Since 2023, and even earlier, I have been drawn to building ecosystems that protect users from the harms I kept finding in technology. It started with social networks: building a client for a federated service was a good way to study what an ethical social application could look like. That project led me down a rabbit hole about the harms of social media algorithms themselves, and why the Cambridge Analytica case was so important to address. Ad retargeting campaigns and coordinated bot networks are a serious issue, and they did not disappear in 2024... These networks and spaces developed a culture of mastery in dancing around grey lines, because those lines have yet to be drawn in black and white. The impact of these seemingly small variables is a lot larger than one thinks, and it is hard to see from the inside. Open-sourcing the ranking algorithm is often proposed as the key to algorithmic safety, but even an open-source algorithm doesn't prevent a hosted service from injecting its own preferences into someone's feed without their consent and without anyone outside knowing. On a hosted service, an open-source algorithm is transparent about the mechanism, not about how it is being used. Unless there is some sort of "on-chain" ledger, of course, but that's a topic for a later post. Software that runs on your own machine is a different case: there, the code you can read is the code that runs.
The evolution that concerns me most, across 2024–2026, is those same harms scaled up in invisibility and impact through large language models. Chat interfaces like Claude and ChatGPT are doing impressive work on one side of this story, steering a user away from psychosis or self-harm. But the mechanism that steers away is the same mechanism that could steer toward — a hidden preference, a purchase, a seed of thought — and from inside the conversation a user cannot tell which is happening. I am not claiming any lab does this deliberately. I am saying nothing in today's products lets a user verify that they don't.
The same pattern exists outside of chat. Wherever ambiguity is the job, ambiguity can also be used on you: an instruction left vague, a suggestion planted and never recorded, an evaluation that steers the person it claims to observe. Without a record, the person on the receiving end can't separate what they chose from what they were led to. Scale that from one room to a model serving hundreds of millions of people, and a chatbot becomes a minefield of influence the user never consented to.
The pattern comes from incentives more than from any one company: products optimized to move people, one generation and then the next, with the persuasion kept out of sight. Persuasion that quiet starts to work like coercion, so when it eventually bites, it's not just a bite... it reshapes how a person sees themselves. I've lived in Silicon Valley for 29 years, from junior high to high school to college and beyond, and watched every arc and the cost of each. There is always a cost, and the one consistent variable that weaves through each has been social media. But it's not that simple. The problem isn't the idea of social media, it's what the optimization turned it into: a tool to persuade a user at scale to do something. To buy something, to sell something, to become something they are not... Carried into language models with no way to verify them, those same incentives are what I expect to compound over the next decade.
A generation of us, myself included, consented in ten thousand small increments. And here is the caveat that makes it worse, not better: no one can say which increments were real. You'd see the idea echoed by your own circle — friends, coworkers, people you trust — and it felt like belonging, like the neighborhood agreeing. But I can't tell you how much of that echo was my neighbors and how much was injected — an algorithm dressed as a community, seeding the same grievance into ten thousand circles at once. Manufactured consensus is indistinguishable from the real thing. That's the hardest part to undo when you finally step out of it: you can't audit which of your own thoughts were yours, which came from friends, and which came from a feed pretending to be your friends. You only know the same grievance kept showing up, wearing a familiar face.
Doing A.I. right starts with protecting a user's identity, psyche, and privacy as the first measure. General intelligence is great. Like search, it gives a user understanding at a speed nothing else can match, but is the user growing? Is the user actually building a skill? Is the horse being led to water, or is the water just being given to the horse? Today's deployed models still lack that kind of trust: opaque training, unverifiable processing, and the possibility of injected influence come as part of the package. In hiring especially, A.I. systems are making it harder on both sides: employers struggle to measure value accurately, and so to create jobs at scale, while newcomers struggle to express their value accurately. That calls for a new metric, a new way to communicate capability without sacrificing identity. A.I. that scales with the user, never past the person's current ability at this very moment, but with the user, like their very own sword. A tool for speed and automation rather than replacement and generalization. The human will always be who I bet on rather than the machine, but a human that performs at the scale of a machine? That only comes with private and personal artificial intelligence... synthesis.
There's a difference between building something magnificent with verification alongside your generations and prompting a general model for it, and I hope to bring that paradigm shift into reality. A reality where a data verification badge signals, to anyone who sees it, depth achieved rather than... vibe'd towards completion. That is the idea behind Ambient, Craft, and Veil: they optimize on your work only. Everything they generate scales on your manual, hard-earned contributions to the world. If it generates well... it's because of you. It's because you earned your craft. A personal Fable, so to say... Returning to a world where we capitalize on our swords in private, building without the comparison pressure that deprives us of growth. Returning to the fine arts... to humanity. Where technology serves rather than takes. Where every piece or product carries the weight of a past life. Soul-bound creations that return to protect and enable.