Why I started SAIPAL

SAIPAL is a research lab I run. This post is the honest version of why it exists.

I want humanity to prosper. I know how that sounds — every lab website says something like it. But it’s the actual reason, so I’m going to say it plainly and then try to earn it. AI is going to have a hand in nearly everything that determines how the next few decades go: what work is worth, what security means, who gets to verify what’s true. If I can spend my working years pulling the expected amount of suffering in the world down — even slightly, even just by measuring things carefully and saying what I find — that beats every alternative use of a career I can name.

The second reason is independence, and this one is more concrete. Most evaluation of AI systems happens inside the companies building them. Some of it is excellent. All of it is, structurally, marketing — even when the people doing it are honest, you can’t easily check their work, and the incentive gradient only points one way. I wanted there to be one more place, however small, with no product roadmap upstream of the questions and no investor thesis the results need to confirm. SAIPAL is self-funded. That caps how fast it grows. It also means nobody gets to decide what I look at except me, and nobody gets to soften what I publish.

My domain is cybersecurity, so that’s the lens this lab starts from. I think safety people and security people are working the same problem from opposite ends — one asks whether the system does what its operator intends, the other asks what happens when someone hostile shows up. You want both answered before you trust anything important to a model.

Which brings me to the first project: HighGround, a king-of-the-hill cybersecurity competition with a public leaderboard. The format matters — king-of-the-hill rewards holding ground, not just breaching it, which is much closer to what security actually is than a one-shot capture. The name is the same idea from the other side: in security, as in Dota, you don’t fight uphill if you can help it. How do current models really perform on offensive and defensive tasks, measured the same way, in public, with methodology open enough that you can disagree with it precisely? That’s under construction now. It gets its own post when there’s something real to show, and the blog will carry the build notes along the way, including the parts that don’t work.

About the name. Saipal is a Himalayan peak in far-western Nepal. It’s also an acronym — Safe & Aligned Intelligence Pal — and the last word is the point. The goal isn’t AI to be survived. It’s AI worth having around.

Right now SAIPAL is one person, a website, and a benchmark taking shape. If any of this is your kind of problem, I’d like to hear from you: [email protected].