osdf (@plastiq) profile picture

Machine Learner, Programmer. PhD in Deep Learning. Opinions are my own.

München
Joined March 2007
Photo by osdf (@plastiq)Photo by osdf (@plastiq)
IMO the essential difficulty of malleable software now boils down to “fork maintenance”. Let’s unpack that a bit. If you want to make your own custom tools or fork existing OSS, the cost of that is headed to zero. But… often we don’t want to start from scratch! We want to remix tools others have created. We want someone else to help us maintain stuff. We want to combine our ideas with others’. Now the moment you have multiple parties both updating the software… you hit gnarly problems! In a nutshell: happens if I modded something and it no longer meshes well with the latest updates from you? This is “fork maintenance”. (Traditionally this would be a centralized model where a “dev” is maintaining the software and you, the “user” might adjust settings or patch things. But even in a more decentralized setting we want to jam with our peers on our tools.) A classic approach here is to expose a stable plugin SDK. Works great because the dev can keep the surface maintained and plugins hopefully compose reasonably…. But only works as long as the plugin surface has what you need. Sometimes you end up hitting walls. If it’s OSS you can also literally fork. Maintaining a fork and merging edits from others is a huge pain. But maybe with enough smart cheap LLMs this is more viable now. You also need to consider data, not just code. Schema migrations are already a pain in the ass with normal software; when every user is doing customizations things get extra gnarly. I don’t think that there’s a perfect solution to this yet. Ideally I’d want something much deeper and broader than a typical plugin API - eg a language or framework designed for layered additive modification (think Harel’s behavioral programming as one reference.) I find it quite poetic that collaboration and working with others is where malleability gets really good but also where it gets hard. Same as many other things in life! For more on this, we go into quite some depth on this topic here: inkandswitch.com/essay/malle…
18
10
2
165
10,648
New high quality blog post: "Curiosity-Driven Tree Search" Here I derived from scratch a completely new alternative to AlphaZero-like algorithms, and it's scalable!
11
41
7
446
32,365
Predicting the answer to interventional "what if?" questions — the outcome of an action you never took — need a *mechanistic* model, not a curve fit. And you can only learn one by *experimenting*. Experiments are costly, so the real game is **data efficiency**. Meet the Model Discovery Agent (MDA). 🧵
28
172
28
1,195
208,810
osdf retweeted
The central problem of biology is the origin of biological organization. In our preprint, “Information-theoretic Limits on Programmatic Specification of Biological Systems,” we show, using information theory, that an organism does not contain enough organism-specific information to specify its own fully functioning microscopic organization. For over 70 years—since the discovery of the double helix and the cracking of the genetic code—biology has been dominated by the metaphor of the genome as an executable computer script or structural blueprint. Yet, it has remained a central mystery how the genome organizes the entirety of the microscopic, fine-scale molecular biological machinery in space and time. We wish to put forward a radically different formulation. We show, as an analogy, that the genome is not a figure-skating program choreographing every molecular movement, but a hockey coach’s game plan : roster, tactics, line assignments. The exact trajectory of the puck is computed by physics at runtime. The game looks like hockey because of the specified ensemble, not because every coordinate was pre-programmed. The organized machinery of life is not the execution of a fully prewritten organism-specific program under favorable conditions. Rather, it is the compilation of a coarse organism-specific specification by a shared physical background that is constitutive of biological organization. We establish the threshold as a family of maximal capacity-compatible coarse-grainings, distinguish the proven impossibility below the threshold from the empirically realized coarse mappings above it, and locate the threshold empirically through worked cases of protein folding, E. coli, Drosophila early development, and C. elegans, together with computational verification using AlphaFold-2 (analyzing tertiary structures of 555,846 proteins across 46 AlphaFold-2 proteomes), the JCVI-syn3A 4D whole-cell simulation, and canonical stochastic gene network models. We further show that no known naturally realized environmental channel can close the gap. The result rules out programmed microstate determinism while leaving physical determinism untouched, reframes the genome as a generator specification rather than a trajectory program, and unifies gene-centric, developmental, and field-theoretic (bioelectric, morphogenetic, and related continuum) views of biological specification under a single coarse-graining framework. It has not escaped our notice that this framework has fundamental implications for understanding evolutionary search, heritability, thermodynamic order, biological modeling, and artificial intelligence in biology, but these will be published elsewhere. biorxiv.org/content/10.64898…
49
113
27
691
71,800
AI won’t just change how we assess students. It can also transform how we teach them. Interactive visualizations can make abstract concepts more intuitive and engaging. Check out this epipolar geometry visualization created with Codex. Link to interactive webpage in thread.
Experimenting with Codex’s UI capabilities. Impressive! csprofkgd.github.io/conferen…
5
14
1
156
32,005
Together with UC Berkeley we are announcing the laser phase plate - a breakthrough in atomic resolution imaging. This is the brightest continuous wave laser in the world, 100 million times the intensity of the surface of the sun. Phase contrast plays an important role in microscopy, but it was thought close to impossible for electron microscopy, where it would require interfering with an electron beam. Holger Mueller and Robert Glaeser proposed exactly this using a standing wave laser. It has taken over 15 years to make this a reality. Biohub partnered with UC Berkeley and Mueller to support this work and to engineer and build the technology. Contrast has been the critical barrier to achieving atomic resolution imaging of the cell. In cryo-electron tomography, a cellular imaging technology that uses electron microscopy, the low contrast makes it impossible to resolve anything but the largest proteins within their cellular context. The laser phase plate removes that barrier. With advances in AI this breakthrough in contrast will start to open up a new frontier in structural biology, that will allow us to see the molecular machines of the cell, and how they assemble into far more complex and dynamic systems, and understand how they work.
91
564
112
3,955
664,353
Relics from the prehistoric era of AI
60
226
31
7,893
432,750
1/?) As promised to Sander Dieleman (@sedielem), we’re finally excited to share: Towards Closing the Autoregressive Gap in Language Modeling via Entropy-Gated Continuous Bitstream Diffusion We show that continuous diffusion can achieve very strong language modeling performance when operating directly on bitstreams, outperforming masked and uniform diffusion baselines, and essentially matching autoregressive models under our evaluation settings.
6
37
5
239
36,437
LLMs can hide a text in another text of the same length. I'll explain how, it is very simple, you'll understand before I finish, and smile. That's what I noticed during my #ICLR2026 poster session in Rio! 🇧🇷 Too bad you missed it, but let me remedy now
21
75
25
672
105,421
Replying to @rronak_
Our CATS workshop @icmlconf should serve a great overview of recent work on continual learning! cats-icml.github.io/.
2
10
1,621
For 40+ years, building a robot that could rally with an elite human table tennis player at full speed was an unsolved problem. Sony AI's Ace research project set out to change that—and the results are now accepted for publication in @Nature and featured on the cover.
13
89
23
360
157,081
I work on the JAX team. If you're new to the field ignore this bait. The things you should focus on are understanding the math and how to program accelerators - really master your hardware and your methods. We try to make a great tool but don't obsess over tools early.
When looking at deep learning profiles, one of the most obvious tells between a mediocre and great candidate is whether they list PyTorch or JAX.
18
80
5
1,791
108,139
I trained an LLM from scratch on pre-1900 text to see if it could come up with quantum mechanics and relativity. While the model is too small to do meaningful reasoning, it has glimpses of intuition. When given observations from past landmark experiments, the model can declare that “light is made up of definite quantities of energy” and even suggest that gravity and acceleration are locally equivalent. I’m releasing the dataset + models and leave this as an open problem to the research community. I also include what this project has taught me about intelligence in a mini essay linked below. 🧵(1/n)
114
257
76
2,015
318,915
I'd like to thank @daniel_rossett for his help in my recovery from the POTS version of Long COVID. Daniel was key in bringing me back from highly disabled and suffering to being able to do what I want to again. This X account is mostly focused on ML / AI. From that point of view, many of you know that in December 2024, I wasn't able to do the test of time award talk at NeurIPS, even by video call. Daniel started working with me in March 2025. By April, I started to have days of no POTS symptoms, by June I was off all heart rate lowering medications, by September I was back to work. I'm back to full exercise, running, lifting weights, mountain biking, and have even done things I hadn't done before I got sick, like riding Whistler Mountain Bike Park. I'm now getting the word out to help Daniel build a company that will bring this approach to more people.
168
80
25
2,670
211,164
after a bit of digging, first author @a1zhang has a GREAT blog-like description of the work + a nice minimal implementation. the blog is like 1000x better in describing the work than the paper. why do we still bother with papers.
i finally read the Recursive Language Models paper by Alex L. Zhang, Tim Kraska and @lateinteraction. First, I really really liked it. Fantastic idea, framing and execution. Very cool approach. Read it. But, and this is an issue with so many "LLM systems" papers today, I really wish it was more detailed. There are just not enough details for me to understand the core mechanism, which is the REPL and the implementation of recursive calls. I could of course *imagine* what they did, but not know for sure, which is a pitty, because I suspect that is where a lot of the heavy lifting is. We (as a field) must get better at describing our systems more accurately at these levels.
9
11
1
218
31,195
Really interesting stuff happening on that other social network thanks to the open protocols and hackable algorithms. E.g there's Personalized Academic Recommendations and a whole paper about it. arxiv.org/abs/2601.04253
6
24
5
226
27,370
New blog post w @pawtrammell: Capital in the 22nd Century Where we argue that while Piketty was wrong about the past, he’s probably right about the future. Piketty argued that without strong redistribution of wealth, inequality will indefinitely increase. Historically, however, income inequality from capital accumulation has actually been self-correcting. Labor and capital are complements, so if you build up lots of capital, you’ll lower its returns and raise wages (since labor now becomes the bottleneck). But once AI/robotics fully substitute for labor, this correction mechanism breaks. For centuries, the share of GDP that goes to paying wages has been 2/3, and the share of GDP that’s been income from owning stuff has been 1/3. With full automation, capital’s share of GDP goes to 100% (since datacenters and solar panels and the robot factories that build all the above plus more robot factories are all “capital”). And inequality among capital holders will also skyrocket - in favor of larger and more sophisticated investors. A lot of AI wealth is being generated in private markets. You can’t get direct exposure to xAI from your 401k, but the Sultan of Oman can. A cheap house (the main form of wealth for many Americans) is a form of capital almost uniquely ill-suited to taking advantage of a leap in automation: it plays no part in the production, operation, or transportation of computers, robots, data, or energy. Also, international catch-up growth may end. Poor countries historically grew faster by combining their cheap labor with imported capital/know-how. Without labor as a bottleneck, their main value-add disappears. Inequality seems especially hard to justify in this world. So if we don’t want inequality to just keep increasing forever - with the descendants of the most patient and sophisticated of today’s AI investors controlling all the galaxies - what can we do? The obvious place to start is with Piketty’s headline recommendation: highly and progressively tax wealth. This might discourage saving, but it would no longer penalize those who have earned a lot by their hard work and creativity. The wealth - even the investment decisions - will be made by the robots, and they will work just as hard and smart however much we tax their owners. But taxing capital is pointless if people can just shift their future investment to lower tax countries. And since capital stocks could grow really fast (robots building robots and all that), pretty soon tax havens go from marginal outposts to the majority of global GDP. But how do you get global coordination on taxing capital, when the benefits to defecting are so high and so accessible? Full automation will probably lead to ever-increasing inequality. We don’t see an obvious solution to this problem. And we think it’s weird how little thought has gone into what to do about it. Many more thoughts from re-reading Piketty with our AGI hats on at the post in the link below.
204
232
158
2,249
1,528,831
How does this compare to berkeley sky RL tinker library?
1
1
10
2,403
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.
19
173
13
1,072
167,127