Sean Arietta

Today I’m sharing CVP Spin Overlay (Curvature-Variable Propagation)—an integrity overlay for NVIDIA GPU workflows that treats data propagation as a geometry-dependent process rather than a static bitstream. CVP Spin Overlay is designed to complement existing ECC/RAS by adding an algorithm-aware integrity plane with: • Real-time anomaly detection using curvature/invariant checks • Stale-state suppression for asynchronous, multi-stream and multi-GPU execution • Deterministic replay + Proof Bundles (TEVR-aligned) to support auditability, replication, and high-assurance validation This is built for the workloads where “plausible but wrong” is unacceptable: CFD and multiphysics simulation, long-horizon AI training, and validation-grade compute. We’re also targeting pilots with DoD, DOE, and NSF test labs where reproducibility and verification are mandatory. If you’re working on GPU integrity, deterministic replay, or validation pipelines at scale, I’d welcome a technical conversation. #206Innovation #CVP #CurvatureVariablePhysics #GPUComputing #NVIDIA #CUDA #HPC #AIInfrastructure #AIFactory #DeterministicComputing #ReproducibleResearch #ModelTraining #Simulation #CFD #GPUDirectStorage #NCCL #DataIntegrity #CyberResilience #NationalSecurity #DoD #DOE #NSF #GPU #NVIDIA #CUDA #HPC #AI #DataIntegrity #DeterministicComputing #Cybersecurity #206Innovation #NationalSecurity

1

Apple just released SHARP(Sharp Monocular View Synthesis in Less Than a Second), it turns 2d images into photorealistic 3d images using 3D Gaussian Splatting. It uses consumer GPU and it's very very fast. You may wonder what photorealistic is. Imagine history documentaries where they take an old photo and free objects from the background and move them round giving the illusion of parallax movement. This software does that in less than a second, creating a 3D model that can be accurately moved (or the camera for that matter) in your video editor. It's not new, but this one is fast and "sharp". But there's a lack of human portraits in the example cases. My experience with all these solutions to date is that when viewed stereoscopically the people end up looking like 2d cutouts against the background. I haven't seen this particular model to use stereoscopically so I can't comment as to its effectiveness, but the lack of a human face in the example set is likely a bit of a tell. Here's the paper link: https://lnkd.in/gBrcne-s

1

Gemini 3 is not just an algorithm — it’s a hardware story. Years ago, during my PhD on haptic interfaces (technology that lets you feel virtual forces like weight or texture), I learned something interesting about human perception: If the force-feedback loop dropped below 500 Hz, the illusion collapsed. The surface stopped feeling smooth — it turned into a “buzzy” vibration. To fix it, we needed raw compute. That’s when I first crossed paths with CUDA (Compute Unified Device Architecture). I realized developers were using this powerful platform, bypassing the graphics pipeline of NVIDIA GPUs — ignoring pixels entirely — and using the GPU as a massively parallel math engine. And for the last decade, GPUs ruled: Crypto mining (BTC), physics, deep learning — all benefited from their versatility. But Gemini 3 hints at something interesting. Public reports suggest that Google trained the model primarily — perhaps entirely — on its custom TPUs (Tensor Processing Units, built with Broadcom). Whether all training phases used TPUs or GPUs were involved in certain pipelines is hard to verify externally — large-scale training is rarely that transparent. So I’m not claiming an absolute “GPU → TPU revolution.” Reality is more nuanced. What is more clear: TPUs are specialized for matrix math, stripping away graphics hardware and legacy GPU overhead. This specialization unlocks huge performance-per-watt gains, which is now the true bottleneck in scaling AI. GPUs, meanwhile, remain incredibly powerful and flexible — indispensable for countless workloads beyond deep learning. In a sense, and maybe: GPUs were a clever hack. TPUs are engineered optimization. Both have a place. But in opinion, the industry’s direction of travel is becoming visible: The biggest players are no longer hacking performance — they’re designing silicon around the workload. The takeaway: AI breakthroughs today are not only about better algorithms — they’re about better compute. The future belongs to whoever controls the most efficient silicon. #AI #Gemini3 #Hardware #TPU #GPU #NVDA #GOOG #AVGO #DeepLearning #Haptics #Engineering IREROBOT LTD

43

AI is starting to do something we’ve never seen before in go-to-market. It’s collapsing whole layers of manual work and turning small teams into machines. The thread that really hit me this week was from Jeanne DeWitt Grosser (ex-Stripe, now COO at Vercel). Three examples stood out because they point to the same theme: First. The exact GTM idea that Stripe could not get working in 2017 suddenly works today. Back then, Jeanne tried to auto-personalize outbound emails using deep company data. Even with world-class data scientists, the error rate was too high. Today, that exact same workflow is not only possible but reliable. The models caught up to the ambition. Second. At Vercel, a single GTM engineer built an AI agent that replaced the output of ten sales reps. Lead qualification, outbound prospecting, and deal-loss analysis. All automated. It costs about $1k a year to operate and the human team moved to higher-value work. This is not a story about replacing people. It is a story about removing the repetitive parts of selling. Third. Their AI deal-loss bot is better than humans at diagnosing what really happened. A salesperson blamed pricing for a big lost deal. The bot reviewed every email, call, and Slack thread. It uncovered the real cause: no budget owner engaged and the value story didn’t land. Now the agent flags these problems in real time. It’s like having a sales coach running silently behind every rep. Here’s the pattern I’m seeing. The winning teams aren’t “adding AI.” They’re rebuilding their GTM engine from the inside out. They are removing the busywork that used to take entire teams. They are letting humans focus only on high-leverage conversations, strategy, and judgment. Most companies think AI will make their teams faster. The truth is AI is changing what teams actually do. If you want to see the full thread, check it out here: https://lnkd.in/gfuwjsgG

13

Anysphere (Cursor), that I profiled recently appears to be launching Composer 2. The shocking part? They claim it’s a frontier-level coding model that matches or beats the performance of flagship models like GPT-5.4 and Opus 4.6 at roughly 1/10th the cost. Composer 2 scores 60.0 on CursorBench and 59.8 on Terminal-Bench 2.0, outperforming Opus 4.6 and Sonnet 4.5. Cursor has priced the model at $0.50/M input and $2.50/M output tokens - drastically lower than any comparable frontier model. Launched a new RL process called self-summarization that allows the model to master complex tasks requiring hundreds of individual actions. Bloomberg and The New Stack also have pieces out on this today.

7

5 Comments