😺 🎙️ This company taught AI how to smell

Welcome, humans.

Imagine sending someone a photo of fresh coffee—and they could actually smell it through their screen.

Sounds like science fiction, right? But a company called Osmo just figured out how to digitize scent using AI, and it's one of the most fascinating things we've covered this year.

In our latest podcast episode, we sit down with Alex Wiltschko, Founder & CEO of Osmo, who told us all about his company’s AI tech and how Osmo became the first company to successfully “teleport” a smell from one room to another. No joke—they recreated the exact scent of a fresh-cut plum using AI, sensors, and a molecular printer.

Here's what makes this so incredible:

• The Science: While vision uses just 3 color receptors (RGB), smell uses 300+ olfactory receptors—making it 100x more complex. Osmo cracked the code by building the world's first “Primary Odor Map” using AI.

• The Breakthrough: Osmo achieved "scent teleportation" by reading molecules in one room, uploading the data to the cloud, and recreating that exact smell in another room. Alex describes the moment they teleported a plum as “one of the wildest days of my professional life.”

• Real Applications: Osmo Studio now lets anyone design custom fragrances in one week instead of the traditional 18-24 months. They've even launched 3 AI-designed molecules that have never existed in nature… and major brands are already using them.

• The Future: This isn't just about perfume. Dogs can smell cancer and Parkinson's earlier than any diagnostic. Osmo's long-term vision? Building sensors that can detect diseases through smell before symptoms even appear (and eventually, via sensors at a small enough size that could fit on your iPhone).

We also dive into why smell is wired directly to emotion and memory (your brain literally touches the world when you smell), how Osmo created a signature scent for the Museum of Pop Culture from just a photo, and Alex's hot take on why AI progress follows S-curves, not exponential growth.

TBH, we learned a ton about the science of smell during this interview; a very educational one if we do say so ourselves!

Watch and/or Listen now: YouTube | Spotify | Apple Podcasts

Dive deeper with these resources:

• Learn more about Osmo 

• Scent Teleportation Update

• Build your own scent on Osmo Studio

• Learn more about the sponsor of this video, Flora.

ICYMI: Three More Episodes You Should Watch

Here are three other conversations packed with actionable insights:

Ely Greenfield, CTO of Adobe Creative Products

What you'll learn: Adobe's strategy for making AI additive, not subtractive—and why that matters for your workflow.

The biggest reveal? The feature that got the loudest applause at Adobe Max wasn't flashy AI magic—it was a simple voice command: "rename these layers for me." Ely breaks down why Adobe focuses on solving the boring, time-consuming work that drives creators crazy (like rotoscoping objects frame-by-frame).

You'll also learn:

• Adobe's 3-part model strategy: They're building their own models, integrating partner models (OpenAI, etc.), AND letting you train custom Firefly models on your own artistic style

• How to train a Firefly model on your unique work—perfect for illustrators, agencies, and brands who want consistent AI-generated assets

• Why Adobe Express is the new on-ramp for creators who don't need the full Creative Cloud but want professional AI tools

Practical takeaway: If you use any Adobe tools, watch Ely explain the "additive not subtractive" philosophy—it'll change how you think about integrating AI into your creative process without losing control.

Logan Lawler, Dell Technologies (Dell Pro Max + NVIDIA)

What you'll learn: The actual hardware requirements for running AI locally—and how to avoid wasting money on marketing hype.

Logan cuts through the "AI PC" confusion and gives you the real specs that matter. His core advice? Under-buying is the bigger risk—because three years from now, you'll wish you had more VRAM.

You'll also learn:

• NPU vs GPU explained: What's the actual difference, and when does each matter? (Spoiler: Most "AI PC" marketing focuses on NPUs, but you need a GPU for real AI work)

• The 24GB GPU sweet spot: Why Logan recommends investing in an RTX Pro Blackwell with 48GB VRAM—it'll run local models like ChatGPT's o1 and give you years of future-proofing

• Practical local AI demo: Watch Logan transcribe a 38-minute podcast episode in 15 seconds using NVIDIA Parakeet—then turn it into a blog post with a local LLM

• Mind-blowing use case: Turning 2D wedding photos into explorable 3D memories using Gaussian splatting (perfect for photographers looking for new revenue streams)

Practical takeaway: If you're buying a new computer in the next 6 months, watch Logan's hardware requirements breakdown before you buy. He explains exactly what specs you need for Comfy UI, LM Studio, and local model inference—and why the 96GB Blackwell GPU is basically an H100 in a desktop.

Kari Briski, VP of Generative AI at NVIDIA

What you'll learn: How to actually run NVIDIA's open-source Nemotron models—from laptops to enterprise servers—and when to choose local vs. cloud AI.

Kari walks through NVIDIA's open-source Nemotron family and gives you the blueprint for deploying AI on your own hardware. The key insight? Businesses want to specialize AI with their proprietary data—and you can't do that by uploading everything to OpenAI's API.

You'll also learn:

• Hardware requirements for each tier: Nano runs on an 18GB GPU (perfect for laptops), Super fits in a single data center GPU, Ultra needs eight GPUs but stays within one box

• When to use local vs. cloud: Kari explains why enterprises are moving from cloud-first to local-first—it's about data privacy, specialization, and not giving away your "intelligence"

• The "data flywheel" strategy: Start with a larger model (Super or Ultra), collect real-world data, then distill down to a faster Nano model for production

• Why smaller models are getting better: Kari breaks down distillation, neural architecture search, and reinforcement learning “gyms” that make 9B models nearly as good as 70B models

Practical takeaway: If you're a developer or enterprise considering local AI, watch Kari's explanation of reinforcement learning environments—she describes how NVIDIA trains models by spinning up "fake Salesforce interfaces" to teach tool use. It's the future of model training, and it's happening now.

Stay curious,
The Neuron Team

P.P.S: Love the newsletter, but don’t want to receive these podcast announcement emails? Don’t unsubscribe — adjust your preferences to opt out of them here instead.