Reposted by Byron Galbraith
@dtemkin.bsky.social 's Rivulet esolang prompts ideas around code and #datavis comprehension. How do we make sense of visualizations when we are unfamiliar with the 'grammar' of their design?
An introduction Rivulet:
observablehq.com/@jwolondon/r...
An editor:
observablehq.com/@jwolondon/r...
An introduction Rivulet:
observablehq.com/@jwolondon/r...
An editor:
observablehq.com/@jwolondon/r...
June 15, 2025 at 7:50 AM
@dtemkin.bsky.social 's Rivulet esolang prompts ideas around code and #datavis comprehension. How do we make sense of visualizations when we are unfamiliar with the 'grammar' of their design?
An introduction Rivulet:
observablehq.com/@jwolondon/r...
An editor:
observablehq.com/@jwolondon/r...
An introduction Rivulet:
observablehq.com/@jwolondon/r...
An editor:
observablehq.com/@jwolondon/r...
Reposted by Byron Galbraith
![5-panel comic. (1) [teacher with long hair next to whiteboard] TEACHER: I’m supposed to give you the tools to do good science. (2) [teacher addressing students] But what *are* those tools? Methodology is hard and there are so many ways to get incorrect results. What is the magic ingredient that makes for good science? (3) TEACHER: To figure it out, I ran a regression with all the factors people say are important: [embedded list in sub-panel, cut off at end] Outcome variable: correct scientific results. Predictors: collaboration; skepticism of others’ claims; questioning your own beliefs; trying to falsify hypotheses; checking citations; statistical rigor; blinded analysis; financial disclosure; open data (4) TEACHER: The regression says two ingredients are the most crucial: 1) genuine curiosity about the answer to a question, and 2) ammonium hydroxide. (5) STUDENT: Wait, why did *ammonia* score so high? How did it even get on the list? LONG HAIR: ...And now you’re doing good science!](https://cdn.bsky.app/img/feed_thumbnail/plain/did:plc:cz73r7iyiqn26upot4jtjdhk/bafkreigzaphhj4mfgfrxddlvdmw47335gnwzqbfqvrnff5x4oovgpf36nq@jpeg)
June 12, 2025 at 8:28 PM
Reposted by Byron Galbraith
If a seven-year-old can’t walk to the store the fault lies not with the parents but with the community.
In order to eliminate any inconvenience to cars, we’ve made it illegal for children to be children
June 6, 2025 at 3:44 PM
If a seven-year-old can’t walk to the store the fault lies not with the parents but with the community.
Reposted by Byron Galbraith
The Neuromatch #ImpactScholars Program is a global initiative that provides early-career researchers with hands-on mentorship, collaborative project work, & next steps to grow as
computational scientists.
🤓 Berilsu Öner shares her #MyNeuromatchStory
⬇️
buff.ly/MBxDOJf
#Neuromatch #Neuroscience
computational scientists.
🤓 Berilsu Öner shares her #MyNeuromatchStory
⬇️
buff.ly/MBxDOJf
#Neuromatch #Neuroscience
June 6, 2025 at 3:02 PM
The Neuromatch #ImpactScholars Program is a global initiative that provides early-career researchers with hands-on mentorship, collaborative project work, & next steps to grow as
computational scientists.
🤓 Berilsu Öner shares her #MyNeuromatchStory
⬇️
buff.ly/MBxDOJf
#Neuromatch #Neuroscience
computational scientists.
🤓 Berilsu Öner shares her #MyNeuromatchStory
⬇️
buff.ly/MBxDOJf
#Neuromatch #Neuroscience
Reposted by Byron Galbraith
Finally finished reading Eshraghian et al guide to training SNNs, and wow. It's one of those review papers that feel so generous. Full of insights that could each generate entire research programmes, but rather than hoarding, they shared it with us all.
arxiv.org/abs/2109.12894
arxiv.org/abs/2109.12894
Training Spiking Neural Networks Using Lessons From Deep Learning
The brain is the perfect place to look for inspiration to develop more efficient neural networks. The inner workings of our synapses and neurons provide a glimpse at what the future of deep learning m...
arxiv.org
June 4, 2025 at 9:08 AM
Finally finished reading Eshraghian et al guide to training SNNs, and wow. It's one of those review papers that feel so generous. Full of insights that could each generate entire research programmes, but rather than hoarding, they shared it with us all.
arxiv.org/abs/2109.12894
arxiv.org/abs/2109.12894
Reposted by Byron Galbraith
Would it hurt people in tech to go and *talk* to someone with a degree in education? Most people don't understand what schools do. TL;DR: it's a lot more than info dumping into student's heads and it is certainly a lot more than daycare.
Ok well this sentiment from Luis von Ahn is absolutely gross. I will not be giving Duolingo any more money. AI first: 🤮 fortune.com/2025/05/20/d...
May 21, 2025 at 1:42 PM
Would it hurt people in tech to go and *talk* to someone with a degree in education? Most people don't understand what schools do. TL;DR: it's a lot more than info dumping into student's heads and it is certainly a lot more than daycare.
Reposted by Byron Galbraith
Was at a conference dinner a while back and one guy said "The future is AI. It doesn't replace senior developers, but no more need for juniors."
To which i replied:
"Where do you think senior developers come from. Straight from the womb?"
He went quiet, then wandered off to a different huddle.
To which i replied:
"Where do you think senior developers come from. Straight from the womb?"
He went quiet, then wandered off to a different huddle.
Yesterday it was a middle aged woman in my gym doing the same thing. Talking a young man’s ear off about how much chatGPT helps her do her work. Charts and writing — “i’m not good at that and now i don’t need someone in the org flow to do it”.
On the C train in Manhattan, two women — middle age, business look, clearly lawyers from the conversation — going on and on about ChatGPT. Helping them do research, write briefs, even plant a garden at home! So quick and easy!
Ugghhhhhhhhhhhhhhhhh
Ugghhhhhhhhhhhhhhhhh
May 9, 2025 at 3:25 PM
Was at a conference dinner a while back and one guy said "The future is AI. It doesn't replace senior developers, but no more need for juniors."
To which i replied:
"Where do you think senior developers come from. Straight from the womb?"
He went quiet, then wandered off to a different huddle.
To which i replied:
"Where do you think senior developers come from. Straight from the womb?"
He went quiet, then wandered off to a different huddle.
Reposted by Byron Galbraith
Bring physical buttons back on absolutely everything. Clicking, snapping, tactile friction. Chunky single-purpose electronics. Retreat from the gesture-based app-connected hellworld of the 2010s into the warm sensory embrace of the retrofuture 1980s
Amazingly, reaction times using screens while driving are worse than being drunk or high—no wonder 90 percent of drivers hate using touchscreens in cars.
Finally the auto industry is coming to its senses. Real buttons are sooooooo back baby!
Finally the auto industry is coming to its senses. Real buttons are sooooooo back baby!
Rejoice! Carmakers Are Embracing Physical Buttons Again
Amazingly, reaction times using screens while driving are worse than being drunk or high—no wonder 90 percent of drivers hate using touchscreens in cars. Finally the auto industry is coming to its sen...
www.wired.com
May 5, 2025 at 9:19 PM
Bring physical buttons back on absolutely everything. Clicking, snapping, tactile friction. Chunky single-purpose electronics. Retreat from the gesture-based app-connected hellworld of the 2010s into the warm sensory embrace of the retrofuture 1980s
Reposted by Byron Galbraith
One weird trick for better diffusion models: concatenate some DINOv2 features to your latent channels!
Combining latents with PCA components extracted from DINOv2 features yields faster training and better samples. Also enables a new guidance strategy. Simple and effective!
Combining latents with PCA components extracted from DINOv2 features yields faster training and better samples. Also enables a new guidance strategy. Simple and effective!
1/n Introducing ReDi (Representation Diffusion): a new generative approach that leverages a diffusion model to jointly capture
– Low-level image details (via VAE latents)
– High-level semantic features (via DINOv2)🧵
– Low-level image details (via VAE latents)
– High-level semantic features (via DINOv2)🧵
April 25, 2025 at 1:03 PM
One weird trick for better diffusion models: concatenate some DINOv2 features to your latent channels!
Combining latents with PCA components extracted from DINOv2 features yields faster training and better samples. Also enables a new guidance strategy. Simple and effective!
Combining latents with PCA components extracted from DINOv2 features yields faster training and better samples. Also enables a new guidance strategy. Simple and effective!
Today I was able to have an LLM generate a midi bass line, have it loop that on my synth via MCP, and then jam on that on my drum kit. This is the start of something exciting
April 18, 2025 at 3:28 AM
Today I was able to have an LLM generate a midi bass line, have it loop that on my synth via MCP, and then jam on that on my drum kit. This is the start of something exciting
Reposted by Byron Galbraith
As long as universities continue to have their main role be "ticket to middle class," it'd be insane to expect them to serve classical liberal education functions. It's too late, but I think things would be a lot better if job-prep and preserving the Great Conversation were separate institutions.
April 15, 2025 at 1:09 PM
As long as universities continue to have their main role be "ticket to middle class," it'd be insane to expect them to serve classical liberal education functions. It's too late, but I think things would be a lot better if job-prep and preserving the Great Conversation were separate institutions.
Reposted by Byron Galbraith
1. LLM-generated code tries to run code from online software packages. Which is normal but
2. The packages don’t exist. Which would normally cause an error but
3. Nefarious people have made malware under the package names that LLMs make up most often. So
4. Now the LLM code points to malware.
2. The packages don’t exist. Which would normally cause an error but
3. Nefarious people have made malware under the package names that LLMs make up most often. So
4. Now the LLM code points to malware.
LLMs hallucinating nonexistent software packages with plausible names leads to a new malware vulnerability: "slopsquatting."
LLMs can't stop making up software dependencies and sabotaging everything
: Hallucinated package names fuel 'slopsquatting'
www.theregister.com
April 12, 2025 at 11:43 PM
1. LLM-generated code tries to run code from online software packages. Which is normal but
2. The packages don’t exist. Which would normally cause an error but
3. Nefarious people have made malware under the package names that LLMs make up most often. So
4. Now the LLM code points to malware.
2. The packages don’t exist. Which would normally cause an error but
3. Nefarious people have made malware under the package names that LLMs make up most often. So
4. Now the LLM code points to malware.
Reposted by Byron Galbraith
It turns out to be hard to evaluate natural language with natural language. What should we take away from the conundrum of LLM evaluation? www.argmin.net/p/evaluation...
Evaluation or Valuation
The infinite regress of evaluating large language models
www.argmin.net
April 10, 2025 at 3:03 PM
It turns out to be hard to evaluate natural language with natural language. What should we take away from the conundrum of LLM evaluation? www.argmin.net/p/evaluation...
Reposted by Byron Galbraith
Good presentation that makes sense of a messy research landscape.
I gave another talk last friday at the USC ISI on some thoughts on RLVR and what to try with it. Not a ton new in here, but going over the challenges and tradeoffs we are working through right now (along with more discussion on some of those new GRPO variants). Slides in YT desc.
Experimenting with Reinforcement Learning with Verifiable Rewards (RLVR)
Here's the latest talk I gave, last friday at the USC Information Sciences Institute. It's a slightly more technical version of the RL talks I've been giving...
buff.ly
April 8, 2025 at 8:13 PM
Good presentation that makes sense of a messy research landscape.
Reposted by Byron Galbraith
This Github ticket is the kind of proposal that inexperienced engineers tend to make, because they think it's about code, and not about the humans who use the system.
In any competent organization, these junior folks would get some very direct feedback and it would be a growth moment.
In any competent organization, these junior folks would get some very direct feedback and it would be a growth moment.
April 5, 2025 at 3:07 PM
This Github ticket is the kind of proposal that inexperienced engineers tend to make, because they think it's about code, and not about the humans who use the system.
In any competent organization, these junior folks would get some very direct feedback and it would be a growth moment.
In any competent organization, these junior folks would get some very direct feedback and it would be a growth moment.
Reposted by Byron Galbraith
Stop everything you are doing and review the best Open Data chart you will see all year.
Credit: @nkgarg.bsky.social's lab
Credit: @nkgarg.bsky.social's lab
2) The dog name/baby name scatter plot: Max and Charlie are more dog name than human name. Dogs aren't called Ethan though.
April 2, 2025 at 2:59 PM
Stop everything you are doing and review the best Open Data chart you will see all year.
Credit: @nkgarg.bsky.social's lab
Credit: @nkgarg.bsky.social's lab
Reposted by Byron Galbraith
We went through this in the 2010s. The Turing Test is a thought experiment for Turing to grapple with the fact that no one could define “intelligence” so he tried to operationalize it and gave the Imitation Game as as example. It was never meant to be conducted. The method is under-controlled.
Rumor has it the Turing Test has been solved. Don’t believe everything you read.
ChatGPT hasn’t really just passed the Turing Test, and it would scarcely matter if it did.
I explain why, here:
open.substack.com/pub/garymarc...
ChatGPT hasn’t really just passed the Turing Test, and it would scarcely matter if it did.
I explain why, here:
open.substack.com/pub/garymarc...
AI has (sort of) passed the Turing Test; here’s why that hardly matters
Don’t panic
open.substack.com
April 2, 2025 at 3:15 PM
We went through this in the 2010s. The Turing Test is a thought experiment for Turing to grapple with the fact that no one could define “intelligence” so he tried to operationalize it and gave the Imitation Game as as example. It was never meant to be conducted. The method is under-controlled.
Reposted by Byron Galbraith
20 years ago, Nine Inch Nails “The Hand That Feeds,” the first music video I directed. To celebrate this anniversary, I have constructed this immersive 2005 simulator to transport you to an authentic media viewing experience of the distant past.
March 28, 2025 at 7:19 PM
20 years ago, Nine Inch Nails “The Hand That Feeds,” the first music video I directed. To celebrate this anniversary, I have constructed this immersive 2005 simulator to transport you to an authentic media viewing experience of the distant past.
Reposted by Byron Galbraith
Accepting “the bitter lesson” and embracing the brain’s complexity by Dyer and Richards. Great article! www.thetransmitter.org/neuroai/acce...
» Accepting “the bitter lesson” and embracing the brain’s complexity Close menuMenuClose menuSpectrum logoYouTubeBlueskyMastodonYouTubeBlueskyMastodon
www.thetransmitter.org
March 26, 2025 at 4:09 PM
Accepting “the bitter lesson” and embracing the brain’s complexity by Dyer and Richards. Great article! www.thetransmitter.org/neuroai/acce...
Reposted by Byron Galbraith
This is fun - training NNs using only biases works well!
1/7: Super excited to share our new paper! This one should be of interest to neuroscientists and deep learning theory folks. This paper was a collaboration with Alexandre Payeur, @averyryoo.bsky.social, Thomas Jiralerspong, @mattperich.bsky.social, Luca Mazzucato, @glajoie.bsky.social
March 25, 2025 at 4:27 PM
This is fun - training NNs using only biases works well!
Reposted by Byron Galbraith
Rivulet is a language that prioritizes flow. Its script was influenced by mazes, space-filling algorithms, Anni Albers' Meander series, calligraphy-friendly conlangs and natlangs
This program calculates the first set of Fibonacci numbers, stopping at 21.
It can be found at github.com/rottytoot...
This program calculates the first set of Fibonacci numbers, stopping at 21.
It can be found at github.com/rottytoot...
March 25, 2025 at 1:34 PM
Rivulet is a language that prioritizes flow. Its script was influenced by mazes, space-filling algorithms, Anni Albers' Meander series, calligraphy-friendly conlangs and natlangs
This program calculates the first set of Fibonacci numbers, stopping at 21.
It can be found at github.com/rottytoot...
This program calculates the first set of Fibonacci numbers, stopping at 21.
It can be found at github.com/rottytoot...
Reposted by Byron Galbraith
A friend pointed out that the quasi-religious way so many tech guys talk about AI is in part because they don't want to grapple with the guilt of what they do; AI is a God, AI will fix it. No need to be responsible for what you put in the world.
I can't stop thinking about that.
I can't stop thinking about that.
The most dangerous thing about AI chat bots is how many men with political power treat them *exactly* and unquestioningly like actual Ancient Greek oracles.
March 22, 2025 at 12:44 AM
A friend pointed out that the quasi-religious way so many tech guys talk about AI is in part because they don't want to grapple with the guilt of what they do; AI is a God, AI will fix it. No need to be responsible for what you put in the world.
I can't stop thinking about that.
I can't stop thinking about that.
Great crowd in Pittsburgh to #standupforscience2025
March 7, 2025 at 4:41 PM
Great crowd in Pittsburgh to #standupforscience2025
Reposted by Byron Galbraith
We're probably a little too obsessed with zero-shot retrieval. If you have documents (you do), then you can generate synthetic data, and finetune your embedding. Blog post lead by @jacobianneuro.bsky.social shows how well this works in practice.
www.databricks.com/blog/improvi...
www.databricks.com/blog/improvi...
Improving Retrieval and RAG with Embedding Model Finetuning
Fine-tune embedding models on Databricks to enhance retrieval and RAG accuracy with synthetic data—no manual labeling required.
www.databricks.com
February 26, 2025 at 12:48 AM
We're probably a little too obsessed with zero-shot retrieval. If you have documents (you do), then you can generate synthetic data, and finetune your embedding. Blog post lead by @jacobianneuro.bsky.social shows how well this works in practice.
www.databricks.com/blog/improvi...
www.databricks.com/blog/improvi...
Reposted by Byron Galbraith
James Webb Satellite took a fantastic shot of the Planet Saturn with it's Near Infrared Camera (NIRCam) and the Mid-Infrared Instrument (MIRI). This is just an Amazing shot that shows the planet and the rings in brilliant detail.
February 24, 2025 at 6:16 AM
James Webb Satellite took a fantastic shot of the Planet Saturn with it's Near Infrared Camera (NIRCam) and the Mid-Infrared Instrument (MIRI). This is just an Amazing shot that shows the planet and the rings in brilliant detail.