Hey makers 👋

Another week, another batch of mind-bending releases:

• Baidu open-sources ERNIE 4.5 and declares price war on token bills.

• ChatGPT (o3) with web search quietly becomes the best “research intern” you own.

• Codex, Lovable, Base44 & Replit square off in our first Vibe Builder showdown.

Take a break, grab a slice

🚀  Ship Faster (and cheaper!)

Baidu just open-sourced ERNIE 4.5 - 10 Models, 424 B Params

Just yesterday, Baidu took its flagship language-and-vision model family, ERNIE 4.5, and released all ten versions on Hugging Face with a 129K context window. Here’s what’s up:

• Pay $0 per prompt. Because the full “weights” are public, you can run ERNIE on your own laptop or a cheap cloud box instead of feeding coins into someone else’s API.

• It’s no slouch. The biggest ERNIE beats DeepSeek V3 on 22 of 28 common tests for coding, reasoning, and maths, even though it’s half the size.

• Sees text and pictures. A new training trick - heterogeneous multimodal Mixture-of-Experts (MoE) - lets the model learn language and images together, so it can caption screenshots or explain charts without extra plugins.

Dive Deeper | Try it out on Hugging Face

🧠 Level Up Your Game

Which chatbot should you trust for web research?

A fresh benchmark from FutureSearch⁠/⁠LessWrong put 12 big models and 11 “Deep Research” products through 89 tough, multi-step web-research tasks. Winner: the plain ChatGPT (model o3) + built-in web search beat every dedicated research tool—and did it faster and cheaper.

Key pizza take-aways

• Default weapon: open a normal ChatGPT tab, toggle Search the web, and go. On FutureSearch’s Deep Research Bench it scored higher than OpenAI’s own “Deep Research” mode and every rival service.

• Iterate, don’t over-engineer. Regular chat + search outperformed most “Deep Research” modes because you can ask quick follow-ups, sanity-check answers, and steer the search yourself.

• PDFs still trip up Claude. If your workflow involves white papers or SaaS T&Cs, skip Claude’s web modes for now—its agents can’t reliably read embedded PDFs.

• Building your own agent? FutureSearch’s API tests show Claude 4 Sonnet/Opus once you wrap them in a ReAct pipeline. That’s good to know especially if you’re baking research into a product.

• DeepSeek R1 is the budget pick. Self-hosting DeepSeek gives respectable accuracy for pennies when you need hundreds of quick queries

How I do it

1. I Ask → but Verify. After a result, prompt: “Double-check each claim and give me the source URLs only.”

2. Link-check. Click the links yourself—LLMs still hallucinate ~20-30 % of the time on number questions…

🛠️ Tools

Vibe-coding Platforms Showdown 2025 (so far)

Here’s my comparison of the four hottest “vibe” platforms: OpenAI Codex, Lovable, Base44, and Replit.

Yes, the back-and-forth rabbit holes are real (I’ve been down a few myself), but remember: we’re living through a once-in-a-generation shift in how software gets made—might as well enjoy the ride!

• Lovable → design-first speed run

  • One prompt = landing page + DB + auth with nice defaults.

  • Best when UI polish matters and the whole app fits in <30 screens.

• Base44 → all-in-one peace of mind

  • Built-in database, auth, file storage, email, analytics—no extra APIs to wire up.

• Replit + AI Agent → code-curious flexibility

  • I use the app wizard scaffolds or the remixed project, then drop into an IDE for fine-grained control of the code.

• OpenAI Codex → autonomous teammate for dev-founders

  • It feels more like a junior engineer: edits your repo, runs tests, pushes PRs, etc ..

The Snapshot

🌶️ The Funny Slice

🎧️ Heard on X…

Keep shipping! 🚀

P.S. - Got an AI tool that's moving the needle for your business? Building an AI tool that helps fellow hackers? Hit reply and tell me about it. Always looking for real-world wins from our community.