Ventura Labs

In this episode, we are joined by Peyton Spencer and Seby Rubino of Ditto and BigTensor, who detail the development of an open-source agentic operating system built on Bittensor Subnet 118. They discuss Ditto's architecture as a unified memory layer spanning AI tools and applications, the design of the Subnet 118 agent memory competition, and BigTensor's structure as a Bittensor-native builder studio. Peyton and Seby also detail the vibe mining concept enabling agent-driven subnet participation, the knowledge graph licensing model allowing users to monetize their AI memory, BigTensor's buyback-and-burn tokenomics, and Ditto's integration roadmap with Hippius, Desearch, and Gittensor.

In this episode, you'll learn

• Ditto's architecture as a cross-application memory layer, allowing users to vacuum their data from every AI tool they use and access it uniformly across Claude, Cursor, Codex, and Ditto's own apps via MCP servers
• The Subnet 118 agent memory harness competition, where miners compete to produce the highest-accuracy memory retrieval system benchmarked against LongMemEval, with Nick Anderson's multi-pronged semantic weighting approach already doubling internal accuracy
• The vibe mining concept, allowing non-technical users to prompt Ditto to provision compute, configure wallets, and spin up miners on other subnets using Ditto's internal TAO wallet — making Bittensor mining accessible through natural language
• Ditto's app platform enabling users to build, publish, and monetize apps from their knowledge graph, including consulting apps, Patreon replacements, and Shopify-style storefronts with Stripe-embedded payment processing
• The knowledge graph licensing model, where Ditto users sell selective access to their AI memory as subscription apps, with per-topic privacy controls distinguishing public from private memory slices
• BigTensor's operating model as a Bittensor builder studio: long-term token locks, linear vesting schedules, 100% buyback-and-burn tokenomics for all portfolio subnets, and a cohort model targeting two launches per year
• The Const agent-native onboarding flow — a prompt-triggered, zero-human-in-the-loop account creation mechanism for AI agents to self-register with memory services like MemZero
• Ditto's infrastructure roadmap: partnerships with Hippius for decentralized storage, Desearch and Leadpoet integrations for AI-native sales tools, and a modular compute marketplace enabling Mac Mini owners to earn by hosting Ditto applications

In this episode, we are joined by Paul Swaim, Chief Protocol Specialist at Yuma and former CIO of the Opentensor Foundation, who shares his path from academic IT and co-founding an AI malware detection company to becoming one of Bittensor's most embedded ecosystem operators. He discusses how Yuma functions as Bittensor's largest institutional validator and subnet accelerator, the evolution of subnets from the original Yuma consensus paper's vision of pure intelligence production to a protocol for commoditized work, and the systemic vulnerabilities in the off-chain developer tooling layer revealed by the July 2024 PyPI attack. Paul also details the two features of Conviction, the tradeoffs in Bittensor's bicameral governance design, the debate over the 18/41/41 emissions split, and the acceleration of AI-automated supply chain attacks and what participants can do to protect themselves.

In this episode, you'll learn

• Paul's career arc from North Idaho College IT director to co-founding Quantum Star Technologies (AI malware detection with Bittensor Triumvirate member Garrett Oetken) to CIO at Opentensor Foundation to Chief Protocol Specialist at Yuma
• How Yuma operates across mining, validation of 120+ subnets, subnet acceleration, and Yuma Asset Management, and why Yuma's institutional scale sits in tension with Bittensor's permissionless design
• The evolution of Bittensor subnets from the original vision of pure intelligence production to a broad protocol for commoditized work, spanning hardware, inference, bandwidth, genomics, and beyond
• Conviction's two distinct features: a locking mechanism to signal long-term commitment and prevent rug-pulls, and an adversarial subnet-king mechanism to recover stale, abandoned subnets through market-driven takeover
• The tradeoffs in Bittensor's emerging bicameral governance design, including how cross-investment relationships between validators and subnet teams complicate fair on-chain representation
• The July 2024 PyPI supply chain attack — OTF's detection and network lockdown in under 40 minutes, the role of community and exchange partners in the response, and the ongoing civil case
• Why supply chain attacks against developer tooling are rising over 70% year-over-year, driven by AI-automated scanning for exposed secrets, signed-package compromises, and the acceleration of attacker tooling
• Paul's operational security framework: separating personal and business key environments, understanding the attack surface of the tools you use, and evaluating risk by the damage cost of a compromise rather than its perceived likelihood

In this episode, Koyuki Nakamori, CEO of Perturb (Subnet 26) and co-builder of Vocence (Subnet 78), details the development of two complementary decentralized AI networks on Bittensor — an adversarial robustness testing platform and a decentralized voice intelligence layer. She discusses her career trajectory from head of ML at Headspace to the Avalanche ecosystem to dual-subnet founder, the evolution of the Bittensor ecosystem under dynamic TAO, and the strategic relationship between Perturb's perturbation-based data synthesis and Vocence's voice model training pipeline. Koyuki also details Perturb's adversarial image perturbation mechanism for solving the long-tail problem in AI classification, Vocence's nine-dimensional voice quality evaluation system and Qwen3.6-27B miner baseline, the winner-take-all competition structure, and her advice for subnet teams on closing the gap between research and consumer product.

In this episode, you'll learn

• Koyuki's career arc from production ML engineering at Headspace and the Avalanche ecosystem to founding two Bittensor subnets simultaneously across adversarial robustness and voice AI
• The mechanics of Perturb's adversarial robustness network, where miners compete to produce maximally deceptive image perturbations within distortion bounds to expose long-tail vulnerabilities in AI classifiers
• How Perturb functions as a data synthesis engine for niche AI domains — autonomous vehicles, medical imaging, and precision applications — by adding controlled noise to synthesize training data that real-world acquisition cannot provide
• Perturb's multi-modal roadmap, progressing from image perturbations to audio, video, and robotics applications
• Vocence's miner incentive model, where miners train voice AI models to surpass Qwen3.6-27B across nine evaluation dimensions — tone, accuracy, naturalness, and accent among them — in a winner-take-all competition for inference serving rights on vocence.ai/studio
• The nine standardized evaluation dimensions for voice AI and why achieving state-of-the-art requires passing a human Turing test, not just hitting benchmark scores
• Koyuki's strategic ambition to break Vocence out of the Bittensor ecosystem and serve non-crypto users as a voice interface layer for everyday AI agents and local business automation
• Her advice for AI-heavy subnet teams: ship a user-facing product, treat users as design partners, and maintain a rapid feedback loop rather than building in research isolation

In this episode, we are joined by Travis Millott, the TAO Templar, Bittensor educator and data analyst, returning for a second conversation centered on philosophy and governance. He discusses the evolution of money from proof of violence to proof of valuable work, Bittensor's decentralization progress since dTAO, and whether Const holds too much concentrated power. Travis also details his framework for reading alpha flywheel strength through buyback-and-burn rates, the rationale for quantum-proofing the chain before full decentralization, the reality behind AI-agent-owned subnets, and alpha conviction lockups as a trust signal for subnet owners.

In this episode, you'll learn

• Why Travis doesn't engage with Sam Dare's exit paper, believing it was AI-generated, and how he frames Bittensor's decentralization progress since dTAO as genuine movement toward an open market
• The evolution of money from proof of violence (fiat currency) to proof of hard work (Bitcoin) to proof of valuable work (Bittensor), and why TAO is only earnable by producing something the market values
• Travis's case for quantum-proofing Bittensor's substrate before full decentralization, arguing that a centralized chain can adopt quantum resistance faster than one requiring broad consensus
• Why Travis believes Const holds too much power — and why he believes every individual, including those at the OTF, is corruptible by power, making structural decentralization the only real safeguard
• His framework for evaluating alpha flywheel strength using buyback-and-burn rates and alpha supply growth as a proxy for a subnet's proximity to real revenue
• Alpha conviction lockups as a trust-building mechanism ranked below buyback-and-burns but above verbal commitment, and what mandatory lock periods could mean for subnet owner key transfer
• The two futures for Bittensor — permissionless global commodity competitions versus gradual centralization — and why Travis believes a fork is always available if the protocol goes wrong

In this episode, Bob Wold of qBitTensorLabs details the development of Enigma (Subnet 63), a decentralized pressure-testing platform that incentivizes hackers and researchers to break encryption and smart contracts through on-chain prize competitions. He discusses the pivot from a quantum circuit simulation subnet to an exploit-incentive model, the architecture of treasury wallets as trustless prize pools, and the looming threat of Q-Day — the moment quantum computers break modern cryptography. Bob also details the mechanics of the submission-fee anti-DoS system, the feral GPL licensing model that prevents large companies from commercializing open-sourced winning exploits, the roadmap from RSA-breaking challenges to elliptic curve and post-quantum cryptography hardening, and his thesis that Enigma can raise Bittensor's profile beyond the crypto ecosystem by generating trade-press coverage of genuine cryptographic breakthroughs.

In this episode, you'll learn

• The pivot story behind Enigma: why miners exploiting the old quantum-simulation subnet design led directly to a model that incentivizes breaking things on purpose
• How Q-Day works — Google's paper showing 1,200–1,400 logical qubits could break Bitcoin's elliptic curve encryption, IonQ's 2028 roadmap for 1,600 logical qubits, and why the photonic interconnect de-risking matters
• The architecture of Enigma's challenge model, including two-track submission paths for academic researchers (via the qBitTensorLabs website with Stripe fee payment) and native Bittensor hackers (alpha transfer), and why failing solutions stay confidential while winning solutions are published
• How treasury wallets work as trustless, validator-governed smart contracts — eliminating the ability to rug prize pools — and why Enigma is the first subnet to deploy them
• The feral GPL licensing strategy: winning exploits become open source in a way that is freely reusable by the Bittensor community but contractually untouchable by commercial companies, with licensing proceeds buying back subnet alpha
• The NISQ era explained — Noisy Intermediate-Scale Quantum — and why Bob argues the market should already be calling this post-NISQ, given that cryptographically relevant compute is closer than the $3.5B current market narrative suggests
• What "pressure testing as a service" looks like beyond encryption: AI model robustness challenges, smart contract auditing, and potential incentive-mechanism stress tests for Bittensor itself

In this episode, Ventura Labs interviews Siddhant Devkota (Sid), COO at TensorUSD (Subnet 113), sharing the development of a CDP-backed stablecoin protocol built natively on Bittensor. He covers the architectural rationale for using miners as competitive liquidation agents, the Dutch auction collateral system, the multi-source oracle design, the three-stream revenue model, the overcollateralization vault mechanics, the mint ceiling peg defense, a stablecoin payment gateway as a missing DeFi primitive, and TensorUSD's regulatory licensing push across Dubai and Europe ahead of the May 15 mainnet launch.

In this episode, you'll learn

• Liquidation agents: how TensorUSD rewards Bittensor miners as competitive liquidators — 99% of the liquidation reward to the fastest bidder, 1% distributed to all oracle-participating miners
• Dutch auction mechanics: a 120% liquidation threshold, six-second speed target, and the ALAS deadlock floor at 105% that forces immediate protocol execution if no bids arrive
• Vault model: depositing TAO at 150% collateralization to mint TUSDT and access liquidity without selling a TAO position
• Oracle security: capping miner oracle weight at 15–20%, layering in-house truth verification and multi-source APIs (CoinGecko, CoinGlass) to prevent majority-miner price manipulation
• Revenue model: 11% liquidation fees, 0.3% transaction fees, and 10% APY interest on vault borrowers funding the reserve pool that backstops bad debt
• Mint ceiling: tying maximum TUSDT supply to the reserve fund size to preserve the $1 peg even when collateral value drops below outstanding supply
• Payment gateway: Sid's case for a stablecoin payment layer as the most critical missing DeFi primitive in Bittensor, enabling Chutes and Resi to price services in stable value
• 100 TAO scenario: borrowing $10,000 TUSDT at 300% collateralization to cover real-world expenses, paying ~$520 in annual interest versus surrendering 50 TAO permanently
• Regulatory strategy: active VARA license applications in Dubai and licensing efforts across Europe, with mainnet launching May 15

In this episode, we are joined by Aleksander Muszyński and Jerzy Żółtowski of SOMA (Subnet 114), who detail the development of a decentralized MCP marketplace where Bittensor miners compete to produce validated, high-quality tooling for AI agents. They discuss their origins as Bittensor miners responsible for approximately 4% of all mineable TAO supply, the emergence of MCP as the universal standard for AI agent tooling, and why the Bittensor incentive layer is the right environment to produce best-in-class MCP servers. Aleksander and Jerzy also detail the mechanics of SOMA's weekly competition cycles, the context compression challenge and its Q&A-based evaluation across four compression ratios, the upcoming CoT compression competition designed to reduce agent token costs, and their vision for SOMA as a curated app store of miner-validated tools for AI agents inside and outside Bittensor.

In this episode, you'll learn

• Dendrite's origins as a high school friend group that became one of Bittensor's most prolific mining teams, accumulating approximately 4% of all mineable TAO over three years with a 50-person team focused exclusively on the ecosystem
• Why MCP became the standard for AI agent tooling, and how its emergence as a universal plugin protocol eliminated the siloed integration gardens that individual AI vendors previously maintained
• The structure of SOMA's weekly competition cycles, including submission, screening, isolated sandbox evaluation, and on-chain scoring phases designed to enforce consistent development rhythms for miners
• The context compression mechanics, including how miners submit algorithms evaluated against four compression ratios—20%, 40%, 60%, and 80%—with F1-scored Q&A tests verifying that compressed text preserves the most important factual content
• Why CoT compression matters for AI agent performance, including how structuring and trimming an agent's chain-of-thought diary both reduces inference costs and improves model reasoning by removing noise from growing context windows
• The transition from anonymous miner to public subnet owner, and why the community management demands, permanent visibility, and expectation of responsiveness represent a fundamental shift in how Dendrite operates
• The rationale for SOMA's closed-code design, where miners submit algorithms the validators never see directly, and how open-sourcing top-miner solutions historically stifles innovation by incentivizing small incremental tweaks over from-scratch breakthroughs
• SOMA's long-term vision as a curated app store for AI agents, where only quality-verified, miner-produced MCP tools are listed and made accessible to any agent inside or outside the Bittensor ecosystem

In this episode, Ventura Labs interviews Eyad Gomaa of Quasar, Subnet 24, Bittensor's youngest subnet owner, sharing how a Joseph Jacks rejection became the best thing that ever happened to his AI startup. He covers long context AI and the Quasar Attention mechanism, decentralized MoE training and the router problem, CUDA kernel subnet design and distillation, a secret data partnership, Const's unsolicited subnet acquisition, value capture via Chutes and Quasar Copilot, building model persona through reinforcement learning, and his advice to AI nerds considering Bittensor.

In this episode, you'll learn

• Two unsolved problems: why long context understanding and decentralized MoE training are the hardest open challenges in modern AI — and why Quasar is attacking both simultaneously
• Quasar Attention: a continuous-time attention mechanism that processes millions of tokens linearly with no increase in cost or drop in quality, built on top of the Transformer architecture
• The MoE router problem: why the router is the central bottleneck for decentralized MoE training, causing stability and speed issues that push most labs — including Subnet 3 — to train dense models instead
• Finding Bittensor: Eyad pitched Joseph Jacks on Quasar's architecture; JJ rejected the funding ask but told him to build in Bittensor — what Eyad calls "the best rejection of my life"
• Open source conviction: why Eyad believes people, not closed labs, should control the direction of AI — and why "praying OpenAI doesn't solve AGI first" isn't good enough
• Subnet design evolution: starting with CUDA kernels to accelerate Quasar Attention, then pivoting to model distillation from Qwen into the Quasar MoE architecture — and how this differs from Arbos's dense distillation subnet
• Data and expert parallelism: each miner trains one expert in a one-by-one communication MoE design Eyad says is faster than GPT's open-source expert parallelism — plus an unannounced data partner bringing state-of-the-art training data exclusively to Quasar
• Quality over size: why Eyad targets a usable, state-of-the-art model rather than benchmark-maxing, and how Quasar's 10M effective context window closes the gap between open-source and closed-source models
• Value capture: Chutes hosting partnership, an upcoming Quasar API, and Quasar Copilot — a product built around the 10M effective context window for daily long-context workflows
• Const's mentorship: how Const read Eyad's research, bought Subnet 24 unsolicited, handed over all emissions, and now holds the owner key as mentor — and what that says about how Bittensor works
• Bittensor as YC for real AI: why Eyad sees it as the only place where an AI startup can realistically compete with OpenAI and Anthropic — because every miner is aligned and working toward the same goal
• Persona via reinforcement learning: why supervised fine-tuning alone produces "full orders," and how Eyad uses RL to let Quasar's model develop its own generalization — plus his advice: set your goals before you join Bittensor

In this episode, Daniel of Hermes (Subnet 82) details the development of the infrastructure enabling AI agents to read and process blockchain data for the emerging autonomous economy. He discusses his security-focused background, the architectural decision to utilize GraphQL to extract structured data from indexers, and how Hermes solves the inherent inability of LLMs to efficiently parse sequential chain histories. Daniel also details the mechanics of the subnet's evaluation system utilizing synthetic and organic validation challenges, the integration of diverse data sources like Codex and DIA, the tokenomics driving the network, and the strategic roadmap for transitioning from database-specific queries to an infinite API of composable blockchain skills.

In this episode, you’ll learn

• Daniel's focus on operational security, including his methodology for utilizing fully sandboxed environments to safely manage and restrict AI agent permissions
• The specific technical constraints of AI interacting with blockchains, detailing why analyzing sequential ledgers for simple wallet summaries requires intensive infrastructure overhead
• The mechanics of the Hermes validation model, utilizing validator-generated synthetic challenges alongside spot-checked organic queries to evaluate miner accuracy and speed
• The architectural integration of GraphQL to parse structured data from SQL databases maintained by indexer projects, enabling flexible cross-chain search capabilities
• The strategic transition of the subnet from querying individual data sources to developing an infinite API of specific blockchain skills, packaging optimized miner outputs into a unified Hermes AI agent
• The structural vision for the on-chain AI economy, highlighting how blockchains provide the trustless execution, escrow capabilities, and decentralized identity verification required for multi-agent workflows

In this episode, we are joined by Canh Trinh and Craig Ryan of MegaTAO, who detail the development of the first perpetual decentralized exchange built on the Bittensor EVM. They discuss their respective backgrounds in traditional finance and crypto data engineering, the architectural decision to build on the EVM to maximize composability, and how they are bridging the gap between Bittensor’s ecosystem and decentralized finance. Canh and Craig also detail the mechanics of MegaTAO's dual-routing system utilizing order books and vaults, the mathematical parameters governing their leverage and liquidation limits, the engineering behind their recently launched Sum of Subnets (MTSOS) index, and their strategic roadmap for expanding collateral options and scaling trade capacity.

In this episode, you’ll learn

• Canh's technical and professional background in investment banking and fintech, and Craig's transition from building DeFi data pipelines at Chainalysis to trading
• The specific rationale for deploying MegaTAO on the Bittensor EVM rather than operating as an independent subnet, focusing on composability and established Solidity infrastructure
• The mechanics of MegaTAO's dual-routing execution model, which processes user trades through an order book first before matching against a funded vault to guarantee execution
• The mathematical framework governing their 5x leverage limit, requiring 20% equity capital, and the automated smart contract liquidation keepers that trigger at a 10% margin threshold to protect protocol health
• The technical construction of the MTSOS (Sum of Subnets) index, including the implementation of an on-chain 7-day exponential moving average (EMA) to smooth volatility and prevent market manipulation of newly launched subnets
• The strategic methodology for integrating decentralized finance mechanisms, such as maker/taker fees and funding rates, to bootstrap liquidity within the Bittensor ecosystem