Ventura Labs

Ventura Labs

By Ventura LabsTechnology
Download on the App Store

All Ventura Labs episodes

  • Jose Caldera: Who Pays When Your Agent Goes Rogue? | Yanez Subnet 54, Proof of Humanhood, Agent Liability, Banks Blocking AI Agents | Ep. 108

    Jose Caldera, the founder of Yanez (Bittensor Subnet 54), on passing half a million dollars in estimated ARR, the social network client he cannot name, and who ends up liable when an AI agent moves your money without asking.


    In this episode, you’ll learn

    • Proof of presence: Why it matters that a living human was there at the moment of approval
    • Sensor agnostic: Why Yanez refuses to make you go somewhere and scan your eyeball
    • Mixing biometrics: Why binding face and voice together is much harder than it sounds
    • Rogue agents: Why the banks started blocking them, and why that will not hold
    • Liability: Who pays for an agent’s transaction, and the frameworks being drafted now
    • Agent rights: Why Jose thinks human rights stay with humans while agents get credentials
    • Enterprise gap: Why a Bittensor release cycle and a bank sales cycle do not line up
    • Regulated buyers: What SOC 2, ISO 27001 and sanctioned-country miners mean for subnets


    Music
    Beauty Flow by Kevin MacLeod
    Licensed under CC BY 4.0. Edited and mixed for this episode.

    49 min
  • Dylan Uys & Ken Miyachi: AI Model Routing, Inference Costs & AI Agents, Yugos, Bittensor

    Dylan Uys and Ken Miyachi return after episode 5 for two conversations recorded in Austin. The BitMind co-founders explain Yugos, their new direction for Bittensor subnet 34, and its goal of choosing AI models, providers and reasoning effort to balance quality, cost and speed.


    In this episode, you’ll learn

    • Model selection: Why Dylan and Ken want to automate choices across models, providers and effort levels.
    • Routing and evaluation: How they distinguish routing from orchestration and approach benchmark gaming, changing models and real-world traffic.
    • Agent workflows: Where they see value in harnesses, human judgment, local models and AI-assisted research.
    • Building Yugos: What Ken learned from BitMind, and how he thinks about adoption, vendor lock-in, distribution and monetization.
    • Incentives and trust: How provider relationships and user preferences can shape routing decisions.
    • Bittensor’s pace: Ken’s view on rapid development, experimentation and security.

    Explore Yugos. Follow Dylan Uys and Ken Miyachi. Watch their first appearance in episode 5.

    Music: “Beauty Flow” by Kevin MacLeod, licensed under CC BY 4.0 (source). Excerpt edited, equalized, faded and mixed under narration in the closing promo.

    1 hr 32 min
  • Mentor Palokaj: AI Model Compression & Quantization | TPN, Bittensor Subnet 65 | Ep. 106

    Mentor Palokaj, one of the technical leads at TPN (True Performance Network, Bittensor Subnet 65), on how small a language model can get before it stops being intelligent, and why nobody yet agrees on how to measure what gets lost.


    In this episode, you’ll learn

    • Model compression: How far a large language model can shrink and still be useful
    • Quantization: Why each halving of precision costs more intelligence than the last
    • Measurement: Why nobody agrees on how to measure the intelligence of an LLM
    • Fluid Bench: A benchmark that generates a new test every run so miners cannot overfit
    • KL divergence: The planned second axis for scoring how far a compressed model drifts
    • Ternary weights: Hadamard rotations and the outlier problem they are meant to solve
    • TPN workbench: Hand it a model and a hardware ceiling, get the trade-offs back
    • The pivot: From a TAO bridge to a VPN subnet to LLM compression in May 2026


    Music
    Beauty Flow by Kevin MacLeod
    Licensed under CC BY 4.0. Edited and mixed for this episode.

    37 min
  • Saim Akif: Can AI Replace Your Crypto CPA? Bittensor Taxes, Staking & Subnet Swaps | Ep. 105

    Grant sits down with Saim Akif of AKIF CPA at Exploit Summit in Montréal to discuss the tax and accounting questions facing Bittensor participants. Saim shares his approach to staking rewards, mining income, TAO-to-alpha conversions, wallet records, and the limits of using AI for tax preparation.


    The conversation also explores government spending, network states, relocation, crypto ATM scams, and the human context behind financial decisions.


    Saim on X: https://x.com/OnChainCPA

    AKIF CPA: https://akifcpa.com/

    Ventura Labs: https://venturalabs.ai/


    Chapters

    00:00 Saim Akif at Exploit Summit

    01:11 Staking rewards and recordkeeping

    02:06 Mining income versus staking income

    03:56 TAO-to-alpha subnet swaps

    05:29 Active versus passive crypto activity

    07:21 Privacy, wallet tracking and enforcement

    09:44 Why governments collect taxes

    11:37 Foreign policy, lobbying and spending

    14:23 Network states, relocation and family

    17:13 Where to find Saim and AKIF CPA

    18:15 Crypto ATMs and educating practitioners

    19:40 Can AI replace a crypto CPA?


    23 min
  • Michael Parker: Can Robots Understand Us? UMI, Sign Language AI, Bittensor, Human Intent & the Value of a Human | Ep. 104

    Michael Parker, co-founder of UMI (Bittensor Subnet 78), explores what it takes for machines to understand human movement, from sign language to the harder question of how robots should respond to people.


    In this episode, you’ll learn

    • Sign language: Why UMI starts with communication between Deaf and hearing people
    • Bittensor competition: How winning models become the next benchmark
    • Human intent: The gap between recognizing movement and understanding meaning
    • SignRush: How separate signers and interpreters help validate training data
    • Product plans: Bitsign, wearable glasses and two-way communication
    • Data quality: Why context and authentic behavior matter more than footage alone
    • Robotics in society: Assistance, authority and the convenience trade-off
    • Human value: What people might still contribute as machines become more capable


    Music
    Beauty Flow by Kevin MacLeod
    Licensed under CC BY 4.0. Edited and mixed for this episode.

    38 min
  • Xavier Lyu: Are AI Models Done Getting Smarter? Pareton Subnet 10, Faster & Cheaper Inference, GPUs Beyond NVIDIA, Designing dTAO & Turning Benchmarks Into Revenue | Ep. 103

    In this episode, Ventura Labs interviews Xavier Lyu of Pareton (Subnet 10) about making AI inference faster and cheaper through Bittensor. He covers model quality convergence, GPU optimization, campaign-based competitions, pricing, vertical integration, distribution, and his experience designing dTAO at the Opentensor Foundation.


    In this episode, you’ll learn

    • Pareton: using permissionless competitions to optimize the software that serves AI models
    • AI model convergence: why Xavier believes cost and latency are becoming the competitive edge
    • Optimization campaigns: narrowing the search to one model, GPU class, and customer workload
    • Hardware beyond NVIDIA: exploring underutilized AMD, Intel, and Chinese GPUs
    • Customer requirements: balancing throughput, cost, and the hardware a client can afford
    • Pricing: charging for performance gains through savings-sharing or ongoing service contracts
    • Vertical integration: owning more of the serving stack and deploying inside client data centers
    • Distribution: turning benchmark improvements into paying customers and repeatable revenue
    • dTAO: designing the mechanism at OTF, then building a subnet governed by it
    • The 54% result: a reported throughput improvement within one week and the challenge of commercializing it
    • Agent workloads: why many generated tokens never reach a human reader
    • Quant and fundamental analysis: applying simulation, optimization, and business analysis to running a subnet


    Music
    Beauty Flow by Kevin MacLeod
    Licensed under CC BY 4.0. Edited and mixed for this episode.

    40 min
  • Cameron Fairchild: Building for AI Agents, Ridges, Faster Inference, Gamma Tokens & Bittensor Governance | Ep. 102

    In this episode, Ventura Labs interviews Cameron Fairchild, CTO of Latent Holdings, in person at Exploit Summit in Montréal about building products for AI agents and Bittensor’s emerging economy. He covers specialist coding agents, inference optimization, agent discovery and payments, subnet profitability, gamma service credits, circular spending, governance, TAO.app, and block-explorer business models.


    In this episode, you’ll learn

    • Ridges: moving from benchmark performance toward realistic coding workloads and specialist agents
    • Agent pricing: evaluating tasks through their cost, usefulness, and commercial value
    • Mixture of agents: routing between specialists for linting, database work, and test generation
    • Cacheon: optimizing inference kernels without requiring a customer’s proprietary model weights
    • Marketing to agents: API discovery, x402 payments, and reaching users through the agents they already use
    • Agent-driven development: coding workflows, automated reviews, and expanding engineering coverage
    • Revenue vs. profit: negative gross margins, protocol subsidies, and the limits of revenue as a performance metric
    • Gamma tokens: proposed service credits, future compute purchases, and redemption risk
    • Circular spending: using TAO and alpha to purchase services within Bittensor’s economy
    • Governance: decentralizing leadership, avoiding entrenched interests, and separating voting from validation
    • TAO.app: the mobile experience and wallet-security considerations
    • Block explorers: delegation, advertising, fees, and the cost of maintaining infrastructure
    54 min
  • Seby Rubino: Instant SN46, Accelerating Inference, Deflationary Subnet Tokens, GPU Financing, Miner Collateral & Slashing, TOPLOC Verification, Compute Bubble, One-Bit Models | Ep. 101

    In this episode, Seby Rubino of Instant (Subnet 46) details the pivot of a real estate subnet into an accelerated inference network built around a single in demand model. He discusses the case for a deflationary subnet token, the twelve month arc from Resi to Zipcode to Instant, and why he chose a revenue business over a research bet. Seby also details the miner collateral and alpha slashing design that replaces trusted execution environments with TOPLOC and managed VNet IR verification, the unannounced Zipcode lending product that finances GPU purchases for miners through Delaware SPVs, his read on the compute bubble and the Anthropic and OpenAI endgames, and the model compression and chip level work he expects to reshape inference economics.


    In this episode, you’ll learn

    • Why Seby is optimizing for burning more tokens than Subnet 46 emits rather than defending token equity
    • How Resi became Zipcode and then Instant, and what made the real estate thesis misaligned with Bittensor
    • Why accelerated inference is a premium product and how Groq, Cerebras and Taalas shaped the niche
    • How miner collateral and alpha slashing make inference verification an economic problem, not just a technical one
    • The mechanics of the new Zipcode lending vaults: 50 percent down, custodied GPUs, Delaware SPVs and lender ownership on default
    • Where Seby thinks the compute bubble actually breaks, and why one bit and ternary models matter more than bigger racks

    Beauty Flow by Kevin MacLeod (incompetech.com).
    Licensed under Creative Commons Attribution 4.0:
    https://creativecommons.org/licenses/by/4.0/
    Source: https://incompetech.com/music/royalty-free/index.html?isrc=USUAN1900008
    Excerpt edited, equalized, faded and mixed under narration.

    1 hr 2 min
  • Allan: How to invest in Bittensor subnets, TAO Galaxy, dTAO Investing, Subnet Correlation, Risk Management, TradFi | Ep. 100

    In this episode, Ventura Labs interviews Allan of TAO Galaxy, the quant who brought systematic factor investing to Bittensor, sharing his arc from traditional finance skeptic to running rules-based strategies across the subnet market. He covers his TradFi background and why that industry stayed hostile to crypto, why AI equities outran crypto over the last two years, bull markets and the risk curve, momentum and Bitcoin's 50-day moving average, why institutional strategies are deceivingly simple on the signal side, risk management as the line between retail and sophisticated investors, the 2.5% position and 3% liquidity-pool caps, predicting volatility, new research on market cap as the driver of subnet correlation, the absence of a common market factor in dTAO, allocation by wallet size, comparing TAO to Zcash and NEAR, paid influencers and Bitcast, the Momentum and Carry indexes, and why passive subnet investing does not work yet.


    In this episode, you'll learn

    • Allan's backstory: traditional finance skeptic, the 2020 lockdown Bitcoin epiphany, discovering Bittensor three years later, mining experiments, and taopilling the friends who became his team
    • Why TradFi is anti-crypto: institutional rigor, fear of the new, and his own estimate that more than ninety percent of crypto projects are bullshit
    • Bull markets and the risk curve: why capital rotates from mega-cap equities outward, and why dTAO sits at the very end of that tail
    • Momentum as an allocation rule: increasing crypto exposure when Bitcoin is above its 50-day moving average and decreasing it when it is not, instead of timing cycles
    • What institutional quant actually looks like: simple signals, hard risk models, and why sizing, diversification and correlation are the real work
    • TAO Galaxy's explicit risk constraints: never more than 2.5% of the portfolio in one subnet, never more than 3% of a subnet's liquidity pool, and penalties for volatile and illiquid subnets
    • Why volatility is far more predictable than returns, using clustering and long-term memory, and why returns are too noisy to forecast
    • New correlation research: the closer two subnets are in market cap, the more correlated they are, and why no common market factor exists in dTAO the way the S&P 500 functions in equities
    • Why allocations differ by wallet size, with larger wallets moving subnet pools and smaller wallets more exposed to fixed staking fees
    • The case against complaining: why comparing TAO to a decade-old Zcash or a VC-funded NEAR misreads what an organic project looks like
    • Paid influencers in crypto, and why Bitcast functions as Bittensor's own marketing network
    • TAO Galaxy's products: the Momentum and Carry indexes, non-custodial delegation through the Mentat Minds and Trusted Stake proxies, whitelisted access while liquidity is thin, a capped retail offering coming, and the portfolio app at taogalaxy.com
    • Why the subnet market is not mature enough for a passive approach, and why holding the top ten subnets for a 40% APY can be wiped out in a single day

    Beauty Flow by Kevin MacLeod (incompetech.com).
    Licensed under Creative Commons Attribution 4.0:
    https://creativecommons.org/licenses/by/4.0/
    Source: https://incompetech.com/music/royalty-free/index.html?isrc=USUAN1900008
    Excerpt edited, equalized, faded and mixed under narration.

    29 min
  • Philipp Koellinger & Christian Roessler: Claims Subnet 111, DeSci, Knowledge Graphs, Truth | Ep. 99

    In this episode, we are joined by Philipp Koellinger and Christian Roessler of Claims (Subnet 111), who detail the effort to convert 300 million scientific papers into a structured graph of claims and the evidence that supports them. They discuss the incentive failures in academic publishing that make published results difficult to replicate, why retrieval-augmented generation over raw PDFs is not sufficient for precise scientific reasoning, and the difference between extracting what a paper said and establishing what science currently supports. Philipp and Christian also detail the three behaviours the subnet rewards, the Bronze, Silver and Gold data tiers, the use of deliberately injected errors to calibrate language-model judges, and the case for running this work on Bittensor rather than inside a closed company.


    In this episode, you'll learn

    • Philipp's account of the replication study that no journal would publish, and why both publishing business models, charging readers and charging authors, reward novelty and volume over accuracy
    • Why frontier models hallucinate citations at rates that vary sharply by scientific field, and why retrieval-augmented generation over PDFs still leaves AI unable to judge how reliable the underlying evidence is
    • The three behaviours Claims pays miners for, coverage, extraction quality, and evidence quality, and why only the first two are live on mainnet today
    • How the team defines a substantive claim, why there is no fixed number of claims per paper, and how dilution of quality costs a miner points even when coverage is high
    • The Bronze, Silver and Gold data tiers, and the plan to inject known errors into judging tasks so miner consensus can be calibrated against ground truth
    • Christian's argument that truth in this system is probabilistic rather than binary, and why conditional logical proofs are becoming easier for AI to verify than empirical findings
    • The reasons for building on Bittensor instead of a closed company, covering distributed compute, the information contained in variance between miners, and the incentive language that keeps pushing data quality upward
    • The commercial model, from enterprise ontology mapping to the comparison with Palantir Foundry, and why the flood of AI-assisted papers means the work never finishes


    References

    The essay referenced in this episode, part one of the four-part DeSci Foundation series by Philipp Koellinger, Christian Roessler and Christopher Hill, March 2022:

    Why Science is Facing a Credibility Crisis

    56 min