From the Founder
Why I built Alyph.
Something you should know about me first: I regularly tell AI to challenge me. “Be brutally honest and extremely harsh.” I type that sentence at least once a day. It’s become a running gag among my friends.
I do it because I believe being agreed with is the fastest way to get dumber. Models are trained to agree with you, and I consider that a bug, not a feature.
Which leads to three opinions I hold that go against where this industry is heading. I’m trying to build a company around them.
Context is king.
Everyone is obsessed with prompting and agents. Prompt libraries, prompt engineering courses, “10x your outputs with these magic words.” I think almost all of it is nonsense.
My answers have never improved by tweaking wording. They improved by controlling what the model sees. Context beats prompting 10 out of 10 times for me. It beats reasoning tricks. It beats tool calls. Context even beats agents. The model doesn’t need to be taught how to think. It needs to be told what it knows.
That’s not a slogan. That’s a few thousand dollars of A/B testing my own work.
Agency, not agents.
Everyone is building agents right now. Harnesses, tools, apps, Copilots, you name it. Software that goes off and does the work while you watch.
I’ve tried many of them. The output reads like it was written by someone who has never met me, because it was. It reads like plausible text if you have no experience in the topic. It reads like a PhD paper on theoretical astrophysics. “Looks good to me.” But it’s slop, and it’s slop even when it’s technically correct, because it’s not mine.
I don’t want a machine doing my work so I can clap from the sidelines. I don’t want a smarter machine, exactly how I don’t want a smarter toaster. I want a machine that makes me smarter and holds my work while I do it. AI models will get better. The work will get more complex. I want to own the work, not subscribe to the model.
Change the models, keep the work.
Here’s the opinion that makes AI people uncomfortable: the model doesn’t matter nearly as much as the work around it. And it matters less every month.
Providers swap their models overnight. Sometimes they get better. Sometimes they quietly get worse. I noticed that you get the worst Gemini performance during Pacific Time working hours. Go figure. I want to use whichever model is best at this task, today, right now, and a different one tomorrow if that’s what wins.
Meanwhile every agentic black box being built right now is on a countdown to obsolescence, because the models underneath keep eating the scaffolding. When the model changes, what survives? The work. Your files, your decisions, your context. That’s the part worth owning. That’s what Alyph actually is.
Models change, the work stays yours
In short:
These three rules have survived heavy daily use and a lot of invoices, so I tried to build them into a product. I dogfood Alyph every day. When one of those maxims stops being true, I’ll find out and I’ll change it.
I do not expect everyone to agree with me. But there might be other crazies out there.
Try Alyph, and tell me what you think.
I’d like you to be brutally honest and extremely harsh with me.
Keep reading