10 Comments
User's avatar
Drew Meister's avatar

I am making agents build fully functioning apps in 2 hours that used to take me a year. Everyone, not just wrappers are in trouble. I wish I could paste pictures here, but I just built Monday.com for finance types is about 2 hours. Here's the LinkedIn URL. https://www.linkedin.com/feed/update/urn:li:activity:7352176010357846016/

Expand full comment
Dave Friedman's avatar

Maybe they'll use their $100 mln fundraise to implement picture capabilities for comments. Anyway, what you posted at LinkedIn looks super interesting--you built that in ChatGPT?

Expand full comment
Drew Meister's avatar

A blend of Replit with Grok 4 Heavy.

Expand full comment
Drew Meister's avatar

Replit still can’t build enterprise ready apps. I’m having a fun back and forth with Jason Lemkin on Twitter about this. But Grok 4 Heavy can help with Replit code reviews. It ironic the AI coding IDE still needs a coder to drive.

Expand full comment
Dave Friedman's avatar

Yeah I've seen Lemkin's struggles as of late. Reinforces my view that enterprise adoption of AI, beyond areas that tolerate high variance like sales and marketing, will be much slower than many startups & VCs expect.

Expand full comment
Dave Friedman's avatar

Cool. I haven’t played around with Replit in a few months but was impressed with it. Have never used grok.

Expand full comment
Jon Rowlands's avatar

Since the MCP protocol opposes this, by letting people swap out foundation models, model vendors might apply Embrace and Extend to MCP to lock agents to their model. Not that hard, since MCP's function call parameter inference is already only weakly defined. Best would be for open foundation models to gang together to support a common and capable profile of MCP. Best outcome for the world would be for the foundation models to become smaller and more reliant on MCP, and ideally run locally, since agents are less expensive, don't suffer from hallucinations, etc.

Expand full comment
Dave Friedman's avatar

Yeah that’s a good point. I think there might be value in a kind of orchestration layer which allows people to choose whichever model they want, though I’m not sure how big the market is for that.

Expand full comment
Jon Rowlands's avatar

Orchestration as the direct caller of MCP, so that the orchestrator focuses on reasoning and the LLM on language? That would be a great split. It'd force the issue of knowledge representation, the biggest barrier to going full neurosymbolic. Let the orchestrator itself be an agent, recursively, like in real life. Huh. Talk about shipping the org chart.

Expand full comment
Dave Friedman's avatar

Yeah, something like that. I'm just not sure that there's much of a market for this kind of thing. For most people ChatGPT is AI is ChatGPT.

Expand full comment