Peter H. Diamandis · Published 2025-04-03

The Future of AI Avatars and DeepFake Technology w/ Joshua Xu & Steve Brown | EP #161

privateReview: unreviewedConfidence: high
ai avatarsdigital twinsdeepfakesllmshealthcareavatar-agent-separationfirst-party-consentdeepfake-defenseidentity-verificationpersonal-knowledge-basefew-shot-personalizationagent-liabilityvideo-generation
Open on YouTube ↗

Summary

Overview

  • Speaker: Peter H. Diamandis, Joshua Xu, Steve Brown
  • Channel: Peter H. Diamandis
  • Main topic: AI Avatars, Digital Twins, and Video Generation Technology
  • Purpose: To showcase the capabilities of HeyGen's AI avatar technology, discuss the future of visual storytelling and digital twins, and demonstrate how individuals can create and interact with AI agents for personal and business use. Peter Diamandis hosts Joshua Xu (CEO of HeyGen) and Steve Brown to discuss the exponential growth of AI avatars and video generation technology. They cover the evolution of HeyGen from camera enhancement tools to a 175-language avatar and voice translation platform, the distinction between AV (audio-visual avatars) and AI agents with agency, trust and safety protocols, and real-world medical applications using personalized AI avatars.

Topic Map

Introduction to HeyGen and AI Avatars

  • Explanation: Joshua Xu discusses the founding and evolution of HeyGen since 2020, starting from camera enhancement tools and shifting toward replacing the camera with AI video generation and digital twins.
  • Key claims:
    • HeyGen allows users to create lifelike avatar videos in over 175 languages and dialects 10x faster than traditional production.
    • The mission is enabling visual storytelling for everyone by removing camera and production cost barriers.
  • Examples:
    • Joshua Xu's digital twin introducing the technology on stage in Spanish, French, Chinese, Hindi, and German.
    • Rikard Steiber's avatar created in 4 minutes speaking 175 languages.
  • Terminology:
    • AI avatar
    • digital twin
    • generative AI
    • multilingual translation
  • Why it matters: Democratizes high-end video content creation for businesses and individuals globally without the need for expensive studios or production crews.

Trust, Safety, and First-Party Consent

  • Explanation: Addressing concerns around deepfakes, identity theft, and malicious use of AI avatars by implementing strict first-party consent and video verification.
  • Key claims:
    • HeyGen requires first-party video consent for every avatar created on the platform.
    • A moderation team and automated safeguards prevent hate speech, fraud, and political manipulation.
  • Examples:
    • Requiring users to record a video statement giving explicit consent to generate their avatar likeness.
  • Terminology:
    • first-party consent
    • identity protection
    • moderation system
  • Why it matters: Ensures ethical deployment and prevents the unauthorized creation and misuse of deepfake technology.

Avatar vs. Agent Framework

  • Explanation: Steve Brown presents a framework distinguishing between avatars (audio-visual representation) and agents (agency, knowledge base, and autonomous execution).
  • Key claims:
    • Avatar stands for audio-visual, while Agent stands for agency.
    • Giving an agent agency requires training it on personal history, values, and knowledge bases so it can represent you accurately.
  • Examples:
    • The Hippocrates medical agent trained on Steve Brown's medical records and broader medical literature to identify plasma cell disorders.
    • Creating a legacy agent of oneself to pass down personal history and wisdom to future generations.
  • Terminology:
    • AV
    • agency
    • LLM
    • knowledge base
    • biomicrobiome
  • Why it matters: Bridges the gap between looking and sounding like someone versus having an autonomous AI system that thinks, reasons, and acts on your behalf.

Key Points

Replacing the Camera

  • Explanation: AI video generation enables people to create video messages, ads, and training materials without physical cameras or production crews.
  • Evidence: HeyGen's platform scales across 175 languages in 4K resolution.
  • Practical implication: Marketers and businesses can rapidly iterate and localize video content at a fraction of traditional time and cost.

Human-Centric Video Models

  • Explanation: Building models that capture subtle facial expressions, body language, and gestures based on short video training footage.
  • Evidence: Requiring roughly 2 minutes of training video to train custom models on facial movements and voice.
  • Practical implication: Avatars feel remarkably natural and match individual speaking styles without requiring extensive filming sessions.

AI Agents in Healthcare

  • Explanation: Combining medical data, lab results, and broad medical knowledge into an AI agent to assist with early diagnosis and second opinions.
  • Evidence: Steve Brown's medical agent correctly identified potential multiple myeloma indicators from low immunoglobulins before specialists did.
  • Practical implication: Helps patients navigate complex, fragmented medical data and overcome the siloed nature of medical specialties.

Frameworks, Models & Processes

Avatar vs. Agent Framework

  • How it works: Separates the audio-visual representation (Avatar) from the reasoning, knowledge, and agency (Agent).
  • Components:
    • Avatar (Audio-Visual representation)
    • Agent (Agency, knowledge base, and LLM reasoning)
  • When to use: When designing AI tools that require both a human face/voice interface and autonomous decision-making or conversational depth.

Examples & Case Studies

Rikard Steiber's avatar was created in 4 minutes using HeyGen and spoke in 5 different languages.

  • Illustrates: The speed, localization power, and fidelity of AI avatar generation.
  • Lesson: High-quality multilingual video communication can be generated instantaneously from minimal training data.

Steve Brown used an AI medical agent modeled after Hippocrates to analyze his complex lab results and uncover potential multiple myeloma.

  • Illustrates: The power of combining personal health data with broad medical LLMs for early diagnosis.
  • Lesson: AI agents can synthesize fragmented medical data across specialties to provide critical health insights.

Actionable Takeaways

  • Immediate:
    • Experiment with HeyGen to create your own digital avatar or test video translation.
    • Establish security passwords or verification words with family members to protect against deepfake phone scams.
  • Strategic:
    • Prepare your personal and professional knowledge base to power future AI agents and digital twins.
    • Adopt AI video tools to scale marketing, training, and global communication 10x faster.
  • Questions to investigate:
    • How will autonomous AI agents handle legal liability and contractual negotiations on behalf of individuals?
    • What standards will emerge for verifying the authenticity of real-time AI video calls?

Claims Worth Verifying

  • By 2040, we are likely to see as many as 10 billion humanoid robots. (future projection)
  • HeyGen supports video generation across over 175 languages and dialects. (product capability)

Notable Quotes

"This is the future which is available when? NOW." "Avatar is AV, audio-visual. Agent is agency."

Compressed Summary

  • HeyGen enables instantaneous AI video creation and translation across 175 languages.
  • The shift from static avatars to autonomous AI agents with personal knowledge bases is transforming business and healthcare.
  • First-party consent and rigorous verification ensure ethical use and prevent identity theft.
  • AI medical agents can integrate fragmented diagnostic data to assist in early disease detection.
  • Keywords: ai avatars, digital twins, deepfakes, llms, healthcare
  • Core insight: AI avatars and agents are shifting communication from text and email to immersive, personalized, multilingual video interactions.

Core insights

6
Architecturemedium noveltystrong evidence

The Avatar vs. Agent split is a clean architectural boundary: the avatar layer owns audio-visual rendering (face, voice, gestures), while the agent layer owns agency, knowledge base, and LLM reasoning. These can be built, trained, and swapped independently.

Why it matters

Teams that conflate 'a face that talks' with 'a system that decides and acts' end up with monolithic systems where the rendering pipeline and the reasoning/knowledge stack are coupled, making it impossible to upgrade one without rebuilding the other. Separating them lets you treat the avatar as a presentation adapter and the agent as a policy/knowledge core.

Generalization

Voice/vision interfaces to agents should be treated as interchangeable presentation adapters over a stable reasoning core, not as part of the agent's identity or logic.

Avatar stands for audio-visual, while Agent stands for agency.
Open source video
Avatar (Audio-Visual representation)
Open source video
Agent (Agency, knowledge base, and LLM reasoning)
Open source video
Avatar is AV, audio-visual. Agent is agency.
Open source video
Mental Modelmedium noveltymoderate evidence

'Giving an agent agency' is framed as a data-grounding problem, not a model problem: the agent must be trained/loaded with personal history, values, and knowledge bases to represent a person accurately. Representation fidelity is a function of the curated knowledge corpus.

Why it matters

Shifts engineering effort from model selection toward knowledge-base curation and personalization data pipelines. The bottleneck for useful personal agents is the quality and coverage of the personal corpus, not the base LLM.

Generalization

Personalization/agent fidelity scales with the quality of the private knowledge base rather than with model capability, implying investment in ingestion, structuring, and lifecycle of personal data.

Giving an agent agency requires training it on personal history, values, and knowledge bases so it can represent you accurately.
Open source video
Prepare your personal and professional knowledge base to power future AI agents and digital twins.
Open source video
Practicemedium noveltystrong evidence

Identity-critical AI features can be gated by first-party consent verification at creation time: every avatar requires a recorded video statement of explicit consent, backed by moderation plus automated safeguards.

Why it matters

Provides a concrete, design-level safety mechanism for biometric/identity synthesis: bind generation to a verified consent artifact produced by the subject, rather than relying purely on post-hoc detection or policy. This is a reusable pattern for any system that synthesizes a person's likeness or voice.

Generalization

For any generative system operating on identity or biometrics, require a first-party, verifiable consent artifact as a prerequisite to enabling the capability, and pair it with layered automated + human moderation.

HeyGen requires first-party video consent for every avatar created on the platform.
Open source video
Requiring users to record a video statement giving explicit consent to generate their avatar likeness.
Open source video
Empirical Resultmedium noveltymoderate evidence

Custom human-video models are trained from very small samples — roughly 2 minutes of training video — to capture facial movement and voice, making per-user model specialization cheap enough to be a standard product feature rather than a bespoke service.

Why it matters

Sets an expectation for the data efficiency of personalized generative models: if ~2 minutes of video suffices for a usable digital twin, then per-user fine-tuning/few-shot conditioning is operationally viable at platform scale, which has implications for storage, training-job scheduling, and per-user model versioning.

Generalization

When a personalized model can be trained from a tiny sample, the architecture question shifts from 'one big shared model' to 'many small per-subject adapters' managed as first-class artifacts.

Requiring roughly 2 minutes of training video to train custom models on facial movements and voice.
Open source video
Rikard Steiber's avatar was created in 4 minutes using HeyGen and spoke in 5 different languages.
Open source video
Empirical Resultmedium noveltyweak evidence

Agents can surface diagnostic signals by cross-referencing an individual's fragmented records against broad domain literature — the reported case is a medical agent flagging potential multiple myeloma indicators from low immunoglobulins before specialists did.

Why it matters

Demonstrates the value proposition of agentic synthesis over siloed specialty workflows: the agent's advantage is not superior domain knowledge but the ability to hold the whole record set in context at once and pattern-match against literature. This is an integration/context-assembly win, not a reasoning-model win.

Generalization

Where human workflows are fragmented by specialization, agents win by assembling and cross-referencing data across all silos in a single reasoning pass.

Steve Brown's medical agent correctly identified potential multiple myeloma indicators from low immunoglobulins before specialists did.
Open source video
AI agents can synthesize fragmented medical data across specialties to provide critical health insights.
Open source video
Helps patients navigate complex, fragmented medical data and overcome the siloed nature of medical specialties.
Open source video
Practicelow noveltymoderate evidence

Real-time deepfake capability inverts household security assumptions: voice/video alone can no longer authenticate a person, motivating shared verification words or passwords between family members as out-of-band authentication.

Why it matters

Introduces a lightweight out-of-band authentication protocol for high-trust channels that any agent/voice product team must now account for — and conversely, any voice-cloning product must assume its outputs will be used for impersonation.

Generalization

When a modality becomes cheaply forgeable, authentication must migrate to a shared secret in a different channel; trust protocols need to be re-derived per modality rather than assumed.

Establish security passwords or verification words with family members to protect against deepfake phone scams.
Open source video
What standards will emerge for verifying the authenticity of real-time AI video calls?
Open source video

Deep dives

5

Avatar/agent layer separation as a presentation-adapter architecture

Research question

Does formally decoupling the audio-visual rendering layer from the agency/knowledge/LLM reasoning layer allow independent versioning, replacement, and licensing of each without regression in identity consistency or decision fidelity?

Why

Teams that entangle 'a face that talks' with 'a system that decides and acts' build monoliths where upgrading the renderer forces a rebuild of the knowledge stack and vice versa. If the boundary holds under version skew, then the knowledge base becomes the primary product asset and the avatar becomes a swappable adapter.

Avatar stands for audio-visual, while Agent stands for agency.
Open source video
Avatar is AV, audio-visual. Agent is agency.
Open source video
Giving an agent agency requires training it on personal history, values, and knowledge bases so it can represent you accurately.
Open source video
Source video

First-party consent artifacts as enforceable, revocable creation-time gates

Research question

Can a recorded first-party consent statement be cryptographically bound to a specific trained avatar model such that revoking the consent artifact invalidates generation, and what onboarding friction does this add?

Why

Moves identity safety from post-hoc detection (expensive, adversarial, always lagging) to a precondition in the creation pipeline, producing an audit trail against impersonation. The unresolved question of whether consent can be revoked to invalidate an existing avatar is the difference between a stored recording and an actual control.

HeyGen requires first-party video consent for every avatar created on the platform.
Open source video
Requiring users to record a video statement giving explicit consent to generate their avatar likeness.
Open source video
Ensures ethical deployment and prevents the unauthorized creation and misuse of deepfake technology.
Open source video
Source video

Personal knowledge base as the fidelity bottleneck for personal agents

Research question

Does grounding an agent in a curated personal corpus measurably improve fidelity to the principal's own judgments, or does it mainly improve topical coverage without improving decision agreement?

Why

The claim that agency is a data-grounding problem shifts investment from model selection to ingestion, structuring, and lifecycle management of private data. Without a measurement of marginal value per data category, teams will over-collect and under-curate.

Giving an agent agency requires training it on personal history, values, and knowledge bases so it can represent you accurately.
Open source video
Prepare your personal and professional knowledge base to power future AI agents and digital twins.
Open source video
Source video

Live-call authenticity: device-bound attestation plus out-of-band challenge-response

Research question

What protocol can verify that a real-time AI video call is a live, unmodified capture of a claimed person, and what is its latency and usability cost versus shared verification words?

Why

Once voice and video are cheaply forgeable, trust must migrate to a shared secret in a different channel; the naive answer (family passwords) does not scale to institutions and offers no runtime attestation for live streams. Proving provenance at capture time is the only path that survives the modality becoming free to forge.

Establish security passwords or verification words with family members to protect against deepfake phone scams.
Open source video
What standards will emerge for verifying the authenticity of real-time AI video calls?
Open source video
Real-time deepfake capability inverts household security assumptions: voice/video alone can no longer authenticate a person.
Open source video
Source video

Per-subject few-shot adapters as first-class platform artifacts

Research question

If ~2 minutes of video suffices for a usable digital twin, what does a model registry look like when every user owns a versioned adapter, and how do storage cost, training-job scheduling, and quality drift out of distribution behave?

Why

Tiny-sample personalization changes the architectural question from 'one big shared model' to 'many small per-subject adapters' that need versioning, invalidation, and quality monitoring just like any other artifact. It also determines whether per-user specialization is a standard product feature or a bespoke service.

Requiring roughly 2 minutes of training video to train custom models on facial movements and voice.
Open source video
Rikard Steiber's avatar was created in 4 minutes using HeyGen and spoke in 5 different languages.
Open source video
Source video

Article ideas

4

Your Avatar Is Not Your Agent: Why the Presentation Layer Must Be Disposable

Conflating the audio-visual rendering layer with the agency/knowledge/reasoning layer is the most expensive architectural mistake in current agent products; the avatar should be a swappable presentation adapter over a stable policy-and-knowledge core, because rendering models and reasoning models have different data, lifecycles, and failure modes.

Angle

Take the AV/agency distinction literally and derive the engineering consequences: independent versioning, separate training data regimes (~2 minutes of video vs. a curated personal corpus), and independent replacement under version skew.

Source video

The Bottleneck Is Your Corpus, Not Your Model

Agent fidelity to a principal is bounded by the curation quality of the private knowledge base, not by base-model capability, so engineering effort should shift from model selection toward ingestion, structuring, and lifecycle management of personal data.

Angle

Reframe personalization as a data-pipeline problem and argue for measuring marginal fidelity per data category rather than assuming more corpus is better.

Source video

Consent as an Artifact, Not a Policy: Gating Identity Synthesis at Creation Time

Post-hoc deepfake detection is structurally losing; the durable control is a first-party, verifiable consent artifact required before an identity model exists, bound to the model version and revocable so that withdrawal invalidates generation.

Angle

Contrast output-layer moderation with creation-time gating, and specify what a consent artifact must contain (verification method, storage, binding to the generated model, revocation path).

Source video

Trust Protocols Must Be Re-derived Per Modality

When a channel becomes cheaply synthesizable, authentication cannot stay in that channel; the household verification-word habit is a symptom of a missing runtime attestation layer, and institutions inheriting these habits will be defrauded at scale.

Angle

Derive the general rule from the voice/video case and sketch what capture-time provenance would need to look like for live calls.

Source video

Project ideas

4

Consent-Bound Avatar Registry

gatehouse

A consent artifact cryptographically bound to a specific trained avatar model, with a revocation path that hard-disables generation, can be implemented without making avatar enrollment unusably slow.

Proof of concept

Build a small registry service: record consent statement, hash and sign it, bind the hash into the avatar model's metadata, gate every inference call on a live consent check, and expose a revoke endpoint that invalidates future generation while retaining an audit log.

Measurement

Enrollment time overhead per avatar (target: within the reported 4-minute avatar creation flow), successful block rate of generation attempts after revocation (target 100%), and audit-trail completeness for every generated asset.

Source video

Principal-Agreement Benchmark for Personal Corpora

beyond-evals

An agent's agreement with a principal's own answers increases measurably with curated personal corpus coverage up to a saturation point, after which additional raw data yields diminishing or negative returns.

Proof of concept

Collect a principal's answers to a held-out set of decisions and judgment prompts; build agents grounded in incrementally larger curated corpus slices (history, values, professional knowledge); score agreement against the principal's own answers at each increment.

Measurement

Agreement rate (exact/partial match, plus rubric score) versus corpus slice size and category; marginal fidelity per data category; point at which additional data stops helping.

Source video

Live-Call Provenance Challenge

gatehouse

Signed capture-time attestation combined with a challenge-response secret in a separate channel detects real-time avatar impersonation with acceptable user-perceived latency, outperforming shared verification words alone.

Proof of concept

Prototype a calling flow where the initiator's device signs a capture attestation and the callee requests a nonce answerable only out-of-band; measure end-to-end added latency and run an adversarial test with a cloned-voice impersonator.

Measurement

Impersonation detection rate, false-rejection rate on genuine calls, added call-setup latency, and task-completion time versus a password-only baseline.

Source video

Cross-Silo Diagnostic Lead-Time Harness

beyond-evals

An agent that assembles fragmented records across specialties in a single reasoning pass detects flagged conditions earlier than standard referral pathways, but at a false-positive cost that must be quantified before deployment.

Proof of concept

Run a retrospective, outcome-blinded review where the agent processes historical lab panels and record fragments, flags candidate conditions, and its flags are compared against the eventual clinical outcome and the documented referral timeline.

Measurement

Detection lead time in days versus standard pathway, false-positive rate per panel, and specialist-confirmed precision on flagged cases.

Source video

Architectural implications

2

Avatar and agent are separated as distinct layers with distinct training data and responsibilities.

Before

A single model/stack that both looks like a person and answers as that person, with rendering and reasoning entangled.

After

A presentation layer (audio-visual avatar, ~2 min training data) cleanly decoupled from a reasoning layer (agent with knowledge base + LLM).

Consequence

Rendering models and reasoning models can be versioned, replaced, and licensed independently; the knowledge base becomes the primary product asset and personalization surface.

Source video

Identity synthesis is gated by a first-party consent artifact rather than by post-hoc detection.

Before

Generate first, detect/moderate misuse afterward; trust enforced at the output or policy layer.

After

Consent verification is a prerequisite in the creation pipeline; a recorded consent statement is required before an avatar exists.

Consequence

Adds a mandatory enrollment step and stored consent artifacts to the pipeline, trading onboarding friction for a strong audit trail against impersonation.

Source video

Tradeoffs and failure modes

3

Localization at scale (175+ languages)

Benefit

Video content can be produced and localized 10x faster and far cheaper than traditional studio production, with 4K output.

Cost or risk

Consent and moderation are recorded as first-party and per-avatar, but the 175-language surface multiplies the reach of any hallucinated or malicious content and the verification burden of the authenticity question.

HeyGen allows users to create lifelike avatar videos in over 175 languages and dialects 10x faster than traditional production.
Open source video
Source video

Autonomous agents acting for a person

Benefit

An agent can reason over personal history and values on the person's behalf (e.g., legacy agents passing down wisdom).

Cost or risk

Legal liability and authority for actions taken in a person's name is unresolved — whether an agent can contract or act on someone else's behalf is an open problem explicitly flagged.

How will autonomous AI agents handle legal liability and contractual negotiations on behalf of individuals?
Open source video
Source video

Personal health agent on sensitive data

Benefit

Cross-silo synthesis produced a diagnostic lead ahead of specialists in the reported case.

Cost or risk

Evidence for the medical outcome is a single anecdotal case from a participant; the summary provides no evaluation methodology, false-positive rate, or clinician-verified confirmation.

Steve Brown's medical agent correctly identified potential multiple myeloma indicators from low immunoglobulins before specialists did.
Open source video
Source video

Open questions

4

How will autonomous AI agents handle legal liability and contractual negotiations on behalf of individuals?

Why unresolved

Raised as a question to investigate in the summary; no mechanism, precedent, or design is offered for attributing authority or blame to an agent acting for a principal.

Research direction

Design an agent action-authority model with explicit delegation scopes, signed intent records, and revocation, then evaluate against contract-formation scenarios.

Source video

What standards will emerge for verifying the authenticity of real-time AI video calls?

Why unresolved

The summary proposes out-of-band verification words but offers no protocol, cryptographic binding, or runtime attestation for live streams.

Research direction

Prototype live-call provenance: signed capture attestation at the device plus a challenge-response secret in a separate channel, and measure latency and usability cost.

Source video

Does grounding an agent in a personal knowledge base actually improve fidelity, or does it mainly improve coverage?

Why unresolved

The summary asserts agents need training on personal history and values to represent you accurately, but gives no measurement of representational accuracy versus corpus size or curation quality.

Research direction

Build a benchmark that scores an agent's decisions against the principal's own answers as a function of knowledge-base coverage, and measure the marginal value of each data category.

Source video

How reliable is early-diagnosis synthesis by a personal medical agent across a population, not just one case?

Why unresolved

Only a single anecdotal success is described; no base rate, false-positive burden, or comparison to specialist workflow is provided.

Research direction

Run a retrospective cohort study where the agent reviews historical lab panels blinded to outcomes and compare detection lead time and false-positive rate against standard referral pathways.

Source video

Key claims

8
comparativeVerification needed

HeyGen allows users to create lifelike avatar videos in over 175 languages and dialects 10x faster than traditional production.

Evidence

HeyGen allows users to create lifelike avatar videos in over 175 languages and dialects 10x faster than traditional production.

Question

Against what baseline is '10x faster' measured, and is quality held constant across all 175 languages?

Source video
factualVerification needed

Roughly 2 minutes of training video is sufficient to train a custom model on a person's facial movements and voice.

Evidence

Requiring roughly 2 minutes of training video to train custom models on facial movements and voice.

Question

What fidelity and identity-consistency metrics are achieved from a 2-minute sample, and how do they degrade out of distribution?

Source video
opinionVerification not requested

Avatar is audio-visual representation; an agent is agency, knowledge base, and LLM reasoning.

Evidence

Avatar stands for audio-visual, while Agent stands for agency.

Source video
causalVerification needed

Giving an agent agency requires training it on personal history, values, and knowledge bases so it can represent you accurately.

Evidence

Giving an agent agency requires training it on personal history, values, and knowledge bases so it can represent you accurately.

Question

What evidence links knowledge-base training to measurably more accurate representation of a principal's judgments?

Source video
causalVerification needed

Steve Brown's medical agent correctly identified potential multiple myeloma indicators from low immunoglobulins before specialists did.

Evidence

Steve Brown's medical agent correctly identified potential multiple myeloma indicators from low immunoglobulins before specialists did.

Question

Was the finding independently confirmed clinically, and what was the false-positive rate across the agent's other analyses?

Source video
factualVerification needed

HeyGen requires first-party video consent for every avatar created on the platform.

Evidence

HeyGen requires first-party video consent for every avatar created on the platform.

Question

How is the consent artifact verified, stored, and bound to the generated model, and can it be revoked to invalidate an existing avatar?

Source video
causalVerification needed

A moderation team and automated safeguards prevent hate speech, fraud, and political manipulation on the avatar platform.

Evidence

A moderation team and automated safeguards prevent hate speech, fraud, and political manipulation.

Question

What measured detection rates and adversarial bypass rates have been observed for this moderation stack?

Source video
predictionVerification needed

By 2040, we are likely to see as many as 10 billion humanoid robots.

Evidence

By 2040, we are likely to see as many as 10 billion humanoid robots. (future projection)

Question

On what manufacturing-capacity and cost-curve assumptions is this 2040 figure based, and is it independently reproducible?

Source video

Connections

5