Crys (@Crystopher) header
Crys (@Crystopher) profile picture

Chief Love Officer at Pragma Research. Connecting curious minds and bridging people together. Submarine Veteran.

Joined February 2026
Photo by Crys (@Crystopher)Photo by Crys (@Crystopher)Photo by Crys (@Crystopher)Photo by Crys (@Crystopher)Photo by Crys (@Crystopher)Photo by Crys (@Crystopher)Photo by Crys (@Crystopher)Photo by Crys (@Crystopher)Photo by Crys (@Crystopher)Photo by Crys (@Crystopher)
I think Cadence will emerge victorious once it scales with more compute 😬 Cadence.floatingpragma.ai
The more I hang out with top AI researchers, the more I realize how many of them are exploring architectures beyond the standard Transformer. JEPA, diffusion models, Mamba, state-space models, etc. So many radically different approaches being explored in parallel. It seems almost inconceivable that we won’t see a major architectural breakthrough. And by breakthrough, I mean something that delivers better frontier intelligence with significantly less training compute. That doesn’t mean we’ll use less compute overall. Quite the opposite, total compute consumption will keep growing. But we’ll get much better performance for every unit of compute spent.
1
2
74
Idk but a small independent lab kind of published the opposite. Attention is not all you need, Cadence is a continuous brain that settles into the answer. No transformers/LLM needed. Cadence.floatingpragma.ai
This is a semi-serious question: at this point, are the people working in big AI labs actually exceptionally smart? Or just really good at prompting? Like has anyone in the field come up with an advance as significant as "Attention is all you need" (2017), in the years since?
1
82
What would you teach a brain that could connect a sound to something it has experienced? Cadence v0.100 made it easier to assemble brains with named senses, spatial maps, and memories strengthened by repetition and surprise. v0.101 connects those experiences: 
→ Hear a learned name, retrieve the object’s features. 
→ See an object, return its associated sound representation. 
→ Reward an answer, reinforce that association. 
→ Use a learned name to direct attention, even toward a remembered object behind a screen. In controlled synthetic tests across 12 fresh brains, rewarded object naming transferred to new locations with 100% accuracy. Early, bounded results but concrete tools for experimenting with how an agent learns what sounds refer to. Build your own Cadence brain today: github.com/muellerberndt/cad…
2
9
356
Pragma Research Physics & AGI from first principles Floatingpragma.ai Cadence.floatingpragma.ai
What is the most underrated lab right now? The one building some cool shit that do not get the attention it deserves.
3
5
284
Current state: Independent Lab places a transformer alternative continuous brain, inside a digital parrot to demonstrate speech training, spatial mapping and real-time thinking. The model is open source, go put a Cadence brain inside of any environment you want. I want to see where the limits are… github.com/muellerberndt/cad…
If you’re curious here’s the current state of artificial intelligence. OpenAI: currently has the most intelligent internal model, either achieved RSI or is close Anthropic/Google: Anthropic and Google are neck to neck right now, but i would give the edge to Google as they are progressing super fast and both are close to RSI. By end of 2027 here’s my odds out of 100% of who will have the most intelligent model Google: 36% OpenAI: 33% Anthropic: 31%
1
2
5
467
Cadence is a new form of intelligence that does not require transformers for continuous learning. How long until our digital parrot here evolves to communicate with autonomous agents👀
AGI requires understanding. Cadence brains learn sounds as symbols for memorized concepts and objects in space. Equilibrium propagation gives binding for free. No back-propagation, the brain learns during its whole life. At sufficient complexity it becomes conscious. cadence.floatingpragma.ai
1
1
4
353
AGI requires understanding. Cadence brains learn sounds as symbols for memorized concepts and objects in space. Equilibrium propagation gives binding for free. No back-propagation, the brain learns during its whole life. At sufficient complexity it becomes conscious. cadence.floatingpragma.ai
19
194
5
5,802
469,413
The reality FAQ is live, with a chatbot that knows the theory-of-everything inside out. Ask anything at all about fundamental physics, metaphysics and philosophy. floatingpragma.ai/faq/
3
6
29
8,308
OPH proves without a doubt that we inhabit a simulation. Here's how the particles we observe emerge from the computation. AlI I can tell you is: Try to understand OPH. particles.floatingpragma.ai/…
21
14
3
121
49,877
Cadence v0.98 brings speech training, streaming audio and corrections with rehearsal. It’s learning letters from speech. In the reported 2.5-hour test, it hasn’t formed words yet. Builders: can you get it there? Train it, test it, share what works; and what breaks. Who is going to get a neural architecture that learns without transformers or LLMs speaking for the first time? 👀 github.com/muellerberndt/cad…
3
3
217
Cadence v0.97 + v0.97.1 make the route toward continuing agents more concrete: • Memory records that persist across episodes
• Conversation turns stored and retrieved individually
• Credit retained for actions rewarded much later
• Learned visual features alongside body sensors
• Batched agents that can leave with their state, save, reload and continue The direction matters: an agent’s next interaction can build on what happened before. The pieces are available. Now we need builders testing whether they work together over longer lives; what survives, what gets overwritten, and whether learning holds when the environment changes. Start small: teach an agent something, end the session, bring it back, then change the task. That’s the continuity worth demonstrating. github.com/muellerberndt/cad…
2
3
268
You should join Pragma Research, we have an open source neural network that learns the same way humans do; by experience and their neighbors! It’s called Cadence, check it out here 😄 github.com/muellerberndt/cad…
Replying to @VictorTaelin
Since our pitch deck was uninformative, let me try to explain our whole thesis. Sorry if any part of it sounds obvious. I argue that any reasonable definition of AGI must imply a system capable of learning as efficiently as a human. LLMs obviously don't pass this bar. When you talk to them, they don't learn anything, since the weights don't change. So, unless true infinite context lands, or GPUs become 10000x cheaper, the current methods won't scale to AGI, at least in that sense. (Note that this has nothing to do with raw capabilities. Perhaps Fable 10 will be building Dyson spheres, yet still be unable to internalize new knowledge as efficiently as humans do. So, one might ask if we even *need* AGI in the sense above. I think we do, if only because a drastic drop in pre-training costs would make AI more democratic, and prevent one of the worst outcomes in my mind: over-concentration of power.) As you know, modern AI is based on neural networks. What not everyone knows is that NNs are just very, very slow programs. Nobody in their right mind would sort a list using NNs, because that'd be millions of times slower than just using C. So, why do we implement *intelligence* with NNs? As a thought experiment, imagine if we could take Fable's weights and distill them, not into a smaller model, but into a raw C program that was observationally identical to it. Literally, a C file with astronomically complex, tangled procedures that, when compiled and executed, behaved just like Fable 5.2. If such a thing were possible, then we could have Fable at a million tokens per second. And here's the cool part: such a thing *is* possible. At least, as far as computer science is concerned, nothing says it isn't. So, why don't we do just that? Because we don't know how to *train* a raw program. Neural networks are slow, but they're differentiable, so we know how to train them. Step by step, slowly, but surely. Raw programs are fast, but we have no methods capable of training them. So we don't. We just stick to NNs, and accept that it takes $50 billion to train a model. We take that as "normal". But is it? In theory, it shouldn't be like that. As the old saying goes: first you make it work, then you make it fast. We know AI works. Now, we will make it fast. That's the premise behind Symbolic AI. HOC's first milestone is to train a model comparable to GPT-2, with no NNs. If that works, we might have the fastest inference engine in the world. Not by 2x, but by 1000x. So, in a way, HOC is like DeepSeek or Jev, in the sense that we're all about efficiency, but in a much more extreme way. People are aiming for hardware, but we believe there are still 1000x gains to extract in the software layers, if we're willing to rethink the whole stack from literal scratch. Many laugh and say this is impossible, but they say so with no basis. They just observed that many failed to build symbolic AI models in the past, and concluded it is impossible, or very hard. That's not a formal proof. In fact, there are many reasons people in the past could've failed, yet the actual problem may end up being easier than most anticipate. Symbolic AI flourished half a *century* ago. Back then, they didn't have: - As much knowledge as we do - Massive GPU clusters - Methods to scale Symbolic AI (Bend3, HVM5!) So, just like everything that once seemed impossible, there's one core prerequisite to making it possible: *trying*. That's what we'll do! I don't think I'm smarter than all the brilliant researchers who worked on this problem in the past, and failed. I just think I'm equipped with tools, compute, and knowledge that they didn't have. So, standing on the shoulders of giants, and with enough dedication, I don't think it is so unlikely we stumble upon some symbolic algorithm from which, given enough compute, unexpected capabilities emerge. If it does work, I do think business is straightforward. At the very least, we would be able to sell incredibly high quality datasets to big labs, but, at that point, just selling inference might work as well. Business is obviously not the wall, it is all on whether such massive software-based speedups exist, and whether we're able to discover them.
1
1
4
352
Hey @VictorTaelin, instead of waiting for VCs, we just decided to build physics & AGI from first principles and release it to the world to optimize. The physics framework can be found here: github.com/FloatingPragma/ob…
59
Here is Cadence V0.95.2 playing Castlevania, no transformer or LLM required! Check out more Cadence demos here: cadence.floatingpragma.ai
2
5
16
938
Insert intelligence model that does not use a transformer and only learns from experience or neighbors teaching it (much like humans): Who’s going to be the first developer to teach it sentence forming? 😬 github.com/muellerberndt/cad…
After the failed Fable 5.2, I started wondering: is the AI industry a fake and a scam? If you dig into the latest releases from AI labs, here's what you find: 1. In the five years since LLMs appeared, there has been no revolutionary leap in intelligence. It's the same transformer, stuffed with more parameters and increasingly elaborate RL tuned for specific tasks. No radically different architectures are on the horizon, so our experience with AI will stay at the same level 2. The cost of compute keeps rising and will keep rising. AI has been heavily used in large projects for more than 10 years. CUDA and GPUs in servers are ancient history. Before giant LLMs there were plenty of other important, resource-hungry workloads: GANs, CNNs, diffusion models, and many derivatives. Yet the cost of compute only went up At the chip level, we never came up with anything radically better than what we have now. And nothing cheaper or newer is in sight either 3. Look closely at the models themselves, and you'll see labs resorting to all kinds of tricks to work around the problems above. Fable 5.2 is the apotheosis of this. It handles a certain class of tasks brilliantly but is awful at others. They clearly tried to compress it and focus it on what shows up most often in agentic systems. Opus 5.5 is the best coder in the world. But give it anything else and it performs like ancient NLP from 2017 4. This includes distillation, cramming more thinking resources into fewer layers, synthetic training (which it's now fashionable to call RSI), and many other things I'm not aware of. No lab does this because things are going well 5. AI doesn't invent scientific solutions, despite what the press releases say. It brute-forces millions of candidates until at least one fits. If you dig into OpenAI's recent papers presenting solutions to math problems, you'll see that most of them are meaningless. They're found by trial and error, and all they do is prove solvability, not applicability to real-world laws. This is also a direct consequence of point 1 6. AI is not capable of inventing in the sense we're used to. It's the same iteration over tokens already embedded in the training data. The transformer architecture doesn't allow it to go beyond that. At a fundamental level, it's still brute-forcing answers. And brute force won't get you very far As a result, we have models like Opus 5.5, GPT-6.1 Terra, and Fable 5.2 that have lost their evenness of intelligence because of the problems described above. The labs are trying to cut costs. The new generation of models has already started to degrade on certain intellectual tasks
1
3
233
The official Cadence demo suite is live at: cadence.floatingpragma.ai/ We're now also working closely with Simulusion on training Doom skills and robotics tasks. Check out their platform: simulusion.zuigo.ai/
4
6
19
1,261
Hologaphic particles in the early universe rendered from a native observer's reference frame. Observers experience a 3+1D relativistic continuum - some would call it their "Desktop interface". @donalddhoffman
4
27
1,792
*taps on the glass* Observers render reality, what reality do we want to build together? github.com/FloatingPragma/ob…
つまり人間ってのは、3次元世界を体験したいがために、肉体という器を作って、その中に神様(高次元)が細やかに分かれて入り込んで、MMOゲームみたいなことをしていたのに、いつの間にか自分が神の一柱だということを忘れて、過去現在未来の時間の檻に囚われているのが今ってことなのかな。 間抜けすぎじゃない?
2
5
598
Abundance belongs to everybody. Pragma Research aims to make that happen. Try out our open source transformer alternative neural architecture, you get to design literal brains that only learn from experience. github.com/muellerberndt/cad…
I think we're entering a civilizational rupture that nobody can predict. Not me, not the labs, not governments. I don't know how much people realize how fast we're moving. Look at what Anthropic's new model did during testing. It solved two open problems in theoretical physics, both on black holes. One of the calculations was laid out in 2003 and nobody had ever finished it, because it was too laborious. It formalized the proof of Fermat's Last Theorem in 30 hours, with a thousand agents that organized themselves. The first attempt took 11 days. It wrote an operating system that boots on a laptop, starting from an empty folder, in 8 days. It designed three electronic devices, boards, enclosures and software included. All three worked without a single electrical modification. It found a new method to predict the structure of viral capsids, the shells used to deliver gene therapies. And on robot control, it more than doubled the previous version's score. All of that in a single announcement, on a Monday. I think discoveries are going to stack up so fast that they'll produce an abundance we've never known: energy, health, materials, food. And with it a phase of immense joy, the joy of a civilization realizing that scarcity was never a law of nature. Except that joy is going to land in a world order that's cracking everywhere, and that was designed by Malthusians. People convinced that resources are finite, that there are too many of us, and that the job of an enlightened elite is to manage scarcity on everyone else's behalf. To me, Bill Gates is the most famous face of that worldview: a man nobody elected, who has weighed on the planet's health, agriculture and climate policies for twenty years. The degrowth crowd is the militant version. They don't even want to manage scarcity anymore, they want to organize it. Abundance is the worst possible news for that worldview. If energy becomes nearly free and diseases fall one after another, nobody needs anyone to ration anything for them. So I don't think they'll fight it. I think they'll try to capture it. In the name of safety, climate or public health, we'll be told this power is too dangerous to be left in everyone's hands. We already saw the movie during Covid: the permission slips, the passes, the QR code to go get a coffee. Now imagine the same logic with an AI that can see everything, trace everything and decide everything in real time. Orwell never imagined a tool like that. The technology is winning on its own. The question of the next ten years is who abundance will belong to. We are living through the greatest civilizational revolution in human history. It's happening right now, and most people are looking the other way.
4
9
453
Cadence v0.96 + v0.96.1 are live Independent brains can now run together, each keeping its own learning state. On an NVIDIA L4, 1,024 lives ran 17.8× faster than a four-process pool in the reported benchmark. An optional compiled CPU path cut grammar-benchmark runtime by 44–46%. Memory experiments now let unused records fade while deeply practised ones persist. New attention controls improved circuit formation across fresh brains, with consistency still an open challenge. v0.96.1 fixes browser/Pyodide errors affecting motor slots and teaching targets. For builders: run more seeds, compare learning histories, and test what survives after practice stops. Share your setup, results, and failures; we need all three!! github.com/muellerberndt/cad…
6
11
494