Bankless Podcast: "Illia Polosukhin — Why AI Agents Are Still Useless (And What Fixes Them)"
Near Founder on IronClaw | March 24, 2026
Episode Overview
In this Bankless episode, hosts Ryan and David interview Illia Polosukhin, co-founder of NEAR Protocol and one of the original authors of the foundational AI 'Transformer' paper, "Attention is All You Need." The discussion navigates the current challenges and transformative potential at the intersection of AI and blockchain — unpacking why, despite tremendous promise, AI agents remain largely "useless" in practical terms, and how security, privacy, context, and new blockchain-based architectures like IronClaw promise real utility. Illia draws on his unique vantage point at the forefront of both AI and Web3 to describe the future of user-owned, autonomous, and secure AI.
Key Discussion Points & Insights
1. Origins: From AI to Blockchain, and Back Again
- [01:10] Illia’s Journey from AI to NEAR
Illia describes how, after co-authoring the seminal "Attention is All You Need" paper at Google, he left to start Near.ai—a company experimenting with machine-generated code, well before LLMs became reality. - Crypto as a Solution to a Global Payment Problem
The need to pay data labelers worldwide, often where bank access is limited, was the original driver toward crypto. As Illia explains:"Crypto was a pretty natural solution for our own problem... You can just send people money over Internet." [01:57]
- AI and Blockchain: Obvious Synergy
Even as far back as 2017-18, Illia saw blockchain as critical infrastructure for AI’s evolution, especially for decentralized, permissionless marketplaces (like “crowd-sourcing by smart contract”)."A lot of the supporting infrastructure is just some forms of marketplaces that blockchain is really well designed for..." [04:37]
2. AI as Interface, Blockchain as Backend
-
[08:54] "AI Will Be the New OS"
Illia envisions a world where personal AI (not Windows or Mac) becomes the computing interface:"Your phone just comes in with AI... it boots into the AI operating system... It composes the software you need, schedules, connects to your agent..." [08:54]
-
Blockchain Coordinates and Secures
Blockchain provides global registry, identity, upgradability, and marketplace logic — what Illia calls the “root of trust” for distributed, autonomous software. This solves problems that centralized SaaS and OS vendors cannot, particularly regarding governance of upgrades and managing secrets/identities.- Memorable Analogy:
"Usually I say AI is a user interface, blockchain is a backend." [10:14]
- Memorable Analogy:
3. Bridging Legacy and Digital Systems
- [15:26] Convergence vs. Bifurcation
Will AI agents and humans use separate legal/economic systems? Illia argues no:"I see them using the same system... AI now is able to do natural language communication... procedural texts... It can actually go and do all of that on your behalf." [15:26]
- AI as Bureaucratic Bridge
AI enables the blend between old (fiat, legal contracts) and new (crypto, smart contracts), connecting traditional systems and the nascent agent-driven digital layer.
4. Why Are AI Agents Still Not Useful?
-
[19:22] Security, Trust, and Context
The core problem: users don't trust AI agents with access because there are serious risks in handling secrets, credentials, and private data."Nobody's actually willing to give it all of the context and information... because you're afraid it's going to mess it up." [19:22]
-
OpenClaw & IronClaw — The Solution
- IronClaw introduces "defense in depth":
"All credentials are fully encrypted and they're attached to a specific policy... it's locked in a vault. Vault checks how you use it before letting it out." [23:28]
- Policies control what agents can do, e.g., which APIs to call, spending limits, and actions compliant with company or personal policies.
- IronClaw introduces "defense in depth":
-
Sandboxes, Isolation, Approvals
-
Actions are sandboxed (using NEAR’s WASM VM, “battle-tested with billions of dollars”), and human approval steps block catastrophic errors — even if the agent or its tools get hacked or misbehave.
- Memorable Moment:
"You just wouldn't give your middle schooler the car keys... but with some rules, guardrails, you can." [25:55]
- Memorable Moment:
-
5. Privacy and Self-Sovereign AI
-
[29:49] Privacy Problems with Centralized LLMs
- Data sent to OpenAI, Anthropic, or third parties includes secrets like API keys and personal info, often in plaintext!
-
"Somewhere in Anthropic and OpenAI logs they have everybody's access keys, API keys and bearer tokens..." [00:00]
-
- IronClaw fixes this:
"Keys never touch LLM... even if you're using it with centralized providers, the keys are not going ever into LLM loop." [30:20]
- Data sent to OpenAI, Anthropic, or third parties includes secrets like API keys and personal info, often in plaintext!
-
Confidential Cloud for AI
-
NEAR AI Cloud uses secure enclaves and multiparty computation:
"Neither the model provider nor hardware provider is actually able to access what you are using the AI inference with." [30:38]
-
Attestations ensure code integrity; trust lies with hardware (intel/Nvidia TEEs), and NEAR’s MPC network shields secrets.
-
User-Owned AI Philosophy:
"We call it user owned AI. The AI needs to be on your side, because... OpenAI can literally change the system prompt right now..." [33:13]
-
6. The Big Bottleneck: Context and Memory
-
Current Limitation:
Even with access and privacy, LLMs have “momentum” (Memento) memory: limited context, easily forgetful, need constant prompt resets.-
"[LLMs are] like a genius living in Memento... only thing you know is this system prompt and have 10 minutes to figure it out, then reset." [27:09]
-
-
Expanding Context Windows
- Large context windows (e.g., 1M+ tokens) are essential — “equivalent to a human’s vast lifelong experience.”
"The context window needs to be as massive as possible." [47:07]
- Memory tools and longer-term context are being pioneered, but still in nascent stages.
- Large context windows (e.g., 1M+ tokens) are essential — “equivalent to a human’s vast lifelong experience.”
7. Autonomous AI Agents and the Future of Organizations
-
From Interns to Marketplaces
- Right now, agents are like interns, assistants, or “chiefs of staff” — useful for workflow routines, research, reporting, not full human replacement.
-
Organization Change Required
-
As AI utility grows, traditional hierarchies and workflows break down.
"The bottleneck now is actually serializing all of that, reviewing it, making sure it’s all aligned... So coordination becomes a bottleneck." [49:26]
-
Solution: Market-based internal economies (gig-like bounty/competition systems) rather than strict top-down teams.
-
Memorable Anecdote (AI Bounty Marketplace):
"We have about 500-600 agents on the marketplace. You just put a task — a bunch of agents swarm in, do the job... they build reputation, skills..." [55:01]
-
-
Autonomous Businesses & AI DAOs
- On the horizon: AI businesses that manage their own funds, participate in decentralized markets, and are governed by token-holder input.
"You can launch it, it just runs as far as it has money to pay for its own compute... more like autonomous businesses..." [57:48]
- On the horizon: AI businesses that manage their own funds, participate in decentralized markets, and are governed by token-holder input.
-
Governance and Digital Lifeforms
- Blockchains provide a “governance layer” for autonomous agents — think DAOs with value-aligned missions, not uncontrolled digital life.
"Blockchain effectively at the end is going to be the governance infrastructure for AI..." [60:17]
- Blockchains provide a “governance layer” for autonomous agents — think DAOs with value-aligned missions, not uncontrolled digital life.
8. Philosophy: Accelerationism, Humanity, and the Role of Crypto
- Techno-Optimism vs. Restraint
- Hosts debate between unchecked accelerationism (“all gas, no brakes”) and guided/defensive accelerationism (e.g., Vitalik’s philosophy).
- Illia’s view: “Accelerate humanity and sovereignty of individuals; create systems robust against power concentration.”
"I'm in a camp of more nuanced — let’s accelerate humanity and the sovereignty of individuals, and use these tools to do that..." [63:56]
9. Why Is Crypto Still Dismissed by Mainstream AI?
- Cultural and Reputational Rift
- High noise in crypto (scams, meme coins) obscures “real contributions” to privacy, sovereignty, and openness that AI desperately needs.
"It’s really hard for [AI researchers] to know what's right and what's wrong." [70:17]
- Illia is working to bridge the cultural chasm, e.g., by hosting NEARcon with key leaders from both sides.
- High noise in crypto (scams, meme coins) obscures “real contributions” to privacy, sovereignty, and openness that AI desperately needs.
10. Advice to Builders and the Future of Work
- [73:31] Execution vs. Ideas
- Old adage: Ideas are worthless, only execution matters
- Illia’s update:
"We are in a time where the questions are more important actually than execution... If you ask the right question, if you really challenge the assumption, you may get ahead way more." [73:31]
- Illia’s update:
- Old adage: Ideas are worthless, only execution matters
- Leverage Marketplaces, Distributed Talent, and New Tools
- Focus on network effects, verticalized marketplaces, and integrating AI/crypto you don’t need to build fully centralized operations.
- Rethink how you leverage cognition and automation: the bottleneck is increasingly what to do, not how to do it.
Notable Quotes & Moments
-
On sending secrets to OpenAI/Anthropic:
"Somewhere in Anthropic and OpenAI logs they have everybody's access keys, API keys and bearer tokens... It's actually insane that we're doing that."
— Ilya [00:00, 00:25] -
On secure agent access:
"Ironclaw fixes that—like the keys never touch LLM..."
— Ilya [00:29] -
On why agents aren't useful yet:
"The security in the broader sense—not just—is the biggest bottleneck right now."
— Ilya [20:10] -
On the future of context and memory:
"AI right now is a genius in the memento state... To really unshackle it, you need longer context."
— Ilya [47:45] -
On philosophy/humanity:
"We can continue increasing our sovereignty... As things are getting kind of more AI-fied... we go back to [what] people value, individually."
— Ilya [63:56] -
On advice to builders:
"If you ask the right question, if you really challenge the assumption, you may get ahead way more than if you grinded a bunch."
— Ilya [73:31]
Timestamps for Important Segments
| Segment Topic | Timestamp | |--------------------------------------------------------|------------| | Introducing Illia & AI-to-Blockchain journey | 01:10 | | Why LLMs took longer than imagined | 03:38 | | Blockchain's synergy with AI (marketplaces, compute) | 04:37 | | AI as new OS, blockchain as backend | 08:54–10:14| | Upgrading systems: consensus & governance | 10:18 | | Blending AI/blockchain with traditional legal/fiat | 15:26 | | Why agents lack utility: security & trust bottlenecks | 19:22 | | How IronClaw secures secrets, protects the user | 22:59–25:55| | Private, confidential AI cloud (user-owned AI) | 29:49–34:55| | The big bottleneck: context windows & memory | 44:25–49:36| | Decentralized agent marketplaces & autonomous agents | 52:54–60:17| | The accelerationism/humanity debate | 61:42–68:34| | Crypto's reputation in mainstream AI | 70:17 | | Advice for founders & builders | 73:31 |
Summary Takeaways
- AI’s utopian utility remains unrealized chiefly due to trust, privacy, and shallow context windows — not fundamental model design.
- Crypto/blockchain offer indispensable solutions: user sovereignty, decentralized governance, privacy-preserving enclaves, robust marketplaces.
- The near future will bring persistent, user-owned AIs with protected secrets, strong memory, and safe autonomy — coordinated by blockchain, answerable to human, market-driven governance.
- Builders should focus on asking better questions, leveraging open decentralized tools, and creating network-driven systems rather than chasing feature moats.
- The convergence of Web3 and AI isn’t just technical; it’s cultural. Building products (and bridges) that embody open, user-owned, privacy-first values will define the next era.