Rangeley, ME in peak fall foliage. It's lovely year-round.
What got me thinking about this: 3 different collections of 2K-ish people
When I interviewed the the delightful @gabboman.xyz and @djara.dev of Wafrn I asked them if Wafrn felt like a small town where you kinda know everybody. At interview-time Wafrn had 4606 total and 846 monthly active users, and numbers reminded me of the small town where I grew up (Rangeley, ME...with ~1,000 people year, jumping to ~2,000 in the summer). And I thought of this yet again after Steve Yegge's The Anthropic Hive Mind and then hearing that Anthropic has 2,500 employees. Somehow I don't imagine working at Anthropic feels like living in a quiet small town.
Why is Anthropic so small?
We won't dwell on this because it's no mystery:
2,500 people = 2,500 * (2,500 - 1) / 2 potential one-to-one relationships, growing quadratically.
And yet...you gotta know the employees are stretched incredibly thin. I have a couple friends who have Blue Chip software jobs. They are always incredibly busy. Their salaries are more than double what you could comfortably live on, so why can't you just split their job in half and give it to two people like Jerry and George in that episode where they team up to handle all the admin/tasks necessary to be one competent boyfriend?
I wish we lived in the world where two-people working half-as-hard = 1 person. It would be a world where more people could legitimately say that their work had impact, and a world where the load/burden could be shared better.
Could I actually hack it as an Anthropic employee in 2026? I honestly don't know. But could I do the work of half-an-Anthropic-employee? 100% yes. And I'd joyfully do it for half-an-Anthropic-salary + stock options.
Context Window Maxxing with Ralph
I have been building jklb.social using Geoffrey Huntley's Ralph method (& blogging as I go). When you Ralph, you break your work into small tasks and then assign those tasks to their own Ralph loop, every Ralph getting its own fresh context window.
On the face of it this is kind of nuts. Every single thing on your codebase was done by an agent who, effectively, was on their first day of the job.
This is a feature, not a big. This means that my code has to be well organized, and my docs (in specs/ * .md) must be immaculate so that every first-day / fresh-context-window Ralph will get their detailed assignment, find just the right files, and make just the right changes.
Do the Ralphs bat 1.000? Of course not. But a second Ralph can fix the mistakes of a first Ralph better than a Ralph can self-correct. The method continues to surprise me.
A Context Window is kind of like a one-to-one relationship
The significantly-more-than-2,500 one-to-one relationships at Anthropic are arguably the soul of Anthropic. That's where the stuff that's not written down lives.
When you're working with a Claude Code instance, the context window is a kind of like a one-to-one human relationship. You go back and forth, and between you and the machine something subtle gets built up over time.
Maybe the back-and-forth is around some technical decision that needs to get made. You share your priorities. Claude Code pushes back with technical constraints. Over time, if it goes well, you come to a kind of agreement.
The Ralph method says, "Hey that's bad practice, actually. You can't inspect that, the subtle understanding/flow of a context window. So 1: no way to know if it's trash or not and 2: it can't be reproduced and 3: even if it's good it's fragile as all hell & can be ruined by a single bad token (which, let's remember...generated statistically)." So the Ralph method is all about finding agreement/insight with LLM assistance, and then immediately getting it down in markdown. And then organizing that markdown so future LLMs--all of which effectively are on their first day of the job--will find just the thing they're looking for.
~steps~onto~soapbox~ We gotta find a way to give more people agency!
A bit of a tangent, but this is something I feel strongly about. The current status quo serves NO ONE. We are sleepwalking towards a world of two classes, both getting a raw deal.
Class 1: The agency-deprived class
Joe Rogan five-hour podcasts are for people who have jobs that require their bodies to be in a place for 5+ hours but ask nothing of their minds. The subtext of every Candace Owens monologue is "Oh, you think I'm stupid? Well would a stupid person be able to find a youtube clip from ten years ago of a girl from Arizona calling their Swedish grandparent something different than what someone who actually speaks Swedish & lives in Sweden would call their grandparent?"
As humans we need to feel like we matter and have agency to be happy and healthy.
Even without LLMs this is a massive problem. If you live below the API line, if your boss is an algorithm exposed to you via an app? That's soul crushing. That's going to make you feel like you don't matter. And since (my opinion) we kinda by definition all walk around thinking we understand the world, if that is your life your "understanding" of the world is going to trend conspiratorial and bleak.
LLMs supercharge this.
Class 2: The drowning-in-agency class
The people at Anthropic, with the Blue Chip software jobs, are busier than people should be. Too busy to stop and smell the roses. Too busy to enjoy the free amenities at the office. Feeling the (actual) weight of the world on their shoulders not because they're particularly smart (though they are smart), but because one-to-one relationships in an organization grow quadratically, and they happen to be the cooks in the kitchen that can't have too many cooks. They steward, in the delicate, ineffable, deeply human whatever of their relationships with their colleagues, the secret sauce of a billion dollar company that is charting the future course of humanity. I can only imagine how exhausting that must be.
Speculative Fiction: A Day at the Ralph Corporation (of humans and LLMs)
During Alice and Bob's weekly one-on-one an LLM silently adds to (& sometimes removes things from) the company-wide specs/*.md whenever they come to anything like a conclusion or point worth remembering.
"Alice and Bob talked about X and decided Alice will look into Y when she finishes with Z"
"Alice reminded Bob that Z is not exactly blocked exactly by A, but her approach to Z will vary greatly depending on what the B team decides about C"
"Alice and Bob talked about D, and realized that neither of them really understands D, despite Bob's job title literally being 'Senior D Engineer'! They had a nice chuckle about this."
"12 minutes of Alice and Bob's 30 minute checkin today were unrelated to work (Alice checked in on Bob re: his brother's recent car accident. Bob appeared to genuinely appreciate Alice's asking). This is acceptable & should NOT be flagged to HR (despite general best practices...that one-on-ones should be at least 20 minutes work related). In fact, this "breaking" of one rule is actually a great example of Alice demonstrating Company Value E."
There's no privacy at Ralph Corp. Like ATProto, private data is something they talk a lot about but haven't solved.
The Ralph Corporation's specs/*.md are baroque. Human-readable, but far too big for any human to read in their entire lifetime (or for any LLM to put entirely in their context window). There are literally people who's jobs are entirely just keeping the specs/*.md as up to date as possible!
Some companies are Open Spec. This means their specs/*.md are public. This is universally acknowledged as insane, but it means you get less pings from vendors asking, "what's the status of such and such." Most companies consider theirs specs/ directory as their most important asset. Copying the specs/ onto your personal machine can get you fired & sued into oblivion.
Humans at the Ralph Corporation judge their daily effectiveness by the length of the diff they auto-generate in specs/*.md. When you have a really productive day you say you "wrote a novel," though it's also acknowledged that if every one of your engineers is generating that kind of change every day that's probably not a great sign.