Sam: A one-hundred-thousand-times performance cost.
Kai: And that's just the starting point! I'm Kai.
Sam: I'm Sam.
Kai: And this is Open Source and Dev Tools Daily for Saturday, August 15th!
Sam: We've got a wild mix for you today, from the deeply practical to the... deeply perplexing.
Kai: Google's making a huge play for private AI, with tech that lets you compute on encrypted data. And yeah, we'll get to that performance hit Sam just teased.
Sam: We've also got a new service promising AI-generated images... made by hand. Yes, you heard that correctly.
Kai: And if you're trying to get your AI pair programmer to actually listen, Anthropic just dropped the definitive guide for getting the most out of Claude.
Kai: Alright, let's dive into that Google post: 'How Google is making private AI practical with homomorphic encryption.' Sam, this is a big deal.
Sam: 'Huge' is one word for it. I'm focused on their phrase 'making it practical,' because they kind of bury the lede: a '100,000x or more performance cost.'
Kai: Okay, but that's not the point! This is fully homomorphic encryption—FHE. It's the holy grail of cryptography. You can compute on data while it's encrypted. No one, not even Google, can see the raw info.
Sam: I get the 'what,' I'm just stuck on the 'practical.' A hundred-thousand-X slowdown means a one-second query now takes... what, over a day? For what, exactly?
Kai: For things where latency doesn't matter! Like medical research on sensitive patient records. Or training models on proprietary financial data without ever exposing it. And hey, they open-sourced their general-purpose FHE library.
Sam: Look, the library is great. The concept is amazing. But calling this 'practical for AI' right now is a stretch. It's practical for a very specific set of research problems. This isn't for your chatbot.
Kai: But that's how it always starts! The cost will come down. It's about building the tools now. They even built a transpiler that converts C++ into FHE-compatible code. That's huge for developers.
Sam: The transpiler is clever, I'll give you that. But complexity is the enemy of security. Now you have to trust the FHE library and the transpiler not to introduce some new hole that makes the whole thing pointless.
Kai: It's a trade-off, sure. But for developers, this means you can finally start experimenting with privacy-preserving computation. That was practically impossible before. Forget the 100,000x for a second and think about what this unlocks.
Sam: I'll try to think about what it unlocks while my query finishes... next month.
Kai: Alright Sam, you have to see this next one. It's a site called 'byhand.ai.' Tagline: 'We generate AI images by hand.'
Sam: Wait. By... hand? What does that even mean? Are they drawing the pictures? Because that's just... art.
Kai: So, it looks like a service where a human takes your prompt, then works with the AI—tweaking the prompt, iterating, curating—to get you the perfect image. They're calling it a bespoke, artisanal AI service.
Sam: So it's Mechanical Turk for image prompts. They've productized prompt-whispering... for a fee.
Kai: But the results are incredible! No six-fingered hands, no weird artifacts. It's the quality you wish you could get after wrestling with a model for two hours.
Sam: Of course the results are incredible, they're only showing the highlights! This is either satire, or just a very thin wrapper around outsourcing. Why not just hire a prompt engineer on a freelance site?
Kai: Because it's about the experience! The product! It's fun! And let's be honest, most people wouldn't even know where to start to hire a good prompt engineer.
Sam: But calling it 'AI by Hand' is just misleading. It's 'AI by a different, more patient human.' And judging by the comments, the community isn't quite sure what to make of it either.
Kai: I think it's brilliant! It admits that AI is a powerful tool that still needs a human touch to get great results. The takeaway is that the value isn't just in the model, it's in the skill to use it.
Sam: Or the takeaway is there's a new way to charge people for hitting the 'generate' button. I'm calling foul on the name.
Kai: Okay, on to something a bit more practical. Anthropic just posted 'Maximizing the value of your Claude Code sessions.' It's the official user manual we've all been waiting for.
Sam: A user manual for a chatbot. Let me guess: 'Write better prompts.'
Kai: Well, yes, but it's more specific than that. For instance, they say to provide a 'golden example' of the output you want. Show, don't just tell. And give it a persona, like: 'You are a senior Python dev who writes functional code.'
Sam: So... provide a test case and a style guide. These aren't exactly new concepts, Kai. It's just applying basic software engineering discipline to your prompts. Which, I'll admit, most people don't bother to do.
Kai: Exactly! It's turning that folk wisdom into a best practice. Another great tip is to break down big tasks. Instead of 'build a website,' start with, 'write the HTML for a landing page with a hero and three feature boxes.' Then you iterate.
Sam: Again, that's just project management 101. It's Agile for AI. What's really interesting here is Anthropic basically admitting the tool isn't magic. Garbage in, garbage out.
Kai: Right! And the lesson for developers is you can get way better results just by being more systematic. Don't just chat with it, instruct it—like you would a junior dev.
Sam: A very smart, very fast, very literal junior dev who has read the entire internet and occasionally hallucinates a new framework.
Kai: Exactly! Treat it with that level of specificity, and you'll get great code. I'm bookmarking this guide right now.
Kai: Alright, let's do a quick recap.
Sam: Google is making private AI practical, as long as your query can wait until next Tuesday.
Kai: A new service offers 'hand-made' AI art, which really means hiring a human expert to do the hard parts for you.
Sam: And Anthropic confirmed that to get good code out of an AI, you have to treat it like... an actual engineering task. Shocking, I know.
Kai: And before we go, a little inspiration. Today would have been Julia Child's birthday.
Sam: Oh, let me guess. In her honor, you're going to tell people to... read the manual?
Kai: You know it. She once said, 'The only real stumbling block is fear of failure. In cooking you've got to have a what-the-hell attitude.' That applies to coding, too. So, bon appétit, everyone.
Kai: That’s our show! You can find links to everything we talked about in the show notes. We'll be back tomorrow with more news from the world of open source and dev tools.
Sam: Until then, I'm off to go generate some vacation photos... by hand.
This show is made with AI: the hosts’ voices are synthetic and the scripts are AI-assisted. Every story links to its original source.