Airbnb, DoorDash Deploy Synthetic Data for Cold Starts, Safety Gaps
Ai4
Synthetic data provides vital scale for privacy-restricted workflows and cold starts, but enterprises cannot eliminate real-world human data when modeling complex user behavior and adversarial abuse.
## Background
Published by Ai4 on YouTube on August 21, 2026, this panel session explores whether and how synthetic data can augment or outperform real-world datasets across industries like finance, healthcare, and autonomous systems. Moderated by Jonathan Shieber of Resilience Media, the discussion features machine learning practitioners and executives from Airbnb, DoorDash, Tonic.ai, and Wirestock detailing performance tradeoffs, regulatory compliance, and practical production pipelines.
## TL;DR
The panel dismantles the notion that synthetic data serves as a total, binary replacement for real data, arguing instead that its true power lies in augmenting edge cases, preserving privacy under strict regulations, and refining noisy pre-training corpuses. Industry leaders from DoorDash and Airbnb illustrate practical implementations in cold-start recommendations, trust-and-safety reviewer training, and abuse detection, while highlighting critical boundaries where human data remains irreplaceable. External vendor leaders from Tonic.ai and Wirestock outline methods for evaluating realism, such as self-play and synthetic benchmark generation, while warning against distribution drift and privacy leakage. Ultimately, the panelists emphasize a hybrid, human-in-the-loop approach where synthetic abstraction accelerates scaling while expert human input grounds evaluation and novel frontier learning.
## Atmosphere & style
The session opens on a relaxed, playful note with a Las Vegas card game icebreaker before quickly settling into a grounded, pragmatic technical deep dive. The dynamic between enterprise platform practitioners and synthetic data vendors is candid and collegial, stripping away generative AI hype in favor of hard engineering trade-offs like inference unit economics and multi-sided auction dynamics. Listening to the panel feels like sitting in on an honest, highly informed practitioner roundtable that balances optimistic technical unlocks with the messy realities of production systems.
## Key takeaways
- [02:09](https://www.youtube.com/watch?v=s6lAx_yiwII&t=129s) Synthetic data exists along a broad spectrum, ranging from purely model-generated outputs to AI-assisted data augmentation and synthetic annotation pipelines.
- [04:10](https://www.youtube.com/watch?v=s6lAx_yiwII&t=250s) Demonstrating the value of synthetic datasets requires rigorous evaluation frameworks, including LLM judges, held-out real-world validation splits, and tracking progress on public benchmarks.
- [05:47](https://www.youtube.com/watch?v=s6lAx_yiwII&t=347s) Synthetic data faces structural limits at the frontier of human creativity, including evolving adversarial abuse patterns and emergent second-order dynamics in recommendation engines.
- [10:07](https://www.youtube.com/watch?v=s6lAx_yiwII&t=607s) Generating rare safety edge cases safely requires extracting abstract behavioral attributes rather than directly cloning sensitive user data, preventing privacy leaks.
- [13:02](https://www.youtube.com/watch?v=s6lAx_yiwII&t=782s) The build-versus-buy decision hinges on IP protection and economic structure: per-query vendor inference adds recurring marginal costs, whereas purchasing training data is a sporadic capital expense.
- [18:53](https://www.youtube.com/watch?v=s6lAx_yiwII&t=1133s) In heavily regulated sectors under frameworks like GDPR and HIPAA, synthetic data's value lies in providing legally compliant, viable approximations when real data is inaccessible.
- [19:56](https://www.youtube.com/watch?v=s6lAx_yiwII&t=1196s) Synthetic data provably outperforms real data in closed-loop reinforcement learning self-play and in curating noisy web scrapes into textbook-quality pre-training corpora.
- [21:21](https://www.youtube.com/watch?v=s6lAx_yiwII&t=1281s) Enterprises leverage synthetic data to bridge cold-start problems in high-risk product launches and to train human Trust & Safety reviewers on rare violation policies.
- [23:51](https://www.youtube.com/watch?v=s6lAx_yiwII&t=1431s) As foundational AI models mature, basic synthetic assets commoditize while the demand for nuanced, complex human expert data for RL and fine-tuning increases.
- [30:17](https://www.youtube.com/watch?v=s6lAx_yiwII&t=1817s) Synthetic evaluation benchmarks like LegalAgentBench allow direct embedding of programmatic ground truths, though expert human oversight remains vital at validation boundaries.
- [33:43](https://www.youtube.com/watch?v=s6lAx_yiwII&t=2023s) Vector similarity search serves as a mathematical guardrail for synthetic fraud datasets, balancing between memorizing private records (too close) and generating ungrounded hallucinations (too divergent).
- [33:43](https://www.youtube.com/watch?v=s6lAx_yiwII&t=2023s) Framing synthetic data as a binary replacement for real data is a false dichotomy; the real engineering task is identifying specific pipeline stages where synthetic generation optimizes scalability and privacy.
## Detailed walkthrough
### [00:04](https://www.youtube.com/watch?v=s6lAx_yiwII&t=4s) Introductions and Favorite Card Games
The panel opens with moderator Jonathan Shieber welcoming the audience to the pre-lunch session on a foundational question facing the AI industry: whether synthetic data can truly replace real data. Setting a relaxed tone for the Las Vegas venue, he invites each panelist to introduce themselves and name their favorite card game. Karl Hanson, co-founder of Tonic.ai, describes his company's mission of transforming sensitive data into safe resources for building and training software and AI agents, naming blackjack as his game of choice. Prashanth Srinivasan, who works on DoorDash's search, advertising, and retrieval systems, selects hearts along with some poker. Aniket Tushar Ajagaonkar, a software engineer working on trust and safety at Airbnb, admits he prefers board games over card games and names *Exploding Kittens*. Mikayel Khachatryan, co-founder and CEO of Wirestock, introduces his company as a multimodal data provider focused on human-created visual content for training frontier models, choosing poker. Jonathan adds that his personal favorite card game is *Skip-Bo*.
- **Karl Hanson** – Co-Founder & CPO at Tonic.ai; favorite game: Blackjack.
- **Prashanth Srinivasan** – Senior ML Engineer at DoorDash (Search & Ads, retrieval models); favorite games: Hearts and Poker.
- **Aniket Tushar Ajagaonkar** – Staff Software Engineer in Trust & Safety at Airbnb; favorite game: *Exploding Kittens*.
- **Mikayel Khachatryan** – Co-Founder & CEO at Wirestock; favorite game: Poker.
- **Jonathan Shieber** – Journalist at Resilience Media (Moderator); favorite game: *Skip-Bo*.
---
[02:09](https://www.youtube.com/watch?v=s6lAx_yiwII&t=129s) Defining Synthetic Data
Before addressing whether synthetic data can replace real data, Jonathan asks the panelists to establish what they mean by the term. Mikayel explains that synthetic data spans a broad spectrum: direct model outputs (AI-generated content), AI-assisted editing and enhancement of human-created data, and synthetic curation or annotation pipelines. Karl adds a mental model where generation involves two adjustable pedals: the model itself and the input data fed into it. One can generate purely model-driven data with zero source inputs, or feed real-world baseline data into a model to produce structured synthetic variants.
```mermaid
flowchart TD
accTitle: Dimensions of Synthetic Data Generation
accDescription: A diagram illustrating the inputs, models, and forms of synthetic data outputs.
subgraph Inputs["Data Inputs"]
R["Real World Data"]
E["Empty / Prompt Only"]
end
subgraph Engine["Generative Engine"]
M["Generative Model / LLM"]
end
subgraph Outputs["Synthetic Data Forms"]
O1["Direct AI Output"]
O2["Augmentation & Enhancement"]
O3["Synthetic Labels / Annotation"]
end
R --> M
E --> M
M --> O1
M --> O2
M --> O3
```
---
[04:10](https://www.youtube.com/watch?v=s6lAx_yiwII&t=250s) Proving Synthetic Data Realism Through Evals
Jonathan asks Karl how synthetic data providers prove to enterprise customers that generated datasets offer genuine underlying insights rather than merely plausible distributions or surface-level facsimiles. Karl stresses that everything starts with a rigorous evaluation framework. Teams can design specific test suites using automated LLM judges or human evaluation rubrics, train models on synthetic sets and evaluate them against held-out samples of real data, or track model performance against standard external benchmarks. If post-training or fine-tuning with synthetic data reliably hill-climbs on public benchmarks, vendors have a grounded case that their data produces real utility.
- Establish structured task suites evaluated by human reviewers or LLM judges.
- Benchmark synthetic-trained models on small, gold-standard real-world validation splits.
- Track score improvements on standardized public benchmarks to confirm effective hill climbing.
---
[05:47](https://www.youtube.com/watch?v=s6lAx_yiwII&t=347s) Frontier Models, Human Creativity, and Second-Order Simulation Traps
Jonathan turns to Mikayel to ask whether real data teaches models things that synthetic data fundamentally cannot. Mikayel explains that training data represents a vehicle for transferring real-world capabilities and knowledge into a model. While synthetic data works well across the downstream application layer to help existing models catch up, it hits steep walls at the frontier of creative, scientific, and programming knowledge. In foundation modalities like video, synthetic data struggles severely with real-world physics, lighting transitions, and temporal causality.
Aniket and Prashanth reinforce the structural limits of synthetic data from the perspective of enterprise production platforms:
- **The Human Edge in Platform Abuse**: In trust and safety, Aniket notes that synthetic data is bounded by the engineer's imagination, whereas real bad actors show boundless ingenuity: "Humans are really creative in how to abuse policies and platforms."
- **Second-Order Recommendation Dynamics**: In recommendation systems, Prashanth points out that simulating end-to-end user journeys misses critical emergent dynamics. If an objective function optimizes for raw click likelihood on a user who likes Red Bull videos, a simulator will simply flood them with pure Red Bull content. In reality, users require content diversity (a few Red Bull videos, then something else, then a return) to avoid burnout—a behavior discovered only when real users interact with the live product.
---
[10:07](https://www.youtube.com/watch?v=s6lAx_yiwII&t=607s) Rare Edge Cases and Privacy-Preserving Abstraction
Jonathan asks how much synthetic data Airbnb and DoorDash deploy in production. Aniket explains that the ratio shifts depending on the frequency of the harm vector: common violations (such as spam, basic nudity, or straightforward misinformation) have ample real training data, whereas rare, high-severity trust and safety incidents rely much more heavily on synthetic augmentation.
To generate rare safety scenarios without overfitting or leaking sensitive information, Aniket highlights an abstraction-based pipeline rather than direct sample cloning:
```mermaid
flowchart LR
accTitle: Privacy Preserving Synthetic Data Generation Flow
accDescription: Workflow showing how real abuse patterns are abstracted into behavioral attributes before LLM synthesis.
subgraph Extraction["Pattern Extraction"]
A["Real User Interactions"] --> B["Extract Behavioral Attributes"]
end
subgraph Generation["Synthesis Layer"]
B --> C["LLM Prompting with Attributes"]
C --> D["Generalizable Synthetic Harm Data"]
end
subgraph Safety["Guarantees"]
D --> E["Protects Private Messages"]
D --> F["Extensible to New Vectors"]
end
```
- Direct copying ("here are 10 samples, generate 1,000 like them") creates narrow blind spots and threatens privacy on sensitive streams like private user messaging.
- Extracting behavioral attributes into an abstraction layer allows generative models to explore novel variations while preventing adversarial reverse engineering of private user messages.
---
[13:02](https://www.youtube.com/watch?v=s6lAx_yiwII&t=782s) The Build vs. Buy Equation for AI Infrastructure
Jonathan shifts the conversation to the calculus of building synthetic pipelines in-house versus purchasing external vendor tools like Tonic.ai or Wirestock. Aniket notes that large tech firms naturally lean toward building internally to guard proprietary IP, manage user privacy risks, and avoid passing sensitive internal data outside company firewalls.
Prashanth breaks down the economic and operational dimensions of build-versus-buy:
- **Inference vs. Training Economics**: Buying a vendor-hosted solution for live search inference incurs a metered marginal cost on every single user interaction, drastically impacting unit economics. Buying training data or fine-tuning infrastructure, by contrast, is a sporadic capital expense.
- **Team Specialization & Maintenance**: Sustaining bespoke in-house tools demands ongoing hiring, motivation, and engineering overhead. Unless the capability forms a core competitive differentiator, maintaining internal tooling can become a long-term resource sink.
Addressing why enterprise clients choose external vendors, Karl argues that vendors must demonstrate they are "faster, better, and cheaper" through published artifacts and proof-of-concept tests (POCs) that overcome simple DIY scripts written with tools like Codex or Claude. Mikayel explains that clients come to Wirestock because certain modalities—notably raw, authentic human video—cannot be synthetically faked without corrupting foundational model quality.
---
[18:53](https://www.youtube.com/watch?v=s6lAx_yiwII&t=1133s) Reframing "Replacement": Compliance, Privacy, and Viability
Jonathan introduces the nuanced question of what it actually means to "replace" real data: must synthetic data perform strictly better than real-world samples, or is it sufficient to provide a viable path forward when legal or cost constraints block access to real records?
Karl points out that in highly regulated domains governed by frameworks like GDPR or HIPAA, raw real data is legally off-limits for model development. In these environments, the objective is not to surpass perfect empirical reality, but to construct de-identified or synthetic approximations faithful enough to solve the target business task before deploying the resulting models safely to production.
[19:56](https://www.youtube.com/watch?v=s6lAx_yiwII&t=1196s) Where Synthetic Data Outperforms Real Data: Self-Play and Pre-Training Pipelines
Karl Hanson explains that while the utility of synthetic data depends heavily on the specific use case, there are proven scenarios where synthetic data actually exceeds real-world human data. A foundational historical benchmark is reinforcement learning through self-play, exemplified by AlphaGo. By playing countless games against itself, AlphaGo generated high-volume synthetic game trajectories as a byproduct, resulting in a system far superior to any model trained purely on human matches.
Beyond self-play, synthetic generation has become essential across pre-training, mid-training, and post-training pipelines at major AI research labs. While private foundation model labs keep proprietary techniques close to the chest, published research demonstrates synthetic data's efficacy in data curation. In pre-training, raw web scrape data from sources like Common Crawl is inherently messy and unstructured. Karl highlights Microsoft's work on the *Phi* model series (the "Textbooks Are All You Need" paper), which demonstrated that using language models to filter, clean, and synthesize noisy web data into coherent, textbook-quality synthetic data yields significantly stronger pre-training performance.
```mermaid
flowchart TD
accTitle: Synthetic Data Pipeline in Pre-Training
subgraph Ingestion["Raw Data Sourcing"]
A["Common Crawl Web Scrapes"] --> B["Noisy / Unstructured Text"]
end
subgraph Synthesis["Synthetic Refinement"]
B --> C["LLM Data Synthesizer"]
C --> D["Textbook-Quality Synthetic Corpus"]
end
subgraph Training["Model Training Pipeline"]
D --> E["Pre-Training & Mid-Training"]
E --> F["High-Performing Foundation Model"]
end
```
[21:21](https://www.youtube.com/watch?v=s6lAx_yiwII&t=1281s) Practical Enterprise Applications: Cold Starts, "One-Way Doors," and Trust & Safety Training
When asked where enterprises see the highest return on investment for synthetic data, Prashanth Srinivasan points to high-stakes product decisions that function as "one-way doors." Launching a major public feature carries cross-functional risk because teams cannot always predict user reception, yet models for retrieval and ranking must perform reliably on day one. Live real-world experiments cannot always be run due to adverse side effects on user experience. In these cold-start scenarios, DoorDash engineers extrapolate existing traffic patterns onto new user surfaces, working backward from an ideal product vision to generate synthetic training sets. Furthermore, when internal LLM usage logs are too sparse or restricted to analyze, synthetic data allows engineers to simulate model capabilities and stress-test failure modes before deployment.
At Airbnb, Aniket Tushar Ajagaonkar outlines two distinct operational use cases: generating machine learning training data to detect platform harm, and training human reviewers in Trust & Safety. When listings or users are removed from Airbnb or when users appeal enforcement actions, human agents make the final determinations based on internal platform policies. Because safety policies constantly evolve and reviewer turnover requires regular onboarding, real-world examples of rare violations are often insufficient. Airbnb generates realistic synthetic violation scenarios to train human reviewers continuously, keeping accuracy high and preventing erroneous user bans.
[23:51](https://www.youtube.com/watch?v=s6lAx_yiwII&t=1431s) Model Maturation, Complex Human Data, and Diminishing Returns
Addressing whether foundation models will eventually absorb enough real data to render real-world collection obsolete, Mikayel Khachatryan describes an ongoing cycle rather than a terminal end state. As generative models mature, demand for simple visual assets (such as generic stock photography) declines, but the requirement for intricate, highly nuanced human input surges. Wirestock observes this across pre-training, mid-training, supervised fine-tuning (SFT), and reinforcement learning (RL) workflows. The modern pipeline operates as a human-AI symbiosis: synthetic data excels at automated post-training evaluation, benchmark auto-rating, and annotation assistance, but models continuously require expert human knowledge to master complex edge cases.
Jonathan Shieber raises the question of whether real-world applications hit diminishing returns where models must return to the "well of real data" to progress:
- **Airbnb's Trust & Safety:** Aniket notes that while platform safety constraints prevent disclosing operational specifics, the team is still early in deployment and has not yet hit a plateau of diminishing returns with synthetic data.
- **DoorDash Marketplace & Search:** Prashanth emphasizes that "the rubber has to meet the road"—if synthetic simulations fail to move business metrics, real-world grounding is mandatory.
- **Marketplace Dynamics:** While user taste evolution is solvable and widely modeled in industry literature (such as research published by Pinterest), multi-sided marketplace dynamics remain notoriously hard to simulate. In ad auctions where multiple advertisers compete across dynamic daily budgets, tweaking one parameter creates adverse selection and unpredictable cascading shifts across the entire system.
[30:17](https://www.youtube.com/watch?v=s6lAx_yiwII&t=1817s) Audience Q&A: Generating Evals, LegalAgentBench, and Human Grounding in RL
An audience member asks about using AI models to translate human intent directly into unstructured synthetic evaluation data with embedded ground truths, effectively bypassing manual human annotation loops.
Karl Hanson confirms that this is a major unlock for building evaluation benchmarks and reinforcement learning environments. By generating synthetic data backward from known programmatic ground truths and introducing controlled noise, developers can embed labels directly into the data without post-hoc manual labeling. Karl cites Harvey's release of *LegalAgentBench* as a prime example: because legal AI companies cannot use confidential client data to build evaluations, Harvey synthesized realistic legal benchmarks entirely from domain principles and embedded verification labels directly into the generation pipeline. However, Karl warns that a human in the loop is still necessary at validation boundaries to ensure unintended artifacts or distribution shifts were not inadvertently injected during synthesis.
Mikayel Khachatryan reinforces that while automated pipelines create the evaluation framework, domain-expert human input remains the single most valuable and expensive element in reinforcement learning. Real-world human knowledge is essential for establishing "golden outputs," designing evaluation rubrics, and executing quality control reviews across complex operational tasks.
[33:43](https://www.youtube.com/watch?v=s6lAx_yiwII&t=2023s) Guardrails Against Bias, Drift, and Final Conclusions
An audience member working in fintech asks how to establish guardrails when training fraud detection models on synthetic data. Specifically, if real-world fraud examples are sparse, synthetic data risks silently skewing toward developer assumptions—creating an illusion of high precision and declining fraud rates while missing novel fraud patterns.
Aniket explains that vector similarity search provides a vital mathematical guardrail to ensure synthetic datasets stay within a realistic distribution:
| Synthetic Data Proximity | Associated Risk | Actionable Mitigation |
|:--- |:--- |:--- |
| **Identical to Real Data** | Leaking private, sensitive, or personally identifiable information (PII). | Add noise, differential privacy constraints, or re-synthesize. |
| **Completely Divergent** | Hallucinatory, artificial, and detached from real-world probability. | Discard as invalid distribution drift. |
| **The Middle Segment (Target)** | Captures realistic structural behavior without exposing real private records. | Deploy LLM-as-a-judge and human-in-the-loop sampling for verification. |
Jonathan Shieber concludes the panel by summarizing the core takeaway: framing synthetic data as a binary replacement for real data is a false dichotomy. The real engineering question is determining the specific niches, pipelines, and use cases where synthetic generation provides the optimal trade-off in scalability, privacy, and performance.
## Notable quotes
> [05:47](https://www.youtube.com/watch?v=s6lAx_yiwII&t=347s) Humans are really creative in how to abuse policies and platforms. — Aniket Tushar Ajagaonkar
> [10:07](https://www.youtube.com/watch?v=s6lAx_yiwII&t=607s) here are 10 samples, generate 1,000 like them — Aniket Tushar Ajagaonkar
> [13:02](https://www.youtube.com/watch?v=s6lAx_yiwII&t=782s) faster, better, and cheaper — Karl Hanson
> [19:56](https://www.youtube.com/watch?v=s6lAx_yiwII&t=1196s) Textbooks Are All You Need — Karl Hanson
## Video description
Aniket Tushar Ajagaonkar - Staff Software Engineer - Trust, Airbnb Inc
Prashanth Srinivasan - Senior ML Engineer, DoorDash
Karl Hanson - Co-founder and CPO, Tonic.ai
Mikayel Khachatryan - Co-Founder & CEO, Wirestock
Jonathan Shieber - Journalist, Resilience Media
This panel examines where and how synthetic data can outperform or augment real-world datasets. Panelists will share case studies from industries like finance, healthcare, and autonomous systems where synthetic generation has filled critical data gaps. Expect an honest discussion about performance tradeoffs, regulatory scrutiny, and what synthetic data pipelines actually look like in production.
Industry practitioners argue that synthetic data should not be viewed as a wholesale replacement for real-world data, but rather as an operational tool optimized for privacy compliance, rare edge cases, and pre-training refinement where empirical data is inaccessible or noisy.
Proving Utility in Training and Benchmark Evaluations (4:10)
Synthetic data spans direct model generations, data augmentations, and automated annotation pipelines. Karl Hanson notes that synthetic generation can outperform real data in specialized environments, citing reinforcement learning through self-play in AlphaGo and Microsoft’s Phi model series, which demonstrated that using language models to synthesize messy web scrapes into textbook-quality corpora improves pre-training performance.
To demonstrate that synthetic datasets provide authentic utility rather than superficial plausibility, vendors rely on structured evaluation frameworks. These include automated LLM judges, human evaluation rubrics, and benchmarking models fine-tuned on synthetic data against gold-standard real-world validation splits and public performance baselines.
Mitigating Privacy Constraints and Cold-Start Challenges (10:07)
In highly regulated environments governed by GDPR or HIPAA, raw real data is legally restricted, making synthetic approximations essential for building and testing models prior to production. Synthetic pipelines also address enterprise scenarios where empirical data is structurally limited. At Airbnb, Aniket Tushar Ajagaonkar explains that while common platform abuses have ample real data, rare, high-severity trust and safety violations rely heavily on synthetic augmentation. To avoid leaking private user communications or creating narrow blind spots, Airbnb abstracts real behavioral attributes into an intermediate layer before prompting generative models.
Prashanth Srinivasan highlights synthetic data's utility in high-stakes "one-way door" product launches at DoorDash. When launching retrieval and ranking features where live experiments could harm user experience, engineers extrapolate existing traffic patterns to synthesize cold-start training sets. In addition to training algorithms, Airbnb deploys synthetic violation scenarios to train and onboard human trust-and-safety reviewers against evolving platform policies.
Limits in Frontier Knowledge and Dynamic Marketplaces (5:47)
Despite downstream successes, synthetic generation encounters structural limits at the frontier of domain knowledge. Mikayel Khachatryan points out that synthetic data struggles with complex human creativity, advanced programming, and foundational physical realities—such as lighting transitions and temporal causality in video generation.
Enterprise platforms face further barriers in adversarial and systemic simulation. Ajagaonkar notes that synthetic harm data remains bounded by engineers' imaginations, whereas real platform abusers exhibit continuous, unpredictable ingenuity. In recommendation systems, Srinivasan explains that synthetic simulations struggle with emergent user dynamics; an objective function optimizing purely for engagement will over-index on single topics without accounting for the content diversity real users need to prevent burnout. Similarly, multi-sided marketplace dynamics, such as ad auctions with competing dynamic budgets, introduce adverse selection and cascading effects that resist realistic synthetic simulation.
Guardrails Against Drift and Infrastructure Economics (13:02)
When generating synthetic data for sparse events like fraud, models risk drifting toward developer assumptions. Ajagaonkar outlines vector similarity search as a mathematical guardrail: synthetic data that is identical to real records risks leaking private information, while completely divergent data constitutes invalid hallucination, requiring teams to target a middle segment verified by LLM judges and human sampling. Karl Hanson notes benchmarks like Harvey’s LegalAgentBench, which synthesized legal evaluation tasks with embedded programmatic ground truths to bypass client confidentiality constraints.
On infrastructure strategy, Srinivasan points out that build-versus-buy decisions depend on workload profiles: external vendor services for live search inference introduce metered marginal costs on every user interaction, whereas purchasing synthetic training data constitutes a periodic capital expense. Hanson and Khachatryan add that external vendors must continually prove superior domain specialization, as certain complex modalities, like authentic human video, cannot be synthesized without degrading foundation model quality.
The panel concludes that viewing synthetic data as a binary replacement for real data is a false dichotomy. Instead, modern machine learning relies on a symbiotic loop where synthetic data scales evaluation, privacy preservation, and cold starts, while expert human data remains indispensable for establishing foundational truths, evaluation rubrics, and frontier capabilities.
## Background
本内容由 Ai4 于 2026 年 8 月 21 日发布在 YouTube 平台上。视频记录了一场由行业专家参与的圆桌讨论，探讨在金融、医疗和自动驾驶等领域中合成数据如何填补关键数据空白、何时能够超越或增强真实数据，以及合成数据流水线在生产环境中的性能权衡与监管考量。
## TL;DR
来自 Tonic.ai、DoorDash、Airbnb 和 Wirestock 的四位行业专家围绕“合成数据能否取代真实数据”展开深入探讨。嘉宾们指出，合成数据与真实数据并非非黑即白的二元对立，而是呈现共生互补关系：真实数据在前沿大模型能力突破、复杂对抗性安全风控及多智能体推荐系统中具有不可替代的锚定价值；而合成数据则在冷启动预判、隐私合规受限场景、罕见长尾危害扩充以及自动化基准评测（Evals）中展现出巨大优势。团队在落地中需结合抽象特征提炼、相似度校验与人工在环机制，以平衡泛化能力、偏差防范与隐私安全。
## Atmosphere & style
整场圆桌讨论务实而严谨，开场以拉斯维加斯桌面游戏的轻松破冰迅速拉近距离，随后迅速切入高信息密度的工业界硬核实践辩论。主持人的提问敏锐直接，嘉宾们摒弃了盲目的技术炒作，在分享各自工程落地经验时既坦诚指出合成数据的能力边界与模拟困境，又生动展现了各自在解决冷启动与合规难题时的技术洞察，呈现出理性、专业且富有启发性的行业前沿对话氛围。
## Key takeaways
- [02:09](https://www.youtube.com/watch?v=s6lAx_yiwII&t=129s) 合成数据并非单纯指端到端生成的全假数据，而是模型与输入数据协同作用的产物，形态涵盖纯模型输出、对真实数据的编辑扩增以及 AI 自动标注与策展。
- [04:16](https://www.youtube.com/watch?v=s6lAx_yiwII&t=256s) 建立客户对合成数据信任的核心是搭建严谨的客观评测体系（Evals），通过在合成数据上训练模型、但在真实留存集上测试，或在公开基准上持续打榜来验证其泛化效果。
- [05:47](https://www.youtube.com/watch?v=s6lAx_yiwII&t=347s) 真实数据在黑产对抗和前沿认知上具有不可替代性：人类在规则钻空子方面的创新性远超工程师的预想想象力，且前沿大模型仍需靠高质量真实数据来突破能力天花板并避免偏见放大。
- [05:47](https://www.youtube.com/watch?v=s6lAx_yiwII&t=347s) 在复杂双边市场与推荐搜索系统中，用户长周期注意力衰减与多智能体动态竞价会产生复杂的“二阶效应”，端到端模拟极难逼真还原，离线合成数据必须依靠真实物理世界数据进行最终锚定。
- [10:07](https://www.youtube.com/watch?v=s6lAx_yiwII&t=607s) 生产环境中使用合成数据的比例取决于危害发生频率：高频违规以真实数据为主，而罕见边缘危害严重依赖合成数据；生成时应抽象底层行为属性而非简单粗暴地克隆样本。
- [12:04](https://www.youtube.com/watch?v=s6lAx_yiwII&t=724s) 合成数据提取高维属性可以在赋予泛化能力的同时保护隐私，但在实践中并不存在标准化的平衡点，工程团队必须反复通过逆向工程测试来验证泄露风险。
- [13:02](https://www.youtube.com/watch?v=s6lAx_yiwII&t=782s) 企业在自研与采购合成数据方案时，需综合权衡敏感数据外流的合规风险、长期推理调用对业务单位经济模型（unit economics）的冲击，以及团队自研的组织维护成本。
- [15:53](https://www.youtube.com/watch?v=s6lAx_yiwII&t=953s) 在时间关联性、光影连续变化与物理因果逻辑等前沿视频多模态领域，合成数据尚无法满足训练标准，高质量真实人类创作者数据仍是绝对核心资产。
- [18:53](https://www.youtube.com/watch?v=s6lAx_yiwII&t=1133s) 合成数据不仅能用于模型训练，还能在法律合规受限场景下打破准入壁垒、利用自我博弈（如 AlphaGo）超越人类上限，以及通过提纯重写（如微软 Phi 系列）大幅升维预训练语料。
- [21:21](https://www.youtube.com/watch?v=s6lAx_yiwII&t=1281s) 合成数据在产品冷启动与人工审核培训中具备高投资回报率：在缺乏线上交互日志时可外推流量训练检索排序模型，亦可在不暴露真实隐私的情况下动态培训信任与安全（Trust & Safety）审核人员。
- [30:18](https://www.youtube.com/watch?v=s6lAx_yiwII&t=1818s) 自动化模型评测（Evals）利用合成数据可直接内嵌“事实真值”（Ground Truth）以降低人工标注成本，但仍需在最终阶段设置人工审核（Human-in-the-loop），防止静默偏差与过度合成。
- [34:51](https://www.youtube.com/watch?v=s6lAx_yiwII&t=2091s) 防止反欺诈等场景中合成数据产生静默偏差的核心手段是相似度搜索校验：数据必须处于“避免隐私泄露”与“避免脱离现实”的黄金平衡区间，并辅以人工抽检与 LLM as a Judge 进行验证。
- [36:02](https://www.youtube.com/watch?v=s6lAx_yiwII&t=2162s) 合成数据能否替代真实数据不是非黑即白的二元对立，核心在于识别在哪些具体业务场景下，采用合成数据替代或增强真实数据才具备最高的性价比与执行效率。
## Detailed walkthrough
### [00:04](https://www.youtube.com/watch?v=s6lAx_yiwII&t=4s) 暖场介绍与破冰：拉斯维加斯的最爱纸牌游戏
作为午餐前的最后一场圆桌讨论，主持人 Jonathan Shieber（来自 Resilience Media）在开场时感谢了观众在这一最具挑战的时段坚持出席。他指出，“合成数据能否取代真实数据”不仅是一个引人入胜的技术议题，更是当前整个人工智能产业中最具基础性、最核心的焦点之一。
在正式切入硬核技术辩论之前，考虑到会议地点位于“赌城”拉斯维加斯，Jonathan 邀请四位嘉宾进行简短的自我介绍，并顺便分享自己最喜欢的纸牌或桌面游戏：
- **Karl Hanson**：Tonic.ai 联合创始人兼首席产品官（CPO）。Tonic.ai 专注于对敏感数据进行脱敏与合成处理，使其可安全用于构建软件、训练模型以及打造 AI Agent。他的最爱是经典的二十一点（Blackjack）。
- **Prashanth Srinivasan**：DoorDash 高级机器学习工程师，长期深耕搜索与广告系统中的检索算法及机器学习模型架构。他喜欢扑克（Poker），同时也非常热衷于红心大战（Hearts）。
- **Aniket Tushar Ajagaonkar**：Airbnb 信任与安全（Trust & Safety）团队的资深软件工程师。他直言自己不常玩传统扑克牌，但极度喜欢桌游，尤其是《爆炸猫》（Exploding Kittens）。
- **Mikayel Khachatryan**：Wirestock 联合创始人兼 CEO。Wirestock 是一家多模态数据平台，专注于汇聚人类创作者生成的高质量视觉数据，以支持前沿多模态大模型的训练。他的选择同样是扑克（Poker）。
- **Jonathan Shieber（主持人）**：则分享了他个人极力推荐的一款经典数字牌类游戏《Skip-Bo》。
---
### [02:09](https://www.youtube.com/watch?v=s6lAx_yiwII&t=129s) 什么是合成数据？行业专家的多维定义
在探讨合成数据是否会彻底取代真实数据之前，主持人首先抛出了核心前置问题：在工业界和学术界语境下，当大家谈论“合成数据”时，究竟在指代什么？
Mikayel 指出，合成数据绝不能被单一化理解，它涵盖了多种层次和应用形态：
1. **纯模型生成内容**：大模型端到端直接输出的 AI 生成内容（AIGC）或全新数据点。
2. **AI 编辑与增强数据**：利用 AI 模型对现实世界采集到的、人类创造的原始真实数据进行修改、重采样或画质/信息增强。
3. **标注与数据策展**：借助模型自动完成对原始数据的打标、分类、过滤与结构化组织。
```mermaid
flowchart TD
accTitle: Synthetic Data Conceptual Model
accDescr: A dual-input framework showing how models and data interact to generate different types of synthetic data.
subgraph Inputs["输入维度 (调控双翼)"]
M["AI / 生成模型"]
D["输入数据 (真实种子数据 / 无输入)"]
end
subgraph Generation["生成与处理管道"]
M --> Proc["合成生成引擎"]
D --> Proc
end
subgraph Outputs["合成数据产物形态"]
Proc --> O1["纯模型端到端输出"]
Proc --> O2["真实数据增强 / 编辑"]
Proc --> O3["AI 自动标注与数据策展"]
end
```
Karl 对此表示完全赞同，并提出了一个简洁的“双踏板”心智模型：合成数据的本质是“模型”与“输入数据”协同作用的产物。在实践中，工程师可以同时调节这两个控制维度——有时是以大量真实数据为种子（Seed）通过特定模型进行扩增；有时则是输入为零，完全依托大模型的先验知识自主生成。不同的生成路径决定了合成数据在下游任务中的适用场景与优劣势。
---
### [04:16](https://www.youtube.com/watch?v=s6lAx_yiwII&t=256s) 合成数据如何自证其真：评估体系与真实洞察的建立
针对 Tonic.ai 等深耕合成数据领域的企业，主持人提出了一个尖锐的落地痛点：企业客户往往担心合成数据只是生成了一种“看起来合理”的近似统计分布，而非反映业务实质的“真实洞察”（Real Insights）。团队如何向客户证明合成数据能够提供真实的价值？
Karl 强调，建立信任的关键在于搭建严密客观的评估体系（Evals）：
- **评测基准与裁判机制**：构建包含具体任务问题的评测集与评分准则（Rubrics），既可以引入“大模型作为裁判”（LLM-as-a-judge），也可以结合人类专家进行盲测评审。
- **合成训练与真实测试验证**：在模型研发链路中，采用合成数据进行训练或微调，但测试集与验证集严格保留高质量的真实数据样本，以此检验模型的泛化能力。
- **外部公开基准打榜（Hill Climbing）**：利用合成数据对模型进行后训练（Post-training），如果在行业公认的公开基准测试上能够持续取得指标突破与性能爬升，这就是合成数据有效性的最强客观证据。
---
### [05:47](https://www.youtube.com/watch?v=s6lAx_yiwII&t=347s) 真实数据独有的壁垒：前沿模型训练、人类作恶创意与推荐系统二阶效应
当被问及“真实数据能教会模型哪些合成数据无法传授的东西”时，Mikayel 从前沿模型（Frontier Models）训练的视角切入。他认为，训练数据本质上是在向模型灌输人类世界已有的知识与能力。在科学计算、高级编程以及复杂的视觉创意等前沿领域，合成数据往往存在严重的能力天花板；过量使用合成数据甚至容易引发模型幻觉与偏见放大。在模型能力追赶或特定垂直应用层，合成数据效果显著；但若想突破前沿能力上限，人类创造的高质量数据依然是不可替代的基石。
```mermaid
flowchart LR
accTitle: Real Data Advantages Across Domains
accDescr: Breakdown of the unique irreplaceable aspects of real data in frontier models, trust & safety, and recommender systems.
subgraph Frontier["前沿大模型 (Wirestock)"]
F1["高级科学 / 编程知识"]
F2["避免偏见累积与模型崩溃"]
end
subgraph TrustSafety["信任与安全 (Airbnb)"]
T1["人类恶意行为的高度创新性"]
T2["突破工程师预设的想象力局限"]
end
subgraph RecSys["推荐与搜索 (DoorDash)"]
R1["复杂的长周期二阶效应"]
R2["用户真实的探索与多样性需求"]
end
```
随后，工业界两位工程师从具体业务场景补充了真实数据的核心壁垒：
- **信任与安全领域（Airbnb 视角）**：Aniket 坦言，真实数据的不可替代之处在于“人类在钻空子、破坏平台规则方面的创造力是无限的”。合成数据的生成往往受限于研发人员自身的想象力边界，而现实中黑产与违规分子的作恶手法远比工程师在白板前预想的更为离奇和狡黠。这一观点引得主持人感叹这是个“略显灰暗但无比真实的人生真相”。
- **推荐与广告系统领域（DoorDash 视角）**：Prashanth 指出，在推荐系统的全链路中，多环节仿真（如竞价模拟、重排模拟）极其困难，因为各模块联动会在时间维度上产生复杂的“二阶效应”（Second-order effects）。例如，系统如果单纯根据点击率模型预测 Prashanth 极度喜欢看红牛极限运动视频，模拟系统可能会把页面全排满红牛视频；但真实人类的心理是看几条之后就需要新的刺激，随后才会回访。这种用户对内容多样性（Diversity）与探索性的微妙心理动态，如果仅靠端到端纯数据模拟，往往很难在系统上线前被精准捕捉。
---
### [10:07](https://www.youtube.com/watch?v=s6lAx_yiwII&t=607s) 生产环境中的配比与长尾危害的合成数据生成策略
关于生产环境中真实数据与合成数据的日常使用比例，Aniket 表示业务中并不存在一个放之四海而皆准的固定百分比（如五五开或七三开），而是完全取决于危害类型在真实世界中的发生频率：
- **高频违规行为**：如常见的虚假信息（Misinformation）或违规涉黄内容（Nudity），真实样本量极其充足，合成数据的引入比例相对较低。
- **罕见长尾危害（Edge Cases）**：针对欺诈、罕见暴力或极其离奇的极端安全事件，真实样本极度匮乏，此时合成数据的占比会大幅提高，主要用于扩充正样本数据以训练检测模型。
面对主持人关于“如何在合成罕见事件时避免过拟合或产生认知盲区”的追问，Aniket 透露了 Airbnb 团队的核心方法论：**切忌将 10 个真实样本直接扔给模型并简单粗暴地下达‘帮我生成 1000 个类似样本’的指令。** 真正有效且健壮的工业级做法是向上一层抽象——深入分析违规行为背后的人类行为模式，提取出高维特征与行为属性，再基于这些抽象属性设计提示词与生成逻辑，从而合成出具有结构多样性与泛化能力的训练样本。
- **高频场景 vs. 极端罕见场景**：高频违规以真实数据为主，罕见边缘危害严重依赖合成数据补充稀缺性。
- **高阶属性抽象合成法**：从长尾行为中提炼底层行为特征，而非在样本表层做直接克隆，是避免合成数据失真与盲区的重要工程准则。
### [12:04](https://www.youtube.com/watch?v=s6lAx_yiwII&t=724s) 抽象化合成数据与隐私保护的平衡难题
在利用大语言模型（LLM）生成合成数据时，提取高层属性（attributes）能够赋予模型足够的泛化能力，使其可以应对不断涌现的新型危害与攻击向量（abuse vectors）。更关键的是，这种方法能极大降低数据泄露风险、保护用户隐私。例如在处理私密性极高的用户聊天信息时，团队必须确保没有任何人能够通过合成数据逆向工程出（reverse engineer）原始消息的出处。
然而，如何掌握这种“抽象的度”是一个极具挑战性的技术难点。业界并没有一个标准答案或“银弹”（silver bullet）来明确规定抽象到何种程度才是最佳平衡点。工程团队通常只能不断实验，一方面评估合成数据与真实数据的相似度，另一方面反复验证是否能够从中逆向推导还原出真实数据。
```mermaid
flowchart TD
accTitle: 合成数据抽象与隐私保护闭环
subgraph Generation["生成与抽象阶段"]
A["真实私密数据<br/>(例如用户私信)"] --> B["提取高维抽象属性<br/>(Abuse Vectors / Patterns)"]
B --> C["LLM 生成合成数据"]
end
subgraph Validation["评估与反向验证"]
C --> D{"逆向工程测试"}
D -- "存在泄露风险" --> B
D -- "具备高泛化度 & 隐私安全" --> E["投入训练与生产使用"]
end
```
---
### [13:02](https://www.youtube.com/watch?v=s6lAx_yiwII&t=782s) 企业自研与采购外部服务的权衡考量（Build vs. Buy）
主持人 Jonathan Shieber 转向一线实操人员，询问在合成数据和数据解决方案上，企业如何做出“自研还是采购”（Build vs. Buy）的抉择，何时会寻求 Tonic.ai 或 Wirestock 等外部供应商的协助。
来自 Airbnb 的 Aniket Tushar Ajagaonkar 指出，这一决策高度取决于公司内部的技术储备、工程团队的带宽以及资金投入力度：
- **头部科技巨头（Big Tech）的倾向**：通常倾向于全部在内部自研。核心原因在于保护自身极具商业价值的核心数据资产，避免将敏感数据暴露给第三方。
- **隐私与法律风险考量**：虽然并非指责外部服务商会做坏事，但数据出境或外流本身就存在被滥用及引发法律合规纠纷的潜在风险。
DoorDash 的 Prashanth Srinivasan 进一步从成本结构与商业周期对自研与采购进行了深度拆解：
- **按次调用与持续成本（Inference / Ongoing Cost）**：如果采购的是全套机器学习解决方案或推理层（inference），每一次搜索或终端调用都需要付费，这会从根本上改变业务的经济模型（unit economics）；而如果购买的是用于训练模型的辅助工具，其开销往往是偶发性的（sporadic），成本更可控。
- **团队建设与维护成本**：在内部组建一支具备相应专业知识的团队并保持其长期自驱与稳定，需要耗费巨大的组织成本。
- **核心业务归属性**：如果某项技术不是业务立足的核心壁垒，且自研周期过长，直接采购现成方案更为合理。当然像 AWS 这种基础设施属于例外——因为底层云服务过于庞大，几乎没有普通企业会去重新发明轮子。
---
### [15:53](https://www.youtube.com/watch?v=s6lAx_yiwII&t=953s) 供应商视角：真实人类数据 vs. 合成数据的立足点
当客户决定签约时，真实数据与合成数据服务商各自的核心卖点截然不同。
Wirestock 首席执行官 Mikayel Khachatryan 阐述了真实人类创作者数据的不可替代性：
- **视频与物理时空理解的局限**：Wirestock 专注于人类创作者内容。前沿模型在视频生成与视频理解（video understanding）方面仍存在巨大短板，对时间关联性（temporal relationships）、光影连续变化以及因果逻辑（cause and effect）的认知非常有限。在这一领域，合成数据根本无法达到训练要求，完全无法替代真实数据。
- **合成数据的实际定位**：在实际业务中，合成数据更多被 Wirestock 用作数据扩增（augmentation）或数据标注（annotation）的辅助与补充手段（supplement），而非彻底替换。
Tonic.ai 联合创始人兼首席产品官 Karl Hanson 则从合成数据供应商的角度回应了“自研 vs. 采购”的竞争：
- 供应商必须拿出“更快、更好、更便宜”的硬核证据。如今借助 Claude Code 或 Codex 等辅助编码工具，企业内部工程师搭建简易原型非常容易，但随之而来的是高昂的维护成本与 Token 消耗。
- Tonic.ai 通过公开发布的白皮书与基准评测（benchmarks）以及客户实际的概念验证（POC），证明专业供应商的方案远胜于企业内部临时拼凑出来的系统，“事实胜于雄辩”（the proof is in the pudding）。
---
### [18:53](https://www.youtube.com/watch?v=s6lAx_yiwII&t=1133s) 重新定义“替代”：合成数据何时超越真实数据？
对于合成数据是否必须“在质量上媲美甚至超越真实数据”才能算作替代，Karl Hanson 进行了深入辨析：
| 替代类型 / 应用场景 | 核心驱动力与实现机制 | 典型案例 / 代表论文 |
|:--- |:--- |:--- |
| **合规与准入驱动的被动替代** | 受限于 GDPR、HIPAA 等严苛法规，企业在特定场景下依法禁止接触真实数据，必须使用去标识化数据或合成数据构建训练流水线。 | 医疗健康与金融受控环境中的模型预演与离线训练 |
| **自我博弈超越人类上限** | 模型通过自我对弈（self-play）产生海量合成数据，突破人类专业知识的局限。 | DeepMind 的 **AlphaGo**：自我对战生成的棋谱数据远超人类棋手历史数据 |
| **预训练语料提纯与升维** | 针对原始 Common Crawl 等网络抓取数据中充斥的混乱杂质，利用模型将其重写合成优质文本。 | 微软的 **Phi 系列论文**（"Textbooks Are All You Need"）：用模型合成教科书级别的高质量数据来优化预训练 |
Karl 指出，尽管顶尖前沿模型实验室对其合成数据管线严格保密，但在预训练（pre-training）、中训（mid-training）和后训练（post-training）流程中，合成数据已被极为高效且广泛地应用。
---
### [21:21](https://www.youtube.com/watch?v=s6lAx_yiwII&t=1281s) 真实落地场景：冷启动预判与安全审核员培训
在具体业务场景中，合成数据正在为企业带来高回报的实际价值：
DoorDash 的 Prashanth Srinivasan 分享了合成数据在产品“单向门”决策（one-way doors）中的关键作用：
- **新品发布冷启动（Day 1 Readiness）**：当推出一个全新的用户交互界面或产品特性时，企业无法预知市场反响，直接线上测试（A/B testing）可能会带来负面副作用。此时无法获取真实线上数据，团队通过将现有的真实流量外推至新界面，并结合产品预期目标进行逆向推导，利用合成数据训练出检索与排序模型，确保新功能上线首日即具备成熟的可用性。
- **LLM 内部任务探索**：在缺乏用户真实交互日志的情况下，利用合成数据探索并外推大模型在不同内部任务中的能力边界。
Airbnb 的 Aniket Tushar Ajagaonkar 则补充了另一大核心价值领域：
- 除了生成用于防范恶意行为（harm detection）的机器学习模型训练数据之外，合成数据还被极富创造性地用于**信任与安全（Trust & Safety）团队的人工审核员培训**，使审核人员在无需接触侵犯用户隐私的真实敏感信息的前提下，熟悉并掌握各类违规判别标准。
### [23:05](https://www.youtube.com/watch?v=s6lAx_yiwII&t=1385s) 合规审核中人类培训的困境与合成数据的介入
在内容审核与信任安全（Trust & Safety）领域，每当用户申诉被下架或违规判决时，往往需要人工审核员参与裁决。然而，平台政策并非一成不变，而是一份不断演化的“活文档（live document）”，这就要求审核人员必须接受持续的动态培训。与此同时，由于审核团队人员流动频繁，仅靠有限的真实违规案例来进行培训显得捉襟见肘。合成数据在此处发挥了关键作用——通过生成多样化且合规的模拟案例来培训审核人员，从而确保平台在执行下架等严厉惩处措施时具备极高的准确率与一致性。
### [23:51](https://www.youtube.com/watch?v=s6lAx_yiwII&t=1431s) 视觉多模态与基础模型：对真实数据的需求演进与人类-AI共生
随着基础模型（Foundation Models）不断吸收海量数据，关于“真实数据是否存在需求上限”的讨论愈发激烈。Wirestock 的 Mikayel Khachatryan 指出，数据需求呈现阶段性的生命周期：传统的简单视觉数据（如常规库存图库照片 stock photos）需求逐渐饱和，但模型在向更高层级演进时，反而对高度复杂、充满细微差别（nuanced）的人类高质量数据产生了更强烈的渴求。
在预训练（pre-training）、中程训练（mid-training）、监督微调（SFT）以及强化学习（RL）等不同阶段，合成数据与真实数据形成了紧密的“人类与 AI 共生（Human-AI symbiosis）”工作流。例如，合成数据在数据标注辅助、数据集策展以及后训练（post-training）阶段针对不同基准的模型输出自动打分（auto rating）上表现极其出色。然而，二者之间并非简单的替代关系，而是一个动态权衡的连续谱（continuum）。当模型能力在缺乏新数据的情况下遭遇边际效益递减（diminishing returns）时，系统仍需回归人类真实数据的“源头活水”，汲取全新维度的知识与认知。
```mermaid
flowchart TD
accTitle: Human-AI Symbiosis in Model Pipeline
accDescr: Pipeline showing the iterative lifecycle of real and synthetic data across training stages.
subgraph InputStage ["数据生成与策展"]
A["复杂人类专业知识 / 真实数据"] --> B["人类-AI 协同标注与策展"]
S["合成数据引擎"] --> B
end
subgraph TrainingStage ["模型训练阶段"]
B --> C["预训练 & 中程训练 (Pre/Mid-training)"]
C --> D["监督微调 (SFT) & 强化学习 (RL)"]
end
subgraph PostTrainingStage ["后训练与评估"]
D --> E["模型输出"]
E --> F["合成基准自动打分 (Auto-rating)"]
F --> G{"是否遭遇边际效益递减?"}
G -- "是" --> A
G -- "否" --> H["生产环境部署"]
end
```
### [27:16](https://www.youtube.com/watch?v=s6lAx_yiwII&t=1636s) 双边市场与复杂系统的模拟困境：当理论遭遇现实
在被问及 Airbnb 与 DoorDash 是否已触碰到合成数据的边际效益瓶颈时，来自 Airbnb 的 Aniket Ajagaonkar 幽默地表示，出于信任安全合规限制虽无法透露细节，但目前技术落地仍处于极早期阶段，远未达到该瓶颈。
DoorDash 的 Prashanth Srinivasan 则从外卖与零售搜索平台的视角切入，指出“关键在于能否经受实践检验（the rubber has to meet the road）”：
- **可建模的单点维度**：例如用户的偏好转移（changing tastes），业界（如 Pinterest 公开的技术文献）已经证明这是完全可以通过真实数据建模甚至高保真模拟来解决的确定性问题。
- **难以模拟的多智能体博弈系统**：在复杂双边市场管道中，系统包含多方动态交互（例如不同广告主在相同查询词下设置每日预算并展开实时竞价）。调整单一变量可能触发恶性选择（adverse selection）及连锁反应，这在合成环境中极难逼真还原。无论离线模拟生成了多少合成数据，如果无法驱动线上核心业务指标的提升，就必须依赖底层真实物理世界的数据锚定（real grounding）。
### [30:18](https://www.youtube.com/watch?v=s6lAx_yiwII&t=1818s) 现场问答：自动化评估（Evals）与内置基准的构建
现场观众提问：能否利用合成数据将意图转化为非结构化数据（如图像），并由另一模型验证系统是否正确解析，从而在 Evals（模型评估）环节彻底剔除人工介入？
Tonic.ai 的 Karl Hanson 对此表示肯定并展开阐述：合成数据的核心优势在于从生成之初就掌控了“事实真值（Ground Truth）”。开发者在构建强化学习环境或评测集时，可以先定义模拟目标并向其中注入可控噪声，在生成合成数据的同时自动嵌入标签（labels），省去了后端繁重的人工标注。
- **Harvey 的法律评测案例**：法律 AI 公司 Harvey 发布的 Legal Agent Bench 便完全基于合成数据构建。受限于严格的客户隐私保护，Harvey 无法直接接触真实案例，但他们凭借对法律场景概念与挑战的深刻理解，合成出全套评测基准并内嵌金标，完成了此前无法实现的大规模评测。
- **人工在环的必要性**：虽然自动化评测前景广阔，但仍需在最终阶段设置人工审核（Human in the loop），防止生成过程中无意间引入预期之外的副作用（side effects）或过度合成。
- **高阶 RL 环境的核心资产**：Mikayel Khachatryan 补充强调，在强化学习环境构建与黄金标准（golden output）审核中，人类专家在工作流设计、质控及专业认知层面的输入，依然是整个技术栈中最具价值、同时也最昂贵的战略资源。
### [33:43](https://www.youtube.com/watch?v=s6lAx_yiwII&t=2023s) 现场问答：合成数据中的偏见注入与静默偏差隐患
在下半段问答中，一名来自金融科技（Fintech）领域的观众提出了一个尖锐的安全隐患问题：在缺少充足真实世界样本的情况下，使用合成数据训练反欺诈（fraud detection）等机器学习模型时，系统是否会悄然向“迎合人类预期”或“虚假拉高 Precision/Recall”的特定方向发生静默偏移（silent skew）？该如何设置护栏机制（guardrails）来防止此类人为设计偏见在合成数据中无形放大？主持人 Jonathan Shieber 与观众确认了问题的核心，即在反欺诈等极端对抗场景下，如何防范设计偏差对合成数据有效性的腐蚀。
### [34:51](https://www.youtube.com/watch?v=s6lAx_yiwII&t=2091s) 合成数据的偏差防范与相似度评估机制
在模型训练与风控反欺诈场景中，如果合成数据生成时产生某种偏差，可能会导致看似“欺诈率下降”但实际上只是数据倾斜制造出的虚假高准确率。为了确保合成数据尽可能贴近真实世界数据且不失真，核心手段之一是引入**相似度搜索（similarity search）**进行对比校验。
```mermaid
flowchart TD
accTitle: Synthetic Data Similarity Search Validation
subgraph S1["相似度校验区间"]
A["合成数据与真实数据对比"] --> B{"相似度水平"}
B -->|"完全相同 / 极度接近"| C["隐私泄露风险<br/>(泄露个人敏感信息)"]
B -->|"过度偏离 / 毫无关联"| D["虚假失真风险<br/>(完全人造脱离现实)"]
B -->|"适度平衡区间 (Golden Mean)"| E["可用高质量合成数据"]
end
subgraph S2["质量与安全性验证"]
E --> F["Human-in-the-loop 人工抽样校验"]
E --> G["LLM as a Judge 自动化评审"]
end
```
评估合成数据的质量与安全性需要在两极之间寻找平衡：
- **过于接近真实数据**：存在隐私泄露风险（leakage of personal or private data），可能无意中复制了真实敏感信息。
- **与真实数据完全脱离**：数据可能沦为脱离实际的完全虚构产物，在现实业务场景中根本不可能发生。
- **黄金平衡区间（Middle Segment）**：既不直接暴露真实隐私，又没有完全脱离现实逻辑，这才是投入生产使用的理想数据段。
为了进一步确认该区间内数据的有效性，可以引入 **Human-in-the-loop（人工在环）** 抽样核验，亦可采用 **LLM as a Judge（大语言模型作为裁判）** 进行自动化质量判定。
### [36:02](https://www.youtube.com/watch?v=s6lAx_yiwII&t=2162s) 讨论总结与结语：合成数据替代真实数据的适用边界
圆桌主持人 Jonathan Shieber 在时间告罄之际向各位嘉宾与观众致谢并进行总结。他指出，探讨“合成数据能否替代真实数据”并不是一个非黑即白的二元对立问题（not a binary question），真正的核心在于：**在哪些具体场景与业务用例中，采用合成数据替代或增强真实数据才是最具性价比和最高效的最优解（most optimal）**。这一答案必将随着具体应用场景（Use Case）的不同而有所差异。最后，主持人向主办方 Ai4 及台上分享真知灼见的专家嘉宾致谢，圆桌讨论在轻松热烈的氛围中正式结束。
## Notable quotes
> [00:04](https://www.youtube.com/watch?v=s6lAx_yiwII&t=4s) “合成数据能否取代真实数据”不仅是一个引人入胜的技术议题，更是当前整个人工智能产业中最具基础性、最核心的焦点之一。 — Jonathan Shieber
> [04:16](https://www.youtube.com/watch?v=s6lAx_yiwII&t=256s) 如果在行业公认的公开基准测试上能够持续取得指标突破与性能爬升，这就是合成数据有效性的最强客观证据。 — Karl Hanson
> [05:47](https://www.youtube.com/watch?v=s6lAx_yiwII&t=347s) 在模型能力追赶或特定垂直应用层，合成数据效果显著；但若想突破前沿能力上限，人类创造的高质量数据依然是不可替代的基石。 — Mikayel Khachatryan
> [05:47](https://www.youtube.com/watch?v=s6lAx_yiwII&t=347s) 人类在钻空子、破坏平台规则方面的创造力是无限的。 — Aniket Tushar Ajagaonkar
> [10:07](https://www.youtube.com/watch?v=s6lAx_yiwII&t=607s) 切忌将 10 个真实样本直接扔给模型并简单粗暴地下达‘帮我生成 1000 个类似样本’的指令。 — Aniket Tushar Ajagaonkar
> [15:53](https://www.youtube.com/watch?v=s6lAx_yiwII&t=953s) 在这一领域，合成数据根本无法达到训练要求，完全无法替代真实数据。 — Mikayel Khachatryan
> [15:53](https://www.youtube.com/watch?v=s6lAx_yiwII&t=953s) 事实胜于雄辩（the proof is in the pudding）。 — Karl Hanson
> [27:16](https://www.youtube.com/watch?v=s6lAx_yiwII&t=1636s) 关键在于能否经受实践检验（the rubber has to meet the road）。 — Prashanth Srinivasan
> [36:02](https://www.youtube.com/watch?v=s6lAx_yiwII&t=2162s) 探讨“合成数据能否替代真实数据”并不是一个非黑即白的二元对立问题，真正的核心在于：在哪些具体场景与业务用例中，采用合成数据替代或增强真实数据才是最具性价比和最高效的最优解。 — Jonathan Shieber
## Mentioned
Jonathan Shieber, Karl Hanson, Prashanth Srinivasan, Aniket Tushar Ajagaonkar, Mikayel Khachatryan, Tonic.ai, DoorDash, Airbnb, Wirestock, Resilience Media, Ai4, DeepMind, AlphaGo, Harvey, Legal Agent Bench, Claude Code, Codex, AWS, Common Crawl, Phi
## Video description
Aniket Tushar Ajagaonkar - Staff Software Engineer - Trust, Airbnb Inc
Prashanth Srinivasan - Senior ML Engineer, DoorDash
Karl Hanson - Co-founder and CPO, Tonic.ai
Mikayel Khachatryan - Co-Founder & CEO, Wirestock
Jonathan Shieber - Journalist, Resilience Media
This panel examines where and how synthetic data can outperform or augment real-world datasets. Panelists will share case studies from industries like finance, healthcare, and autonomous systems where synthetic generation has filled critical data gaps. Expect an honest discussion about performance tradeoffs, regulatory scrutiny, and what synthetic data pipelines actually look like in production.
[Source](https://www.youtube.com/watch?v=s6lAx_yiwII)
Source quotes
[0:04] Jonathan Shieber: “合成数据能否取代真实数据”不仅是一个引人入胜的技术议题，更是当前整个人工智能产业中最具基础性、最核心的焦点之一。
[4:16] Karl Hanson: 如果在行业公认的公开基准测试上能够持续取得指标突破与性能爬升，这就是合成数据有效性的最强客观证据。
[5:47] Mikayel Khachatryan: 在模型能力追赶或特定垂直应用层，合成数据效果显著；但若想突破前沿能力上限，人类创造的高质量数据依然是不可替代的基石。
[5:47] Aniket Tushar Ajagaonkar: 人类在钻空子、破坏平台规则方面的创造力是无限的。
[10:07] Aniket Tushar Ajagaonkar: 切忌将 10 个真实样本直接扔给模型并简单粗暴地下达‘帮我生成 1000 个类似样本’的指令。
[15:53] Mikayel Khachatryan: 在这一领域，合成数据根本无法达到训练要求，完全无法替代真实数据。
[15:53] Karl Hanson: 事实胜于雄辩（the proof is in the pudding）。
[27:16] Prashanth Srinivasan: 关键在于能否经受实践检验（the rubber has to meet the road）。
[36:02] Jonathan Shieber: 探讨“合成数据能否替代真实数据”并不是一个非黑即白的二元对立问题，真正的核心在于：在哪些具体场景与业务用例中，采用合成数据替代或增强真实数据才是最具性价比和最高效的最优解。
Source quotes
[5:47] Aniket Tushar Ajagaonkar: Humans are really creative in how to abuse policies and platforms.
[10:07] Aniket Tushar Ajagaonkar: here are 10 samples, generate 1,000 like them
[13:02] Karl Hanson: faster, better, and cheaper
[19:56] Karl Hanson: Textbooks Are All You Need
[21:21] Prashanth Srinivasan: one-way doors
[23:51] Jonathan Shieber: well of real data
[23:51] Prashanth Srinivasan: the rubber has to meet the road
[30:17] Mikayel Khachatryan: golden outputs
Original transcript
[0:04] Thank you all very much for coming out to this
in this is going to be a really fun panel. This
[0:09] topic is fascinating and it's it's so relevant
right now to I think everything that one might
[0:14] say foundational uh to everything that's going
on um in the industry. Uh we have a great panel
[0:21] who's going to be talking about this. So I'm
going to shut up uh and let them talk very
[0:25] soon. But I did want to say thank you very much
to y'all for coming out and bearing with us on
[0:29] the last panel before lunch, which is always a
tough one. Um, [clears throat] all right. So,
[0:34] can synthetic data replace real data? Before we
get into that, um, I'd love for y'all to very
[0:41] briefly introduce yourselves and then since we're
in Vegas, tell us what your favorite card game is.
[0:50] I'll go first, please. U, hi, I'm I'm Carl.
Thanks for having us here. Thanks, John,
[0:54] for hosting. Uh, I'm Carl Hansen. I'm one of the
co-founders of Tonic AI, which is a company that
[1:00] takes sensitive data and makes it available
for agent building agents, building software,
[1:05] and agent training. And one of the techniques that
we use is creating synthetic data for that. By far
[1:10] my favorite game is uh, Blackjack. Fair. Uh, hi.
Yeah, I'm Prashant. I work in Door Dash primarily
[1:17] in the search and ads systems, right, around
retrieval and ML models. Um yeah and I like a
[1:23] bit of poker but also a lot of hearts. Uh my name
is Anik. Uh I'm an engineer in Airbnb. Uh interest
[1:31] in safety. I don't have a card game but I do like
playing board games and exploding kittens is my
[1:38] favorite. Nice. Hi Mikail um co-founder and CEO
of Wiretock. We um a multimodal data company uh
[1:47] focusing on human uh created um visual uh data uh
and training um frontier models. Uh I'm going to
[1:58] have to go with poker for my favorite card game.
Nice. Mine's skipo. I don't know if any of y'all
[2:04] have ever played that, but it's an phenomenal
game. Um okay, we're going to get into it now. Um,
[2:09] but there's another little table setting I want to
do before we talk about synthetic data versus real
[2:15] data and whether one will replace the other. What
do we mean by synthetic data? What do when when I
[2:21] say synthetic data, what does that mean for each
of you? How do you define it? And we'll start with
[2:28] you again, Carl. Actually, you know what? We'll
flip it up. Kyle, let's start with you. Yeah,
[2:32] I was thinking about that too before the panel.
It's like uh it's really important to define what
[2:39] what we're talking about uh in terms of synthetic
data. It could be uh we could be talking about
[2:44] like direct model outputs, AI generated content or
or uh data. It could be um using AI models to edit
[2:54] or enhance um like real world or human created
data or it could be used for annotation, curation
[3:03] and all sorts of other use cases. Um, so, um, so
I think all of those things would be classified as
[3:11] as synthetic data. And depending on which one of
these types of synthetic data we're we're talking
[3:18] about, there could be different uh pros pros or
cons of of human vs synthetic. So, I'm excited
[3:26] to dive into more of this topics. I see a lot of
headnodding on the panel. Are we all in agreement
[3:32] that this is how we want to think about synthetic
data? Is there anything that you would add to that
[3:38] definition or this this classification or sounds
directionally right? If it sounds directionally
[3:44] right, we can just move on. So Carl, I was just
going to add some sometimes when I think about
[3:49] synthetic data, it's it's best thought of like
the you have a model and you have data that goes
[3:54] into it and the output is then synthetic and you
can vary both of those. You use different models
[3:58] to create it. You can use different types of
data. Sometimes you use real data to create
[4:01] the synthetic data. Sometimes you use nothing
to create the synthetic data. You just use a
[4:05] model to create it. And so it's sort of these two
these two pedals that you have to to create that
[4:10] synthetic data. So when you think about sort of
the output that comes out of all of this, how do
[4:16] you prove to your customers since y'all are are
so deeply involved in synthetic data? How do you
[4:21] pro prove to your customers that that synthetic
data is actually presenting real insights versus
[4:27] like sort of convincing distributions or or
reasonable faximile of what a real insight may
[4:33] look like? What where's the truth claim coming
from? How do you make that to folks? Carl,
[4:40] that's for you buddy. Yeah. Um to do that well you
have to start with uh your evals. You have to set
[4:47] up a really good framework to judge the output
of the synthetic data and you know that can be
[4:54] uh you know it can be simply like setting up a
bunch of uh so problems for it to work through
[5:02] and evaluation rubrics whether it's through an
LLM judge or human judge on the other side to
[5:05] to figure out whether it's it's realistic or not.
Um it can also be comparing it to real data. You
[5:10] could train a model on synthetic data and then
do your test and eval on a small sample of real
[5:15] data. Um that's one way to do it. It can also be
um in you know in the case of if you're trying
[5:21] to climb a hill and there's external benchmarks
that you're going after you can you can apply you
[5:26] can you know post-train a model with synthetic
data and if there's a public benchmark that you
[5:30] want to see if it improves the model on you can go
try it out on the benchmark and if it hill climbs
[5:34] properly then you have a pretty good reason
to think that your synthetic data did a good
[5:37] job. Is are there things miky that uh that real
data can teach models that synthetic data can't?
[5:47] Um it's it's not yeah I'm it's not it's not so
much about things but um in terms of like we we
[5:55] like to think of [snorts] of of training data as
a set of capabilities that you're passing onto
[6:02] the model and uh and and um all sorts of kinds
of like knowledge that exists in the world and
[6:11] um I think it's much more the is at the
frontier of of frontier models where
[6:20] um synthetic data falls short of what's needed
as you're training for more advanced knowledge
[6:27] and more advanced capabilities whether it's
in science whether it's in programming or
[6:32] uh in creative domains where we operate
where wiretock does a lot of work in
[6:37] um sort of absorbing that knowledge and like
infusing that that knowledge and then passing
[6:43] it on to to the models through whether it's u RL
data or it's SFT data or pre-training is much more
[6:51] scarce and you see a lot more biases in the in the
model outputs on the synthetic data side and so
[6:58] I think there's a lot of like I've seen a lot of
good good use of synthetic data at the application
[7:04] layer uh and and you know kind of for models
that are trying to catch up and and push towards
[7:10] the frontier but I think it's it's it's much more
difficult to to absorb that knowledge and and and
[7:17] and that kind of data uh at the frontier. So so
frontier is more real data and at the application
[7:26] layer you can find more synthetic data that
can work effectively. I think yeah I've I
[7:31] think it requires a lot more of human input at the
frontier although you we we've seen we we operate
[7:37] at the frontier training like foundation
models whether it's like image video or or
[7:43] um multimodal uh models and we see we we have we
have seen good use of synthetic data but it's more
[7:52] on the uh annotation sort of like augmentation
of of human data but not not in terms of uh full
[8:02] replacement or substitution of it. So Anakad and
Pashant what you Yeah. Can I can I add one point?
[8:09] Uh one more place where uh especially in trust
and safety where real data kind of does better is
[8:16] is kind of humans are really creative in how to
abuse policies and platforms. Uh and you get to
[8:24] learn what people are thinking. uh something that
synthetic data sometimes doesn't because synthetic
[8:30] data generation is also limited by our imagination
of what uh people like how people could abuse
[8:37] the the capacity for abuse in humanity is much
more more widespread than than you could imagine
[8:43] as an people are really creative that's that's
disappointing that's that's kind of a bleak bleak
[8:50] statement on life before I I yeah please um so in
the recommended systems place where I've seen you
[8:56] know real data really stand out is you can try to
simulate a whole lot of outcomes and you know how
[9:01] your page should look right um you can simulate
parts of the auction you can simulate how ranking
[9:06] might be but all of them together have second
order effects over time and that's pretty hard
[9:11] to um model right just computationally but also
like you never really know how it's going to work
[9:18] until the rubber hits meets the road right and um
yeah sometimes users might just not like seeing
[9:25] like part of your objective function might be
you know I want to to surface say a bunch of
[9:29] videos that the user is most likely to click on
now a model like that will probably pick you know
[9:35] like let's say if I like watching Red Bull videos
right it might just say every Red Bull video that
[9:42] Prashant sees you know we're going to think it's
you're going to rate it very highly but in reality
[9:48] what I'd want to see is like a few Red Bull videos
and then something else and then come back right
[9:53] you wouldn't people wouldn't have discovered like
hey we need to inject diversity into this without
[9:57] you know if they just simulated the whole thing
and tried to you know set up a system end to end
[10:02] no spoilers guys we're going to get to all of
this a little later in our our discussion but
[10:07] um I I think for right now like I would love
to get a snapshot from from Anukette you and
[10:12] Pashant both on on what percentage of synthetic
versus real data y'all are using day-to-day right
[10:20] now just a like rough is it 50/50 is it 70
synthetic, 30 real. How how do you break it
[10:27] down? Uh actually it depends. Uh there is no one
answer to it. Uh the way mostly in trust sensity
[10:35] center data is used for augmenting uh data for
training models for detecting harms and there
[10:43] are certain harms that are kind of rare. We do
not have a lot of training data for it. Uh so
[10:48] there the percentage of sensory data is higher uh
as compared to harms that are fairly common like
[10:55] say misinformation nudity is fairly common versus
yeah I don't know what's what is really you don't
[11:01] have to go I can get very dark very quickly so
yeah maybe maybe we don't go there but I I I mean
[11:06] I think now I'm going to say that we should uh in
trust and safety events are are rare right there
[11:13] they're there are sort of fraud which is more
widespread, but then there's significant sort of
[11:20] rare weird edgecased trust and safety issues. When
you're looking at simulations in those instances,
[11:27] how do you balance like what's improving safety
versus what's either creating blind spots or or
[11:34] maybe overindexing on certain things once
something appears that that's unexpected.
[11:42] I think the what has worked is not using um real
data to just generate synthetic data directly like
[11:50] these are 10 samples generate thousand samples for
me but rather getting kind of a a level above and
[11:57] try to understand the the human behavior try to
extract attributes from it and then pass those
[12:04] attributes to the LLM to generate synthetic data.
So that way you get an enough generalization that
[12:10] you can potentially even extend as new harms and
new abuse vectors come up. Um you also have lesser
[12:18] risk of uh exposing basically protecting user
privacy. If you are trying to uh run some model
[12:25] on say user messaging which is fairly private uh
you don't want to you don't want anyone to reverse
[12:32] engineer which message this has come from. Uh so
that is helpful there. The tricky part is how do
[12:39] we get to this level of abstraction? There's no
correct one answer. There's no silver bullet that
[12:44] tells you okay this is the level of abstraction
that you need to get to. Um you just experiment
[12:49] and try to see how similar it is to the real data
and how and whether you can reverse engineer real
[12:54] data from it. Uh yeah I and one question I I think
you know again sort of for the the operators who
[13:02] use this stuff. How do you think about buy
versus build when it comes to these kinds of
[13:08] of formulations, right? Like when would you want
to use either Aonic or um Myle's um services?
[13:18] Like what what makes what makes that decision
for you whether you can do it internally or
[13:22] you want to rely on folks externally? I think it
depends on what internal capabilities available.
[13:28] Uh again it varies depending upon engineering band
band bandwidth how much investment a company can
[13:35] put into it. Um big tech is generally tries to do
a lot of things internally. Uh one is to protect
[13:43] their own data as well. They don't want to give
out data to to third party services because data
[13:48] is really critical here. Um and second is also
user privacy is like u uh we have all seen how
[13:55] what happens when data is sent out. Um I'm not
saying that that the external services would
[14:01] be doing anything bad but there is a potential
for abuse there right and creating more legal
[14:06] challenges. So uh it depends on company's budget.
Uh some companies prefer to just do it outside and
[14:12] get it off the box and some companies want to
build it themselves. How do you all think about
[14:17] it and about the build versus buy decision? Um
yeah so I think there's usually like a short-term
[14:23] medium-term long-term view like you know factoring
right. So um yeah just to take like some examples
[14:31] right if you want to buy a whole ML solution
right whatever kind um if it's something that
[14:37] you're buying like say inference you're going to
pay for it in every search right and that changes
[14:41] the economics of like a lot of the business itself
versus like you know buying a solution that helps
[14:47] you train models better right which is which tends
to be more sporadic right so that's one axis right
[14:52] in terms of like whether it's onetime cost versus
like an ongoing um The second is like yeah you
[14:59] know you have it you you you have to build up
a team with that expertise and you know retain
[15:04] them and yeah you know keep them um well motivated
and all of that right so that can also affect the
[15:10] build versus buy and you know what you need in
the short term and what helps in the long term
[15:15] um yeah [clears throat] and so these are primarily
the main factors like how quickly can you know can
[15:20] someone build a viable solution in house is it
worth keeping it alive in house right right um
[15:26] or you know just bet. Is it core to the business?
Is it something that you know you're going to pay
[15:30] like is it going to be metered for like basically
every usage versus uh every end user usage versus
[15:36] like something sporadic. Right. Right. So all of
this uh obviously there are exceptions to this
[15:41] right like everything I say here applies to AWS
[laughter] but uh yeah those are so fundamental
[15:46] that most companies don't want to right right
build that out. So, but broadly Carl and Mle when
[15:53] you when you have a a you know a customer that
gets to yes with y'all what is the argument so
[16:00] for real data mle and Carl for for synthetic data
on your side? Okay. Yeah. I mean um for a lot of
[16:09] demand that we see for a lot of asks and requests
from customers um Wireto is known as you know
[16:18] as as a as a human creator um oriented focused
company that focuses on sourcing uh human created
[16:27] content specifically in the creative domains but
also beyond that and um typically uh that's you
[16:37] naturally the type of demand we see and I think
the reason for that is like I was saying earlier
[16:43] for a lot of data that the the model builders
need synthetic data is just is not going to
[16:50] uh hit the mark like for example if you take
video video content as training data I mean
[16:57] um the video understanding and video
generation of this of of of the current
[17:03] um models at even at the frontier are very like
limited uh there is limited knowledge of like
[17:11] our understanding of temporal relationships, you
know, changing lighting and cause and co cause and
[17:17] effect like relationships and all these things.
Um and so there is really no replacement or really
[17:24] even not a lot of place for for synthetic
data for specifically in that modality.
[17:30] Um so and then and then in some other modalities
you know um I think we've seen customers be okay
[17:37] uh with some synthetic augmentation or potentially
annotation of the data which we have used like if
[17:45] something you know a lot of things could be done
much more efficiently uh with with with you know
[17:51] um synthetically generated data then and then
you know that that's that's where you you use
[17:56] it but in our case it's been more of as a as as
a supplement and a substitution. Carl, when when
[18:04] when folks say yes to you, why are they saying yes
about synthetic data? Um, I mean, your original
[18:09] question, the buy versus build, is is a very
common question for all vendors these days. Um,
[18:14] I think it's really on the vendor to prove that
it's that you're faster, better, and cheaper. Um,
[18:20] with tools, you know, Cloud Code, Codeex, and
the like, it's really easy to spin up something
[18:24] quickly. Um, but there's there's still maintenance
costs. there's still token costs of of doing that.
[18:30] Um, and so, you know, we have white papers, we
have benchmarks showing that our solutions are are
[18:34] faster and cheaper and better than something that
you can roll on your own or are cobbled together.
[18:39] Uh, and it's on us to show that either through our
published artifacts or through Ryan PC's and they,
[18:46] you know, the proof is in the pudding. Yeah. Um,
we talked a little bit about the definition of
[18:53] synthetic data. I'd like to actually talk a little
bit now about the the definition of replace like
[18:59] what do we mean when we say replacement and what's
the benchmark for replacing? Does it have to be as
[19:05] good as the real data when we talk about synthetic
data? Does it have to be better than than real
[19:11] data in some capacity or does it just have to
enable work in places where either a regulatory
[19:18] environment or a lack of data or the cost of
acquiring real world data makes it prohibitive?
[19:27] Can I take this? Yeah. Yeah. I um I think you you
hit it on the nail really uh there are situations
[19:33] where because of things like GDPR or other
regulation you know HIPPO regulation you cannot
[19:37] use real data for certain situations and in those
you know you don't have a choice sometimes you
[19:42] have to use either deidentified data or synthetic
data to get to do what you want to do um for so in
[19:48] those situations it's just about how close can you
get to the real data to solve the use case that
[19:52] you're trying to solve. Um and then maybe when you
push it to production it you can run it on real
[19:56] data but it's it's all in the training pipelines
where you you run it on synthetic data. Um as far
[20:00] as exceeding you know there's a few situations uh
and again you back to what Mikuel has been saying
[20:05] it's like it's very very use case specific. There
are situations where synthetic data can actually
[20:10] exceed real data. Um a classic one is you know if
you're looking at uh selfplay where uh you know a
[20:16] model is playing against itself and it's creating
synthetic as a byproduct of that. You know,
[20:20] it's a little dated now, but the probably the best
reference of this is Alph Go and how, you know,
[20:24] all the games that Alpha Go played against itself
created a lot of synthetic data. They produced a a
[20:27] model that was way better than anything that
would have been trained on human data. Um,
[20:31] there's also examples in the labs aren't very
open about what they're doing with synthetic data,
[20:35] but they do say they use synthetic data in parts
of their pre-post and mid-training pipelines. Um,
[20:39] and in in the pre pre-training pipelines, there's
a lot of pre-processing uh of data. So, you know,
[20:45] there's a common crawl, it's really messy web
scrape data. Um there was a paper done by the
[20:50] Microsoft group fee paper. Um it's a lot
of the subsequent work is actually about
[20:53] in that original paper but showing that uh
if you take a the common crawl caucus that
[20:59] is pretty messy and you you synthesize it you
you use models to synthesize textbook quality
[21:04] paper you actually build a better pre-training
pipeline. Um and so the lot they're public about
[21:09] what they're doing and a lot of other labs are
using that. It's not so obvious what you know
[21:12] some of the the more private foundation labs
are doing. So there is there is synthetic data
[21:16] being used very effectively in some of the
the pre-training steps of of and where are
[21:21] y'all finding the most value in synthetic data
right now when when you are applying it or you
[21:27] are using it where are you seeing sort of the
most return on your investment in that stuff.
[21:34] Yeah. So it might be hard to quantify but like
internally [clears throat] where we've seen it is
[21:39] uh sometimes there are product decisions which
can feel like oneway doors like do you want to
[21:43] enable this product and you know widely announce
it and then you know there might be cross function
[21:50] like they don't know how it's going to be received
right but but you still want to have models that
[21:53] are you know fairly capable on day one right you
want to be able to retrieve certain things you
[21:59] want your ranking to work in certain use cases
and in those cases like um yeah you cannot get
[22:05] a lot of real world data because running those
experiments itself can have other side effects,
[22:11] right? Um so yeah, in those cases, you know, yeah,
it it helps quite a bit to like extrapolate from
[22:17] existing real traffic onto a new new user surface
and you know, use some level of product vision of
[22:23] like I I'm sure this is what I would like
and what I wouldn't like as an end result.
[22:27] Work backwards from that like what the model
should be able to do. So that's been one area.
[22:32] M [cough] um [clears throat] yeah another area
where we've seen is like yeah you you can't really
[22:37] you know capture a lot of like your usage logs
in LLMs right but you want to know what kind of
[22:42] tasks it can do what it can't do and then try to
extrapolate better you know even internally right
[22:47] uh so in those cases yeah that's again a place
where you know you really want to use synthetic
[22:53] and on a kid what about airb I think one I talked
about was generating training data for ML models
[22:58] uh for harm the other interesting use case is for
training human reviewers uh for trust and safety
[23:05] because uh every time somebody kind of uh appeals
a decision or is taken down from a platform,
[23:12] a lot of times there is a human involved in in
either verifying that appeal or taking a decision
[23:18] to take something down and these humans need uh
training uh on the policy. The policy is kind of
[23:25] a live document. It kind of evolves. So the humans
need to be continuously trained. humans also kind
[23:32] of come and go into this review work. So, uh
generating training on real data is sometimes
[23:38] a bit hard because you do not have that much real
data. So, synthetic data can be used for training
[23:43] humans as well to ensure that uh the accuracy is
pretty high when you take down someone. Uh thank
[23:51] you. Thank you very much for that. Mail, one of
the things that I I sort of wanted to touch on
[23:56] with you as as these models get better, as things
get refined, as more and more data gets absorbed,
[24:03] um is there a limit to how much real image data or
real world data these models will need? Is there a
[24:10] point at which there's enough real data that will
render or obviate the necessity of of having more
[24:18] to create um new models? Um I think it's an sort
of an ongoing and um a continuous development is
[24:29] what we're seeing with the with the um foundation
models. So um it's um you know certain categories
[24:37] of visual content you know might be less needed
in a certain peri you know certain uh um instances
[24:45] and then and then and then there you might see
demand for it grow again over time you know
[24:53] um so there are like phases of of of development
that that you know that there's like a life cycle
[25:00] right and and then Um and then I see that um
what we what we observe what we've observed
[25:08] um is that as the the models get better they need
more and more complex input from humans. So it
[25:16] would be the same humans creating the data but
it would be much more complex and nuanced data.
[25:22] So um and so so a lot more simple like visual
for example like stock photos or more simple
[25:28] data might not as be might not be as needed but
a complex version of or may maybe a um so for
[25:35] example like we we help companies in pre-training
like mid-training supervised fine-tuning and also
[25:42] RL sort of pipelines and we've we've depending
on which development stage they are at at any
[25:50] point in time and So um we see demand for all
sorts of different kind kinds of visual data
[25:59] um at different phases of development at different
stages of development. And so uh in terms of
[26:05] answering to your previous question about like
whether where where it can replace or not I think
[26:11] uh like I was saying earlier like it's really good
at like helping annotate or maybe help curate like
[26:17] a lot of pipelines we have we have like a human
AI symbiosis in terms of like how we curate the
[26:23] data and then like in in a lot of post- training
stuff we we use you know you see a lot of good
[26:30] um um a good use of synthetic data in terms
of like auto rating you know the model outputs
[26:36] against different benchmarks and things like that.
Um so it's always like it's more it's always a a
[26:44] delicate kind of a mix you know of both or like a
continuum right there's there's no there's no like
[26:51] one hard answer. Models will improve and improve
and improve and then you'll hit diminishing
[26:56] returns on like how much these models can improve
without new training data. So you need to go back
[27:02] to the well of real data to refine the model
again, right? Feeding it with new capabilities,
[27:09] new knowledge, new understanding as it's and a
lot of it is coming from humans. Is that the way
[27:16] things have played out in your own deployments
of these things? I know we're still very very
[27:20] early in this process and it's it's um still nent,
but like when you use um these these models at at
[27:31] Airbnb and at Door Dash, do you find there are
diminishing returns at a certain point and you
[27:36] do have to sort of go back and refresh what
you're doing with with more real data versus
[27:42] synthetic? No. Uh I think we're pretty early. uh
we haven't reached there yet where we have fed
[27:50] this need but yeah I I can't reveal much on trust
and safety obvious reasons but uh yeah we haven't
[27:58] we are not nowhere close there yet for sure yeah I
think in at least like the retail and search space
[28:05] uh you know the rubber has to meet the road right
so we can try to simulate and generate as much
[28:10] synthetic data as we want but if it doesn't move
the metrics then right right so you do need some
[28:15] level of real uh grounding there. Um yeah, but
helps in expansion. Is it is it more difficult
[28:22] for a marketplace where there's so much weird
noise that that can distort things? Um, you know,
[28:31] tastes change, delivery options change,
food restaurants go out of business. Like,
[28:38] do you need to use more real data because it's
harder to model in these cases? like because
[28:43] of the vagaries of the industry. Um I think yeah
I'll split that into a couple of parts right so
[28:51] um I mean there has been quite a bit of literature
now by multiple companies where they've been able
[28:56] to model like changing tastes of the user uh among
publicly available ones Pinterest has you know
[29:02] published quite a bit of literature on this right
so I think that part something that you can model
[29:08] you know from real data or you know even simulate
them so that's you know that's like a solvable
[29:14] problem and you know data volume and trust how
much of how much of the data you trust kind of
[29:19] problem. I think the second thing you mentioned
around like all these things in the pipeline and
[29:23] how they play with each other that is pretty
hard to simulate because if you let's say have
[29:29] a have an advertis advertiser they have certain
daily budgets right but there might be multiple
[29:35] advertisers comp you know kind of competing in the
same space of queries and like tweaking even one
[29:42] of them can change like everyone else's behavior
right there's adverse selection there's a whole
[29:46] bunch of dynamics there that can come into play
so Those are like uh it's pretty hard to simulate.
[29:52] You can try your best but we we do try a lot but
uh yeah again the rubber has to meet the road.
[29:58] Yeah. Um we have about 4 minutes left and I can
ask these folks questions all day but I'm sure
[30:05] y'all may have some questions that you'd like to
ask these panelists as well. Um I'm happy to open
[30:11] the floor up for questions if anyone has any or
I can just keep peppering these guys with stuff.
[30:17] Yes sir, you in the front. And then I saw another
question right behind him. Can you do you mind? I
[30:22] don't know if we have a mic. I can actually run
down. I'll run down. Excuse me. All right. Here
[30:29] we go. It uh is it? Yes. Okay. Is um very cool.
Yeah. So this seems like a big unlock for evals
[30:41] and it was interesting like how you guys have
seen this impact. So you know the intention
[30:45] uh using a model potentially to to translate an
intention into like unstructured data an image
[30:52] something like that and then another model then
to take that and make sure that your system is
[30:56] um interpreting that data correctly with the with
the idea that the intention is there and you don't
[31:02] and you can kind of remove the human aspect of
creating that data. So kind of interested to hear
[31:07] what you guys have to say about that. Who wants to
take a crack at that first? Did y'all understand?
[31:14] Did y'all hear the question? Yeah. Um I I can
start if anybody wants to jump in feel free. So
[31:22] um you're absolutely right. So if you if you're
synthesizing the data, you can start with a ground
[31:25] truth. So you can say, you know, this is the model
that I want to simulate and then you generate the
[31:30] data around that. You can add noise to it.
You can do stuff like that. We do that with,
[31:33] you know, it's really useful for creating,
you know, reinforcement learning environments,
[31:38] for building evaluations, um, where you can sort
of work backwards from not having to have a human
[31:44] annotate at the end. You you have the model build
the annotation into into the actual synthetic
[31:49] data. It it it's good. It doesn't always complete
full loop. Sometimes you have to add an annotator
[31:53] onto the top on top of synthetic data in order to
make sure you didn't put something else in there
[31:57] that you by accident, some sort of side effect
that you weren't expecting. Um but it does it
[32:01] does give you um a really good way to sort of this
is a ground truth for the environment that I'm
[32:06] trying to create and uh and then you you hide
that from obviously you know the the eval and
[32:11] what you're doing. And so you we're seeing a lot
like um Harvey published their legal agent bench
[32:16] recently and that was entirely synthetic data and
it's really cool that they did that because you
[32:19] know Harvey can't touch any of their customers
data and they don't have real legal examples but
[32:23] they can create synthetic data based on the the
concepts and the ideas and challenges they know
[32:26] they solve and then they can create this bench
for that publish it entirely synthetic and when
[32:30] they created that bench they could also put the
labels in there as they created it. So it's a
[32:34] it's a really good tool for that. But I I will
say say like you do have to have a human in the
[32:37] loop there sometimes to at the end to make sure
that you don't over synthesize you know go beyond
[32:43] that. Just a quick comment on that. Yeah, I think
it's it's such a good example and a good question
[32:48] in terms of like uh what part of the the you
know simulated world RL um environment building
[32:56] you know humans will play and we see actually the
most valuable part of the RL environments or the
[33:04] RL environment you know building models through
RL uh and and and benchmarking as as being human
[33:10] human knowledge and real world human whether it's
tasks or processes or workflows or thinking about
[33:17] how you do quality control uh and and uh you know
human expert review of the model output or whether
[33:24] it's the golden output across the entire stack. I
think we're still seeing human human input as the
[33:32] most valuable and as a result most expensive uh
resource. So great question and I believe there
[33:43] was one. Hi um kind of related question [clears
throat] with synthetic data if there's not a
[33:51] lot of real world examples can it skew in certain
direction which will be mostly like pleasing what
[33:57] we want to see or like increasing the precision
or recall um in a way that we wanted to see and we
[34:03] think that it is going in positive direction but
it's actually the skew of the data in synthetic
[34:07] sense which is causing it. um what are the guard
rails against it and then how can we if it's kind
[34:13] of like silently doing it like how how do we
make sure that it doesn't happen? question was
[34:20] guardrails on synthetic data. Yeah. How do you how
do you set up guard rails to pre prevent sort of
[34:26] the biases from design biases from expressing
themselves in the data? Right. And did I get
[34:32] that question right sir? Yes. Okay. For example,
let's say um I work in a fintech and if I want
[34:38] to create an ML model which um can look at some
fraud fraud detection. Now instead of using real
[34:45] world example if I create synthetic data which
I know that uh could be created that how humans
[34:51] would behave or how transactions would go what if
it skews in some direction that I wanted to see or
[34:56] like it wanted to u work in one direction rather
than actual problem and then I see oh the frauds
[35:03] are going down but instead like it's just silently
creating that precision for like with this skew of
[35:09] the data that's kind of showing that I I think
one way to ensure that the data is as close to
[35:15] real world data is using similarity search. Uh
essentially what you're trying to do is once
[35:21] you have synthetic data try to compare with real
data. If it's pretty identical then there is a
[35:26] chance that you're leaking something personal or
private. If it's completely away then there is a
[35:32] chance that this is completely made up completely
artificial. There's no chance that something like
[35:36] this can happen. But that middle segment is what
we are looking for where it is not completely
[35:42] exposing private data and not completely synthetic
and that is where you we kind of start using it.
[35:48] We could have humans in the loop to sample and
verify if that is correct or not. Um we can even
[35:54] have LLM as a judge to kind of determine if
uh if that data is good enough or not. Well,
[36:02] we are over time. Uh I have a few more questions
I could ask these folks but I think uh in the
[36:07] interest of getting everyone fed and uh and off
I I will thank our panelists very much for doing
[36:13] it. I think like in closing one thing to keep in
mind it's not exactly a binary of can synthetic
[36:20] data replace real data but rather the question
is where is it most optimal for synthetic data
[36:26] to replace real data and that question the answer
to that question I'm sure will vary by use case
[36:32] as we've we've sort of determined here. Thank you
all so much for sitting around and listening to
[36:38] us paver. Thank you so much to AI4 for having
us all here and thank you to my panelists for
[36:43] being so freaking smart and awesome. Appreciate
y'all. Bye. Thank you. Thanks, Ron. Thank you.
