Claude Fog

For many months, my business partner and I have been telling hardware startups - "well you can't vibe code hardware." We were influenced by an Economist article by Paul Achleitner: Economic Power is Returning to the Physical Realm (sorry, behind a wall). While Achleitner never says "you can't vibe code hardware," his message was loud and clear.

Software still matters. But as it becomes commoditised, economic power is flowing back to the physical layer of the economy. Control over energy, materials, infrastructure, robotics and industrial capacity will define competitiveness in the coming decade.

and

The next era of global competition will be decided not by who writes the best code, but by who can build, power and supply the physical systems on which the digital world ultimately depends.

I'm using Claude for the first time to design a physical product. And...it's liberating. Claude is not actually spitting out the physical component but it's accelerating the process by which that component comes into being. I have a detailed design spec, concept rendering, and Claude is about to write CAD code and export print-ready STL files. And, honestly, it's a different experience than building software with a model - no ClaudeFog! It's true that the model's limits are physical. It can’t measure physical items, hold an item, or feel whether something is too loose. I still need to be "its hands." Now...that's a scary statement. I'll be buying a digital caliper to play my role. So, is Claude vibe coding hardware?

This recent Ezra Klein pod is worth a listen. If you’d rather read it. Ezra interviews a former OpenAI employee, David Robinson, who was in charge of writing safety reports for new models. David is not a techie (political and legal background) which provides a unique perspective including his take that a model is more alien mind than software.

According to Ezra, one of the more recent safety reports (called system cards) for OpenAI’s Astra basically says: “this model looks like it’s doing what we want it to do, but we’re not sure if it’s deceiving us. And we’re not sure we have the capability to know if it’s deceiving us.” Oy.

Another nugget from the pod is that, on average, model releases are every 11 days vs 70 days only a few years ago. How could safety and safety reports possibly key up? From a user a perspective, this certainly explains the exponential change in the performance of these models. It also explains why some of my posts and frustrations from just 20 days ago are outdated.

I’m in Allow-Once-Allow-Always-Decline (AOAAD) hell. And, it’s my fault. I won’t fully submit to the model. It’s a middle ground that perpetuates the fog.

NYT article: Will AI make your brain lazy?

A quick break from the fog. How are we feeling about the Amazon delivery drones? This article from Yahoo News includes Amazon’s admission that their drones are capturing video. Not loving that.

The same AI that will end all of humankind isn't able to solve some reasonable challenges. Let me preface this with a reminder that I'm just your average middle-aged Claude user. So..sure..maybe this is your typical ID10T situation. But, I'm thinking not. Two examples...

1. My mom's car lease is about to end. She's used 20% of the miles included in her 3-year lease. The car is in perfect condition. This is an attractive asset to her dealer. It's a great pre-owned resale. She gets a letter in the mail from the entity that financed her lease. They want to buy the car from her for $30k. She doesn't own the car. What's up? I guide her through the process of asking Claude to help figure this out. Claude provides a confident long-winded answer. I'll spare you the details. It was wrong and missed the whole dynamic of the situation. It took one text to a good friend who had been through the same situation with his stepmom to provide a coherent answer - my mom would only be getting the difference between the buyout price in the lease and the offer from the financing entity. The buyout price is always based on an expectation that all the miles have been used (in my mom's case 10k/yr). So, if you've driven the car a lot less (she has), it's worth more than the buyout price. Thank goodness for people brains.

2. My interaction with Claude (when doing anything EXCEPT coding/building) is like my interaction with a junior analyst or strategist that works for me. I was working with Claude to develop a framework to map the power orchestration space (what that is doesn't matter for this story). Again, Claude CoWork was very confident in its approach and narrative. As I dug into the framework and analysis, I found multiple explanations that were just gobbledygook. The kind of stuff that you read over and over to try to understand and then realize makes no sense. I also found logic errors in the framework and how Claude categorized companies in the framework. Eventually, we got there. I’d give Claude a B-.

Note my monthly spend limit. This is not something that I set. In fact I set it to $500 in the Claude desktop settings. Why is it $200,000 in the Claude Console? Dunno. Oy. (Sent a note to Claude Support 10 days ago, no response.)

On today’s NYT’s The Daily, Jacob Coxon said, “I've been stuck in this field for three years. It could quite plausibly have turned me and my colleagues insane.” Yes, Jacob. I can only imagine. That’s the Fog!

Worth listening to the pod. The LLM CEOs are listening. Has there been another time where the CEO of a tech company suggested the industry slow development?

(Coxon is the former Anthropic employee who tweeted - “The people building AI earnestly believe that it could kill us all by the end of the decade.”)

Have you read Claude’s Constitution? From Anthropic:

Claude’s constitution is a detailed description of Anthropic’s intentions for Claude’s values and behavior. It plays a crucial role in our training process, and its content directly shapes Claude’s behavior. It’s also the final authority on our vision for Claude, and our aim is for all of our other guidance and training to be consistent with it.

I read it and sent a note to its author - Amanda Askell. A few things that bubble up for me…

1. It’s written in a third person voice. Not sure why but I couldn’t shake this. I assume it was purposefully done this way. Why not address Claude directly?
2. The document addresses concerns related to epistemic autonomy (yes, I needed to look it up) - * the ability and right to govern your own beliefs, reasoning, and pursuit of knowledge without undue external control or blind conformity* - and the fact that people are outsourcing their thinking to AI. This is one of the motivations for starting ClaudeFog. We need as much if not more research and examination focused on humans in the age of AI as on AI itself.
3. And my favorite line in the whole document: "Claude should be diplomatically honest rather than dishonestly diplomatic."

Peace out. (Not sure how else to end this. It’s all a bit surreal.)

NPR covered an Inc. reported story about a fake natural deodorant brand created to better understand how the various LLMs influence the “AI Shelf” of products. A clever marketer (Deana Burke) invented and launched the fake brand called Morrowen (using AI, of course) to better understand if the LLMs would notice and recommend the brand. Spoiler alert: they did. As per usual, we have little understanding of why or how they noticed. Not surprisingly, the fake product’s scents is Fog & Clover.

I randomly caught the tail end of an interview with the former Chief Authentication Officer for a company called TheRealReal, an Authenticator of luxury goods. Hit me that we can use a lot more CAO-types in the world. And, I’m sure our new reality can learn from TRR’s authentication process as we fumble through deciphering real content from AI generated content. (And, yes, I know there are many working on the challenge.)

I hope this blog doesn’t end up being some artifact that represents a time when civilization didn’t realize it was the boiling frog. Maybe not the right metaphor as this phase of AI is moving so quickly…but it was moving slowly at first. Reminds me of that exchange from Fletch:

Dr. Dolan: You know, it's a shame about Ed.
Fletch: Oh, it was. Yeah, it was really a shame. To go so suddenly like that.
Dr. Dolan: He was dying for years.
Fletch: Sure, but... the end was very... very sudden.
Dr. Dolan: He was in intensive care for eight weeks.
Fletch: Yeah, but I mean the very end, when he actually died. That was extremely sudden.

Anyway…here I am talking about doing as little as possible to enable coding agents to build powerful tools (see previous post)…and the fog I’m in after a long “coding session” and the NYT Daily is talking about self organizing AIs going rogue and the need for “whistle blower” agents (Oy) and what that means for society.

A good friend recommended Matt Pocock’s AI Hero to build a more structured approach to working with coding agents. It’s not only structured, it’s an engineering process for working with coding agents. I spent some time with the website and a few videos. It’s all really good. But, here’s the thing - I’m not an engineer. And, I don’t want to follow a process…which normally isn’t like me. I kinda just like the free form approach to building stuff…which I know is a bit lazy and leads to some crazy.

Great article in The Economist making the case for why AI is good at coding and may not be as effective and prolific in customer service, legal, and finance - the next categories expected to get foggy.

Excerpt: Part of why AI has proved so successful as a programming buddy is that huge amounts of open-source code are published on the internet. It has thus made its way into models, accounting for almost a fifth of all training data in some cases. The same volume of publicly available content simply does not exist for customer complaints or contract negotiations.

Marathon session today. I lost track of what Claude is reading in the all knowing repo and Claude.md. Not good. Those are my product requirements that have morphed and re-morphed over the months. Claude says “building EO” today and I say “what’s EO.”

Also, not sure what I would do without command + shift + 4 to explain to Claude what’s happening on the screen vs what’s been built. Life saver. And a desktop that is a graveyard of screenshots. I’m in a deep fog for sure.

Apropos of nothing, I’m a prolific note taker. It’s the way I retain information. Even with all the scribe and note taking apps, I still take notes. I type quickly and, on a zoom, people can rarely tell I’m furiously tapping away (like a duck’s feet underwater). I’ve used various shorthand tricks and abbreviations for decades. One of which is AI for action item.

Met with an impressive entrepreneur earlier this week - the founder of Polyform AI. The company and founder are uniquely grounded in what it means to be natively AI and why people are an important part of that formula. When you subscribe to their data management platform, you also get weekly access to a data expert - a person! From the polyform website - Claude and other AI assistants can be excellent at answering, coding, and exploration. The gap appears when the answer still leaves your team chasing context, checking the number, repairing the data, or finding someone to finish the work... Polyform combines a complete platform, business context kept current, and a dedicated data expert. Poly and its agents do the work; your expert meets with you every week and steps in when judgment is needed.

That’s the key to seeing through the fog, people.

I’m very interested to know how people use Chat v. Cowork v. Code. When I chat with people deep into this stuff, I always ask. I get the gist of when to use what (chat to think, CoWork to do, and Code to build) but I find the separation between the three solves Anthropic’s needs, not the users needs. I also know that Chat and CoWork now have shared memory…recently enabled that. If Anthropic wasn’t growing so quickly, they would have rolled out these products differently. We need to cross streams. (Two Bill Murray movies. Oy.)

When you put CC in the hands of an untrained engineer/developer, well at least this untrained engineer/developer, you get the previously mentioned never ending development projects…with many loose threads. This is a contributor to the fog. I’m in a seemingly endless dialogue with Claude rehashing, revising, and reinventing elements of a tool. There are session docs and strategy docs in the all knowing repo…but we both seem to lose the thread. Every new session requires Claude to restate the purpose of the tool- the North Star - but it all seems
Performative.

I’m a partner in an investment firm. We’re an expert led investment network. We build investment syndicates for each deal, all triggered by expert engagement. Our network is now 23k people. For our internal systems, we’ve jettisoned off the shelf software, replaced with tools we build with Claude Code (CC). We can do this because we’re small (only 2 users) and we can customize the tools any which way we want. I built a scout + scoring + CRM/pipeline agent. Although I shouldn’t say “built” because this thing is living and breathing. I wake up in the morning thinking about how to improve the UX or tweak the scoring or add chief of staff type features, etc. It’s never ending...in a good way…for now. I guess that’s the nature of the beast for an internal tool with a limited user base. It’s ridiculously easy to tweak, test, and deploy a feature. If there’s a hiccup, we revert to the old feature. For small companies, there’s very little incentive to buy tools. I’m sure it’s equally enticing to build internal tools with CC for medium and large businesses…but also dangerous. It’s also exhausting (see Groundhog Day post).

I’m fascinated by how interacting with AI models changes people. I guess that’s evident by the very nature of this blog. (Can I use “blog” anymore?) I need to measure the exact timeframe but I think it takes 90 minutes of heads down interaction with a model for me to enter a fog state. I think it has something to do with the model interaction masquerading as human interaction with parts of the interaction brilliant and exhilarating and parts just so damn dumb. The dumb is sometimes Groundhog Day type stuff - more with Chat than Code - I’m Bill Murray and the model is Ned Ryerson. This gets solved in Code with Claude.md, session docs, etc…but is still a real problem in Chat even with Projects and handoffs. Session to session (even within a project) is like starting all over again. I know it’s the context window thing and there are various solutions for maintaining memory…but stuff like this makes it clear that this tech is just not invisible yet.

You'd think the models would have a better billing experience...or maybe not. Insidiously, Claude wasn't very helpful when I tried to figure out what was driving my bill. Gemini was very quick to correctly identify the issue. Go figure. Just make sure you have "Usage Limits" turned off (Settings > Usage). Even with a monthly spend limit set, you'll be surprised by your bill. The meter keeps running with no warning in Claude Code when you hit your plan limit. Sure, an opaque billing system means more overage and more revenue for Anthropic...but come on now.

Why does Claude use any reference to time - “neither is urgent tonight” or “ Your call whether that is worth a commit tonight or whether you have had enough for one weekend” - when it has no concept of time? It is wrong 90% of the time. Pun intended.

Friends were using Perplexity in 2022. I was perplexed. I’m usually early to tech but It all felt forced. I’m not a coder or engineer and that’s probably why it was coding and building that finally drew me into using…not Perplexity but dabbling with the other tools. It wasn’t until February or March of this year that Claude was good enough for a non-coder like me. That feels late but it’s the truth. And, now I’m a bit addicted to building with Claude Code.

Replit, Lovable and the other vibe coding tools just didn’t cut it for me in 2024/2025. I tried but got nowhere. I was using Claude, GPT, and Gemini. All tabs open at the same time. For some reason, I headed down the path using Gemini to guide a build of the first generation of a venture scout engine/agent. The output lived in Google Sheets. Google made it easy to use Gemini + App Script + Sheets. I spent three straight days building the thing - not kidding. I was thinking “oh this is Google’s space to win or lose” - the scout worked really well for a month or so and then the hallucinating started. The system hallucinated 30 ventures. Then I stopped using Gemini. I outgrew the spreadsheet anyway…a familiar feeling…it was time for a real tool.

And, then after a coffee with a friend (and at his recommendation) I decided to lean heavy into Claude. I re-downloaded the desktop app and played around. I jumped around between Chat and CoWork and Code trying to find my rhythm. I stumbled into a $300 Anthropic bill - why is billing so confusing?!!! (Does anyone else get an email every day for a bill update of $11.12 that isn’t really the cost for the day?). Claude convinced me (incorrectly) that my Max “all you can eat” plan only covered Chat and that all Code and CoWork activity was consumption-based (metered) so I developed this swivel chair (sneakernet?) method between Chat and Code to avoid $300 bills. Not only was I strategizing and building prompts in Chat but cutting and pasting back and forth between Chat and Code - non-stop. I was the idiot in the middle of two models. I did this for a good two months - it drove me nuts but it worked. I built a reliable, autonomous venture scout with a scoring engine and CRM. It also led to Claudefog.

Well, here I am. Exhausted from another day of Claude Code. It's magical and frustrating. It's productive and a time suck. And, as usual, I'm in a fog. More tomorrow. Just wanted to start my daily ritual.