Ben Hofferber

Ben Hofferber

@benhofferber.com

Host of stablediscussion.com AI Product Builder hint.services Speaker, D&D Enthusiast

36 followers47 following84 posts

Cowork by Claude is an easier to configure Claude Code. But be warned, I do find it's pretty sketchy to use since it's in beta still. Asked it to organize my downloads (after backing up!) and it deleted files that were too big to copy (so lazy!) claude.com/blog/cowork-...

claude.comIntroducing Cowork | ClaudeClaude Code's agentic capabilities, now for everyone. Give Claude access to your files and let it organize, create, and edit documents while you focus on what matters.

It later deleted a series of .pem permission files I'd downloaded (don't need those - better regenerate them anyway) lol It's definitely a nice interface thought and falls right in line with where I've been wanting to push UI for AI

Smithery has become so unreliable for MCP hosting that I'm moving off of it completely. Their own servers aren't reliable and I'm frustrated they seem to be leaning into being a MCP-powered ChatGPT wrapper instead of becoming the infrastructure company they initially indicated. Oh well...

Feels like every AI release I try this week is broken. Claude Code for Web is awesome but I can't seem to figure out when servers are running or not. They seem to just stop responding at times but sort of work better on the mobile app for some reason...

So far my Cursor 2.0 experience is very underwhelming... Tried Cursor 2.0’s new cloud agent. It pushed changes to a branch fine, but then got stuck looping its own edits and stopped responding to new requests. Web version seems better at this point.. which is funny in it's own way

An AI browser isn’t very high on my list of things I want these days but there’s a lot of perceived need with Atlas coming out. Perplexity is feeling the pain I guess… Meanwhile I’m just over here vibing with Dia :)

Clause Skills are essentially Claude Code Commands that can run within your Claude App. I love doing AI tasks on my phone on the go but this still doesn’t feel like it for me. Definitely cool but I don’t think I’ll be able to step away from Claude Code long enough to adopt

Claude Code works really great until it hit's a problem that is sufficiently difficult. Just hit a wall where it wasn't able to understand why a React modal was kept open and the close buttons weren't working to close it.

The bug actually happened because we earlier added a click handler on the parent component it would immediately re-open the modal. That on its own seemed like a good change and without context for the underlying modal on the clicked element, it didn't seem to impact much.

Gemini 2.5 pro in Cursor was able to figure it out because I could shape the context of what it knew better around the problem rather than just relying on Claude Code to try to figure it out. These tools are cool but still important as ever to know escalation paths to solving problems

Getting tips on where to go next in a video game when I'm stuck is my most common genre of AI conversation. I talk at my computer to dictate things but it's rare that I want to have a conversation with a yes-man. Even longer therapy-lite conversations end up just being a way to collect my thoughts

Ben Hofferber reposted

I had early access to what is Claude 4 (I don't know which mode yetl, they had codenames) & I have been very impressed Fun example, this is what it made in response to the prompt: "the book Piranesi as a p5js 3d space. do it for me" - just that, no other prompting (note the birds, water, lighting)

Feels like ChatGPT is hanging on by a thread these days. Rough prompting happening on 4o, images failing to generate repeatedly, chat unable to edit documents, memory keying in on the wrong details… It’s hard to build an AI experience and stay at the edge of all AI. Someday they may need to focus.

Realizing that some of the things we attribute to AI were already trending before AI… Specifically small mostly senior teams has seemed to be a multi-year trend at startups I’ve interacted with. While AI didn’t help, I think we have the causes wrong on this one but the discussion is on rails now

View full thread

𝗧𝗵𝗲 𝗙𝘂𝘁𝘂𝗿𝗲 ✨ I'm incredibly grateful for everyone who's supported this journey so far. The constant experimentation, the late nights, the uncertainty of finding our niche - it's all been worth it to build this community.

View full thread

I've built a discussion-based channel where: • You'll discover AI content you won't see elsewhere • I showcase practical tools you can actually use • We explore fascinating trends emerging in the field • I share insights from my AI work at Hint Services

The growth to 3.4K subscribers represents something beyond numbers - it's validation that there's genuine hunger for substantive AI discussions that go deeper than the typical tech news cycle.

It's almost hard to believe. Abdella Ali and I launched our AI podcast with fewer than 5 subscribers and a passion for meaningful conversations about artificial intelligence. Fast forward through 52 videos of consistent grinding, experimenting, and refining our approach

𝗧𝗵𝗲 𝗘𝘃𝗼𝗹𝘂𝘁𝗶𝗼𝗻 📈 While Della has since moved on to explore other interests (though our conversations still fuel some amazing content - reunion episode someday? 🤞), I've been focused on carving out a unique space in the AI content world.

Ben Hofferber reposted

Large scale job displacements due to AI are likely to occur more slowly than a lot of people talking about AI and work might suspect. Translators were paid more and in more demand a year after GPT-4 than they were the year it launched. Change is likely to come, but the timing is hard to predict.

Yikes! MCPs completely changed in #Claude Desktop and have pretty bad visibility now… feels like a dev feature escaped to prod Intended @anthropic.com ? I’m seeing the tool calls but no way to click in for details on the calls. I see a Left chevron but no functionality

Ben Hofferber reposted

Replying to @handle.invalid

Model Context Protocol (MCP) client support allows your app to connect to hundreds of pre-built tools: GitHub, Slack, filesystem operations, and more. Your users can also build and connect their own custom MCP servers to extend your application's functionality.

Someone asked me on my YouTube channel if it’s possible to automate grounding of AI Agents. I think hard to ground via interfaces available and I don't think it's a coding problem as much as a people building poor interfaces problem. That makes me think this isn't getting automated anytime soon.

View full thread

That said, there are many companies focused on aggregating access to interfaces. If those interfaces can be sufficiently standardized and the abstractions those interfaces provide are held, then I could see it being possible. I'm not very confident in that given what I've seen of interface design

Someone asked me on my YouTube channel if it’s possible to automate grounding of AI Agents. I think hard to ground via interfaces available and I don't think it's a coding problem as much as a people building poor interfaces problem. That makes me think this isn't getting automated anytime soon.

Grounding is integration based as I showed. Integrations are one of the easiest spots for a breakage to occur and when it comes to confirming if an integration is setup correctly, that's a hard job. There's just a lot of things that can go wrong.

Thinking it might make sense to spin up a Bluesky Likes export to Obsidian. Love the knowledge collection and I feel like this should exist… At least provides something similar to bookmarks for now

Ben Hofferber reposted

I see a lot of (correct) complaints that AGI and agents are badly defined. This problem will not be solved because: 1) AGI and agents inherently rely on comparisons to humans, and we don't have good definitions of human agency or general ability 2) Marketing is incentivized to blur any definitions