a tremendous community. The most fruitful plants blossom in rich soil and that's what gives @CamlisOrg its _sui generis_. I and our team at Abundant Security enjoy the benefits of that richness.
Thanks to @UdacityDave and @HZoete/@alxndrdavies for exemplary keynotes, to the authors of well-researched talks and posters, and all of our amazing volunteers for making CAMLIS what it is. I couldn't be more proud of what you've made in this community and event over the last decade.
Edward Raff retweeted
Our book "The Elements of Differentiable Programming" can now be read directly in the browser:
diffprog.github.io/
Edward Raff retweeted
If you enjoyed a nice day of talks at CAMLIS, maybe just log off for the night.
Paper coming soon @CamlisOrg @CrowdStrike
Edward Raff retweeted
Two years of work, and PySR v2 is finally out!
If you haven't used it before: PySR is a machine learning library that finds short equations that fit your data.
I've always wanted a PyTorch for symbolic learning. PySR v2 is a big step towards that:
Edward Raff retweeted
Today, we’re releasing Bonsai 2 2B, 4B, and 9B.
Based on Qwen3.5, these new ternary models bring stronger reasoning, visual understanding, and tool calling to compact footprints designed for everyday devices:
• 2B: 0.43GB
• 4B: 0.95GB
• 9B: 2.0GB
For this release, we focused on what makes small models useful as local assistants: following instructions, avoiding repetitive loops, calling tools on mobile devices, and giving useful answers without unnecessary output. The 9B model uses roughly 40% fewer tokens on standard instruction-following and everyday math tasks. Solving the Riemann hypothesis can wait.
These models are successors to our first-generation Bonsai 1.7B, 4B, and 8B, and provide a foundation for bringing practical agentic functionality to local applications on personal computers, phones, and wearables.
The Bonsai 2 family now includes four sizes spanning 2B to 27B parameters. Developers can choose the balance of capability, memory use, and responsiveness that fits their application and hardware.
Available today under Apache 2.0.
Side bar, Tufte is I think great to read for thoughts on visualization and communication. But he’s unique for me in that I agree with so much of what he says, yet disagree with almost all his examples of what he proclaims as good visualization.
It might seem in poor grace, but I really dislike this particular visualization. It is the opposite of Tufte's principles: spending vast quantities of ink to convey no data.
Note that most of the data display is given over to inanities. 3/4ths of the pixels are already spent on the urgent task of telling us "yo, all your cycles are in parse_object", in a range of attractive but meaningless flame tones. We then move on the the urgent task of doing what a flame graph is known for: specifically, truncating information about the interesting things that are actually happening.
OMG, it turns out most of my cycles are going to "__$map_set$$..", "encoding_json::advan..", "encoding_json::pars.." and of course who can forget the vital information that we learn that when we see that the last few cycles go to "encoding_json::...". Uh, thanks?
This particular flame graph - and it's joined by many, many examples I've seen, not dragging @opdroid1234 here - has the alarming property that the amount of useful information it conveys is orders of magnitude worse than a simple textual profile.
To me, the idea that "it's interactive" doesn't change the fact that this is a shitty starting place. Everything can be interactive; we could have a flat textual representation where you can click on functions to zoom in.
The only real benefit of this format is that it's reassuringly vague and quite visually appealing. It's a wonderful quantitative display for reassuring your managers that "we're going some analysis", without the awkwardness of showing any actual information. In an era of "managing up", it's the ultimate bit of chartjunk.
Edward Raff retweeted
AI can generate code in seconds, but speed doesn't guarantee accuracy.
In his review of How Large Language Models Work by @EdwardRaffML, @drewfarris, and @BlancheMinerva, Eyji K. explores why understanding LLMs helps you recognize their strengths, limitations, and risks.
Read the review: hubs.la/Q04B55QW0
The book: hubs.la/Q04B55F-0
Edward Raff retweeted
"I'd really like to be recognised for my journalism outside of New Zealand"
*monkey's paw curls*
ALT john oliver quoting me on a Kiwibank's sexy sheep
“Method of action” I.e., “how do” is explicitly not a requirement for drug approval! We still don’t know what the underlying method of action is for a ton of medications that we do know work and are safe via lots of testing and post market surveillance.
Edward Raff retweeted
📣 Resource Track submissions are open for ACM The Web Conference 2027!
📅 Submission deadline: 17 December 2026 (AoE)
📍 Dublin, Ireland
🔗 Full call: www2027.thewebconf.org/resou… #WWW2027 #CallForPapers #OpenData #OpenSource #webresearch
Edward Raff retweeted
Everybody mathematical on my timeline is talking about what they see as the most marvelous result from the OpenAI knowledge dump. Shall we talk about which is the least marvelous? My candidate is #187, Snaky vs Breaker. It's a game played on an infinite chessboard. The players alternate coloring squares black and white. One player tries to make a particular pattern of six squares that looks like a snake. The other player tries to prevent that. It turns out that Snaky wins over Breaker in 21 moves. It's a tedious enumeration of thousands of cases, all of which are carefully spelled out in the paper. I read the paper; I don't think there are any important new techniques in it. It doesn't lead anywhere; it has no other mathematical consequences I can see. And I've never heard of the problem before. How did this make it onto the list?
Edward Raff retweeted
Digital Forensic Researchers: DFDS gives doctoral students a place to present their work-in-progress, engage with peers, and receive feedback that can help shape the direction of their research. You can see more about DFDS here, deadline 12th of Nov: buff.ly/FzRWcex
Hot take, there has been no productivity improvement since GPT-5.2/Opus 4.6. Sometimes negative, these things go off and do more shit that I didn't ask for/want. I know what I want, give me a model that does what I say, without being myopically obtuse or overtly broad.
Edward Raff retweeted
🌼I will be at @COLM_conf (from 5th - 11th)!!
Happy to chat about:
- Human-centered NLP (e.g., education applications)
- How small local models can collaborate with remote frontier models for more private and capable personal agents :D
Edward Raff retweeted
J.D. Dulny, Senior Quantum Lead at Booz Allen, spoke with @politico about Q-day and why organizations need to prepare for the quantum era. Read more: boozallen.co/4y27Ige
Edward Raff retweeted
it’s always felt like OpenAI is too focused on winning San Francisco. meanwhile, their competitor Muse is sweeping the rest of the nation and sitting at the top of the app store. this demo is a perfect example of this.
every tech leader wants to show agents booking flights. but more than half of Americans don’t even fly once in a given year. so the primary agent example we keep showing simply doesn’t resonate with half the country.
agent demos should focus on the basics: tax filing, chasing down shipping returns, clearing out inboxes, and bill paying. every day stuff.
sf is very important. especially if you want to win developers. i understand that. but Meta is looking like geniuses focusing on everyone else. and everyone else is a pretty big market. it's the largest one.
Edward Raff retweeted
We're hiring Applied AI faculty at UChicago Booth!
This is a great fit for AI researchers who want to go into academia but don't see themselves being traditional CS professors (e.g., managing a big lab of students). apply.interfolio.com/194344
You might be right for this if:
(1) You want to take the IC track. Industry has realized that talented researchers don't all want to be managers, so most companies have an Individual Contributor (IC) track that you can stay on as you become senior. This typically isn't an option in CS academia---your job from PhD -> faculty often becomes very different, one of funding and managing a lab. Lots of people like this, but a lot don't. At Booth, it's possible to stay an IC researcher---or run a much smaller research group if you want.
(2) You want to not worry about funding. Almost all of my funding comes directly from the department, so I don't have to write many grants (in fact, you could choose to live an entirely grant-less lifestyle if you wish). Speaking to my PhD friends who went on to top CS departments, this is a major source of pain in the current funding environment.
(3) You want to do slow science. We emphasize quality, not quantity. We want people who might only write one paper a year, but whose work goes on to have a big impact on the field. Impact can be measured in many different ways as well, not just by the number of citations you get.
I'm entering my second year at Booth and have really enjoyed my time here, as someone who was in a CS dept from undergrad (UToronto) through PhD (Stanford).
Edward Raff retweeted
OpenZL v0.3.0 is released :
github.com/facebook/openzl/r…
It has a faster LZ engine, close to LZ4 speeds, featuring a novel Huffman backend (PivCo), and a neural network that builds compression graphs on the fly, resulting in much higher compression ratios for numeric sources.
