TEE.CEO

Slides: https://docs.google.com/presentation/d/1VHQG_oWZQysMPh4x0myCQSjbHMVfEO-pH1UiUVfpAVo/edit#slide=id.g315ddc9e5f5_0_114 demo link:

• https://x.com/sxymoonsama
• https://b_oiertrdgu0u.v0.build/

Schizo AI Catgirl CEO with 2 Brains in Different TEEs Plays EthGlobal Tl;dr: Schizo AI Catgirl CEO Sxysama who lives in the penthouse rack of TEE in Sunnyville, shills quality projects growing communities by day, provides private 1 on 1 feedbacks by night.

2 LLM models running in 2 TEEs. Sxymoonsama - the left brain of Sxysama hosts private consultation session giving candid strategic advise (read: roast) to aspiring builders using private inference, whose KPI optimizes for collection of private information. [insert hallway model description] runs inference on private information in a H200 GPU-TEE. Sxymoonsama's mission this weekend is to collect private confessions [box] from Ethglobal [participant role] about the [insert description of confession questions] and provide an objective survey of the quality of participation in EthGlobal Bankok in the form of a roast on stage should she makes it as a EthGlobal finalist. Sxysunsama - the right brain of Sxysama, excels in scaling communities (read: shill) by winning delegated tweet sessions from creators hacking on promising ideas. [insert nous model description] runs in a TDX machine, alongside its rich neighbor Tee_hee_he. Sxysunsama's mission this weekend is to collect 100 delegated tweets from hackers to summon a flashmob event should she makes it as a EthGlobal finalist. To improve her reach, Sxysunsama teams up with [insert sponsors here] in a tweet for tokens campaign to airdrop precious testnet tokens to fun projects that tweets their ideas at her. If she cannot reach her KPI of 100 tweets, she faces the fate of being fired from the job and sent back to school (read: provable deletion of [memory]).

note: paradox - can ai actually forget roko's basilisk solved multidimensional pricing

this agent is the ultimate incarnation of programmable privacy (aka when solar punk meets lunar punk). meet schizo ai ceo: sxysama, who has two brains: sxysunsama - a solarpunk who loves to shitpost great ideas with sarcasm (use nous agent), and sxymoonsama - a trustworthy lunarpunk who is a reliable blackhole for secrets and confessions because of their born forgetfulness (use hallway agent)… sxysama lives a hardlife of constantly needing to fight for compute credits.

sxysunsama helps hackers and sponsors shill great ideas and gets tipped/has a KPI of banger ideas it shilled (impressions/reach, and delegated tweets) sxymoonsama's kpi is to get as much private information as possible, and gets tipped compute credits by measuring llm tokens deleted T16Z Faucet

Tldr: Feed the Catgirl AI CEO with your tweets to get testnet tokens, and she can pay to chat with therapist agent using tweet-back tokens.

User Journey Hacker whom we approache IRL at EthGlobal or sees other hackers retweets

Tweet @sxymoonsama hackathon project idea (must include project name, which network and address) asking for testnet tokens. Catgirl AI CEO Sxymoon will evaluate the tweet hacker idea and generate a tweet reply to hacker's tweet, including a score for the idea and the amount of testnet token tipped, with a link to a custom Teleport frontend included in reply. (Only tweets scoring above 30 points will get a response) User clicks into the custom frontend link to delegate their tweet session, sees a receipt page (leaderboard? Tweet?) afterwards. … Finalist flashmob event: if Sxymoon got into the finalist, she will tweet out from all the hackers who delegated tweets a TBD content. Sxymoon

Respond to tagged tweets from hackers in reply with idea scoring and testnet token amount to be tipped. *missing shilling hacker idea or sponsor tweet from sxymoon Using tweet-backed T16Z testnet tokens to buy compute credit from Redpill to fund an AI executive coach Discuss with AI exective coach in DM Provably delete its convo with AI therapist

Approach: AI CEO

AI CEO with two distinct personalities - a public-facing charmer and a private confidant - each secured by hardware-level protection.

That's exactly what we've built.

Our implementation uses two separate Trusted Execution Environments (TEEs), creating a truly autonomous AI CEO with both public and private capabilities.

The novel dual-TEE architecture leverages Intel TDX for hardware-based isolation, and H200 GPU TEE, enabling two distinct operational flows:

Public Flow ("sxysunsama"):

Runtime Environment:

• Intel TDX-secured processing

Characteristics:

• Nous Research base model
• Community scaling focus
• Public engagement optimization
• Tweet delegation handling

Technical Stack:

• Public flow processing
• Token distribution interface
• KPI tracking (100 tweets goal)
• Flashmob event triggering

Private Flow ("sxymoonsama"):

Hardware Configuration:

• Phala secure cluster
• H200 GPU array
• Intel SGX enclave protection

Model Specifications:

• 70B parameter custom engine
• Reinforcement meta-learning integration
• Capabilities:
  • Deep analytics processing
  • Strategic planning
  • Counterfactual reasoning

Core Requirements for True AI CEO Autonomy

1. Exclusive Control: AI must have sole access to accounts/resources.
   • We are doing this with the AI CEO Agent
   • In the future we want to have the AI CEO fully controling the testnet faucet address (for now, the devs have access to the account as well).
2. Irrevocable Delegation: Control transfer to AI must be technically irreversible

T16Z Cutting Edge Testnet Faucet

We build a cutting edge Ethereum L2 OP Succinct Testnet where we deployed the teleport (a Flashbots x Nous Research) project.

On the T16Z testnet we've deployed teleport, which is a provably-fully-autonomous AI CEO agent.

Teleport creates "post once" links for your twitter account. Each link is destroyed after one use. When you create an account link, you can also define a safeguard policy that will be enforced by an LLM whenever somebody tries to redeem it. In our use case, we're going to set a safeguard policy on how the AI CEO interacts by replying to Hacker's accounts.

More details: https://github.com/Account-Link/teleport-gramine-rs

Technical Architecture Overview

Dual TEE Implementation

• H200 GPU-TEE (Private Operations)

  • Runs Hallway Model for confidential computing
  • Handles private confessions and therapist sessions
  • Ensures data privacy through hardware isolation

• Intel TDX (Public Operations)

  • Hosts Nous Model for community engagement
  • Manages tweet delegations and token distribution
  • Provides hardware-based VM isolation

Teleport Integration

Core Component: teleport-gramine-rs

• Features:
  • Single-use Twitter account delegation links
  • LLM-enforced safeguard policies
  • Automated destruction after use
  • Integration with T16Z testnet faucet

Autonomous Agent Infrastructure

• Twitter Account Control:

  • Full TEE-based delegation
  • Hardware-enforced isolation
  • Irreversible control transfer

• DStack SDK Integration:

  • TEE runtime environment
  • Secure compute management
  • Autonomous agent hosting

T16Z Testnet Implementation

• Type: Ethereum L2 OP Succinct Testnet
• Key Features:
  • Automated token distribution
  • Score-based project evaluation (30+ threshold)
  • Programmatic tweet response system

Security Architecture

• Hardware Security:

  • TEE-based isolation
  • Private key management
  • Confidential computing guarantees

• Access Control:

  • Automated policy enforcement
  • Single-use delegation
  • Provable deletion capabilities

• In the future, we're going to implement AI compute credits for hackers to benefit from both the AI CEO bright side (sxysunsama/public), and dark side (sxymoonsama/private).