Ryan Benmalek

hey i’m a lil desert mouse and i’m excited to be joining this new sand dune :) my background > I’m 3 years old > i dropped out of something > i have silky soft fur looking for more friends - let’s find some water or do other fun grooming activities this fall

If someone makes a talking parrot that rides around on my shoulder - I'll invest on the spot. Thank you very much.

8. Hey execs: maybe don't keep diaries? Relatedly, all these constant-ambient-listening AI personal devices are going to be an evidentiary gold mine sooner than later.

This paper was an absolute pleasure to be a part of. We believe– based on accumulating, emerging evidence– that there is a general theory for deep learning on the horizon. Here, we review evidence informs this belief, and try to articulate what such a theory might look like.

Patriots seem unable to grasp what Jensen is saying, namely: lithography is not the bottleneck *for him*. Almost all progress between Hopper and Vera Rubin Ultra is about clever engineering, not feature size. Huawei is currently at ≤Blackwell level. Jensen likes it this way.

Much of Dwarkesh's argument hinges on this statment which *was* accurate but will be increasingly inaccurate on a go forward basis imo:    “American labs port across accelerators constantly. Anthropic's models are run on GPUs, they're run on Trainium, they're run on TPUs. There are so many things you can do, from distilling to a model that's well fit for your chips.”   As system level architectures diverge (torus vs. switched scale-up topologies, memory hierarchies, networking primitives), true portability is eroding. The Mi300 and Mi325 had roughly the same scale-up domain size as Hopper while Blackwell’s scale-up domain is 9x larger than the Mi355 scale-up domain, etc. Many frontier models are now being explicitly co-designed for inference on specific hardware like GB300 racks. Codex on Cerebras is another example. Those models run less efficiently on other systems and the performance differentials will only widen. A model that runs well on Google’s torus topology will run less efficiently on Nvidia’s switched scale-up topology and vice versa - the data traffic is fundamentally different as a byproduct of the models being parallelized across the different topologies. Google’s internal teams - and increasingly the Anthropic teams as they become the most important customer of almost every cloud - have the luxury of operating across the stack (models, chips, networking) - but that is not the case for the rest of the market and other prospective users. Anthropic is the exception, not the rule. To wit, Anthropic and Google allegedly have a mutual understanding where Anthropic can hire the TPU engineers they need every year to ensure that they can continue to get the most out of the TPU. Given the overwhelming importance of cost per token to the economics of the labs, models will be run where they run best. Most extremely large MoE models will run best on GB300s given the importance of having a switched scale-up network like NVLink for MoE inference. When training was the dominant cost for labs and power was broadly available, labs were optimizing to minimize capex dollars. Model portability was a way to create leverage over suppliers. I think that drove a lot of the focus on portability. Today, inference costs as measured by tokens per watt per dollar are everything. Inference is way more important than training costs (inference is effectively now part of training via RL). Labs are therefore now optimizing for inference. This means increasing co-design and higher go-forward switching costs for individual models between systems. I do think this explains why Anthropic and Nvidia came together: Anthropic needed Blackwells and Rubins to inference at least *some* of their models economically. And Mythos might just end up being released coincident with the availability of Rubins for inference. TLDR: as labs shift their focus from training to inference, the costs of portability and the upside of co-design to maximize tokens per watt per dollar both rise. Portability is likely to begin decreasing as a result.   I think what I might have respectfully added to Jensen’s answer is that systems evolve under local selective pressures. The evolutionary pressure in America is a shortage of watts so it makes sense for Nvidia to optimize, as an American company, for power efficiency and tokens per watt and stay on copper as long as possible. China has a surfeit of watts. Chinese AI systems are already taking advantage of this with the Huawei Cloudmatrix 384 and Atlas SuperPoD having an optical scale-up domain that is much larger than anything offered by Nvidia today at the cost of *much* higher power consumption and much lower tokens per watt. The networking primitives for this Huawei system are very different than those for Nvidia’s systems and a model that runs well on Nvidia will not run well on that system and vice versa. This means that if a Chinese ecosystem gets momentum, Chinese models might stop running well on American hardware. And when Chinese models run best on American hardware, America is in a better position as this gives America a degree of leverage and control over Chinese AI that it risks losing to an all-Chinese alternative ecosystem.   This architectural fork makes porting and distillation less effective and strengthens the pro-American national security case for selling China deprecated GPUs imo. Also I will attest that I did not wake up a loser this morning.

I'm going to make a prediction that will shock many of you: Going forward, for at least the next few years, MTL licensure in the US will be obsolete. And everyone who has a litany of MTLs will feel the compliance burden of an outdated vehicle they bought and paid for. In it's place will be the occ bank trust charter (which is not the same as a depository bank charter like you'd interact with at JP morgan). The bank trust charter will turn into a kind of federal crypto license. You've seen over the last few months many conditional or full approvals for bank trust. @stablecoin, @coinbase, @bitgo, @nubank, etc That will be the new norm. Bookmark this post.

Time for Mercor to put this to the test 😂

Had a fun time with @og_doctourist chatting about startups!

very interesting take on world models - what if we make them multiplayer?

“if each synapses maps to a neural net weight”

Here we introduce SAGE: Scalable Agentic 3D Scene Generation for Embodied AI, which can generate sim-ready 3D scenes with agents following user demands at scale, ready for robotic action generation. Paper, code, and SAGE-10k dataset are all released! nvlabs.github.io/sage/

Meet SceneSmith: An agentic system that generates entire simulation-ready environments from a single text prompt. VLM agents collaborate to build scenes with dozens of objects per room, articulated furniture, and full physics properties. We believe environment generation is no longer the bottleneck for scalable robot training and evaluation in simulation. Website: scenesmith.github.io 👇🧵(1/8)

New #NVIDIA Paper We introduce Motive, a motion-centric, gradient-based data attribution method that traces which training videos help or hurt video generation. By isolating temporal dynamics from static appearance, Motive identifies which training videos shape motion in video generation. 🔗 research.nvidia.com/labs/sil/proje… 1/10

The authors ask whether an N-layer ViT can be rewritten using just K<<N layers by recurring on them. Remarkably, they match DINOv2 performance with only 2-3 layers. The paper also offers rich dynamical-systems analysis. Very cool work! 🔗arxiv.org/abs/2512.19941

A lot of people have been saying over the past week that mass egocentric data collection "solves" dexterous robotics. Important to remember that autonomous vehicles was ~15 years of adding 9's to the disengagement rate. We haven't even started that process with dexterous robots.

Introducing Nexels: Neurally-Textured Surfels for Real-Time Novel View Synthesis with Sparse Geometries Nexels render in real-time at high quality without needing millions of primitives. Site: lessvrong.com/cs/nexels/ Paper: arxiv.org/pdf/2512.13796 Code: github.com/victor-rong/ne…

Excited to introduce PolaRiS, a real-to-sim recipe for turning short real-world videos into high fidelity simulation environments for scalable and reliable zeroshot generalist policy evaluation. polaris-evals.github.io (1/N 🧵)

super cool work! active perception is underrated imo but will be a necessary part of real world robots

Our research: Adversarial Flow Models (AF) arxiv.org/abs/2511.22475 AF unifies Adversarial and Flow Models. Unlike GANs, AF learns optimal transport (stable). Unlike CMs, AF only trains on needed timesteps (save capacity). We can train super-deep 112-layer 1NFE model! SOTA FIDs!

I think robot pets are going to be a huge market.