Sanders’ 50% AI Stock Plan Turns Training Data Into a Political Fight

Yoland Yan described ComfyUI as “the polar opposite” of the prompt boxes people know from ChatGPT or Midjourney. In those tools, the user types into a black box, changes a word, and may get an entirely different image. ComfyUI exposes the machinery instead: a node-based interface in which image and video generation are broken into connected components, with parameters visible and editable.

That visibility is the point. Yan said ComfyUI is “very complex” and takes time to learn, but that complexity is why it has become a production tool. He said it is now becoming “a professional skill required in many studios,” including Netflix, because it gives creatives control over variables that consumer tools hide: noise, width, height, the prompt, the model, the seed, and the relationships among those choices.

In the demo, Yan showed a text-to-image workflow using Ideogram V4, which he described as a new open-source image model from a team he believed was based in Canada and the United States. The ComfyUI canvas contained connected nodes, exposed parameters, and a preview of a generated poster-like image of a skateboarder. Jason Calacanis read the interface as something closer to Photoshop layers or production software than a chatbot prompt: many boxes connected by arrows, each governing some aspect of the output.

Yan’s example used not just a single prompt but a structured description of the image. One visible text node specified the word “COMFY” and described it as “Massive 3D puffy, inflatable white typography spelling ‘COMFY’,” with a white and gray palette. Yan explained that Ideogram V4 allowed bounding boxes, so the user could tell the model where on the canvas a particular element should appear. In the output, the puffy “COMFY” text appeared in the designated area, acting as a cloud-like backdrop.

Calacanis called this a set of “micro prompts within a mega prompt.” In ordinary image tools, he said, the model decides where the logo, person, and other elements go. In ComfyUI, the user can specify the logo placement, the person placement, the font, and other details. Lon Harris added that the difference is like a slot machine versus a controlled workflow: using a model directly can mean repeatedly pulling the lever and hoping the image lands; ComfyUI lets the user adjust weights, LoRAs, and other controls to get closer to the desired result the first time.

Yan emphasized reproducibility as a major production requirement. Many prompt-box products hide parameters such as the initial seed, he said, which means the same prompt will not reliably produce the same result. In ComfyUI, a fixed seed can make a generation repeat exactly. That matters for creative teams, because the same image or workflow can be reproduced by the original creator or someone else in a production pipeline.

When Yan changed the seed, the image changed. When he kept the same configuration, the output could be regenerated. The demo showed the diffusion process as it generated the image, making visible what is normally hidden behind a consumer product’s progress indicator.

Calacanis pressed on how the prompt itself is written. Yan said users can do it in several ways, including by asking a model such as Claude to generate a structured prompt from a simpler instruction. That led Calacanis into a broader claim: many users still misunderstand what AI should be doing for them. He argued that “the number one job of AI is to write the prompt,” and described an All-In workflow in which a model produced a highly structured prompt for gathering news aligned with the show’s interests.

Harris agreed that different models are good at different parts of a workflow: Claude may be especially useful for prompt writing, while image models such as Ideogram or “Nano Banana” may be better at producing visuals. The “magic,” Harris said, often comes from combining models rather than treating one model as the tool for everything.

The “wrapper” critique did not fit Yan’s account of ComfyUI. Jason Calacanis asked him to explain why ComfyUI is not merely an interface on top of other people’s models, and how the company thinks about open source as either a weapon or a liability. Yan’s answer was that ComfyUI sits across multiple layers, including a technical layer that is not just interface.

At the core, Yan said, ComfyUI is an inference engine. It performs the calculation needed to run diffusion models. For people who understand the difference between training and inference, he said, there is a “huge gap” between a model and an inference engine that can unlock the model’s value, speed, and capabilities in practical use. He described ComfyUI as “the most popular diffusion model inference engine,” and said the open-source project ranked around the top 70 GitHub projects by stars.

The other part of ComfyUI’s position is extensibility. Yan said anyone can write a custom node. If the baked-in version of ComfyUI does not contain a capability, users can bring their own code or use Claude to write the code for a new node and connect it into the pipeline. In an agentic coding world, he argued, that makes ComfyUI a foundation layer: an “operating system” for visual generative AI.

The demo made that claim concrete without requiring the viewer to understand every node. Yan opened a subgraph inside the workflow — a component that encapsulates functionality so users can abstract away some complexity — and showed the lower-level nodes behind the image generation. The visible workflow included nodes for loading a diffusion model, encoding the positive prompt with CLIP, choosing a sampler, guiding the model, setting latent dimensions, generating random noise, and configuring a parameter called CFG. Yan described CFG at a high level as controlling how much guidance the prompt injects into the model. It can be turned up or down, changing how tightly the output follows the prompt versus how much freedom the model has.

Harris said this kind of tool made him better at generating images because it exposed the relationship between prompt words, weights, and final outputs. A setting as simple as how much fidelity the model should have to the prompt can produce very different images. For Harris, ComfyUI was valuable precisely because it showed what was “under the hood.”

Yan also drew a distinction between local and hosted use. ComfyUI is open source and can run locally, including on a Mac or Windows machine with a sufficiently capable GPU. Some models will run on a high-end Mac, he said, but performance requirements can be large, and he recommended Nvidia chips for many local workloads. Users who do not want to run the software locally can use ComfyUI’s cloud version for access to state-of-the-art graphics hardware. Yan also said some studios run ComfyUI in enclosed environments because of privacy or air-gap requirements.

The business model reflects that split. Yan said the local version is completely free. Users can still use closed-source models through ComfyUI’s API platform and pay as they go. The cloud version has basic tiers at $20, $35, and $100, depending on usage needs. For enterprises, ComfyUI offers customized solutions with the features and security those customers require, sold through a sales-led motion.

Calacanis said ComfyUI had raised $30 million from Craft at a $500 million post-money valuation. Yan confirmed the round and said the company was hiring “cracked engineers” and operators. The most urgent hiring needs, he said, were scaling the sales team and the front-end development team. He also said the company had only been able to process roughly 100,000 inbound requests because of very high demand.

Calacanis’ operating advice was not about the product but about where to build the sales team. He warned Yan against trying to hire a San Francisco sales force in a market where large companies and well-funded startups can overpay experienced sales executives. His suggestion was to scout cities such as Phoenix, Salt Lake City, Austin, Dallas, Houston, or Las Vegas — places with experienced sales talent, lower costs of living, and workers more willing to accept lower base salaries with higher commission upside.

His example was a company that placed sales recruiting ads in several cities, set up in airport Marriotts, interviewed candidates over a compressed schedule, and compared the quality of the candidate pools city by city. He framed it as the kind of founder action that beats consultant-driven planning: pick five cities, spend two days in each, and learn where the talent really is.

The most tangible demonstration of ComfyUI’s production value came in Yan’s explanation of outpainting. Harris defined the distinction: inpainting means putting something into an image that was not already there; outpainting means extrapolating beyond the boundaries of the original frame. If the original image shows Dorothy in a small window of Kansas, outpainting asks what the rest of Kansas would look like outside that frame.

Yan said ComfyUI had been used by studios working with the Google Veo team on the Wizard of Oz Sphere project, though he was careful to say ComfyUI was the tool builder and did not directly participate in the production. The problem, as he described it, was not simply upscaling an old film. A 1920s color film cannot be stretched into 32K video for the Sphere without distortion. If an image is scaled directly into that environment, a face can warp around the globe and look wrong to the audience.

The production task was to create imagery that had never existed in the original film. That could mean expanding the frame, generating background environments, or producing characters reacting to the main action. For that, Yan said, studios used ComfyUI with models such as Veo to run many parameters and find the best outpainted results while maintaining quality and control.

Yan then showed a simpler version using a clip of Calacanis from a prior show. The original was a vertical video. A ComfyUI workflow using an LTX open-source video model outpainted the sides into a wider frame, generating Calacanis’ arm and background beyond the original frame. The result was not perfect — Calacanis noticed it had put him in a short-sleeve shirt — but it made the point. The system was guessing what existed outside the original video and turning a vertical source into a wider shot.

Yan also showed a reference-to-video workflow that transformed Calacanis into “Señor JCal,” with a sombrero and maracas, using a reference image and a style detector. The example was comic, but it showed the same underlying pattern: the interface is not a single magic prompt but a pipeline that can ingest video, reference images, prompts, style information, and model choices.

That production framing also explains why ComfyUI’s complexity is not incidental. Consumer tools optimize for ease and a one-shot answer. ComfyUI optimizes for control, reproducibility, and composability. Calacanis called it “industrial strength” prompting. Yan’s preferred framing was “production level”: a product that enables fine-grain control and quality for AI.

Yan said Coca-Cola ads and recent AI Super Bowl ads were generated using ComfyUI. Harris also mentioned the Wizard of Oz Sphere workflow. The boundary of Yan’s claim matters: ComfyUI was described as infrastructure and tooling used inside professional pipelines, not as the credited producer of those works.

The AI policy argument turned on a sharp tension: the same companies pushing frontier models are also warning about the risks of moving too fast. Lon Harris said Anthropic’s heads of internal research and policy had written that slowing the pace of global AI development would “likely be a good thing.” Alex Wilhelm read another quoted passage from Anthropic: “We believe it would be good for the world to have the option to slow or temporarily pause frontier AI development to enable societal structures and alignment research to keep up with the advance of the technology.”

The concern, as Wilhelm summarized it, was recursive self-improvement: AI models showing the capacity to self-improve without human intervention. Anthropic’s position, as presented by Wilhelm, was not merely that AI needs safety research, but that society may need an option to slow or pause frontier development so institutions and alignment work can catch up.

Calacanis found the timing incoherent. He connected Anthropic’s warning to its own recent conduct around a model he called Mythos, which he said the company had withheld while acting responsibly. He argued that Anthropic already has the ability to slow down and had used that discretion correctly. He cited an interview with Nikesh Arora of Palo Alto Networks, who, according to Calacanis, said the company used Mythos to find internal attack vectors and close them, and that the model found bugs at a level of performance they had not seen before.

His objection was not that Anthropic should ignore safety. It was that Anthropic calling for a global slowdown while operating one of the leading AI businesses creates an obvious political opening. Wilhelm put the critique more bluntly: if Anthropic genuinely believes AI is too dangerous and development must slow, “you first.” Shut down Claude if it is too dangerous; otherwise, Anthropic is one of the actors driving the development it is warning about.

That tension matters because Anthropic’s institutional posture is not isolated. Calacanis immediately connected it to Senator Bernie Sanders’ proposal for public ownership of major AI companies. If a frontier lab tells the public that AI may displace jobs, require pauses, and threaten society’s ability to keep up, Sanders can incorporate that into a broader case: the companies themselves say the stakes are enormous, and therefore the public should share in ownership and governance.

Calacanis described Sanders’ proposal as “completely insane” and “deranged,” but he also said it may be one of the arguments most likely to resonate with the American public. The source played a clip from Sanders’ video, in which Sanders argued that artificial intelligence may be “the most transformational technology in the history of the world,” affecting the economy, democracy, emotional well-being, environment, education, and child-rearing. Sanders also warned that as AI becomes smarter than humans, it could eventually function independent of human control with catastrophic consequences.

Sanders’ central question was ownership: who will own and control the AI future, who will benefit, and who will be hurt? He argued that AI was not created “out of thin air” but was built on “our collective human intelligence”: books, songs, artwork, journalism, code, scientific research, videos, conversations, images, and ideas across generations. He quoted Sam Altman as acknowledging that AI models were trained on the “collective experience, knowledge, and learnings of humanity.”

Calacanis repeatedly interjected that Sanders was right about parts of this. “Training data,” he said. “Where’d they get it from?” When Sanders said big tech oligarchs had fed this knowledge into their models without permission, acknowledgment, or compensation, Calacanis responded, “That’s true.” When Sanders said the creative work of millions had been stolen by the wealthiest people in the world, Calacanis said there was “some truth to that.”

Sanders’ proposed remedy was the American AI Sovereign Wealth Fund Act. As described in the clip, it would give the public a direct ownership stake in the largest AI companies in America through a one-time 50% tax on stock, not profits. Sanders argued this would give Americans a direct role in determining the future of AI and ensure that trillions potentially created by AI improve the lives of all people rather than enriching only the richest.

Calacanis rejected the policy mechanics but took the political argument seriously. His view was that Americans already believe AI companies took intellectual property without permission. He pointed to lawsuits, settlements, and licensing deals over training data and copyright as part of the public record voters will absorb, not as settled proof of liability or an admission by every AI company. The New York Times lawsuit was mentioned as an example. Calacanis said the courts will decide what training uses are legally permissible, and he believed early signals suggest some amount of training on copyrighted content may be allowed.

Wilhelm offered a more specific characterization of the legal picture as he understood it, while stressing that more lawsuits could change the picture. In his summary, there is a distinction between illegally obtained copyrighted material and copyrighted material that was legally accessed. He cited Anthropic’s alleged use of a bootleg archive of books as an example of conduct that, in his view, raises a different problem from buying a book, scanning it, and training on it. He also said AI outputs that infringe copyright are actionable. The discussion treated the legal field as unsettled rather than closed.

Calacanis’ political point was that the legal nuance may not matter to voters. Sanders’ argument combines three grievances: companies took everyone’s content, used it to build trillion-dollar businesses, and now those same companies warn they may take jobs. Calacanis said he did not necessarily agree with that argument, but he believed the public would. He predicted it would appeal across factions: people who believe the rich do not pay enough taxes, people who care about intellectual property, people worried about job loss, parts of the left, and parts of the right.

Wilhelm described a split on the political right. On the left, he said, “basically everybody is against AI.” On the right, he saw a divide between a pro-business, pro-AI faction associated with figures such as JD Vance, Peter Thiel, and David Sacks, and a more traditional MAGA faction associated with Steve Bannon that has moral, ethical, labor, and free-market concerns about AI concentrating wealth and power.

Calacanis predicted that AI companies will try to negotiate before anything like a 50% stock tax becomes real. He suggested Anthropic and OpenAI might sit down with Sanders and offer something like 10% of equity dedicated to public service, job retraining, Invest America accounts, or another social mechanism. Wilhelm called Sanders’ 50% proposal directionally correct as an opening salvo: start at 50%, the companies counter at 10%, and now there is a dialogue. He also argued that giving Americans some ownership exposure could make AI more acceptable to a public that is otherwise skeptical or hostile.

Calacanis also noted that OpenAI has a nonprofit structure that owns some percentage of its future value, though he did not know the exact amount, and suggested that such a structure could provide a vehicle for a public-benefit commitment. He said Anthropic and OpenAI, through licensing deals and settlements, have at least acted in ways that Sanders could use to reinforce the training-data critique, even if the legal conclusions remain unresolved.

The Sanders argument became a larger claim about income, taxation, and social programs under AI-driven labor disruption. Calacanis tied together three positions from major AI figures: Anthropic warning of job displacement and the need to slow development, Sam Altman funding universal basic income research, and Elon Musk advocating “Universal HIGH income” through federal checks as a response to AI unemployment.

A Musk tweet shown on screen said: “Universal HIGH income via checks issued by the Federal government is the best way to deal with unemployment caused by AI. AI/robotics will produce goods & services far in excess of the increase in the money supply, so there will not be inflation.” The screen showed it as an Elon Musk post with a visible timestamp of April 16, 2024 and 69.4 million views, but no direct post URL was provided in the source.

Wilhelm then summarized the Altman-backed UBI study. According to Wilhelm, the three-year study gave 1,000 low-income adults $1,000 per month, while a control group of 2,000 people received $50. Researchers expected to see whether steady cash reduced motivation to work. Instead, participants reported valuing work more. Their belief in the importance of work rose slightly, and many agreed with statements such as “work is a duty towards society” and “people who don’t work turn lazy.” Participants worked fewer hours on average, but did not stop working. They used the financial cushion to make career decisions: returning to school, pursuing certifications, switching to jobs with greater long-term potential, or accepting temporary pay cuts for better growth.

Calacanis interpreted the study as evidence that a modest cash cushion gives people flexibility rather than making them idle. A thousand dollars per month, he said, is not enough for most people to quit working, but it might cover half of rent or let someone take a lower-paying role with better long-term upside. His example was a waiter or bartender making strong nightly income but lacking a career path, compared with joining an associate training program at a lower initial salary but with learning and access.

From there, Calacanis made his central political prediction: the 2028 presidential election will be the AI job-displacement election. Not inflation, not wars, but whether AI is a net positive or negative for Americans. The attack line, he said, will be that AI companies became worth trillions while jobs declined and a small group of Silicon Valley founders, employees, venture capitalists, and equity holders benefited.

His proposed response was a “grand bargain.” The first part was tax simplification and relief for lower earners. He floated thresholds rather than a finished plan: people making under $50,000 might pay nothing; under $75,000 might pay a token $750 flat amount; under $100,000 might pay something like $2,000. The purpose would be to address what he called a K-shaped recovery, in which equity holders and the wealthy benefit while lower-income and non-equity-holding workers fall behind.

The second part was broader ownership through Invest America, also referred to as Trump accounts, which he described as giving children a stake in equity. The third and most ambitious part was replacing complex means-tested welfare programs with direct cash.

Calacanis listed food stamps, unemployment, SNAP, SSI, housing, Medicare, Medicaid, and other programs as examples of systems that are complicated, paperwork-heavy, and administrator-heavy. His idea was to calculate the total cost of these programs and distribute a percentage of that pool directly to people under income thresholds, with benefits phasing out as income rises. He called it “kind of” UBI, but framed it as a simplification of existing social spending rather than a wholly new entitlement.

Harris agreed that both Democrats and Republicans could accept the principle because few people are attached to “elaborate bureaucratic nightmare systems.” He pointed to Los Angeles homelessness spending as an example where, in his view, money meant to help people can be lost to waste or graft. He argued that if people simply received checks, the government would not need as many complicated eligibility systems. He also criticized merit-based and means-tested programs as full of criteria, applications, and exceptions that confuse the people they are meant to serve.

Calacanis guessed that waste, fraud, abuse, and administration may mean less than 50 cents of each dollar reaches recipients, though he framed that as a guess. Harris supported the general claim that social welfare programs often spend large shares of their budgets on determining eligibility, paperwork, offices, and staff.

The argument was not presented as a finished policy platform. It was a forecast of where AI politics may go if the public becomes convinced that automation-driven wealth is being captured by a narrow ownership class. Sanders’ 50% stock proposal is extreme in Calacanis’ view, but the underlying pressure — public ownership, direct cash, job retraining, and a simplified safety net — is likely to grow.

Brian Chesky’s new AI lab raised a familiar startup question: when is an AI product merely a wrapper, and when is it a vertical product with its own interface, workflow, data, and distribution? Harris said Chesky will found the lab while remaining Airbnb CEO, and that the lab is focused, at least initially, on user interaction and design rather than simply building another chatbot. Chesky’s argument, as Harris summarized it, is that AI for travel needs a richer interface and better design. People do not just want to talk to a chatbot about a trip; they want something more bespoke, designed, and customized. Harris noted that Chesky studied design before founding Airbnb.

Calacanis connected that to a Launch accelerator company called Roam Around. The company pitched in September 2023 as “an AI concierge in your pocket.” In the demo shown, founder Shai described entering “two days in Barcelona” and receiving a complete itinerary. If the itinerary was not right, the user could say “it’s our anniversary,” and the app would revise the plan with romantic restaurants, couples massage, hot-air ballooning, and similar additions. The user could add a beach day, translate the itinerary to Spanish, and share it by SMS, WhatsApp, or email.

The pitch itself anticipated the wrapper critique. On screen, Roam Around asked, “Aren’t you just a ChatGPT wrapper?” and listed potential moats: reinforcement learning from human feedback based on user responses, bespoke recommendations, fine-tuned models and embeddings, and partnerships.

Calacanis said he had begged the founder to keep going. His argument at the time was that being accused of a wrapper should not end the company. There are many things around a model — workflow, interface, brand, features, distribution, human-in-the-loop services, data, vertical focus — that can become the business. He compared a Cadillac using a Corvette engine to a wrapper: the engine may be shared, but the product, brand, look, feel, and value are different.

Roam Around was eventually purchased by Layla, according to Calacanis. But he still viewed the company as a missed opportunity because it had been early in a large category. If the team had persevered for three years, he argued, they might have built a vertical travel model, features around existing models, or another differentiated product in the space Chesky is now taking seriously.

His broader founder lesson was to persevere in important categories. If the market is big, the company has a lead, and there is cash in the bank, he believes the correct decision is often to pivot within the category rather than shut down and return capital. If a founder no longer believes in the original vision but has a stronger second idea, Calacanis said investors should often support the pivot. The money is already in the company; returning it may send investors back only a fraction of the original amount. His preferred approach is to “play the hand” and see whether the founder can figure it out.

Harris connected this back to his earlier point about multi-model workflows. A travel product may be more useful if it pulls together different tools for different parts of a trip rather than simply asking ChatGPT what to do in Rome. The value could be orchestration, not one model.

Calacanis also raised a governance issue for Chesky. If Airbnb’s CEO creates a separate AI lab that could serve travel, hospitality, or transportation companies, Airbnb shareholders may want exposure. Calacanis suggested Airbnb might put in the first $100 million for 10% and a perpetual license. Otherwise, if the lab succeeds and serves Expedia, Uber, Hilton, Marriott, or competitors, conflicts could arise. His view was that Chesky is a strong enough designer that a design-first AI product is worth watching, but the relationship to Airbnb shareholders needs careful handling.

The late discussion of creator-led films and stunt platforms sharpened the same incentive question in a different market: tools and distribution systems make new behavior possible, and attention markets decide what gets amplified.

Harris pointed to YouTuber-led films such as Obsession and Backrooms as examples of internet-native intellectual property reaching theaters. Obsession, he said, came from YouTuber Curry Barker and collaborator Cooper Tomlinson after they built an audience with sketches and horror shorts. The film’s premise is simple: a man uses a magic-shop device to wish that a woman in his friend group loved him more than anyone else, breaks the device, and finds that the wish worked in a terrifying way. Harris said actress Indi Navarrette’s performance becomes genuinely frightening.

He described Backrooms as another example of internet-native IP: a YouTube and creepypasta phenomenon about people trapped in a yellowed, liminal maze of hallways. Harris said it shows IP does not have to mean Transformers, Star Wars, or a long-running television show. If enough Gen Z viewers care about a YouTube concept, it can become theatrical IP. Harris said Obsession cost $750,000 and had grossed more than $166 million globally at the box office. Calacanis compared the return profile to The Blair Witch Project and said the future may include more one-to-twenty-million-dollar films.

The risk side of attention became clearer when Harris described Pump.fun’s new bounty product, Go. Harris characterized it as a platform where anyone can create or complete bounties for any task, with unlimited rewards. The promotional visuals showed bounties such as creating a game character, holding a meme onscreen at the pyramids, writing “Schillhouse” on a front door, jumping into a public fountain, and rallying for a meme coin at the World Trade Center. Harris compared the premise to dystopian films and shows in which people are paid or incentivized to do dangerous, humiliating, or disturbing things.

Calacanis distinguished this from Fiverr, Upwork, or Rent-A-Human, where people can hire others for legitimate tasks. His concern was that a crypto-native bounty platform will push toward public disturbances and dangerous stunts. Even if some tasks are merely silly, he said, “it always goes too far” and someone gets hurt. He asked whether anonymous users with crypto wallets could put up money for someone to hang off the Golden Gate Bridge. Harris said the creation process, as described, requires connecting an X account and wallet, setting a description, timeframe, deliverables, and escrow payment, with Pump.fun reviewing submissions; he did not see a public-face disclosure requirement.

Calacanis’ proposed alternative for spectacle was not to eliminate risk entirely but to design safety into it. He criticized entertainment built around the possibility that someone might die, using Alex Honnold climbing without a rope as an example. If Honnold climbs a building, Calacanis suggested, put a net 10 stories below that rises with him. The feat remains exciting; the fall is still consequential; but death is not the entertainment product.

The thread echoed the earlier AI debate in cultural form. ComfyUI’s value came from giving professionals more control over powerful systems. The Sanders and UBI arguments were about public control over AI’s economic consequences. The bounty-platform example showed the inverse: when incentives are unconstrained, the market can turn human risk itself into the product.