Enterprises need AI middleware, not just models
Plus, strange and fun finds from the World Wide Web and when RTO workers dream of electric sheep.
AI Stone Soup
One of my co-workers, AI Adib, pointed me towards the AI stone soup, uh…analogy? Parable? Anyhow, the point of it is that the models at the core of all this AI stuff are cool, but it’s the layers and layers of applications and people on-top that make the difference. The app that you put on-top of the AI is what matters.
Indeed, the whole reason we’re talking about AI now is because of the app of ChatGPT, then Gemini, Midjourney, NotebookML’s podcasts, Claude, Perplexity, etc. These apps do a lot of the heavy lifting, workflow coordination, and anyone would suspect pre- and post-processing. Even all that vector database stuff.
As ever, it’s the packaging.
All of this amounts to “AI middleware,” all those other ingredients you put into the soup with models. AI middleware isn’t it all either: you need to actual apps too. I’ve used a lot of these apps and their usability and feature set matters a lot. The most valuable is retrieving and searching the web, uploading files and connecting to Google Docs is good too.
I’m thinking that, for “enterprises” looking to do AI, this AI middleware and the app layer are what will matter the most, be the most affordable (ROI’able?) option, and likely the quickest and easiest to succeed at.
//
Speaking of AI, I switched to using Claude over the weekend. For playing D&D it is amazing, so much better than ChatGPT. Its notion of projects is like Perplexity’s Spaces. ChatGPT does this kind of thing with custom GPTs. To the stone soup point, ChatGPT’s approach feels clunky and kind of forgotten about.
Claude will also lets you change writing style/tone. It has some canned ones, but you can “train” it. So, I uploaded some long newsletter episodes into it to see what it came up with. I think it’s fine?
The Model Context Protocol thing in Claude has a lot of potential. It’s trying to be a generic API for interacting with an AI. This includes, of course, bringing text in, but also “writing” out. For my basic uses, since it works with the GUI, you could have it interact with your local desktop, say like random NPC generators and game trackers for D&D.
Claude lacks web search, which is a major problem. If it had that, I could see giving up on ChatGPT.
In all of these, you can see that what matters isn’t the stone of the lower level models, but the layers and layers of app-stuff. Stone Soup!
Relative to your interests
I asked the AIs to write up this episode’s links in a Harper’s/Matt Levine link round-up style, made some tweaks, and here it is:
Strategy has finally outgrown its Sun Tzu phase, embracing gardening metaphors just as Dell's earnings wilt while customers await the next generation of AI silicon. Speaking of outdated frameworks, it turns out employee performance isn't actually Gaussian—though someone should tell HR—and Wardley Mapping continues its to enchant. I asked various social channels about Wardley Map case studies and stories, and there were a handful enough of replies: people love the stuff.
For those exhausted still lost in strategy-land, delight in some hybrid cloud chatter. Speaking of, Tanzu Platform 10 shipped, promising to transform Kubernetes from arcane yaml-craft into something mere mortals might actually deploy—complete with Cloud Foundry characteristics for those who remember when PaaS was supposed to save us all. Oh, and also just with Cloud Foundry for the discerning.
Meanwhile, modern work drowns in SaaS sprawl—though maybe we should just embrace the #defaultslifestyle of Microsoft 365—while growing organizations rediscover why hierarchy and meetings persist, much to everyone's chagrin.
Wastebook
The AI didn’t know what to do with the wastebook section.
“there are more idiots here.”
“Real leopards ate my face energy here” Reid.
“my employer has never wanted me to share an opinion publicly and i do my best every day to ensure they never will” @maya@occult.institute.
And: “leading voice of the goblin web.”
”I have a scraggly patch of hair on my right calf from when I scraped off a swath of skin in an Ultimate Frisbee tournament.” Jason.
“But you wouldn’t comment on telecommunications now, it’s too normalised” Matt.
“Because no matter how much fun we had, it wasn’t the plan.” RotL #559.
“My entire existence of life.”
“OKR stratosphere.”
“I’ll wait for Burger King counter attack.” On Delta’s burgers in the sky.
“Looking forward out in the 20 mile stare.” Furrierism.
“AI that builds pipe, not hype.”
Logoff
I feel like there was more original content I wanted to share, but I can’t find it. Did I share this interview with the aforementioned “AI Adib” yet? In the second half we talk about how executives can think through their enterprise AI strategy.
Meanwhile, why not enjoy these FY27 Q3 vibes:
I let that guy shaking hands go on a little long, but it pays off.