Sameed Khan

I actually agree with a lot of what @poteto is saying except for the central point. IMO most people don't have actual software factory systems. They've just made it much easier for people to ship with AI, so they assume the category doesn't exist. That's more a lack of vision than anything else. Our conceptualization of a software factory is a system where every piece of work goes through the same stages to standardize the process not the work output. For us, this standardized process begins with what we call a signal. This is anything that tells you how your software should evolve, including the first signal that your software should exist at all. A Slack thread, a Linear ticket, a Sentry alert, a failing CI job, a tweet, someone's idea. There's two ways in for a signal. A person picks it up and builds it with Droid, or a workstream picks it up on its own, triages it against a goal someone set, and has droids write the spec and open the PR. Most work still starts with a person, but a growing share comes from workstreams. Then every change goes through the same checks no matter who or what wrote it. Droid review, security review, a risk judge that can't see who wrote it, CI, and an agent that QAs it on a live preview. Your validation has to be really high quality whether a human or AI built it. The goal of validation is to hillclimb your "regret-free merge rate" until you can automatically merge most changes (or fix them until they can be automerged). Then it ships, and when something breaks in production we trace it back to the PR that caused it. That becomes another signal. It's also how we know if the checks actually work. If the risk judge approved something that later broke, we find out. It's the same way we let self-driving cars on the road. Nobody just flipped a switch. We agreed that once they crashed less than human drivers, they could drive. Do the same thing with code. Once AI-written changes break things less often than human ones, and you can actually measure that, you let them merge. That doesn't mean you turn on automerge tomorrow. Most of our changes still get a human stamp. But once you have automerge, workstreams and a risk score together, you basically have software that improves itself from the signals you choose to track. The human element becomes obvious if you have a system like this. People remove themselves from the actual process line and begin to define the rules by which the process operates. What to watch, goals and owners, skills, review guidelines, and how much autonomy each workstream gets. They're the regulators. I think the 1920s factory that a lot of people picture is the wrong idea. That was people doing the same motion on a line all day. This is closer to a Star Trek replicator. You say what you want and it gets made, because the patterns are already stored and how things get made is the same every time, and checked every time.

Co-sign. The only future that is worth building is one in which AI remains a tool in human hands, used to elevate the human condition.

this should be very obvious but: pointing models at their own brains and getting them to accurately understand what's going on and effectively make changes that cause large shifts in behavior and capabilities is like one of the most dual use technologies imaginable knife's edge

We asked GPT-6 Astra to edit Qwen3.5-0.8B weights to improve math. It directly changed only 0.016% of weights, fixing 0.8B's reasoning loops & improving math up to 6x and even instruction following 3x, without gradients or training. We call this Weight-Space Programming: frontier LLMs can now natively read/write neural network weights! 🧵 Blog: rikiyat.github.io/blog/weigh…

my hypothesis for why AI-generated slides for technical presentations are so awful: people who don't have a lot of experience presenting think that to make a talk you just put all of the stuff you want to talk about on the slides, then you go through each slide and say the stuff the slide says. how you're actually supposed to prepare a presentation: you decide what to say, write it out as a rough script, then put maybe 3% of the text on the slide, plus a bunch of visual aides (e.g. plots that only include data you're planning to discuss). AI-generated presentations are mostly the former. they have a ton of stuff the presenter never mentions, and the stuff that is most useful for the presenter is absent or not highlighted. they include enough content for a presentation 10 times longer than the actual presentation is supposed to be. maybe try doing it the other way: first write the script for what you actually want to say, and make sure you can say all of it in the time you have. then prompt the model to create the slides based on the script. ask it to use as little text a possible, and a minimalist visual style rather than the claude house style. you can thank me later (I accept both venmo and zelle).

this was like 90% of grad school for me, glad claude is doing it now

After co-inventing ChatGPT, I kept asking myself: why have superhuman chat models not led to AGI? I’ve spent the last 2 years in stealth building a new way to train models (RLCD), and a new type of frontier AI model that we are releasing today: Jev • 20-200x faster • 40-400x cheaper (w/ output tokens free) • Frontier composable intelligence optimized for decisions AFAICT the shortest path to AI-based economic revolution