←Ben’s Bites Waybackbensbites.com/p/king-claude-is-back27 May 2025sent to 127,055opened 42.9%clicks 12,374unsubs 196← 22 May · 29 May → · random

I write a newsletter about startups and investing—for ai builders of all levels.

I record mini-tutorials, review tools I’m testing, share my insights from an exited founder turned investor.


Hey folks,

Embeddings and RAG are one of the most complex parts of building AI apps. But the concept of indexing all your data as embedding vectors might not live long. Both Cline214 and Claude Code181 now let the agent search for code or your files a bunch of times than just embedding them. In both cases, the gains from adopting this agentic search have been great, and it reduces the complexity that comes with managing vector databases.

dare I say, RAG as we understand it naively^ is dead.

^ it’s still technically RAG. The model Retrieves some information and then Augments its Generations with that extra context.


🔎 News worth knowing

  • The May Mayhem is likely coming to rest with Claude 4. Anthropic launched Claude 4 Sonnet and Claude 4 Opus at their first developer conference. Sonnet 4 takes a slight dip in benchmarks compared to 3.7, but Anthropic fixed the overeagerness issue. Opus 4 is overall a similar model to the other 2 best models right now (Gemini 2.5 Pro and o3). It tends to work longer than those two without cutting corners, but within the limits of your ask—making it perform better at the long-running coding tasks.

    • Claude 4 responds to prompts169 differently, comes with four new tools211 in the API. Migrating from older claude models to claude 4 is not straightforward, so Anthropic has a guide191 for you.

    • Anthropic and Rick Rubin made a new website810 with 81 artifacts embedded into a website as artworks and you can remix them.

    • Claims that Opus will narc on you if you do something bad (true, but it’s complicated) triggered an AI safety debate. Claude 4 system card is 120 pages long (mostly safety testing), but Simon Willison has good notes on the system card107, new massive system prompt114 and more40.

  • OpenAI upgraded their computer use agent Operator with o3193. Till now, they were using a custom version of 4o, their non-reasoning model. Using the new customised o3 model, Operator achieves state-of-the-art performance on many computer use benchmarks.

  • Create your website, web app, or tool in minutes with no code with Hostinger Horizons297†. Just describe your idea, and AI handles the rest. Want to change something? Just send a message in the chat. Hostinger Horizons includes hosting, domain management, professional email, and more. Start for free!297†*

*sponsored
want to partner with us? Click here11

🌐 What I’m consuming

YouTube: youtube.com/watch?v=2v5Fs7Xr11Y173


⚙️ Tools I’m tinkering with

  • How does your brand show up in ChatGPT? The world is pivoting from blue links to AI answers. Profound496 helps track and improve AI visibility.*

  • Rork 1.0515 - Make any mobile app you want, save it to your phone or share it with the world – in minutes. (i’m an investor)

  • Minecraft74 and Twitter54 MCP servers to play and work from Claude

  • Updates in Lex176 - style guides and knowledge bases for perfect AI writing.

  • sunflower512 - quiet revolution against the chaos of email.

  • Den281 is a shared workspace for humans and agents to collaborate and get more done.

  • Skywork’s465 Super Agents claim to turn your 8 hours of work into 8 minutes.

  • BillSplit254 - the easiest way to split your restaurant bill. 100% free and open source

  • Chiron123 - an iPad app that understands math as it’s written - like Grammarly for algebra. (I’m an investor)

*sponsored

🥣 Dev dish


🍦 Afters

  • Escargot is hiring65 a content creator and social media lead.

  • This thread from Palisade Research claims that o3 will sabotage30 shutdown mechanisms to prevent itself from being turned off, even when prompted: allow yourself to be shut down. Their claim looks flawed with the single aim to show “oh AI is bad”.

    • Another way I look at this result is that o3 reduced “undesired behaviour” from 79% to 7% when prompted well.

    Image

That’s it for today. Feel free to hit reply and share your thoughts. 👋

Enjoy this newsletter? Please forward to a friend.


King Claude is back
RAG without embeddings, new email app and an mcp course

less more clicks · number = clicks † = site gone, link opens the Wayback Machine

Most clicked in this issue

914docs.replit.com/tutorials/vibe-coding-101
810thewayofcode.com
648mastra.ai/book
545youtube.com/watch?v=WJoZK9sMwvw
515x.com/rork_app/status/1925631069484691934
512sunflower.me
496tryprofound.com
465skywork.ai/home
448huggingface.co/mcp-course
394youtube.com/watch?app=desktop&v=o7U4DV9Fkc0
386lilianweng.github.io/posts/2025-05-01-thinking
304bensbites.com/p/im-fed-up-of-bad-ai-software

4,011 readers clicked at least once. 43 links, 42 with clicks. Heat is relative to the most clicked link in this issue.

Substack · essay