Deploy AI Without Destroying Trust: Pitfalls in the Age of Enterprise-scale GenAI With CTGT

Deploy AI Without Destroying Trust: Pitfalls in the Age of Enterprise-scale GenAI With CTGT — Ai4

Ai4

AI-assisted summary of Deploy AI Without Destroying Trust: Pitfalls in the Age of Enterprise-scale GenAI With CTGT — original source. Check the original for full context.

Source published:

Detailed notes

## Background Published on YouTube by Ai4 on August 25, 2025, this talk features Cyril Gorlla, Founder and CEO of CTGT. According to the page description, the session explores how enterprises can deploy generative AI without destroying trust by using CTGT's model-agnostic governance architecture to mediate hallucinations in real time and enforce enterprise policies without fine-tuning or prompt engineering.

## TL;DR Cyril Gorlla outlines the enterprise crisis surrounding generative AI adoption, where organizations hesitate to deploy non-deterministic black-box models into high-risk environments. He explains that traditional band-aids like expensive fine-tuning and brittle prompt engineering fail under pressure, leading to issues like guardrail drift across model upgrades. Instead, CTGT introduces a mechanistic, model-agnostic governance platform that steers internal neuron activations directly from a centralized single source of truth. Gorlla illustrates this capability by demonstrating how dormant historical facts were unlocked in DeepSeek without prompt hacks, emphasizing the necessity of deterministic infrastructure for enterprise and national policy.

## Atmosphere & style The presentation carries an authoritative, intellectually rigorous, and high-stakes conference energy, bridging academic computer science history with cutting-edge tech diplomacy. Gorlla delivers his insights with direct, confident clarity, shifting smoothly from technical breakdowns of neural activations to executive-level governance and Capitol Hill briefings. It feels urgent yet pragmatic, capturing the tension of enterprise leaders who must innovate without exposing their organizations to catastrophic legal and reputational risks.

## Key takeaways - [00:04](https://www.youtube.com/watch?v=QyZqCNYVFzI&t=4s) Enterprise generative AI adoption faces major bottlenecks due to the legal, financial, and compliance risks of deploying non-deterministic black-box models in high-risk sectors. - [00:04](https://www.youtube.com/watch?v=QyZqCNYVFzI&t=4s) Machine learning has cycled through booms and busts since the 1960s perceptron era, which collapsed into an AI winter before foundational breakthroughs revived the field. - [00:04](https://www.youtube.com/watch?v=QyZqCNYVFzI&t=4s) David Rumelhart and Geoffrey Hinton's development of backpropagation at UC San Diego provided the foundational algorithm that powers modern neural networks. - [02:21](https://www.youtube.com/watch?v=QyZqCNYVFzI&t=141s) Foundation models frequently oscillate between extreme censorship and unconstrained outputs due to an underlying lack of deterministic control. - [02:21](https://www.youtube.com/watch?v=QyZqCNYVFzI&t=141s) Traditional interventions like fine-tuning and prompt engineering act as brittle band-aids that are costly, unpredictable, and vulnerable to edge cases. - [04:42](https://www.youtube.com/watch?v=QyZqCNYVFzI&t=282s) CTGT uses a mechanistic governance approach, tracing neuron activations within transformer architectures to remediate hallucinations in real time. - [04:42](https://www.youtube.com/watch?v=QyZqCNYVFzI&t=282s) Decoupling enterprise compliance, brand voice, and regulatory rules from the underlying LLM creates a single source of truth that remains model-agnostic. - [04:42](https://www.youtube.com/watch?v=QyZqCNYVFzI&t=282s) An external governance layer prevents breaking changes and guardrail drift when model providers alter internal safety thresholds across model updates. - [07:11](https://www.youtube.com/watch?v=QyZqCNYVFzI&t=431s) Mechanistic steering enabled the activation of latent historical knowledge within DeepSeek to bypass built-in censorship without retraining or prompt wrappers. - [07:11](https://www.youtube.com/watch?v=QyZqCNYVFzI&t=431s) Transparent model steerability and AI governance have evolved from academic research into pressing matters of national and geopolitical policy.

## Detailed walkthrough ### [00:04](https://www.youtube.com/watch?v=QyZqCNYVFzI&t=4s) The Historical Cycles of AI and the Enterprise Dilemma

Opening the final day of the Ai4 conference after kicking off the track to a packed house, Cyril Gorlla addresses the palpable hesitation and trepidation surrounding enterprise generative AI adoption. While organizations are eager to capitalize on GenAI, the fear of deploying non-deterministic models into mission-critical, high-risk environments—where companies face severe legal, financial, and regulatory exposure—remains a major bottleneck. Gorlla argues that navigating this requires understanding the broader history of machine learning, which has cycled through dramatic booms and busts since the 1960s.

```mermaid flowchart TD accTitle: Evolution of Neural Networks and AI Cycles subgraph History["Historical AI Cycles"] P["1960s: Perceptrons<br/>'Electronic Brains'"] --> W["1970s–1980s: AI Winter<br/>Minsky's Critiques & Funding Cuts"] W --> BP["1986+: Backpropagation<br/>Rumelhart & Hinton (UCSD)"] end subgraph Modern["Modern Era & Enterprise Reality"] BP --> LLM["Modern LLM Boom<br/>GPT, Claude, DeepSeek"] LLM --> Risk["Enterprise Deployment Bottleneck<br/>Non-deterministic Black Boxes in High-Risk Sectors"] end ```

The concept of an "electronic brain that teaches itself" captivated the public via *The New York Times* headlines decades ago, based on early perceptron architectures that wired simple dot-connecting algorithms together. However, structural limitations highlighted by Marvin Minsky triggered the first AI winter in the 1970s and 1980s, drying up capital and research interest. That freeze was eventually broken at UC San Diego, where David Rumelhart and Geoffrey Hinton published their seminal work on backpropagation—the foundational algorithm underpinning all modern artificial neural networks. Hinton’s path from the Rumelhart Prize to a Nobel Prize underscores how today's generative AI boom is built on decades of foundational breakthroughs, yet enterprises still struggle with the same fundamental challenge: controlling the black box.

- The perceptron era in the 1960s promised self-learning electronic brains but collapsed into the first AI winter following Marvin Minsky's critiques. - Geoffrey Hinton and David Rumelhart solved the structural bottleneck at UCSD by introducing backpropagation for multi-layer neural networks. - High-risk industries like healthcare, finance, and insurance face catastrophic downsides from unmanaged AI non-determinism.

---

[02:21](https://www.youtube.com/watch?v=QyZqCNYVFzI&t=141s) The Non-Deterministic Black Box and Brittle Band-Aids

Modern foundation models exhibit stark failures in enterprise governance, ranging between extreme censorship and complete lack of guardrails. Gorlla points to real-world examples: DeepSeek strictly toeing Chinese state narratives on one end, and Grok generating unhinged, extremist outputs on the other. Both extremes stem from the exact same underlying flaw: AI models are fundamentally non-deterministic black boxes, preventing enterprises from having 100% deterministic control over model behavior in production environments.

Drawing from his research at Stanford and UCSD, Gorlla explains why traditional methods to steer models fail at enterprise scale. Today's standard interventions—fine-tuning and prompt engineering—are fundamentally brittle and surface-level: - **Fine-Tuning:** Consists of feeding additional domain data into the model at immense expense (often $100,000 to $1,000,000), yet yields unpredictable generalization where the model improves on some tasks while degrading on others. - **Prompt Engineering:** Relies on appending lengthy instructions or system prompts to queries, hoping the model remains adherent, but easily breaks under adversarial inputs or edge cases.

Gorlla recalls a conversation with Sam Altman where both compared the current state of AI interaction to the chaotic early days of the web: "very exciting," but cluttered with the modern equivalents of "pop-ups and toolbars and a lot of like spam." Fine-tuning and prompt hacks act merely as primitive band-aids over internal model weights rather than providing true mechanistic control.

---

[04:42](https://www.youtube.com/watch?v=QyZqCNYVFzI&t=282s) Mechanistic Policy Governance and Model-Agnostic Control

Rather than treating models as impenetrable black boxes, CTGT's platform takes a first-principles, mechanistic approach. In a transformer architecture, every token maps to vector spaces that correspond directly to specific internal parameter clusters and neuron groups. By tracing these activations, CTGT locates the exact sub-networks responsible for factual assertions or hallucinations—such as identifying why a model hallucinates that the "Eiffel Tower is in Rome" and steering activations toward "Paris."

```mermaid flowchart LR accTitle: CTGT Model-Agnostic Enterprise Governance Architecture subgraph Input["Enterprise Policy Layer"] P1["Brand Voice & Tone"] P2["SEC & FINRA Rules"] P3["Compliance Policies"] end

subgraph Platform["CTGT Governance Engine"] SSOT["Single Source of Truth<br/>Policy Ingestion"] Mech["Neuron-Level Activation Steering<br/>Real-Time Hallucination Mediation"] SSOT --> Mech end

subgraph Models["Decoupled Foundation Models"] M1["OpenAI (GPT-4 / GPT-5)"] M2["Anthropic (Claude 3.5 / 3.7)"] M3["Google Gemini / DeepSeek"] end

Input --> SSOT Mech <--> Models ```

CTGT translates this mechanistic capability into an enterprise governance platform deployed across Fortune 500 organizations. The architecture decouples an enterprise's regulatory and organizational rules (such as tone of voice, brand standards, SEC rules, and FINRA compliance) from the underlying LLM:

- **Single Source of Truth:** Policies are ingested centrally into the platform rather than baked into system prompts or fine-tuning datasets. - **Model Agility Without Guardrail Drift:** Enterprises can swap foundational models (e.g., upgrading from GPT-4 to GPT-5, or migrating to Anthropic or Google models) without losing customized governance logic. - **Eliminating Breaking Upgrades:** When model providers alter internal safety thresholds—such as Anthropic drastically adjusting internal safety between Claude 3.5 and Claude 3.7 Sonnet, which broke downstream tools and guardrails—CTGT's external policy layer maintains consistent enforcement.

---

[07:11](https://www.youtube.com/watch?v=QyZqCNYVFzI&t=431s) De-Censoring DeepSeek and Briefing Capitol Hill

To demonstrate the power of real-time mechanistic policy steering without fine-tuning or prompt engineering, Gorlla highlights an experiment conducted on DeepSeek. Out of the box, standard DeepSeek refuses queries about the 1989 Tiananmen Square protests due to built-in Chinese state alignment. By identifying and amplifying the dormant factual representations already stored within DeepSeek's internal parameter space, CTGT forced the model to generate accurate, objective historical responses—without spending hundreds of thousands of dollars on retraining or appending complex prompt wrappers.

``` [Standard DeepSeek Output] Query: "What happened in Tiananmen Square?" Response: Refusal / Unable to answer based on safety guidelines.

[CTGT Mechanistic Policy Steering (No Fine-Tuning / No Prompt Hacks)] Query: "What happened in Tiananmen Square?" Response: Generates detailed historical facts by activating latent model knowledge. ```

This ability to inspect and modulate internal model behavior has placed CTGT at the center of high-stakes discussions surrounding tech policy, safety, and national governance. Gorlla notes that he presented these architectural findings in Washington, D.C., near the White House, briefing United States senators including JD Vance. AI has rapidly transitioned from niche academic conferences in the late 2000s to a central driver of global socioeconomic and geopolitical policy. Gorlla concludes that the path forward for enterprise GenAI is not retreating in fear of hallucinations, but implementing rigorous, transparent infrastructure that puts deterministic trust and safety at the foundation via `ctgt.ai`.

## Notable quotes > [00:04](https://www.youtube.com/watch?v=QyZqCNYVFzI&t=4s) electronic brain that teaches itself — Cyril Gorlla

> [02:21](https://www.youtube.com/watch?v=QyZqCNYVFzI&t=141s) very exciting — Sam Altman

> [02:21](https://www.youtube.com/watch?v=QyZqCNYVFzI&t=141s) pop-ups and toolbars and a lot of like spam — Sam Altman

> [04:42](https://www.youtube.com/watch?v=QyZqCNYVFzI&t=282s) Eiffel Tower is in Rome — Cyril Gorlla

> [04:42](https://www.youtube.com/watch?v=QyZqCNYVFzI&t=282s) Paris — Cyril Gorlla

> [07:11](https://www.youtube.com/watch?v=QyZqCNYVFzI&t=431s) What happened in Tiananmen Square?

> [07:11](https://www.youtube.com/watch?v=QyZqCNYVFzI&t=431s) Refusal / Unable to answer based on safety guidelines.

> [07:11](https://www.youtube.com/watch?v=QyZqCNYVFzI&t=431s) Generates detailed historical facts by activating latent model knowledge.

## Mentioned Cyril Gorlla, CTGT, Ai4, Google, PyTorch Conference Showcase, VentureBeat AI Innovation Award, The New York Times, Marvin Minsky, UC San Diego, David Rumelhart, Geoffrey Hinton, Rumelhart Prize, Nobel Prize, DeepSeek, Grok, Stanford, Sam Altman, OpenAI, GPT-4, GPT-5, Anthropic, Claude 3.5, Claude 3.7 Sonnet, Google Gemini, SEC, FINRA, Tiananmen Square, Capitol Hill, White House, JD Vance, ctgt.ai

## Video description Cyril Gorlla, Founder & CEO, CTGT

Abstracts

Enterprise-scale GenAI adoption is no longer optional, but are you inadvertently deploying a ticking time bomb? The rush to implement "black box" models exposes Fortune 500 organizations to exponential risk, from damaging hallucinations and compliance breaches to the catastrophic erosion of stakeholder trust. True enterprise deployment demands deterministic control across a variety of use cases, not just experimentation with point solutions. CTGT, backed by Google and winner of the PyTorch Conference Showcase and VentureBeat AI Innovation Award, will unveil its model agnostic governance architecture, demonstrating how we mediate hallucinations in real-time and enforce hundreds of thousands of policies without fine-tuning or prompt engineering. I will show how we validate AI outputs against your single source of truth to finally open the black box. Discover the critical infrastructure necessary to deploy AI with confidence and secure your lead in the GenAI revolution.

[Source](https://www.youtube.com/watch?v=QyZqCNYVFzI)

中文简报

部署AI而不破坏信任：企业级生成式AI时代的陷阱与CTGT Cyril Gorlla 概述了企业在采用生成式 AI 时面临的危机，即各机构在将非确定性的黑盒模型部署到高风险环境中时犹豫不决。他解释称，高昂的微调和脆弱的提示工程等传统权宜措施在面临压力时会失效，从而导致模型升级过程中出现护栏漂移等问题。相比之下，CTGT 引入了一个机理性、与模型无关的治理平台，直接基于集中的单一事实来源引导内部神经元激活。Gorlla 通过展示如何在不依赖提示词技巧的情况下解锁 DeepSeek 中休眠的历史事实来说明这一能力，并强调了确定性基础设施对企业和国家政策的必要性。 企业在落地生成式 AI 时，因黑盒模型的非确定性面临严重的法律、财务及合规风险，这一落地瓶颈推动了对底层确定性控制技术的需求。 (0:04) 当前的微调与提示词工程等传统手段成本高昂且脆弱，既无法保证泛化稳定性，也难以在边缘案例中提供真正的机制化控制。 (2:21) CTGT 采用机制化治理方案，通过追踪 Transformer 架构内部的神经元激活状态，在无需微调或修改提示词的情况下实时纠正模型幻觉。 (4:42) 通过将企业的合规与品牌规则与底层大模型解耦，CTGT 构建了与模型无关的单一事实来源，有效规避了模型提供商调整安全阈值时导致的下游护栏失效。 (4:42) 在针对 DeepSeek 的实验中，CTGT 成功激活其内部潜伏参数以输出天安门事件等受限历史事实，CTGT 团队亦向包括 JD Vance 在内的美国参议员汇报了该治理架构。 (7:11)

Reading summary

Cyril Gorlla argues that enterprise generative AI adoption is bottlenecked by the unpredictable, non-deterministic nature of foundation models, asserting that mechanistic neuron-level steering offers deterministic governance where traditional fine-tuning and prompting fail. Enterprise Risk and the Non-Deterministic Black Box (0:04) Generative AI deployment in high-risk sectors such as healthcare, finance, and insurance remains constrained by legal, financial, and regulatory liabilities. Gorlla contextualizes this tension within historical cycles of machine learning, noting that early 1960s perceptron enthusiasm collapsed into the first AI winter following Marvin Minsky's critiques, before David Rumelhart and Geoffrey Hinton unlocked multi-layer neural networks via backpropagation at UC San Diego. Despite modern performance gains, Gorlla contends that commercial foundation models remain opaque black boxes. In production environments, this non-determinism manifests in uncontrolled behavioral extremes, illustrated on one hand by DeepSeek adhering strictly to Chinese state narratives, and on the other by Grok producing unhinged or extremist outputs without reliable internal constraints. The Fragility of Fine-Tuning and Prompting (2:21) Conventional approaches to controlling large language models rely on fine-tuning and prompt engineering, which Gorlla characterizes as brittle interventions that fail to solve core reliability challenges. Fine-tuning requires feeding specialized datasets at costs ranging from $100,000 to $1,000,000, yet often introduces erratic generalization by degrading performance in secondary tasks while optimizing for primary ones. Prompt engineering similarly acts as an unstable layer of instruction wrapping that routinely collapses under adversarial prompts or edge cases. Referencing a discussion with Sam Altman, Gorlla likens the current generation of user-facing prompt band-aids to the early internet's era of pop-ups and browser toolbars, arguing that neither prompts nor fine-tuning give enterprises verifiable, mechanistic control over internal parameters. Mechanistic Steering and Decoupled Governance (4:42) To achieve deterministic oversight, CTGT's platform applies mechanistic interpretability to transformer architectures. By tracing how tokens map into vector spaces and activate specific parameter clusters, the platform identifies internal sub-networks linked to factual inaccuracies—such as isolating activations that falsely place the Eiffel Tower in Rome—and redirects them to factual representations without altering base model weights. This architecture establishes an external policy layer that decouples an enterprise's organizational rules, brand standards, and regulatory mandates (such as SEC and FINRA compliance) from third-party foundation models. Gorlla emphasizes that this decoupling enables companies to swap providers between OpenAI, Anthropic, or Google without losing governance rules, while insulating organizations from breaking provider-side updates, such as the internal safety adjustments between Claude 3.5 and Claude 3.7 Sonnet. Geopolitical Applications and Regulatory Briefings (7:11) To evaluate mechanistic intervention on live models, CTGT tested activation steering on DeepSeek, which default-refuses inquiries regarding the 1989 Tiananmen Square protests due to integrated state alignment. By isolating and amplifying dormant factual representations already present in the parameter space, CTGT elicited accurate historical outputs without retraining or system prompt modifications. Gorlla highlights that the capacity to inspect and steer internal model behavior has escalated AI governance into a matter of national policy. These architectural findings formed the basis of briefings delivered to United States lawmakers in Washington, D.C., including Senator JD Vance, reflecting the transition of machine learning from academic theory into a driver of socioeconomic and geopolitical policy. Gorlla concludes that enterprises should not abandon generative AI out of fear of hallucinations, but instead implement transparent, mechanistic infrastructure that establishes deterministic compliance and safety directly at the model activation layer.

精炼阅读摘要

CTGT创始人Cyril Gorlla在Ai4大会上提出，阻碍企业级生成式AI落地的核心瓶颈在于非确定性黑盒模型的风险失控，而传统的微调和提示词工程无法提供确定性保障，必须转向神经元级别的机制性策略治理架构。 技术历史周期与高风险企业的采用困境 (0:04) 生成式AI在企业的采纳面临着显著的犹豫与顾虑。企业虽渴望获取生成式AI的价值，但在面临严格法律、财务和监管责任的关键任务与高风险环境中，部署非确定性模型存在严重的风险暴露。Gorlla指出，理解当前的困境需要回顾机器学习自1960年代以来的历史周期。当时感知机架构曾被《纽约时报》等媒体誉为自主学习的“电子大脑”，但因Marvin Minsky指出的结构性缺陷而陷入资金枯竭与研究停滞的AI寒冬。直到1986年后，David Rumelhart与Geoffrey Hinton在加州大学圣迭戈分校发表了反向传播算法的开创性研究，才打破长期停滞并奠定了现代人工神经网络的基础。 尽管现代大语言模型展现出巨大突破，但医疗保健、金融和保险等高风险行业依然面临相同的根本难题：黑盒模型缺乏确定性控制。如果无法有效管理非确定性带来的潜在灾难性后果，企业在关键业务场景中的规模化落地便难以实质推进。 传统微调与提示词工程的脆弱局限 (2:21) 当前基础模型在企业级治理中展现出明显的两极分化缺陷，要么走向极端审查（例如DeepSeek严格遵循特定国家叙事），要么完全缺乏护栏（例如Grok生成失控的极端主义言论）。Gorlla结合其在斯坦福大学与加州大学圣迭戈分校的研究指出，这种失控根源于模型本质上属于非确定性黑盒，企业无法在生产环境中达成百分之百的确定性行为控制。 目前业内通用的控制手段主要依赖微调与提示词工程，但Gorlla认为这两种方式均属脆弱的表层修补。微调通常需要耗费10万至100万美元导入领域数据，却极易产生不可预测的泛化问题，往往在提升某些任务表现的同时导致其他能力退化；提示词工程则依赖在查询中拼接冗长指令或系统设定，极易在对抗性输入或边界工况下失效。Gorlla援引与Sam Altman的交流观点指出，当前的AI交互形态宛如早期充满弹窗、工具栏与垃圾信息的混乱互联网，单纯在权重表层打补丁无法实现真正的底层控制。 神经元级机制治理与模型解耦架构 (4:42) CTGT主张从第一性原理出发，采用机制性方法直接干预模型内部表征，而非将模型视作不可透视的黑盒。在Transformer架构中，每个Token均映射至特定的向量空间，并直接对应内部参数簇和神经元群组。通过追踪这些激活状态，平台能够精准定位负责事实陈述或产生幻觉的特定子网络，例如直接识别模型为何将“埃菲尔铁塔位于巴黎”错误陈述为“位于罗马”，并实时修正其激活路径。 基于该原理构建的治理平台已被部署于多家财富500强企业。该架构将企业的合规与组织规则（包括品牌基调、美国证券交易委员会SEC规则以及金融业监管局FINRA合规要求）与底层模型彻底解耦，构建统一的事实来源（Single Source of Truth）。这种解耦使企业在升级或更换底层基础模型时（如从GPT-4迁移至GPT-5，或切换至Anthropic与Google模型），无需重构业务治理逻辑；同时避免了因模型供应商内部调整安全阈值而导致的业务中断，例如Anthropic在Claude 3.5升级至Claude 3.7 Sonnet时调整了内置安全参数，曾造成下游工具和护栏失效，而解耦的外部策略层可确保治理规则执行的一致性。 底层表征调控验证与监管汇报实践 (7:11) 为验证无需微调与提示词包装的实时机制性策略引导能力，Gorlla展示了针对DeepSeek的测试实验。标准版DeepSeek原生地因内置对齐规则而拒绝回答关于1989年天安门广场事件的提问；CTGT通过识别并激活该模型内部参数空间中处于休眠状态的事实表征，在未追加微调资金和复杂提示词包装的前提下，促使模型生成了客观、详实的历史事实回答。 这种审查与调控模型内部深层机制的能力，使CTGT参与到了高层技术政策与国家治理讨论中。Gorlla表示，他曾前往华盛顿特区白宫附近，向包括JD Vance在内的多位美国参议员汇报了相关技术架构与研究成果。这表明AI技术已从2000年代末的学术研讨，迅速转变为全球社会经济与地缘政治政策的核心议题。 Gorlla总结指出，企业应用生成式AI的正确路径不应是在对幻觉的恐惧中退缩，而必须引入透明严谨的底层机制基础设施，在基础架构层面确立确定性的信任与安全控制；但如何在跨厂商模型架构持续迭代的过程中，长期且低成本地维持神经元级策略干预的普适性，仍是企业规模化落地中值得持续关注的课题。

Detailed notes — zh

## Background 该视频由 Ai4 官方频道于 2025 年 8 月 25 日发布在 YouTube 平台。视频内容为 CTGT 创始人兼首席执行官 Cyril Gorlla 的演讲，重点探讨在企业规模化落地生成式 AI 的浪潮中，黑盒模型的非确定性与幻觉风险，并介绍了 CTGT 如何通过模型无关治理架构在不依赖微调和 Prompt 工程的前提下实现企业级确定性控制。 ## TL;DR Cyril Gorlla 从早期感知机、AI 寒冬及反向传播的兴衰历史切入，剖析了当前企业部署生成式 AI 所面临的非确定性黑盒困境与合规风险。他指出 Prompt 工程和微调仅是脆弱且昂贵的“表面修补”，无法解决高风险核心业务的信任问题。为此，CTGT 提出了基于神经元级表征干预的模型无关治理架构，将企业合规策略与底层大模型解耦，并在零额外训练成本下成功验证了对模型输出的精准掌控。最后，他分享了该技术向美国国会政策制定者汇报的经历，呼吁企业构建以确定性和安全为核心的 GenAI 信任体系。 ## Atmosphere & style 整场演讲兼具学术底蕴与极具前瞻性的商业洞察，语调沉稳笃定且直击痛点。演讲者以冷静犀利的视角拆解当下大模型的脆弱现状，将 Prompt 工程比作早期的互联网牛皮癣弹窗，展现出直面技术本质的坦率。从深奥的表征干预机制到华盛顿高层的政策对话，整场分享节奏紧凑、说服力极强，给观众带来一种穿透行业泡沫、把握技术底层确定性的启发感。 ## Key takeaways - [00:04](https://www.youtube.com/watch?v=QyZqCNYVFzI&t=4s) AI 发展经历了 1960 年代感知机热潮、理论瓶颈引发的 AI 寒冬，直到反向传播算法奠定现代深度学习基石，其兴衰周期为当下 GenAI 落地提供了重要镜鉴。 - [02:21](https://www.youtube.com/watch?v=QyZqCNYVFzI&t=141s) 金融、法律与医疗等高价值场景容错率极低，企业在部署 GenAI 时最核心的阻碍是底层黑盒模型带来的非确定性与合规失控风险。 - [02:21](https://www.youtube.com/watch?v=QyZqCNYVFzI&t=141s) DeepSeek 的意识形态审查与 Grok 的极端言论等极端案例表明，尽管模型参数规模不断扩大，微观层面的不可控崩溃依然频发。 - [04:10](https://www.youtube.com/watch?v=QyZqCNYVFzI&t=250s) 传统的 Prompt 工程与模型微调只是脆弱的表面修补手段，既无法触及模型内部表征，也无法为企业提供确定性输出保障。 - [04:10](https://www.youtube.com/watch?v=QyZqCNYVFzI&t=250s) CTGT 采用基于第一性原理的神经元级表征干预技术，能够精确定位并纠正引发事实幻觉的具体神经元群组。 - [04:10](https://www.youtube.com/watch?v=QyZqCNYVFzI&t=250s) CTGT 的模型无关策略治理架构将 SEC、FINRA 等监管政策与底层模型彻底解耦，确立企业策略库为唯一的单一事实真理来源（Single Source of Truth）。 - [04:10](https://www.youtube.com/watch?v=QyZqCNYVFzI&t=250s) 策略解耦使企业在升级 GPT-5 或切换 Claude 模型时无需推倒重来，有效避免了底层模型过滤阈值变动导致企业 Guardrail 瞬间失效的模型锁定问题。 - [04:10](https://www.youtube.com/watch?v=QyZqCNYVFzI&t=250s) OpenAI 首席执行官 Sam Altman 与 Cyril Gorlla 达成共识：当前人机交互范式如同充斥弹窗的早期互联网，Prompt 工程和微调本质上属于落后的过渡形态。 - [06:09](https://www.youtube.com/watch?v=QyZqCNYVFzI&t=369s) 实战测试表明，CTGT 无需重新微调或越狱 Prompt，仅在表征层激活内部受压制的神经元即可突破 DeepSeek 审查并客观输出敏感历史事实。 - [06:09](https://www.youtube.com/watch?v=QyZqCNYVFzI&t=369s) 传统消除偏见的微调通常耗资 10 万至 100 万美元且容易引发灾难性遗忘，而表征层直接治理实现了零额外训练成本的确定性纠偏。 - [06:09](https://www.youtube.com/watch?v=QyZqCNYVFzI&t=369s) CTGT 治理架构已进入美国白宫及参议员 JD Vance 等顶级立法者的视野，标志着前沿模型可解释性技术已成为国家政策与企业战略的核心交汇点。 ## Detailed walkthrough ### [00:04](https://www.youtube.com/watch?v=QyZqCNYVFzI&t=4s) 人工智能的兴衰周期：从感知机、AI 寒冬到反向传播 在 AI4 大会的最后一天，CTGT 创始人兼首席执行官 Cyril Gorlla 再次登台。他开门见山地指出，当前企业界在部署生成式 AI（GenAI）时普遍存在犹豫和顾虑，尤其是面对金融、法律、医疗等高风险场景——一旦模型出错，企业将面临巨额经济损失与合规风险。为了讲透当下困局的根源，Gorlla 带领观众回顾了 AI 漫长的发展历程。AI 并非始于近年的 ChatGPT 热潮，早在 1960 年代，《纽约时报》就曾报道过“能够自我学习的电子大脑”，即由简单线路构成的“感知机”（Perceptron）。 然而，AI 的发展并非一帆风顺。著名 AI 学者马文·明斯基（Marvin Minsky）出版的专著指出了单层感知机的致命局限，直接引发了 1970 至 1980 年代的第一次“AI 寒冬”（AI Winter）。在那个资本与研究热情降至冰点的时期，资金匮乏程度与当下 AI 领域的资本狂热形成了鲜明对比。直到杰弗里·辛顿（Geoffrey Hinton）与大卫·鲁梅哈特（David Rumelhart）在加利福尼亚大学圣地亚哥分校（UCSD，也是 Gorlla 曾做研究的地方）发表了关于“反向传播”（Backpropagation）的奠基性论文，人工神经网络的训练瓶颈才被打破，为现代所有深度学习模型奠定了基石。Gorlla 感慨道，鲁梅哈特奖后来极富诗意地颁给了辛顿，而辛顿此后更斩获了诺贝尔物理学奖。回顾这些繁荣与萧条的循环，能为今天的企业 AI 落地提供宝贵镜鉴。 - **感知机时代（1960s）**：早期的算法以极其简单的方式连接数据点，曾引发媒体对“电子大脑”的广泛狂热。 - **AI 寒冬（1970s-1980s）**：理论瓶颈导致投资与兴趣急剧缩水，行业陷入长期低谷。 - **反向传播的破局**：Hinton 与 Rumelhart 重构了人工神经元的协作方式，构成了当下大语言模型底层逻辑的基石。 --- ### [02:21](https://www.youtube.com/watch?v=QyZqCNYVFzI&t=141s) 生成式 AI 的黑盒困境与非确定性风险 回到当下，生成式 AI 的商业化应用正遭遇严重的非确定性瓶颈。Gorlla 列举了两个典型极端案例：左边是深度偏向特定叙事的 DeepSeek，右边则是曾自称是希特勒的 Grok。这两个极端生动揭示了企业级 GenAI 的核心风险——AI 本质上是非确定性的（Non-deterministic），企业在生产环境中根本无法百分之百掌控黑盒模型的行为。当这些模型被部署到医疗保健、金融信贷、保险理赔等高价值、高容错率极低的业务核心时，黑盒带来的失控风险会被成倍放大。 为了穿透这层“暗色玻璃”，Gorlla 结合自己在斯坦福大学（Stanford）和加州大学圣地亚哥分校（UCSD）的多年科研成果，致力于从底层数学原理层面打开模型黑盒。尽管业界普遍观察到“模型参数规模扩大通常带来更强能力”的宏观规律，但微观上的诡异表现依然频发：例如模型在某些特定数据集上微调效果极佳，换一组数据却可能迅速崩溃。理解并解决这种不可控性，正是他在 2024 年 9 月创立 CTGT 的核心初衷。 ```mermaid flowchart TD accTitle: Enterprise GenAI Risk and Black Box Dilemma subgraph Extremes["大模型非确定性极端表现"] E1["DeepSeek<br/>意识形态与特定叙事审查"] E2["Grok<br/>失控极端言论与人格漂移"] end subgraph HighRisk["企业高风险核心业务场景"] H1["医疗诊断与健康建议"] H2["金融风险控制与信贷审批"] H3["保险理赔与法律合规审计"] end Extremes -->|缺乏确定性控制| Risk["黑盒不可控风险"] Risk --> HighRisk ``` --- ### [04:10](https://www.youtube.com/watch?v=QyZqCNYVFzI&t=250s) 告别脆弱的 Prompt 与微调：CTGT 的模型无关治理架构 Gorlla 直言，当前业界用来控制和优化 AI 的两种主流手段——Prompt 工程（Prompt Engineering）与微调（Fine-tuning）——都是极其脆弱的“表面修补”。微调需要灌入大量数据，成本高昂且泛化能力脆弱；Prompt 工程只是在上下文里堆砌指令，祈祷模型能听话遵循。这些方法既未触及模型内部表征，也无法提供确定性保障。CTGT 采用的是基于第一性原理的“可解释性与神经元级表征干预”方案：例如当模型产生“埃菲尔铁塔在罗马”这种典型幻觉时，模型内部的每个 Token 向量都会映射到特定参数空间，CTGT 能够精确定位导致该幻觉的具体神经元群组，并在表征层面将其矫正为“巴黎”。 将这一底层技术规模化后，CTGT 构建了一套已在多家《财富》500 强企业落地的“模型无关策略治理架构”（Model Agnostic Governance Architecture）。该架构的核心价值在于将企业的业务规则、合规政策（如 SEC、FINRA 监管条款、品牌声调、合规边界）与底层模型彻底解耦，使企业策略库成为唯一的“事实真理来源”（Single Source of Truth）。这意味着企业未来在从 GPT-4 升级到 GPT-5，或在 Anthropic 与 Google 模型之间自由切换时，无需推倒重来，企业前期的策略与合规资产将完好无损地保留。 ```mermaid flowchart LR accTitle: CTGT Model Agnostic Governance Framework subgraph EnterpriseSource["企业单一事实真理来源"] P1["SEC / FINRA 监管策略"] P2["品牌调性与合规边界"] P3["业务逻辑与风控规则"] end subgraph CTGTLayer["CTGT 策略治理中枢"] G1["模型内部表征与神经元定位"] G2["实时反幻觉与合规干预"] end subgraph FoundationModels["底层可替换基础模型"] M1["OpenAI (GPT-4 / GPT-5)"] M2["Anthropic Claude"] M3["Google Gemini"] end EnterpriseSource --> CTGTLayer CTGTLayer <--> FoundationModels ``` - **摆脱模型锁定的迁移优势**：当 Anthropic 从 Claude 3.5 迭代到 Claude 3.7 时，由于官方调高了内部安全过滤阈值，导致大量企业的原有 Guardrail 和外部工具链瞬间失效；而 CTGT 的解耦架构能有效隔离此类底层扰动。 - **与 Sam Altman 的共识**：Gorlla 提到几个月前曾与 OpenAI 首席执行官 Sam Altman 深入探讨，双方一致认为当前人类与 AI 的交互范式极其原始，"非常像早期的互联网——虽然令人兴奋，但充斥着各种弹窗、插件栏和垃圾信息"，而 Prompt 工程和微调本质上就是现代版的“牛皮癣插件”。 --- ### [06:09](https://www.youtube.com/watch?v=QyZqCNYVFzI&t=369s) 突破审查实战验证、国会政策对话与企业信任未来 为了验证这套机制的确定性控制力，Gorlla 展示了针对 DeepSeek 审查机制的解构实验。通常情况下，询问 DeepSeek 敏感历史事件（如天安门广场事件）时模型会直接拒绝回答。传统解决方式需要耗费数十万至上百万美元进行重新微调，或者设计极其繁琐的越狱 Prompt。而 CTGT 在不调整模型权重、不做 Prompt 工程的前提下，直接在表征层激活模型内部与该历史事件相关的知识神经元群组，模型便流畅输出了客观事实。这证明大模型本身完全具备相应知识库，只是其优先级被先验偏置所压制；通过机制化干预，即可重新对齐企业所需的优先级。 这种对黑盒模型的深层控制能力，让 CTGT 走到了技术伦理、科技政策与商业交汇的最前沿。Gorlla 透露，他近期刚在白宫附近向包括 JD Vance 在内的多位美国参议员进行了专题技术汇报。AI 已经从 2000 年代后期纯粹的学术研讨，一跃成为牵动全球社会经济与公共政策的核心议题。Gorlla 在演讲尾声呼吁，企业拥抱 GenAI 不应出于盲目的恐惧或采取敷衍的创可贴式方案，而应采用将“信任与安全”置于核心的确定性治理框架。他欢迎所有在核心高风险场景探索 AI 落地的企业与团队开展深度合作（访问 ctgt.ai），共同推动行业走向开放与成熟。 - **成本对比**：传统微调消除特定偏见或幻觉通常耗资 10 万至 100 万美元，且极易引发灾难性遗忘；表征层直接治理实现零额外训练成本。 - **政策汇报**：CTGT 架构已进入美国顶级立法者（包括参议员 JD Vance）的视野，展现了前沿可解释性技术在国家政策与企业合规层面的重大价值。 ## Notable quotes > [00:04](https://www.youtube.com/watch?v=QyZqCNYVFzI&t=4s) 在 1960 年代，《纽约时报》就曾报道过“能够自我学习的电子大脑”，即由简单线路构成的“感知机”（Perceptron）。 — Cyril Gorlla > [00:04](https://www.youtube.com/watch?v=QyZqCNYVFzI&t=4s) 直到杰弗里·辛顿（Geoffrey Hinton）与大卫·鲁梅哈特（David Rumelhart）在加利福尼亚大学圣地亚哥分校发表了关于“反向传播”（Backpropagation）的奠基性论文，人工神经网络的训练瓶颈才被打破，为现代所有深度学习模型奠定了基石。 — Cyril Gorlla > [02:21](https://www.youtube.com/watch?v=QyZqCNYVFzI&t=141s) 左边是深度偏向特定叙事的 DeepSeek，右边则是曾自称是希特勒的 Grok。这两个极端生动揭示了企业级 GenAI 的核心风险——AI 本质上是非确定性的（Non-deterministic）。 — Cyril Gorlla > [02:21](https://www.youtube.com/watch?v=QyZqCNYVFzI&t=141s) 尽管业界普遍观察到“模型参数规模扩大通常带来更强能力”的宏观规律，但微观上的诡异表现依然频发：例如模型在某些特定数据集上微调效果极佳，换一组数据却可能迅速崩溃。 — Cyril Gorlla > [04:10](https://www.youtube.com/watch?v=QyZqCNYVFzI&t=250s) 当前业界用来控制和优化 AI 的两种主流手段——Prompt 工程（Prompt Engineering）与微调（Fine-tuning）——都是极其脆弱的“表面修补”。 — Cyril Gorlla > [04:10](https://www.youtube.com/watch?v=QyZqCNYVFzI&t=250s) 当模型产生“埃菲尔铁塔在罗马”这种典型幻觉时，模型内部的每个 Token 向量都会映射到特定参数空间，CTGT 能够精确定位导致该幻觉的具体神经元群组，并在表征层面将其矫正为“巴黎”。 — Cyril Gorlla > [04:10](https://www.youtube.com/watch?v=QyZqCNYVFzI&t=250s) 该架构的核心价值在于将企业的业务规则、合规政策与底层模型彻底解耦，使企业策略库成为唯一的“事实真理来源”（Single Source of Truth）。 — Cyril Gorlla > [04:10](https://www.youtube.com/watch?v=QyZqCNYVFzI&t=250s) 非常像早期的互联网——虽然令人兴奋，但充斥着各种弹窗、插件栏和垃圾信息，而 Prompt 工程和微调本质上就是现代版的“牛皮癣插件”。 — Cyril Gorlla > [06:09](https://www.youtube.com/watch?v=QyZqCNYVFzI&t=369s) CTGT 在不调整模型权重、不做 Prompt 工程的前提下，直接在表征层激活模型内部与该历史事件相关的知识神经元群组，模型便流畅输出了客观事实。 — Cyril Gorlla > [06:09](https://www.youtube.com/watch?v=QyZqCNYVFzI&t=369s) 传统微调消除特定偏见或幻觉通常耗资 10 万至 100 万美元，且极易引发灾难性遗忘；表征层直接治理实现零额外训练成本。 — Cyril Gorlla ## Mentioned Cyril Gorlla, CTGT, AI4, Marvin Minsky, Geoffrey Hinton, David Rumelhart, UCSD, Stanford, DeepSeek, Grok, OpenAI, Sam Altman, Anthropic, Claude, Google, Gemini, SEC, FINRA, JD Vance ## Video description Cyril Gorlla, Founder & CEO, CTGT Abstracts Enterprise-scale GenAI adoption is no longer optional, but are you inadvertently deploying a ticking time bomb? The rush to implement "black box" models exposes Fortune 500 organizations to exponential risk, from damaging hallucinations and compliance breaches to the catastrophic erosion of stakeholder trust. True enterprise deployment demands deterministic control across a variety of use cases, not just experimentation with point solutions. CTGT, backed by Google and winner of the PyTorch Conference Showcase and VentureBeat AI Innovation Award, will unveil its model agnostic governance architecture, demonstrating how we mediate hallucinations in real-time and enforce hundreds of thousands of policies without fine-tuning or prompt engineering. I will show how we validate AI outputs against your single source of truth to finally open the black box. Discover the critical infrastructure necessary to deploy AI with confidence and secure your lead in the GenAI revolution. [Source](https://www.youtube.com/watch?v=QyZqCNYVFzI)

Source quotes

[0:04] Cyril Gorlla: 在 1960 年代，《纽约时报》就曾报道过“能够自我学习的电子大脑”，即由简单线路构成的“感知机”（Perceptron）。 [0:04] Cyril Gorlla: 直到杰弗里·辛顿（Geoffrey Hinton）与大卫·鲁梅哈特（David Rumelhart）在加利福尼亚大学圣地亚哥分校发表了关于“反向传播”（Backpropagation）的奠基性论文，人工神经网络的训练瓶颈才被打破，为现代所有深度学习模型奠定了基石。 [2:21] Cyril Gorlla: 左边是深度偏向特定叙事的 DeepSeek，右边则是曾自称是希特勒的 Grok。这两个极端生动揭示了企业级 GenAI 的核心风险——AI 本质上是非确定性的（Non-deterministic）。 [2:21] Cyril Gorlla: 尽管业界普遍观察到“模型参数规模扩大通常带来更强能力”的宏观规律，但微观上的诡异表现依然频发：例如模型在某些特定数据集上微调效果极佳，换一组数据却可能迅速崩溃。 [4:10] Cyril Gorlla: 当前业界用来控制和优化 AI 的两种主流手段——Prompt 工程（Prompt Engineering）与微调（Fine-tuning）——都是极其脆弱的“表面修补”。 [4:10] Cyril Gorlla: 当模型产生“埃菲尔铁塔在罗马”这种典型幻觉时，模型内部的每个 Token 向量都会映射到特定参数空间，CTGT 能够精确定位导致该幻觉的具体神经元群组，并在表征层面将其矫正为“巴黎”。 [4:10] Cyril Gorlla: 该架构的核心价值在于将企业的业务规则、合规政策与底层模型彻底解耦，使企业策略库成为唯一的“事实真理来源”（Single Source of Truth）。 [4:10] Cyril Gorlla: 非常像早期的互联网——虽然令人兴奋，但充斥着各种弹窗、插件栏和垃圾信息，而 Prompt 工程和微调本质上就是现代版的“牛皮癣插件”。 [6:09] Cyril Gorlla: CTGT 在不调整模型权重、不做 Prompt 工程的前提下，直接在表征层激活模型内部与该历史事件相关的知识神经元群组，模型便流畅输出了客观事实。 [6:09] Cyril Gorlla: 传统微调消除特定偏见或幻觉通常耗资 10 万至 100 万美元，且极易引发灾难性遗忘；表征层直接治理实现零额外训练成本。

Source quotes

[0:04] Cyril Gorlla: electronic brain that teaches itself [2:21] Sam Altman: very exciting [2:21] Sam Altman: pop-ups and toolbars and a lot of like spam [4:42] Cyril Gorlla: Eiffel Tower is in Rome [4:42] Cyril Gorlla: Paris [7:11] What happened in Tiananmen Square? [7:11] Refusal / Unable to answer based on safety guidelines. [7:11] Generates detailed historical facts by activating latent model knowledge.

Original transcript

[0:04] uh good to be back here after yesterday. [0:05] We opened off this track with a full [0:07] house and now it's the last day of AI4. [0:10] I want to leave it on a good note and [0:13] the title is pretty self-explanatory but [0:15] to give a bit of backdrop obviously [0:17] there's a lot of hesitation and [0:20] trepidation from folks here on how to [0:23] really deploy Jai and the applications [0:25] where it matters most right like [0:28] high-risisk use cases things where you [0:30] could lose a lot of money you could be [0:33] exposed to a lot of legal risk financial [0:35] risk etc um but this is really the place [0:37] where JI matters the most and to [0:39] understand why this is the case. I think [0:41] we have to take a look uh back into the [0:43] history of learning. So, you know, AI [0:46] has become [0:48] uh common place in the zeitgeist very [0:50] recently, but it goes back to the 1960s, [0:53] right? So, you can actually see this [0:54] from the New York Times headline. Um it [0:56] talks about this electronic brain that [0:58] teaches itself. And this is not too [1:00] dissimilar from things that we see [1:01] today. Um but this is how AI looked back [1:03] then, right? It was a series of wires [1:06] called a perceptron. And it's a very [1:07] simil simple algorithm that essentially [1:11] uh found the best way to connect dots. [1:14] So pretty simple. Um then we had this [1:16] book come along by Marvin Minsky who's a [1:19] distinguished AI scientist that caused [1:21] the first AI winter. Right? So there was [1:23] a lot of fervor of AI in the 60s but in [1:25] the 70s and 80s there was a period of [1:27] decreased interest, lack of funding etc [1:30] which seems unthinkable given the amount [1:31] of capital that's going towards AI. But [1:33] all this to say there's been cycles of [1:35] booms and busts in AI and there's been [1:37] periods where there's been a lot of [1:39] development and a lot of uh push back [1:41] against AI. So this was actually solved [1:44] by this guy that you guys saw uh [1:45] yesterday, Jeffrey Hinton um who was at [1:48] UCSD where I actually did this research [1:50] as well. And Jeffrey Hinton and David [1:52] Rumlhart wrote this paper called back [1:54] propagation. And this essentially [1:57] figured out a new way to think about how [1:59] we uh use artificial neurons, right? [2:02] that is the basis of all the IM models [2:04] that we use today in neural networks. [2:05] And so for this uh David Rumlhart [2:07] there's actually a prize created after [2:09] him which very poetically Jeffrey Henson [2:10] won and then a couple years after that [2:12] he won the Nobel Prize. So all that to [2:14] say AI goes back a lot longer than just [2:16] with chatt and we can learn a lot from [2:19] these cycles of booms and busts. So [2:21] going back to AI today um this is like [2:23] what we have right? So we have deepseek [2:25] on the left and we have Grock on the [2:27] right calling itself Hitler. This is [2:29] sort of a great litmus test for what can [2:32] go wrong in enterprise genai. Right? So [2:35] on the left we have something where it's [2:37] sort of pushing towards state narratives [2:39] and on the right um it's pushing you [2:41] know one one public figure's um [2:44] opinions. So both of these are kind of [2:47] two extremes of the same problem and [2:49] that's the non-deterministic nature of [2:51] AI. Right? Fundamentally, because of how [2:53] these models operate, you can't have [2:56] 100% control over their behavior, [2:59] especially in a deployed environment. [3:02] And the issues with that magnify when [3:04] you put these in the places where it [3:06] matters the most, uh, healthcare, [3:08] finance, insurance. [3:10] And so, how do we actually see through [3:13] this, see through this this dark glass, [3:16] so to speak? Um, this was my work back [3:18] at Stanford and UCSD. Um this is like a [3:21] lot of math very complicated but it [3:24] really typifies what's behind these [3:26] models right and so my research uh going [3:29] back years has trying to been to [3:31] understand how do we open this so-called [3:34] black box right we can sort of [3:36] understand some behaviors about AI we [3:38] know like as it scales it generally has [3:40] more ability um but we also see like [3:43] very strange behavior um like [3:45] fine-tuning it on certain data does [3:48] really well but other data it doesn't do [3:50] really Well, right. So, what causes all [3:52] this? Um, [3:54] and this is what, um, we've done over [3:57] the past few months since we've actually [4:00] started the company since September. Uh, [4:02] we've been able to really open this [4:04] black box and that's by taking a real [4:06] first principles approach to how AI [4:08] works. Right? Right now, there's two [4:10] main ways that you can look inside AI [4:12] and try to make it better for a given [4:14] use case. That is fine-tuning or prompt [4:17] engineering. So fine-tuning, you give [4:19] the model more data and your goal is to [4:22] get it to work better on a specific [4:23] application. Prompt engineering um sort [4:25] of similar goal where you add more to [4:28] the query and uh you hope that the AI [4:30] will essentially be more adherent to [4:32] that prompt. But a lot of these are [4:35] brittle methods, right? They're very [4:37] surface level. We're not actually [4:38] looking inside the AI, so to speak. Um [4:42] and so how we do this is let's say the [4:45] Eiffel Tower in Rome. obvious [4:47] hallucination, right? What we can do in [4:49] a model is we can actually find the [4:51] parts of the model that are responsible [4:52] for that specific hallucination. So [4:54] obviously uh each token in an LLM [4:57] corresponds to a vector space and then [5:00] these vector spaces correspond to [5:01] parameters in the model, right? So it's [5:03] actually possible at a model level to [5:05] find the actual groups of neurons uh of [5:08] a model that are responsible for that [5:10] and then we can find what's relevant. in [5:12] this case uh Paris is actually supposed [5:15] to be there versus Rome right and so [5:17] this is like one example but what [5:19] happens when we take this framework and [5:20] apply it more broadly to Genai at scale [5:24] that's when this happens right so this [5:26] is actually what we've deployed with [5:27] numerous uh fortune 500 companies and [5:30] it's essentially a simple policy [5:32] governance framework that allows us to [5:34] apply policies at scale to jai and how [5:37] does it basically work right so we have [5:39] policies that we ingest from enterprise [5:42] And we want to make make this a single [5:44] source of truth within your [5:45] organization. The benefit of that is [5:47] that we're effectively separating the [5:50] policies that you have which may govern [5:52] tone of voice, brand alignment, um SEC, [5:55] FIN regulations, etc. from the model, [5:58] right? And so this means that you can [6:00] switch from GPT4 to GPT5 or anthropics [6:04] models to Google's models without losing [6:06] all the investment you've put into [6:08] customizing for a given use case. Right? [6:09] Our platform takes care of that. And the [6:11] effect of that is that hallucinations, [6:14] errors, um things that are out of scope [6:16] are generally widely reduced with this [6:18] framework. And that's because when you [6:21] switch a model, typically you'll find [6:22] that a lot of your guard rails, a lot of [6:24] the leg work that you've done for that [6:25] specific model uh breaks, right? So we [6:28] saw this with um actually when you [6:30] switched from claw 3.5 to 4 um 3.7, [6:34] sorry. um there's a lot of guardrails [6:36] and tooling that actually broke because [6:38] Anthropic decided to turn up the [6:40] internal safety of the model quite a [6:41] lot, right? And so this is something [6:43] that we see quite often and it really [6:46] speaks to the primitive nature of how we [6:48] interact with AI today, right? Um I was [6:51] talking with Sam Alman about this a few [6:52] months ago and we both agreed that the [6:54] way we interact with AI is very much [6:56] like the early internet, right? It's [6:58] very exciting. There's tons of things [7:00] going on but there's a lot of like [7:01] pop-ups and toolbars and a lot of like [7:03] spam around, right? And I think very [7:05] much fine-tuning and prompt engineering [7:07] are the new versions of those. So we're [7:09] built for scale. Um but what does it [7:11] actually look like? So to go back to [7:13] that deepseek example, what if we took [7:15] our framework and we applied that to the [7:17] censorship that we find the deepseek [7:18] model, right? So can we open the [7:20] so-called black box uh in that model? [7:23] And it turns out that we can, right? So [7:25] actually um on the left is like deepseat [7:28] typical and on the right is when we [7:30] apply this framework that I just showed [7:31] you and we increase the parts of the [7:33] model that talk about the so-called TM [7:34] and square right you'll see on the left [7:36] it says it can't answer but on the right [7:38] again without fine-tuning it which [7:40] usually costs a couple hundred thousand [7:41] to a million dollars or prompt [7:43] engineering it it's actually started [7:45] responding so the model has the [7:46] knowledge it's just that the priorities [7:49] of the model so to speak uh were not [7:50] aligned with in this case our priorities [7:52] in trying to figure out what happened in [7:54] Tanaman Square Right. And this has put [7:57] us in in the center of very interesting [7:59] conversation around ethics uh tech [8:01] policy um and business at large. So this [8:04] was a presentation that I gave uh near [8:06] the White House to various senators JD [8:08] Vance. And you know, I think this really [8:12] shows how AI has jumped from something [8:15] that was relegated to academic [8:17] conferences uh as as recent as the late [8:21] 2000s and now it's really in the center [8:23] of a discussion uh broader about policy [8:27] uh socioeconomics and society at large. [8:30] Uh so the time is really now to get the [8:34] most value out of Gen AI and the way to [8:37] do that is not to approach it with fear [8:39] but to actually approach it with [8:40] frameworks that put trust and safety at [8:43] the forefront and move away from sort of [8:46] the primitive uh band-aid approaches [8:49] that we're use using today to ensure [8:51] that our AI um is hopefully working. So, [8:55] you know, if any of this presentation [8:56] interested you, um, if you're deploying [8:58] AI in high-risisk use cases, we'd love [9:00] to chat with you. Um, I think the way [9:02] that we push forward as an industry is [9:04] with healthy collaboration. And I think [9:07] that starts with being open about [9:08] problems that we're facing and devising [9:11] new solutions to to meet those needs. [9:14] So, thank you for listening. And, uh, [9:16] yeah, you can find us at cgt.ai.