Podcast Summary: This Bittensor Subnet Could Cut Drug Discovery Costs in HALF | E2267

This Week in Startups | Host: Jason Calacanis (guest hosted by Alex & Lon Harris)
Date: March 26, 2026

Main Theme

This episode focuses on the innovative projects being built atop the Bittensor decentralized AI network, with founders from three very different subnets: Meta Nova (drug discovery), Bitcast (content/creator economy), and Score (vision models for commercial applications). Alex and Lon explore how these subnets leverage Bittensor’s crypto incentives and distributed compute to attack previously intractable problems, democratize participation, and accelerate innovation.

Key Discussion Points and Insights

1. What is Bittensor? (04:00)

• Bittensor is a decentralized network, using blockchain technology and crypto tokenomics to incentivize the global development and sharing of “useful AI models” through application-specific subnets.
• Subnet owners/operators set problems; “miners” (either humans or agents) compete to solve them; validators score submissions and allocate rewards.
• “Think about a hackathon that never sleeps, that anyone can enter, applied to any problem you can describe.” — Michaela Baso, Meta Nova [05:45]

2. Subnet 68: Meta Nova — Decentralized Drug Discovery (03:40–30:56)

The Problem

• Drug discovery is infamously slow (avg. 10 years, $2.6B per drug), high-risk, and expensive, with much of the process “like shooting in the dark.”
• Most failure points occur early, during the search for promising molecules (“virtual screening”).

The Approach

• Meta Nova decentralizes molecular discovery:
  • Miners submit candidate molecules or chemical search algorithms in ongoing competitions.
  • Two incentive mechanisms: submit molecules for a given target, or submit optimized chemical search algorithms (can be open or closed source).
• “We started with a data set of a billion molecules, layered five combinatorial reactions, and now it's up to about 65 billion possibilities.” — Michaela Baso [10:45]
• Finalists are “heat picked”: best-performing, least toxic candidates advance to “wet lab” validation through contract research organizations (CROs).
• Meta Nova operates as a “virtual biotech,” running lean, maximizing “shots on goal,” and even supports others’ drug development as a gateway or screening-as-a-service platform.
• International partnerships (e.g., Shanghai’s Yalatane) enable expansion into new therapeutic classes, like nanobodies.

Timeline and Impact

• Some AI-developed drugs already in late clinical trials — “We could be seeing these earlier than expected... in the next three to five years.” — Pedro Pena [18:28]
• Cost and time reductions depend on efficiency improvements, global clinical trial arbitrage, and regulatory advances (e.g., FDA starting to recognize international trials) [19:43–22:12].

Unique Dynamics

• Open participation — “You don’t need a background in this field. We’ve reduced it to a search problem,” which has led to “cross-pollination of ideas” and even novel approaches outperforming industry standards. — Michaela Baso [22:58]
• Miners’ “unruliness” helps surface model weaknesses faster; tuning incentives is a continuous process, aligning the emergent behavior with meaningful outcomes. “You're productizing the unruliness.” — Lon Harris [26:33]
• “When you get it right, you can get really, really good things very fast... interesting agents are already helping select molecules, check for patent conflicts, etc.” — Pedro Pena [29:00]

3. Subnet 93: Bitcast — Creator Economy Reinvented (30:56–47:46)

The Opportunity

• Bitcast lets content creators mine crypto on Bittensor by generating engaging YouTube videos from campaign briefs.
• Workflow: Brands submit briefs → creators make videos following those prompts → AI/validators reward based on watch time (not raw views), ensuring engagement and quality.

Key Innovations

• Decentralizes and automates the notoriously high-admin brand deal process, democratizing access to creator monetization.
• “Our miners are YouTubers... people essentially mine crypto with YouTube content.” — Tom Bliers, Bitcast [32:45]
• AI validation system scans any language for brief compliance, unlocking massive global, distributed creator participation.

Market Impact

• Shifts rewards from top 1% “superstar” creators to the “long tail” by reducing brand admin burden and unlocking campaigns from thousands of micro/influencers at once.
• “The creative economy is absolutely booming... about $250 billion worldwide.” — Tom Bliers [40:48]
• “Now you can activate all of these creators at the push of a button... with the same budget.” [41:30]

Challenges

• Ensuring AI validation quality is ongoing; scaling to check 100,000+ videos/day is on the near-term roadmap [45:04].

Growth Metrics

• Q1 2026: 2M subscribers in the network, 50 active creators, accelerating 40–60% month over month [46:11].
• Demand is already scaling beyond Bittensor-native brands to mainstream tech, AI, and crypto products.

4. Subnet 44: Score — Vision Models for Agents & Commercial Applications (55:30–71:41)

The Problem

• Most computer vision models are too bulky/expensive for practical use, or too general-purpose to excel at specialized tasks.

Subnet Model

• Score incentivizes miners to “distill” large vision-language models (VLMs) into tiny, specialized ‘skills’ (modules, e.g., for detecting a person, a car, etc.).
• “If you want to build with vision, not just text, you need small, accurate, cheap models. Our miners turn huge VLMs into expert vision skills anyone can use—even on a CPU.” — Max Sebti, Score [57:05, 67:06]
• Each skill competes in its own “winner-take-all” emission structure — best model for a task gets all rewards [61:49].

Customer Integration (Mamico App)

• GUI abstracts Bittensor mechanics; users can prompt for needed vision skills, have them fine-tuned, deployed, and integrated into their workflow without any ML expertise.
• Example: A gas station chain using a custom-trained model to detect/alert on pump collisions, reducing costly delays [63:30–65:02].
• Two tracks: public open-source models; private client-specific skills [65:56].

Technical Achievement

• Models run on edge CPUs (as low as 50 MB), making vision AI affordable and deployable for businesses with minimal IT infrastructure [67:06].
• “That's what we all need... I rarely need Opus 4.6 for my problems; I need a very specialized model that my agent could use just to help me write tweets.” — Lon Harris [67:35]

Go-to-Market

• “We believe in vision vibe coding” — making CV accessible to non-experts and driving adoption through community and partnerships [69:17].
• “2026 is definitely our commercial year.” — Max Sebti [70:56]

5. AI Bubble Burst? PolyMarket Segment (48:44–54:48)

• Hosts analyze a prediction market about a potential AI downturn: dramatic stock price collapses, bankruptcies, H100 price crash, etc.
• Consensus: “Unlikely this year — the field is only accelerating from the inside, though outsider perception of a bubble remains strong.” [53:35]

Notable Quotes & Memorable Moments

• “Think about a hackathon that never sleeps... that's what a Bittensor subnet is.” — Michaela Baso, Meta Nova [05:45]
• “We’ve reduced [drug discovery] to a search problem...you don’t need a scientific background to participate.” — Baso [22:58]
• “You're productizing the unruliness.” — Lon Harris, on miners exploiting incentive mechanics [26:33]
• “You can now have 100 different languages all on the same campaign. ... as we build, the reach we can do would take multiple teams, all speaking different languages.” — Tom Bliers, Bitcast [43:57]
• “Our miners are YouTubers ... people mine crypto with their YouTube content.” — Bliers [32:45]
• “If you want to build with vision, not just text, you need small, accurate, cheap models. Our miners turn huge VLMs into expert vision skills anyone can use—even on a CPU.” — Max Sebti, Score [57:05, 67:06]
• “That's what we all need...I rarely need Opus 4.6 for my problems; I need a specialized model that my agent could use to help me write tweets.” — Lon Harris [67:35]
• “2026 is definitely our commercial year.” — Max Sebti [70:56]

Important Timestamps

• [03:40] — Intro to Meta Nova, Bittensor overview
• [05:45] — Subnet actors: owners, miners, validators
• [07:04] — Deep dive into Meta Nova’s drug discovery model
• [10:45] — Molecular search scale: 65 billion candidate molecules
• [18:28] — AI-developed drugs in trials: “3 to 5 years” till first approved
• [22:58] — Openness: non-experts can participate, cross-pollination of algorithms
• [26:33] — Miners’ adversarial gaming as feature, not bug
• [30:56] — Intro to Bitcast (creator economy/YT mining)
• [32:45] — “Our miners are YouTubers”
• [40:48] — Creator economy size: $250B, democratizing participation
• [43:57] — “100 languages on the same campaign... AI validation”
• [55:30] — Intro to Score (vision models)
• [57:05/67:06] — Technical breakthrough: “small, specialized models” on CPUs
• [63:30] — Use case: instant property damage detection for fuel stations
• [70:56] — 2026 as commercial push for Score/Manaco

Language & Tone

The episode is fast-paced, enthusiastic, and playful. The hosts and guests are technically deep but prioritize broad accessibility for listeners. Memorable moments include Alex’s nerdy joy at “subnets with numbers” and playful grilling of guests on secret new partnerships, and Lon’s humorous commentary on economics and prediction markets.

Summary Conclusion

This episode provides a rich and accessible tour of how Bittensor enables radically different use cases through tokenized, decentralized “subnets”—from smashing barriers in drug discovery, to democratizing the creator economy, to instant, specialized computer vision apps. Each project reveals not just technical ingenuity, but also creative applications of incentives and community, offering blueprints for what the future of AI infrastructure—and the startups built atop it—might look like.