Sasha Petrov

I tweet about product, design, AI, and LLMs

[Sasha Petrov @deepnavy]

A few minor adjustments to the default refine and summarize chain in @langchain have helped me create better summaries.

🧵:

[Sasha Petrov @deepnavy]

New AI jargon: smoking guns

[Sasha Petrov @deepnavy]

Launch claude with claude --chrome
Now it's filling in my privacy settings...

[Sasha Petrov @deepnavy]

I just told my Claude to fill in Apple Age ratings. I pasted info from the Claude Code CLI to provide details about the features. I know I can do it even directly from cli.. mindblown🤯

[Sasha Petrov @deepnavy]

What??? @AnthropicAI @GitHubCopilot

[Sasha Petrov @deepnavy]

@VibeAlongHQ @EricSimons @lennysan @boltdotnew Interesting. So how is there any supply or demand for it?

[Sasha Petrov @deepnavy]

Vibe coding is cool, but can Bolt, Lovable, or Cursor take your app all the way to production? 🤔

Inspired by @EricSimons on @lennysan Lenny's Podcast, I quickly built a landing page in @boltdotnew to test a marketplace idea:
• Vibe coders get expert help
• Developers help others (and earn $)

Check it out & fill out the short survey👇
https://t.co/IeKJpUFj0U

[Sasha Petrov @deepnavy]

@conormartinai @EricSimons @lennysan @boltdotnew Thanks! Do you have any projects where expert help like that would be helpful to you?

[Sasha Petrov @deepnavy]

...
- 2000s 3D game
- Sin City

What do you think?

[Sasha Petrov @deepnavy]

A lot of avatar stylers are coming with new @OpenAI image capabilities, but I tried some styles on my pfp today:
- Rick and Morty
- Simpsons
- Don't Starve
- Adventure time
- ... 🧵

[Sasha Petrov @deepnavy]

...
- Minecraft
- Starwars, C3PO
- Severance
- GTA
- ...

[Sasha Petrov @deepnavy]

If I had something like this as a kid, I wouldn't need any toys or Legos... I would be building games non-stop

[Ricky @rickyrobinett]

How do 8-year-olds vibe code?

Here are highlights from Fay's first solo session with @AnthropicAI Claude Code where she tries building a theme park simulator using @CloudflareDev Workers and @honojs.

Some observations:
- 💃 100% more dancing and singing than when I vibe code.
- 📖 she read a lot more of the code with Claude Code than Cursor.
- 🤩 somehow it felt like she got a lot more time with the app she was building AND the code when using Claude Code. I can't explain it exactly, but it's pretty magical.
- 💸 Her 45-minute session ending up costing a little over $5. "Thats so good!" she says. (Though I expect she may ask for me to raise her allowance if she keeps using it!)
- 🎓 She's spent a lot of time building with Cursor over the past 6 months, and that set her up to move quickly with Claude Code. I suspect if she had started here there would've been more of a learning curve.
- 🥹 It's amazing how much I learn from her when I watch her try out these tools, and how much fun we have together exploring the world of AI.

P.S. - She wants everyone to know she's doing a STEM fair project on AI :)

P.P.S. -- Mini-disclaimer: These videos are prompted by Fay and unrelated to/unaffiliated with my job 😇. That's why there hasn't been one in about 6 months. I'm not sure if/when there will be one again. She keeps asking for her own YouTube channel, but so far the answer has been "no".

[Gergely Orosz @GergelyOrosz]

Next time you see a large company move sluggish, and ship products that feel poorly thought out: before criticizing it loudly, think about what it means.

It means that in the industry they operate in, there's still opportunity for a fast-moving startup to win by great execution.

Once the large company starts to execute quickly: it becomes much harder for any startup to compete with them. So be careful what you wish for :)

[Ben Lang @benln]

The most obvious non-obvious advice from Jan Koum (Whatsapp founder)

[Sasha Petrov @deepnavy]

@jenny_wen I had the same issue. Nothing helped, so I just went to the Apple Store, and they replaced the AirPod.

[Sasha Petrov @deepnavy]

I have always been a strong advocate for using OpenAI models at the company where I work. However, with the latest news, it seems that convincing leadership will be much more challenging. Open source models now appear to be a safer option. Luckily with @langchain we can switch in no time.

[Sasha Petrov @deepnavy]

The result:

[Sasha Petrov @deepnavy]

This is what you do when you want to return retrieved documents for RAG with LCEL in @langchain

But is there an easier way?

[Sasha Petrov @deepnavy]

@Galileo_AI Sign me up!

[Sasha Petrov @deepnavy]

@Galileo_AI Awesome!