Frank Portman

The through-line

Frank Portman's consistent argument is that privacy is not a tax on capability but an engineering frontier that has barely been explored, and that any company building AI infrastructure which does not treat it that way from the first day will be punished, first by consumers and only later by regulators. He frames Yobi around building "the most expressive foundation models of user behavior with privacy and security built in from day one" [1], and the operational consequences run through everything: consent and payment for training data, whole verticals refused, customers vetted the way data partners vet Yobi, and models deliberately trained to be bad at inferring who someone is [1][6]. The second, quieter through-line is a distaste for AI-by-association: he defines his company first by what it is not, and argues that some problems, intent prediction among them, need more than a large language model [1][3][6].

On what Yobi is not

He opens the description of the company with a negative on purpose, because of how crowded and loosely labelled the space has become: "we are not a large language model product or company. We're not even necessarily in Gen AI per se" [1]. What Yobi builds instead are foundation models grounded in real user signal, and even those are not the product. The behavioural foundation model is "the foundation on which we build different other products," with the business itself being software as a service that gives companies "access to the benefits of this kind of rich differentiated data without any of the privacy risks, legal risks and all that other stuff" [1]. He has made the same case in technical terms elsewhere, arguing that predicting user intent takes more than a large language model [3]. One of the products sits in the advertising vertical [1].

On consent, compensation and clear ownership of data

Yobi's data comes through partners "that can actually get consent from their own users to use their data in this way, which is for model training purposes," and it "typically comes with some form of financial compensation from us" [1]. He treats the resulting clarity as the asset: "we have a very clear distinction of who owns what data, who's getting paid for what data, and what is able to be used in the models" [1]. Data partners run substantial due diligence before working with Yobi, and he applies the same standard in reverse, vetting the companies that buy the products. He is candid that this cuts against commercial instinct: "when you sell products, you want as little friction as possible," but the friction is "totally worth it" and, he believes, containable without becoming overbearing [1].

On the use cases he will not touch

Some applications are simply out of scope, and the two he names are financial underwriting, meaning insurance and loan decisions, and health. The reason is not that they are regulated verticals. "That's not even necessarily what's driving our decisioning there because they're regulated, but you are allowed to operate in them. We do not want to work in these highly sensitive fields," he says, because he does not want "to leave any possibility for bias or unfair outcomes" [1][6].

On training models to be bad at demography

The most distinctive technical commitment he describes is a deliberate inversion of the usual targeting logic: "under no circumstances do we feed demographic information into our models nor do we ask our models to even be good at predicting demographic information." Yobi goes further and adds adversarial objectives "so that they are rewarded for being bad at predicting demographic information" [1][6]. He notes the awkward commercial edge this creates, since marketers and agencies routinely segment by age, household income and gender, both to explain what happened and to decide who to target next, and they ask Yobi for that data as an add-on. The answer he has to give is that the company holds the data "but we only use it to make sure that we are bad at recovering that data," which has led to some funny conversations with customers [1].

On privacy and performance as an up-and-to-the-right problem

He rejects the premise that scale and privacy are inherently at odds, while conceding the trade-off may exist somewhere far ahead. Picture a graph with privacy on one axis and performance on the other: moving straight up is "the pure model R&D tuning problem," more compute, more expressive methods, better training objectives, more willingness to pay for the electricity behind the GPUs. Moving right without moving down comes from better data matching, encryption, and staying at the research frontier of homomorphic machine learning, which he insists are not buzzwords for Yobi, which intends to publish research in the area [1]. The roadmap does both columns at once, delivering customer wins now while moving up and to the right, and he concludes that "we have a very long time before anybody here has to have the difficult ethical philosophical question of what it might mean to trade something off" [1]. The concrete instance of this is a medium-term goal of letting the foundation models learn from customer data without it leaving the customer's environment. It has been specced, prototyped and simulated, and he expects neither a performance loss nor a lowered ceiling on future performance [1][6].

On cloud infrastructure and the trust chain

Privacy engineering fails if any link in the chain is untrusted, and he uses the cloud layer to make the point: two parties can be satisfied with each other's infosec and anti-bias practices, "but if the cloud environment is itself not trusted, then you're at an impasse. It's like, are we driving hard drives over to each other or what's going on here?" [1] The partnership with Microsoft covers both cloud infrastructure and go-to-market scaling, and extends to the more trailblazing pieces of the business, clean rooms and deploying Yobi's models directly inside customer accounts, which he calls "the highest tier of privacy because their data does not even have to leave their servers" [1][6]. He reads Microsoft's interest plainly: AI needs data and AI runs in the cloud, so facilitating this kind of product keeps them at the forefront of the privacy-safe version of the field [1].

On why "move fast and break things" expired faster this time

His diagnosis of the industry is that AI is relearning a lesson software already absorbed, without the grace period. The broader software industry had "about 10, 15 years to figure that out and almost even had a soft landing on the way," whereas "the AI industry, they got to the break things in about a month and then in the next month there's all sorts of regulatory scrutiny, all sorts of consumers demanding accountability" [1]. He expects the market rather than legislation to do most of the enforcing: regulators like GDPR will be there, but the sharper signal is consumers, and his evidence is domestic. His own parents, slow adopters of any gadget, now ask him which chat company to use because they want to know which one will not take all their information and use it as training data. "People are right to be very sensitive to these things," and a company not thinking about it from day one "is going to be in a lot of trouble" [1].

On scraping the open web, and where he declines to plant a flag

He raises the provenance question behind the leading large language models, noting that as far as he can tell the foundation for essentially all of them came from scraping the open internet, from engineering forum posts to the collected works of William Shakespeare to open-domain video [1]. The tension he sits with is that this is content "created by humans for humans," hosted either as a public good funded by donations or monetised through the attention economy via ads and subscriptions, and Shakespeare had no concept of what his plays might feed into. He is explicit that he does not want to appear to hold a strong position, because restricting scraping entrenches "the people that already have all the data and all the money" [1]. He watches it half as a consumer interested in tech and society and half as a CTO whose core value proposition is differentiated data obtained without loading customers up with legal, compliance and privacy risk [1]. On the regulatory geography he is similarly wry, noting the pattern that when the EU does switch to innovation it tends to be to spite whatever the US is doing at the moment, and that the timing of recent European calls to deregulate is "interesting" [1].

On moats, team and first-principles thinking

Asked what protects the business, he and CEO Max name three things: the data, the infrastructure, which is both the integrations and the ability "to properly utilize millions and millions of dollars worth of GPUs," and the team [1]. The hiring bar is not purely technical. He is looking for people who can build the best models and products and "operate with ethics," a character component he ties directly to how sensitive the space is [1]. At roughly twelve or thirteen people spanning software through machine learning, he still gets to write code, which he says he is "holding on to for dear life," and to sit with every person on the team reviewing their code and designs [1]. What he tries to leave them with is first-principles critical thinking: what backs this assertion, why does this hypothesis follow from that one, and "are you doing something reasonable or are you doing the right thing" [1]. The standard he sets is not that every answer be the right one, but that people know exactly what they are doing and why [1].

Takeaways

• Define the company by what it is not: Yobi builds foundation models of user behaviour and is "not a large language model product or company" [1], and he argues intent prediction requires more than an LLM [3].
• Training data comes only from partners who obtain explicit user consent, usually with financial compensation, producing a clear record of who owns what and who was paid [1].
• Financial underwriting and health are refused outright, not because they are regulated but to remove any possibility of biased or unfair outcomes [1][6].
• Yobi adds adversarial objectives so models are rewarded for being bad at predicting demographics, and declines customer requests for demographic data even though it holds it [1][6].
• Privacy and performance are separate axes with plenty of headroom on both; encrypted matching and homomorphic machine learning move privacy right without moving performance down [1].
• The highest privacy tier is deploying models inside the customer's own account so data never leaves their servers, with a medium-term goal of training on customer data in place with no expected performance loss [1][6].
• Accountability will be forced by consumers before legislators; AI reached "break things" in a month where software had 10 to 15 years [1].
• Hiring weights character alongside capability, and leadership means reviewing code and designs personally while demanding first-principles justification for every assertion [1].