Hey — and welcome to Sunday Signal.
Most founders are overpaying for AI because they're solving the wrong problem.
📡The Signal — 3 AI moves worth knowing this week
Simon Willison — one of the most reliable people I follow for cutting through AI noise — compressed the last six months of LLM progress into something you can actually read. Not hype. Just what changed, what it means, and what to watch. If you've been too busy building to keep up with the news, this is your shortcut.
Why it matters for you: You don't need to track every model drop — but you do need to know what the landscape looked like last quarter vs. now. This is that five-minute briefing.
A small 8-billion-parameter model was getting the right answer 53% of the time on agentic tasks. Add structured guardrails via Forge, and it jumps to 99%. Same model, same hardware, wildly different results. The insight is almost uncomfortable: the model was never the bottleneck — the structure around it was.
Why it matters for you: If you're building anything with AI agents, stop defaulting to the most expensive frontier model. Structure and constraints often matter more than raw model power — and they're free.
Karpathy built some of Tesla's most important AI systems, co-founded OpenAI, and has spent the last couple of years making AI education genuinely accessible. Now he's joining Anthropic — the company behind Claude, which is the model I use most in my own stack. Talent this respected doesn't move without reason.
Why it matters for you: If Claude is in your stack — or you've been curious about it — this is worth noting. The people who understand AI best are placing their bets, and this one's hard to ignore.
🌶️Hot Take — Waiting to learn to code is just procrastination with a technical-sounding excuse
I hear it every week. 'I want to build something but I need to learn Python first.' Or React. Or Swift. Pick your flavour.
Here's what I actually think: learning to code before you validate an idea is one of the most expensive delays a founder can make. Not because coding is bad — it's not — but because you're deferring the real question. Does anyone want this? Does it solve something real?
I built Caption Coach AI without starting from a Computer Science degree. I used Claude to write code I didn't fully understand, iterated fast, and put something real in the App Store. Not because I'm special. Because the tools now let you move before you're 'ready.'
The Forge guardrails story this week is a perfect parallel. The model wasn't the problem. The structure was. For most aspiring builders, the code isn't the problem. The starting is.
You can learn as you build. You cannot validate while you wait.
✦ The Build Log — Week of May 18, 2026
Honest version of this week: I don't have a dramatic ship to report. No major update dropped, no viral post, no milestone number to put in a subject line.
What I do have is momentum — which is quieter and more important.
Caption Coach AI is live on the App Store at $6.99/month and I'm in the phase that nobody really talks about: the part after launch where you just keep going. Iterating on the product, figuring out what's actually resonating with users, and not disappearing just because the initial buzz has settled.
This Sunday Signal is part of that. Issue #2 means I actually came back. That sounds small but showing up for week two is where most newsletters die.
What I'm focused on right now is the habit of building in public consistently — not just when I have something impressive to show. The Forge story this week hit close to home because I've been thinking about the same thing for Caption Coach: where are the constraints I haven't built yet? What structure would make the product 10x more reliable for users?
Build Nights are back this week — Monday May 25, Wednesday May 27, and Friday May 29, all at 8pm EDT. Short, practical, 60-second AI tutorials.
If you're using Caption Coach AI, hit reply — I genuinely want to know what's working and what isn't. That feedback loop is the whole game at this stage.
✦ Try This — Replace your prompt, not your model
The Forge guardrails story got me thinking about a simple exercise you can run in the next five minutes.
Take something you already ask an AI to do — write a post, summarise a doc, give feedback on copy — and instead of just prompting it, add a constraint layer.
Here's a paste-ready structure:
---
Task: [what you want it to do]
Format: [exactly how you want the output — bullet list, one paragraph, 3 options, etc.]
Do not: [one thing it usually does that annoys you]
End with: [a specific closing action or line]
---
That's it. Same model, same task, but now you've built guardrails. Nine times out of ten the output improves without touching anything else.
Try it on something you've been getting mediocre results from and notice whether the structure does more work than swapping to a 'better' model would.
✦ Closing
That's Issue #2 done. If this landed for you, forward it to one person who's still waiting to start — they need it more than you do.
Hit reply and tell me what you'd want more of. More build log? More hot takes? Shorter? Longer? I'm reading every reply.
See you next Sunday.
— Kalpesh Mistry
P.S. Caption Coach AI is on the App Store at $6.99/month. If you're a founder or creator who posts on social and wants to get better at it — that's what it's built for. https://apps.apple.com/us/app/caption-coach-ai/id6765720293