Johnny Lee

What a plot twist! Apple & Google just announced a multi-year AI collaboration and yeah, it’s a big signal for where this industry is going 🤖 Yesterday Apple and Google confirmed a multi-year collaboration where the next generation of Apple’s “core AI models” will be built on Google’s Gemini models + cloud tech, helping power future Apple Intelligence features, including a more personalized Siri. Why this is very relevant? 1) Apple has finally decided to integrate existing AI technology rather than continuing to struggle to create its own. Apple's choice of Google's technology over OpenAI or similar will mean truly enormous growth for Gemini. This means that the era of “AI silos” is ending. Even the biggest rivals are realizing that scale and quality matter more than pride. (hey Amazon, check this!) 2) Distribution at insane scale, as Apple has an installed base of 2+ billion active devices. Putting stronger AI into that ecosystem isn’t a feature update, it’s a clear market shift. 3) The money (and momentum) is clearly here. Gartner forecasts $644B in global GenAI spending in 2025, and highlights that a huge part of that growth is driven by AI becoming embedded into consumer devices. That’s exactly the lane Apple plays in and how other competitors are trying to grow (see OpenAI's next potential devices...) 4) Apple is officially “back in the AI race” narrative-wise! For a while, the story was: Apple is late, Siri is overdue, and others are moving faster. This move changes the vibe as they are going all-in with tech they publicly describe as the “most capable foundation” for what they want to ship. 5) Privacy will become a differentiator, as Apple is positioning this as "Gemini-powered intelligence, delivered with Apple’s privacy approach". Let's see what this means at the end of the day... I have to say that I love seeing moves like this. Big companies working together just because it makes sense to make it easy is exactly what we need many times. With these type of market shakes we are entering a phase where AI isn’t a feature anymore. It’s becoming part of the default user experience.

8

1 Comment

🚀 Computing at the Speed of Thought I’m proud to share the formal release of the Dillon Equation Framework—a curvature‑based computation model where geometry rewrites the limits of light speed. This work integrates Quantum Overlay Architecture, VSL cosmology simulations, and predictive lattice geometry, forming the foundation for a new era of symbolic infrastructure. Key highlights: • 📐 Curvature defines computation: The Dillon operator `\( \partial c / \partial G \)` unifies cosmology and computation. • ⚡ Performance breakthroughs: Early pilots show 20–30% latency reductions and up to 40% energy savings in data centers. • 🌌 Cosmological validation: VSL fits relieve long‑standing tensions in `\( H_0 \)` and `\( S_8 \)`, aligning with DESI/Euclid data. • 💡 Business impact: Projected $50–100B market opportunities by 2030 across AI, biotech, and finance. 📜 A humble lineage: • Newton gave us the laws of motion and gravity. • Maxwell unified electricity and magnetism. • Einstein reframed gravity as geometry. • Shannon formalized information theory. • Turing defined computation itself. Each of these breakthroughs rewrote limits by revealing deeper structures. The Dillon Equation seeks to honor that tradition—showing that curvature rewrites the limit, and positioning computation within the same continuum of discovery. Phase I pilots begin January 2026. The journey ahead is about scaling, integrating, and immortalizing this framework across industries and archives. 🏛️ And just as past breakthroughs were etched into textbooks, plaques, and monuments, the Dillon Equation is being preserved as legacy‑grade artifacts—from capsule inscriptions and digital monuments to museum‑grade archives—ensuring this authored physics stands alongside history’s great architectures of thought. 🔗 #QuantumOverlay #PredictiveGeometry #ConsciousCosmos #InnovationLegacy #DillonEquation

5

5 Comments

We've been building for AI that responds. GPT-4o forces us to build for AI that reacts. Its near human-level latency in audio and vision processing, validated by benchmarks where it edged out Claude 3 Opus in real-time interaction, radically changes the interface paradigm. This isn't about better chatbots anymore. It's about seamless, naturalistic interaction that unlocks entirely new workflows. The real challenge now is how fast we redesign our systems to truly leverage this real-time intelligence.

1

We kept getting asked "what does DeviceNexus actually do?" So we built something you can try in 30 seconds. Open a browser, talk to a Pollen Robotics Reachy Mini. It sees, it listens, it moves. A full AI pipeline plus motion control plus physics simulation, all running on a single NVIDIA DGX Spark (or if you are like me and have affinity towards using Jetson) under a desk. No cloud, no API keys, just URL. (Full video in comments)                                                                                                                                                    The interesting part wasn't the demo itself though. It was what it took to make it work. Because a robot isn't just AI models. It's models plus motion control, simulation, health monitoring, audio pipelines, and deployment tooling, all running concurrently with hard latency constraints. Few things we learned: Edge optimization is non-negotiable: The Jetson community has done incredible work pushing AI models onto edge hardware with quantization recipes, inference runtimes, model zoos. We've learned a ton from it. But getting a single model to run fast isn't the same as getting an entire pipeline to share one GPU under real-time constraints. That requires profiling end-to-end and knowing exactly where your latency budget is going before you start optimizing. Simulation must be in the loop, not beside it. MuJoCo runs alongside the real robot (again shoutout to the amazing work done by Pollen & Mujoco together). Every behavior change gets validated in sim before it touches hardware. Bad policy? Caught before it moves a joint. Multi-model pipelines are the real deployment unit. A robot isn't one model. It's speech, vision, reasoning, control, safety, motion, all running concurrently on constrained hardware. Orchestrating that full pipeline is harder than training any single model in it. Continuous improvement is an infrastructure problem. The robot on the floor tomorrow should be better than the one today without sending an engineer to the site. OTA model updates, production data flowing back, and a pipeline that turns operational experience into better behavior automatically. At DeviceNexus this is what we build. Model-agnostic, hardware-agnostic. The same infrastructure running this demo handles fleets. #PhysicalAI #Robotics #EdgeAI #DeviceNexus #NVIDIAGTC

55

3 Comments

Fei-Fei Li just raised $1 billion for World Labs. Largest spatial AI round in history. Autodesk alone committed $200 million. They're betting on 3D world models to automate design workflows. Meanwhile, Yann LeCun left Meta to start a $5 billion world model lab. Two of AI's most influential researchers are betting the next paradigm isn't better language models — it's machines that understand physical space. Pattern emerging: → LLMs commoditized → Spatial intelligence = new frontier → Enterprise buyers moving early (Autodesk proof) Most founders are still optimizing text generation. The big bets are already moving to spatial.

5

1 Comment

Jensen Huang is a visionary. Regardless of your industry you can see a straight line to the future by watching his GTC keynote from yesterday end-to-end. ↔️ Early on he is discussing NVIDIA’s neural rendering technology in their new graphics platform, and says “We combined 3D graphics, structured data with generative AI, probabilistic computing. One of them is completely predictive, the other one probabilistic yet highly realistic. We combine these two ideas... Controlled perfectly and yet generating at the same time." Very useful advice for optimizing your site and your products for AI and search, yeah? #agenticcommerce #aeo #geo #seo

7

2 Comments

There's a strange symmetry in the way AI tools are now babysitting the very code they helped flood into existence. CodeRabbit just locked down a $60M Series B, led by Scale Venture Partners with nVentures, NVIDIA's venture arm, plus repeat conviction from CRV, Harmony Partners, Flex Capital, Engineering Capital, and Pelion Venture Partners. That lifts their total raised to $76M and pegs valuation at $550M. For a startup barely out of its rookie contract, that's velocity. CEO Harjot Gill isn’t new here. He co-founded Netsil, sold it to Nutanix, stayed on as Senior Director of Technology, then built FluxNinja, which CodeRabbit acquired in 2024. That deal brought Gill fully into the driver’s seat, joining Co-founder Guritfaq Singh, who spent a decade leading engineering at Alegeus Technologies. Together, they saw GitHub Copilot ignite a frenzy of auto-generated code and realized the choke point wasn’t creation, it was review. You can crank out pull requests all day, but without oversight, you’re just building castles on quicksand. CodeRabbit turned that friction into fuel. 8K paying customers, with Chegg Inc., Groupon, Life360, and Mercury in the mix. Over 100K open-source projects using Pro-tier features for free. 10M pull requests handled across a million repos. 10K #developers on the platform daily. Revenue scaling 10x in a year, now topping $15M ARR. Month-over-month growth at 20%, headcount more than doubling in a quarter with plans to double again by 2026. Momentum engineered. The product isn’t a static analyzer. Context-aware AI pulls from historical pull requests, code graphs, project docs, and even #Jira and #Linear. Line-by-line feedback with one-click fixes. Pull request summaries that don’t read like spam. Integrations with 40+ linters and security tools. Adaptive learning tuned to each team’s standards. Recent upgrades: #CLI support, auto unit test generation, pre-merge #guardrails, and Model Context Protocol integration. For enterprise: #SOC2TypeII, #GDPR compliance, zero-data retention, no model training on customer code. Trust at the root. This round is about governing Copilot. CodeRabbit is staking its claim as the review layer for AI-powered development, the infrastructure that turns AI #codegeneration from novelty into necessity. And when the noise settles, governance will be the difference between scaling with confidence and crashing in production. Congrats to the CodeRabbit crew. The market just bet $60M that your answer to one question will define the next wave of software: when the machines ship the code, who makes sure it’s any good? #Startups #StartupFunding #VentureCapital #SeriesB #AI #DevTools #Data #CodeReview #OpenSource #Security #Compliance #Infrastructure #Technology #Innovation #TechEcosystem #StartupEcosystem #Hiring #TechHiring If software engineering peace of mind is what you crave, Vention is your zen.

5

Late night sessions and why I’m awake at 1:30AM Canonicalized memory. Vector indexing. Policy-based recall. Context compression. Graph memory. That’s the ideal. That’s the architecture I’m building toward But tonight wasn’t about building the flashy layers. It was about shipping the boring one correctly. File-based memory. Facts stored in plain files. You can open them. You can read them. You can understand exactly what the system believes and why. If the agent forgets something, you don’t need to inspect a vector database to debug it. I saw how powerful that felt in OpenClaw. The simplicity of memory living in files made the system feel grounded. It felt tangible. That stuck with me. So instead of jumping straight to embeddings and semantic search, I rebuilt the foundation first. Because here’s the uncomfortable truth. You can’t build graph memory without canonicalized memory. You can’t implement policy-based recall if you don’t have confidence, provenance, and structure. And you shouldn’t bolt vector indexing onto a rebuild process that isn’t deterministic. If the base layer wobbles, everything built on top of it is performance art. Tonight was about making memory something I can trust when the system boots. It should start fast. It should behave the same way every time. It shouldn’t require a model call just to reconstruct state. The advanced layers will come. Vector indexing is already scoped. Redis for session storage is on the roadmap. Graph memory will happen. But durability comes first. Tests are green. The roadmap is documented. The issues are filed. That’s enough for tonight.

5

1 Comment

AI companies are *finally* starting to pay artists. Udio just announced a deal with Merlin. If you generate AI audio on Udio, it will compensate independent artists if: a) they are under the Merlin umbrella b) they have opted in c) their music is used in the training pool. Merlin represents 15% of all recorded music so it's a big step. We don't know exactly what the deal looks like yet. But Merlin's previous deal with ElevenLabs worked as follows: 1. Artist opts in to AI usage. 2. Payments are calculated based on the song's weight in the training pool. 3. Final distributions are adjusted based on the song's "popularity score" This was seen as a blueprint for future deals so it might look something like this.

30

4 Comments