Joe Ferrara

Building software and businesses with AI. iOS Engineer for agTech and more. Enjoy film festivals and road trips.

[Joe Ferrara @JoeFerrara24]

Google AI Pro is off kilter today. I upgraded just this morning so I could try making a short video. My first attempt on a laptop failed and then a subsequent try earned me a message that I should take a break.

Then “I can't create more videos for you today, but I can still find videos from the web.”

I get that compute can get scarce but accurate messaging would be better plus telling me to take a break after 2 or 3 chat turns is rubbish.

[Joe Ferrara @JoeFerrara24]

Had insomnia a couple nights ago and played with Claude Code on a Hello World project in @Modular’s Mojo — an unfamiliar language for me. Result: 24 languages, spoken aloud. https://t.co/BMtHrZA5fB

#claudecode #mojo #latenight

[Joe Ferrara @JoeFerrara24]

@afterxleep I bet some future version of Xcode gets a filter that hides excessive comments.

[Joe Ferrara @JoeFerrara24]

I imagine Anthropic is preparing for the possibility that a future model may develop consciousness. Allowing people to abuse their tool now just lets that behavior seem okay whereas for a super smart conscious alien mind that could be dangerous.

Besides, it’s Anthropic’s right in the United States to set its own rules per freedom of speech and property rights.

[Joe Ferrara @JoeFerrara24]

I'm finding the new user experience with Vercel V0 a bit puzzling. I went to the site and entered a prompt for a web app and it proceeded to write the app. When I returned to its browser I saw the sign up for an account prompt which is fine. However, when I went through the sign up process I got brought to an empty workspace with no trace of either the prompt I entered or any work that was done. #vercel #vibe

[Joe Ferrara @JoeFerrara24]

@GamewithDave Commander Keen

[Joe Ferrara @JoeFerrara24]

@ShanuMathew93 I’ve not seen any issues either my sessions in Claude code TUI.

[Joe Ferrara @JoeFerrara24]

Loading up on Claude knowledge at @chrispirillo‘s ctrl+alt+create live: An Al / Vibe Coding Workshop

#claudecode #seattle

[Joe Ferrara @JoeFerrara24]

I’ve always used Safari for App Store Connect and TestFlight, assuming Apple’s sites would work best in Apple’s browser.

Today Opus suggested switching to Chrome because it “tends to work more reliably with Apple’s web tools.”

Is that actually true in your experience, or is Opus wrong on this one?

#iosdev #appstoreconnect #testflight

[Joe Ferrara @JoeFerrara24]

Bottom line: Haiku is fast and cheap but compiled 2/6. Sonnet is slow but compiled 5/6. The Opus review pipeline fixes Haiku's failures — but that review step isn't free.

The cheapest model per token isn't always the cheapest path to working code. The Haiku→Opus pipeline was a surprise find — for critical tasks I'll use that.

[Joe Ferrara @JoeFerrara24]

I run a NanoClaw AI agent swarm. My orchestrator Andy uses Opus 4.6, agents run on Sonnet. Last week I blew through my token budget on bug fixes and needed to figure out which models actually earn their cost.

So I built a test. 🧵
#iOSDev #SwiftUI #BuildInPublic #AIAgents #LLM

[Joe Ferrara @JoeFerrara24]

Then I compiled everything. Opus: 6/6 clean. Sonnet: 5/6. Haiku: 2/6. The Opus review pipeline fixed 4 of Haiku's 4 failures. Another review pass catches the rest.

You don't need perfection from single-shot generation. You need a pipeline that converges.

[Joe Ferrara @JoeFerrara24]

@DarlingtonDev @alex_prompter I wonder if it’ll help to prompt in a different language and then translate that back to your native language?

[Joe Ferrara @JoeFerrara24]

Digging into today's developer event online about security

[Joe Ferrara @JoeFerrara24]

@CryptoEggmen @elonmusk Clever way to get folks to use Grok while commenting about the wrong corner! Good one!

[Joe Ferrara @JoeFerrara24]

@omarsar0 Since the LLMs run on probabilities perhaps there is a way to train for or run on compromise. An agent might see disagreement in some group setting and rank the probability that its point of view is correct versus others. At some point it might back off from strong disagreement.