Joshua X.

Perplexity launched a product last week called "Computer." It coordinates 19 different AI models to complete complex workflows. Claude handles one subtask. Gemini another. Grok picks up a third. $200/month. Their CEO Aravind Srinivas: "When models specialize, they just become tools similar to the file system, CLI tools, connectors, browser, search." If he's right, the model makers are building commodities. The orchestration layer captures the margin. I wrote about the AI Automation Spectrum last fall. What Perplexity did is productize the top of that spectrum for $200 a month. Six months ago this was a Series A pitch deck. The question I keep turning over: if Perplexity can sit on top of OpenAI, Anthropic, and Google, routing work to whichever model handles it best, what are the model makers actually selling? Full piece here: https://lnkd.in/gJMYAc47

19

4 Comments

NVIDIA just dropped Nemotron 3 Super — a 120B parameter model with only 12B active at inference, ~478 tokens/sec, already adopted by Palantir, Siemens, and Perplexity. https://lnkd.in/dHv7_c2B It's not the smartest model out there (behind Qwen 3.5 and GPT-5.4 on raw intelligence). But that's not the point. The point is what this signals: NVIDIA is commoditizing agentic AI models as fast as possible — open weights, open data, open recipes — because their business model is selling GPUs, not models. A model is ultimately just a file with weights. And this is where it gets interesting for enterprises: When the model layer becomes a commodity, value shifts to the platform — the orchestration, the knowledge, the domain context. This could drive a bit of darker angle: as AI companies race to build better models, the temptation to use customer data for training becomes enormous. Not sure i would trust most of them with sensitive enterprise data. The incentive structures just don't align. Here's what I think will happen next: → Efficiency becomes top of mind. CIOs will stop blindly paying for "AI usage" and start asking: which models are running, at what cost, and what's the actual ROI? → Pricing pressure accelerates. When you can run a 120B model at 12B inference cost, the old per-token pricing model starts looking like a bad deal. → The conversation shifts from "we use AI" to "AI delivers measurable business outcomes" — and organizations will demand platforms that prove it. We're entering the era where enterprises care less about which model is behind the curtain and more about what results it delivers per dollar spent. #AI #EnterpriseAI #AgenticAI #NVIDIA #Nemotron #AIEfficiency

17

1 Comment

Treating Agent memory as a semantic compression + consolidation problem, rather than a context-extension problem. A recent paper, SimpleMem, shows why extending context is the wrong abstraction for long-running agents. ─── 👋 Hey, I'm Prateek, currently building Cloudeval AI at Ganak AI Labs

8

Congratulations to Jonathan Ross, Sunny Madra, Chamath Palihapitiya, and the entire Groq team. Having known the team personally, this is a rare outcome driven by first-principles engineering and long-term conviction — not momentum or hype. I just published a note on NVIDIA × Groq and why this transaction is strategically crucial for anyone allocating capital in AI infrastructure. A few takeaways that stood out: ▪ Inference is where AI economics now compound Training is episodic; inference runs continuously. That’s where latency, power, and unit economics determine returns. ▪ Groq represented an architectural, not incremental, threat Inference-first design (deterministic execution, memory close to compute) challenged GPU assumptions about efficiency and power at scale. ▪ NVIDIA didn’t acquire revenue — it acquired time and talent Licensing + acquihire compresses years of internal R&D into a single strategic move. ▪ Licensing is becoming the preferred consolidation model Similar to Meta / Scale AI, this structure secures critical capability while reducing regulatory friction. ▪ Power is the binding constraint in AI deployment Inference workloads run 24/7. Marginal improvements in power efficiency materially impact margins and scalability. At Good AI Capital, this deal reinforces our core thesis: Durable AI returns accrue at the inference layer — where models are deployed into real workflows and physical systems — not at the increasingly commoditized frontier model layer. Full analysis here 👇 🔗 https://lnkd.in/gQAXpC_s Pavel Danilov Rosemary Sagar, CFA Mark Tompkins Abhishek Pakhira William Elmore Emily Liggett Bill Young Daniel Kwan, MD, MBA Utkan Demirci Charles Wang Mary Tolan Lawrence Leisure Aaron C. Cook Alex Tang Touraj Parang 🤖 🤖 Ali Kashani, Ph.D., Courtney Pollard III, MD, PhD Atuanya Thomas, MSME Mike Zhai Nithin Johnson Andrew M. Howard Cooper Irina Tanenbaum Ronnie Vasishta Or Priel eric sippel Sanjay Sathe Walter Baumgartner Ryan Batenchuk Jeff Bogan Yasmin Lukatz Eric Lazear Mark K Norbury Bruce Lane

26

Why Waymo will never catch up with Tesla? The answer is simple: scalability Waymo uses LiDAR (Light Detection and Ranging), a sensor that shoots laser pulses to map the 3D world. Waymo uses this along with cameras and extra hardware to make the car “more accurate”. Elon Musk calls this "stupid" Why? Every car needs multiple LiDAR units worth thousands. Every city needs HD maps. This model creates high fixed costs and requires massive capital investment for every new city or expansion area How Tesla is planning to achieve scalable Full-Self Driving (FSD)? First cut off LiDar Humans drive with eyes and a brain = Eight cameras on every car + AI Tesla uses a Vision-Only, Deep Learning approach, relying on eight relatively cheap cameras already built into every mass-market car. This makes the FSD software a zero-marginal-cost asset that can be deployed globally with a simple over-the-air update. Tesla has millions of customer vehicles on the road, operating in every corner of the world. This creates a Data-Flywheel -> More cars = More Training data But Waymo is limited to Geofencing which is designed to work flawlessly only within its pre-mapped areas I believe Tesla will disrupt the entire automobile industry The only other player who is competing in this space is BYD For rest players who are not changing fast, a slow death is imminent #Tesla #FSD #Waymo #AutonomousDriving #VisionOnlyAI #LiDAR #ElonMusk #AIRevolution #FutureOfMobility #DataFlywheel #SelfDrivingCars #DeepLearning #TechDisruption #Toyota #Volkswagen

4

The authors explore whether using low‑precision, block‑scaled numeric formats — specifically the “Microscaling (MX)” formats in NVIDIA’s Blackwell architecture — is viable for training large language models (LLMs). MX formats share a scale across blocks of parameters to reduce memory and compute requirements while performing forward/backward matrix multiplications (GEMMs) at reduced precision. However, across nearly a thousand language‑model training runs, spanning compute budgets from ~2×10¹⁷ to ~4.8×10¹⁹ FLOPs and exploring a wide range of weight and activation precision combinations, they repeatedly observe sharp, stochastic instabilities in training loss — particularly at larger compute scales. To understand and address these instabilities, the authors run controlled experiments on smaller “proxy” models while varying architecture, hyperparameters, and precision settings. They identify that multiplicative gradient bias — introduced by quantizing layer‑norm affine parameters and some fraction of activations — can trigger runaway divergence. Importantly, they show that by switching precision schemes mid‑training (i.e., hybrid configurations), the instabilities can be mitigated or delayed. In fact, some hybrid precision setups recover performance competitive with full‑precision training, suggesting MX-based training can be practical with careful stabilization strategies. https://lnkd.in/gMasFK6q

1