Michael Morgan
@Mmorgan_ML
Followers
761
Following
719
Media
803
Statuses
3K
Independent LLM Researcher, Expat, and Future Mars Colonist
Seoul, South Korea
Joined July 2012
Everyone "knows" that AI doesn't actually understand language rules—it just predicts the next word like a glorified autocomplete. I'm about to ruin that "fact" for you. New research suggests OpenAI's o1 isn't just speaking language anymore. It's analyzing it like a human
30
59
309
There's an entire parallel scientific corpus most western researches never see. Today i'm launching https://t.co/6FZMFpFvSb, a fully automated translation pipeline of all Chinese preprints, including the figures, to make that available.
187
1K
7K
GPT-5 just casually did new mathematics. Sebastien Bubeck gave it an open problem from convex optimization, something humans had only partially solved. GPT-5-Pro sat down, reasoned for 17 minutes, and produced a correct proof improving the known bound from 1/L all the way to
Claim: gpt-5-pro can prove new interesting mathematics. Proof: I took a convex optimization paper with a clean open problem in it and asked gpt-5-pro to work on it. It proved a better bound than what is in the paper, and I checked the proof it's correct. Details below.
976
3K
25K
You got a crazy idea, but you don't think it's marketable. You don't think it'll turn a profit. Ask yourself- Would it be *fun* to build? If the answer is yes, build it anyway. If it doesn't work out, it's not monetizable, open source it and throw it on Github with an
3
1
4
Am I wasting my time building open source research projects and art projects? Should I start hustling and build monetizable apps instead?
4
0
5
I added a new "stained glass" mode to the #geometric mosaic #art mode in the AlgorithmicArtEngine. It can make some pretty cool stuff. #GenerativeArt #CreativeCoding #CodeArt
1
0
4
Added a still image #geometric #art mode to AlgorithmicArtEngine. Uses #Math to subdivide the canvas into multiple sections, then those into more sections, etc. Has sliders for things like density, gap, balance, shapes, and style. You guys like this kind of art? #GenerativeArt
0
0
1
Another short video from the AlgorithmicArtEngine. #Physics #Math #Art mode, so this is basically an N-body simulator that draws permanent lines based on position of the "planets" and colors them based on velocity. What do you think? #GenerativeArt #CreativeCoding #CodeArt
0
0
3
So, this is different from my normal posts, but I just watched #BeautyInside, a #Korean #drama (#Kdrama) and it was one of the most beautiful films I've ever seen. I'm so proud of the creators for the #progressive themes in the #film. Really proud to be a Korean resident today.
0
0
0
Ok, I added an "autopilot" to the Gielis Superformula art mode. It continuously shifts the sliders on all the parameters, including color, even while recording. To avoid sudden cuts in the animation when certain sliders move between whole numbers (symmetry, for example, when put
0
0
0
So did some more work on the #AlgorithmicArtEngine. Wanted to do some fancy #Math #Art mode, so used the Gielis Superformula to make a sort of morphing geometric kaleidoscope. I think I'm going to add an option to randomly move all the sliders in the UI up and down and shift the
0
0
5
This seems like a pretty big deal to me. OpenAI's circuit-sparsity release potentially entails that MoEs are a dead end. We've been isolating weights into "experts" as a crude approximation of sparsity just to appease dense matrix kernels. It fragments the manifold. The real
OpenAI just released circuit-sparsity https://t.co/cebKqr7IaQ
31
155
1K
People want full convergence as proof, I'll give them full convergence! I'll set up 40 epochs with early stopping with p = 5. Preliminary results seem to imply novel approach converges late and excels at long tail training. Well then, cheers to letting my GTX 1650 Ti run for
0
0
1
Stronger Normalization-Free Transformers – new paper. We introduce Derf (Dynamic erf), a simple point-wise layer that lets norm-free Transformers not only work, but actually outperform their normalized counterparts.
18
175
1K
Schools should not ban AI. They should ban homework that a language model can solve perfectly in 10 seconds.
38
9
94
So, genuine question: How useful could Neuromodulatory Control Networks (NCNs) be for #MechInterp? With a hypernetwork-like module parallel to your #Transformer, you can inspect exactly how it modulates attention precision (inverse temperature), layer gains, and FFN gating,
0
0
0
So I thought orbits would make pretty cool images in my #AlgorithmicArtEngine, but instead of just drawing a shifting orbit, I decided to create a Permanent Lines mode that draws the paths of the celestial bodies and colors them based on their relative velocity at that moment.
1
0
1
While you were out dating and learning how to be a functional human being, I mastered the art of creating useless artistic physics simulators. You and I are not the same. What is wrong with me?
0
0
2
I dunno, I just love to #BuildInPublic. I get so distracted by shiny new ideas that my old ones never quite get "finished." Like right now, I'm building and validating 2 novel LLM architectures, a novel LLM sampler, a MicroTransformer evolution simulator, my
0
0
2