How Claude Dreams Became My Nightmares
When innovation feels less like building a unique product and more like feeding Claude Code with ideas.The future of product development, according to Boris ChernyProduct innovation matters.It moves the market forward, beyond organic growth and agreed conventions.It lets small teams make an impact…
When innovation feels less like building a unique product and more like feeding Claude Code with ideas.The future of product development, according to Boris ChernyProduct innovation matters.It moves the market forward, beyond organic growth and agreed conventions.It lets small teams make an impact where giants might fail, forcing the market to become better and more valuable.I have spent most of my career in corporate innovation centers, developing “first-of-a-kind” solutions. So I know how hard and painful it can be.But for the first time in a 20+ year career, I feel there is a special kind of pain in product innovation, specifically in AI-related products.I mean the pain of spending months building something weird, fragile, personal, and maybe actually innovative, only to open your laptop one morning and see a major AI company announce it as their feature.So I ask myself:Do we all just converge on the same ideas, or is there another source of this pain?Is there a reason most AI products just look and feel the same?Let’s start with something directly from the source.On June 18, Boris Cherny, the “father” of Claude Code, was asked about the “next big shift in software development.” In response, around 01:49, he said:“Claude Code is starting to come up with ideas… Claude is looking at feedback, bug reports, looking at telemetry… and Claude is starting to think about bug fixes and things to ship.”https://medium.com/media/bdd19f7979c361954bd32c938d938c70/hrefStop for a second and let that sink in.“Claude Code is starting to come up with ideas.”That sounds magical, AGI-like behavior. But is that really what is happening?Because the next part is the one that made my brain short-circuit:“Looking at feedback, bug reports, looking at telemetry.”Ha! So ideas are not appearing out of nowhere. They are coming from our telemetry logs: the most intimate signal data you can imagine.And then:“…things to ship.”That is the part that hurts.Because if you are using these tools to ship your next killer product, your logs are not abstract. Your bugs are not abstract. Your telemetry logs are not abstract. These are your sweat, thoughts and dreams, slowly shaping and perfecting into a product. A machine like Claude can easily turn those into polished “ideas”, even better than you.So when someone says the system is looking at telemetry and coming up with things to ship, I do not hear “AI magic.”I hear:My most precious data can become your product, without my real knowledge or consent.Around the same time, Alex Karp, the CEO of Palantir, went on CNBC and basically said companies are worried that frontier AI labs get access to their IP while charging them for tokens.“Who owns the data? Are the prompts secure? Is this being transferred to you?” “If it was so valuable, and I can make you a billion dollars, wouldn’t I say I’ll make you a billion dollars and I want 30%? Why are they charging for tokens if it’s so valuable?”Someone finally asked the right question.https://medium.com/media/b9ce090a74e9fc6cdf352019096c5a18/hrefThink about what Alex is saying.If tokens are not really worth that much, then what is?The answer is simple: your data, your workflows, your business processes, and your unique way of solving real problems. That is where the money is.Listen to David Sacks, a key figure in AI who is also known as Trump’s AI czar.https://medium.com/media/7bf8519d23c753f99bcd30eaacd163fa/hrefDavid summarizes the threat well:The pattern is consistent: watch where value is being created, then move in directly. Dominate the model layer, then use that position to capture the most lucrative verticals.In the past, vendors like IBM and Oracle wanted to lock you in. Now they are after your brain.My JourneyBefore the details, a note: the remarks below are based on my own experience, with Git commit hashes available for review.You spend months building long-running agents with small LLMs.You realize standard tool calling is not efficient and breaks easily.You follow the great research done by Dr. Linxi “Jim” Fan on skills two years earlier.Months in, you finally get it working.Bye-bye, full tool-context overload.A few weeks later, skills are suddenly a thing in Claude Code.“No worries,” you say. “There is still much work to be done.”You run deep-research agents with local LLMs, and they all fail after a short period because of long context and wrong planning. So you decide research and learning should be dynamic, governed by simple markdown files and self-evaluation.You build it with Claude Code. It actually works. The agent builds its own curriculum, monitors progress, and even performs self-training. You give these powers special commands: /research, /autolearn.Two weeks later, Andrej Karpathy, well known researcher with deep ties to Anthropic, releases /autoresearch and makes headlines. Shortly after, he starts working at Anthropic.Lost the innovation race again. This is the moment you start wondering whether we are all converging on the same AI slope, or whether the winner was simply playing with everyone’s cards face up.You still have tokens to burn, so you don’t let go.You realize running long, stateful, dynamic tasks is hard, so you create a mechanism with Claude Code that you call a goal. Unlike, previous deep-research agents, it builds its own evaluation and improve itself. It works. You are happy for approximately ten business minutes. Then Claude Code releases /goal.Okay. No worries. Clearly, goals are not enough.You are already a step ahead, because your dream is to have the infrastructure to run managed background agents, extract important signals, and promote recursive learning and insights.A couple of weeks later, Anthropic released something called Dreams. According to Claude, it is almost identical to my product pitch.Dreams? Fricking nightmare, if you ask me.The Trap I Created for MyselfAt some point, out of pure frustration, I decided to lay a trap.Why not add some new features, things never seen anywhere in the AI sphere, let Claude digest them, and wait?The problem and user need were simple: when a user wants to hand over a task to an agent, remote or local, there is no simple way to pass along all the needed context.So I created a special tag for handing over requests to agents, backed by APIs so it could work from any app. Now, you could pin an agent from third-party apps and push context with ease.Well, you guessed it: couple of weeks after and we now have Claude Tag.Is it really the same thing? No.Unlike Claude Tag, I did not care about Slack at all. In fact, I hate Slack. But the main concept, the ability to generalize a remote agent with the right context, is there.The Trap I Did Not SeeImagine this: you have all of Claude Code’s logs, and you wonder what you can do with them.If you are not an AI engineer, this might sound harmless.But if money means nothing to you, and you can leverage an endless number of agents to summarize insights, look for ideas, and research innovative directions, this is a goldmine.Remember what David Sacks said before:“Capture the most lucrative verticals.”So the real trap, the real total token cost and total cost of ownership, is your vertical knowledge. Your intellectual property.Speaking of money meaning nothing, listen to what Boris Cherny said during the opening session of the 25th annual Fortune Brainstorm Tech conference:“Some days it’s … thousands, or tens of thousands.”Anthropic's Claude Code creator says there are days he manages tens of thousands of AI agents at once | FortuneMy morning routine is to have coffee, feed my cats, and hate the world. By 09:00, I’m a happy person.Boris’ health routine is collecting daily Claude Code telemetry, feeding it to tens of thousands of agents, and waiting for them to come up with ideas to ship.This is the definition of an unfair advantage.At least now I understand why Boris smiles all the time.BTW, you cannot afford Boris’ health routine. It could easily cost $1M a day to run thousands of agents.The Cherni EffectThe Cherni Effect is when your MVP wakes up, looks in the mirror, and sees someone else’s release notes.You build and burn billions of tokens. Your research and time let you pinpoint the problems early on. You start with an idea. You add feature after feature. A new workflow. A strange agent that learns. You push it forward one commit at a time. You consult with your cat. You think, “Maybe this is a step forward. Maybe this is truly innovative.”Then the AI machine starts moving.First it becomes a demo on their X feed.Then a blog post.Then a platform feature.Then someone on X says, “Obviously this was inevitable.”Obviously.Yes. Everything is obvious after someone else ships it.Before that, it is “too weird,” “too early,” “too complicated” or “ too hard to explain”.The funniest part is that I am not even sure who I am angry at. The companies? The market? Myself? The fact that AI moves so fast?Maybe this is just what innovation feels like in 2026. You don’t get copied. You get absorbed by the machine. Resistance is futile.So yes, I closed my $200 AI accounts to focus on open-source LLMs and local AI.Not forever, probably.But emotionally, I needed to walk away from the machine that kept eating me for lunch. I have no power here.There is no grand lesson here. No “10 things I learned building in AI”.Just this:It is almost impossible to innovate when your competitor has full access to all your dreams.Good night, sweet dreams.AppendixThis article does not claim that anyone copied anything from me. It doesn’t, and I’m not a lawyer. Most importantly, this article is about you, not me.But to be honest, the burden is not on me. Since Boris Cherny detailed his work routines, Anthropic needs to explain what it means to let Claude come up with “ideas to ship” from our telemetry logs.I have dated local evidence, commit history, working prototypes, and audit backups for the product ideas below. The full commit logs, source bundles, screenshots, and manifests can be shared privately with serious reviewers.This is my shortlist. Please don’t ask me about Claude Managed Agents. I have suffered enough.Last note: I disabled all data sharing with Anthropic on day one. But it turns out it does not opt out of sharing telemetry logs with Anthropic.This story is published on Generative AI. Connect with us on LinkedIn and follow Zeniteq to stay in the loop with the latest AI stories.Subscribe to our newsletter and YouTube channel to stay updated with the latest news and updates on generative AI. Let’s shape the future of AI together!How Claude Dreams Became My Nightmares was originally published in Generative AI on Medium, where people are continuing the conversation by highlighting and responding to this story.Source: Generative AI Pub — Published — Category: Image AI