Why Medium is HIDING from AI | E2179

Medium CEO Tony Stubble on RSL and the Future of Content Monetization in the AI Era

Key Concepts:

• RSL (Really Simple Licensing): A new initiative for monetizing media in the generative AI era, similar in concept to RSS.
• AI Training vs. Inference (RAG): Training involves using content to build AI models, while inference (Retrieval-Augmented Generation) uses real-time data to answer queries.
• Vendor Lock-in: Dependence on a single provider for a technology or service, hindering interoperability and competition.
• Dead Internet Theory: The idea that the public internet will be dominated by AI-generated content, pushing creators to private spaces.
• Poisoning Data: Intentionally introducing errors or misleading information into a dataset to disrupt AI training.
• Attribution and Compensation: Ensuring proper credit and payment for content used by AI companies.

The Problem: AI's "Antisocial" Behavior and the Threat to the Public Internet

• Medium is experiencing increased traffic from AI crawlers, including both established and emerging AI companies.
• Tony Stubble argues that AI companies have broken the "social contract" by taking value from writers and creators without offering adequate compensation.
• He believes this behavior threatens the public internet, potentially leading to a "dead internet" dominated by AI-generated content.
• Writers may retreat to private spaces and paywalls if they are not fairly compensated for their work.
• Stubble emphasizes the need to "fix" the behavior of AI executives and establish a workable model for all parties.

RSL: A Collective Approach to Force AI Companies to the Table

• Medium joined RSL as a way to collectively address the issue of AI companies using content without compensation.
• Stubble had previously attempted to form a coalition but lacked the "heft" of larger companies like Meta.
• He believes that individual deals have been cut, and now a coalition is necessary to exert pressure on AI companies.
• RSL aims to create an internet standard that AI companies must adhere to in order to train on new data.
• Medium was one of the first companies to sign on to RSL, with the decision made rapidly due to prior consideration of the issue.
• Evert Walter, one of the creators of RSS, is also involved in RSL, lending credibility to the initiative.

Medium's Commitment: Passing All Revenue Back to Writers

• Medium intends to pass all revenue generated from RSL negotiations back to its writers.
• This commitment distinguishes Medium from other platforms, which may not share revenue with creators.
• Stubble calls for other platforms to follow Medium's example and prioritize compensating writers.

AI Training vs. Inference: A Nuanced Perspective

• While AI training is a primary concern, Stubble acknowledges the importance of AI inference (RAG) and its impact on traffic.
• RAG often includes citations and sends traffic back to the original source, but the exchange of value is "very weak."
• Traffic from ChatGPT converts to paying Medium members at a higher rate than normal traffic, indicating higher intent.
• However, the overall traffic volume may be lower, potentially negating the benefits of higher conversion rates.
• The long-term concern is that the rise of AI-generated summaries could reduce the need for original content.

Forcing Collaboration: Blocking Crawlers and the Threat of Data Poisoning

• Stubble praises Matt Prince at Cloudflare for advocating a mass blocking of AI crawlers.
• However, he notes that Cloudflare's approach involves vendor lock-in, which is not ideal for protecting the internet.
• Medium has added an addition to the RSL standard that would allow them to do the block on a per page basis, while Cloudflare does it on a per site basis.
• Medium also considered "poisoning" its data by introducing errors or misleading information into the content crawled by AI companies.
• For example, they could insert slanderous statements about OpenAI into Medium articles.
• Stubble hopes to avoid such "warfare" and instead encourage AI companies to engage in fair negotiations.

The Simplest Version of RSL: Blocking Training, Demanding Attribution for Inference

• Medium's initial implementation of RSL prohibits AI companies from using its stories to train their models.
• It allows them to summarize and link back to the writing in AI-generated search results, demanding fair attribution on the RAG side.
• Medium will not make any changes unless an AI company comes to the table with money.
• The announcement is intended to start a discussion with the Medium community about how the company plans to negotiate on their behalf.
• The conversations don't make sense until you have a number, right? Are we saying we're going to pay every story one penny? No one's going to care, right? Is it $5? Well, some people are going to start to care. Is it the best stories that show up in AI results over and over again? Are they going to make a hundred $1,000? Oh, that that's going to catch people's attention.

The Economics of RSL: Aiming for Standard Internet Pay Rates

• Stubble believes that the economics of RSL should align with standard internet pay rates.
• He uses the ad model as a benchmark, suggesting that 1,000 clicks should generate around $5.
• However, he acknowledges that AI-generated summaries may "steal" clicks, requiring AI companies to offer higher compensation.
• He also emphasizes the importance of validation and having ideas spread, which are diminished by AI summaries.
• In order for RSL to create a healthy ecosystem, AI companies may need to offer compensation above standard internet pay rates.

The Enduring Value of Writing and Reading

• Stubble dismisses concerns that AI will replace writing and reading, arguing that "writing is thinking and reading is learning."
• He believes that humans will always value being smarter, and writing and reading are essential for intellectual growth.
• He draws a parallel to Cliff's Notes, which did not replace books, suggesting that summaries will not replace original stories.

Conclusion

Tony Stubble's perspective highlights the urgent need for AI companies to fairly compensate content creators. RSL represents a collective effort to force AI companies to the negotiating table and establish a sustainable model for the future of the internet. Medium's commitment to passing all revenue back to writers sets a positive example for other platforms to follow. While the long-term impact of AI on writing and reading remains uncertain, Stubble remains optimistic about the enduring value of original content and the human desire for knowledge.

Human Native's Pivot: From AI Content Marketplace to Data Preparation for AI

Key Concepts:

• AI Content Marketplace: A platform connecting IP holders with AI companies for content licensing.
• Data vs. Content: AI companies view content as data for model improvement, while rights holders see it as their life's work.
• Fair Use: The legal doctrine allowing limited use of copyrighted material without permission.
• Rag (Retrieval-Augmented Generation): Using real-time data to help serve a query.
• Not Invented Here Syndrome: A reluctance to adopt ideas or technologies from external sources.
• Vector Search: A method of searching for similar items based on their vector representations.

The Initial Vision: Connecting IP Holders and AI Companies

• Human Native initially aimed to build an AI content marketplace, connecting IP holders with AI companies seeking to license content.
• The assumption was that publishers and authors would be eager to license their data and get paid for it.

The Mismatch: Data vs. Content and Conflicting Priorities

• The company encountered a "communication breakdown" between rights holders and AI companies.
• Rights holders viewed content as their "life's work" and emphasized respect, editorial control, and fair compensation.
• AI companies viewed content as "data" for improving models and focused on ROI, specific requirements, and model performance.
• This fundamental mismatch made it difficult to facilitate negotiations and reach agreements.
• Publishers wanted high price and low shared control, while AI companies wanted low price and high levels of control.

The Challenges: Data Structure, Discoverability, and Distrust

• Rights holders often lacked a clear understanding of their own content, making it difficult to fulfill specific AI company requests.
• Many had been "throwing" content into cloud storage systems without proper organization or metadata.
• Philosophical differences regarding fair use and copyright further complicated negotiations.
• There was a history of distrust between media organizations and large tech companies, making it challenging for a startup to mediate.

The Pivot: Focusing on Data Preparation and Internal AI Usage

• After 15 months, Human Native decided to pivot away from the marketplace model.
• The company realized that building a low-friction platform for trillions of transactions required more standardization and precedence.
• The team's expertise lay in building product and technology, not in sales or mediation.
• The new focus is on helping companies make their data assets useful for both internal AI usage and external licensing opportunities.

The Rag Use Case: Still the Same Fundamental Disconnect?

• While the RAG use case has become more important, Human Native believes it still faces the same fundamental disconnect.
• Publishers may be reluctant to participate in RAG systems where they lose control over the user experience and their direct relationship with customers.
• AI companies may question the need to pay for information that is already available on the free internet.
• AI companies often prefer to own the technical infrastructure for RAG systems, making it difficult for third-party solutions to gain traction.

The Solution: Helping Companies Answer Sophisticated AI Buyer Requests

• Human Native aims to help companies answer sophisticated AI buyer requests by making their data assets useful.
• For example, if an AI company wants to license only videos of sunscreen bottles with busy backgrounds, Human Native can help content companies identify and extract those specific assets.
• This involves building an internal search engine and taxonomy generator to extract meaning and value from images, video, and audio.

The Technology: A Blend of Traditional and Generative AI

• The system will use a combination of traditional software engineering, data engineering, and generative AI techniques.
• Generative AI can unlock new abilities to understand content at scale.

Preparation and Exploration: Two Sides of the Same Coin

• Human Native believes that preparation of content enables exploration, and the two are fundamentally linked.
• The company is building an interface for understanding and searching content, based on the work they have done to prepare the data.

A Unified View: Data Location Agnostic

• The solution aims to work regardless of where the data is stored, whether on-premise, in multi-cloud environments, or in services like Box, Dropbox, or SharePoint.

Building Something on Their Own: Facilitating Internal AI Innovation

• Human Native is also helping companies build their own AI products on top of their data.
• This involves taking a third-party or open-source model and applying the company's archive to create a product that deep dives into their content.
• Companies can also use a RAG system to bring in content they don't have, creating an even more compelling product.

The Future: Business and Operational Leadership, Not Just Technical

• The barrier to entry for using AI systems is getting lower, making it possible for non-technical people to leverage them.
• Business and operational leadership may become more important than technical leadership in driving AI innovation.

Conclusion

Human Native's pivot reflects the challenges of building a shared marketplace for AI content licensing. By focusing on data preparation and internal AI usage, the company aims to empower content holders to unlock the value of their assets and build innovative AI products. The future of AI innovation may lie in enabling business and operational leaders to leverage AI systems without requiring deep technical expertise.

TE Trucks: Building a Mini Electric Truck for Urban Environments

Key Concepts:

• Mini Truck: A compact pickup truck designed for urban environments.
• Crew Cab: A truck cab with two rows of seats, accommodating multiple passengers.
• Virtual Validation: Using software simulations to validate vehicle designs before building physical prototypes.
• Unit Profitability: The point at which a company's revenue from each unit sold exceeds the cost of producing that unit.
• K Truck: A small, lightweight truck manufactured in Japan, often imported to the US.

The MT1: A Small, Adorable, and Functional Truck

• TE Trucks is building the MT1, a mini truck with a crew cab and a five-foot bed.
• The MT1 is designed to be smaller than a two-door Mini Cooper, making it the smallest vehicle on US roads.
• Despite its small size, the MT1 is capable of doing "truck stuff," such as hauling cargo and towing.

The Inspiration: Solving the Urban Truck Problem

• Jason Marks, the co-founder and CEO of TE Trucks, is a truck guy who lives in downtown San Francisco.
• He found it difficult to navigate and park his Toyota Tacoma in the city.
• The MT1 is designed to solve the urban truck problem by providing a functional truck in a compact package.

The Unique Positioning: Automotive Safety and Electric Vehicle Expertise

• Marks has a background in mechanical engineering and automotive safety.
• He worked on the safety systems for some of the first electric pickup trucks.
• He understands how to repackage the front of a vehicle to shrink its footprint by removing the large engine block.

Manufacturing in the USA: A Focus on Unit Profitability and Scalability

• TE Trucks plans to design and build the MT1 in Irvine, California.
• The company is taking a phased approach to manufacturing, starting with small volumes and gradually scaling up.
• They are contract manufacturing some components, such as stamped steel structures, while building battery packs in-house.
• This approach allows them to control costs and achieve unit profitability at moderate volumes.

Performance and Capabilities: Towing, Payload, and Range

• The MT1 is expected to have a 152-inch length and be able to tow 6,600 pounds and carry a payload of 2,000 pounds.
• While towing has a significant impact on range for EVs, payload has a less substantial impact.
• The company sees towing as a key use case for the MT1, particularly for urban work trucks.

Target Market: Urban Truck Users and Fleets

• TE Trucks is targeting urban truck users who need a functional vehicle for hauling cargo and towing in the city.
• Fleets have also shown interest in the MT1, particularly for its cargo-carrying capacity.

Pricing and Competition: A Unique Capability

• The MT1 is expected to start in the low $40,000s, depending on accessories, range, and motor options.
• The company is not aiming to be the cheapest option on the market but rather to offer a unique capability that doesn't exist today.
• The main difference between TE Trucks and competitors like Slate is that the MT1 holds four passengers and is more capable as a city vehicle.

Market Size: Three Million Trucks Sold in Downtown Cities Every Year

• Three million trucks are sold in downtown cities every year, indicating a large potential market for the MT1.
• The company also sees latent demand for mini trucks, as evidenced by the popularity of imported Japanese K trucks.

The American Automakers: Incentives and Ground-Up Redesign

• American automakers have been incentivized to make larger trucks due to emissions regulations.
• Building a mini electric truck would require a ground-up redesign, which can take eight years in traditional automotive.

Virtual Validation: Accelerating the Design and Development Process

• TE Trucks is using virtual validation to accelerate the design and development process.
• This involves building and validating vehicles entirely in software before building any hardware.
• The company is also using AI tools to speed up crash testing and other simulations.

Pre-Orders and Production: A Phased Approach to Scaling

• TE Trucks has nearly 12,000 pre-orders for the MT1.
• The company plans to deliver its first vehicles to early access customers in 2026.
• They will then build 500 vehicles, followed by 5,000 if they are happy with the results.
• The company expects to reach unit profitability at 5,000 vehicles and corporate profitability between 10,000 and 20,000 vehicles.

Funding: Capital Efficiency and Future Fundraising

• TE Trucks has raised $6 million to date.
• The company will need to raise at least another $100 million to reach unit profitability.

Conclusion

TE Trucks is building a mini electric truck that is designed to solve the urban truck problem. The MT1 offers a unique combination of functionality, compactness, and electric power. The company is taking a phased approach to manufacturing and is using virtual validation to accelerate the design and development process. With a large potential market and a strong team, TE Trucks is well-positioned to disrupt the truck market.