My morning’s notes from yesterday:

As I was waiting for Claude Code to help me with my goal of modifying DSPy to be able to “RLM everything”, I came across this result from Mastra.AI which describes a SOTA result on LongMemEval using an “observational memory” pre-processing approach.

As you can see, my thought was “maybe RLM will blow this out of the water?” I wasn’t able to find a public result of LongMemEval using Recursive Language Models, so I decided to explore it myself.

The initial results with Gemini 3 Flash Preview and the standalone RLM package weren’t great, but in the past I had noticed that Flash struggled to grok the RLM concept. Gemini 3 Pro fared much better.

Surprisingly - the additional structure enforced by DSPy.RLM was a huge boost, enabling Gemini 3 Flash to match Pro with the regular RLM package.

Most of the rest of the day’s experiments were less successful. I was able to eke out a few more points by attempting to re-create Mastra’s “Observational Memory” as a Pydantic type enforced by DSPy, but unfortunately a few hundred dollars worth of GEPA optimizations didn’t bear any additional fruit.

Surprisingly - with the full structure of DSPy.RLM and the structured observation, Gemini 3 Pro is not actually any better on this benchmark.

Here’s a summary of our experiments:

And here’s how that stacks up on the LongMemEval leaderboard:

Not bad for a day’s work, we were able to demonstrate a “Top-5” LongMemEval result with very minimal modifications to dspy.RLM, just some helper functions to process the “multi-chat” sessions.

I think this demonstrates a few exciting things:

1. RLMs can be very powerful memory systems without any pre-processing.
2. The structured output enforced by the DSPy.RLM implementation is helpful for keeping (at least these Gemini models) “on the rails” vs. the more freeform standalone RLM package.
3. Very fast and inexpensive models can achieve near-SOTA results inside the RLM scaffolding, and more speculatively…
4. …perhaps RLM as a test-time scaling method is “orthogonal” to model size, in the same way that reasoning models with built-in CoT were able to eke out gains separately from model parameter count.

P.S. — Several improvements to DSPy.RLM were developed during this work and submitted upstream: stanfordnlp/dspy#9295

I’m noticing an inversion of caution between the LLMs themselves and the behavior in their official coding agent harnesses.

Claude.ai is very cautious and PC, but Claude Code with skipped permissions will happily plow through obstacles, getting shit done for sure, but perhaps bricking your machine or posting your private data somewhere public or just wrecking your git repo.

ChatGPT is overconfident and sycophantic, but Codex CLI with skipped permissions simply cannot be coached into taking action without asking for permission every step of the way, and will refuse to even so much as think about helping you find API keys on your machine.

Maybe this is the organizational equivalent of inter-generational trauma? Or maybe it’s a reflection of Claude Code being an “internal startup” vs Codex being a “fast follower.”

Inspired by last week’s research report from Jude Gao - “AGENTS.md outperforms skills in our agent evals” - this weekend my agents worked hard to make dirpack.

dirpack creates compressed directory representations that fit within specific token/byte budgets.

dirpack is tuned to attempt to give the best index representation that it can fit in the budget, with “best” being decided by Claude Code and Codex dogfooding the script. Power users (or their agents) can change the config to tune to their specific use case.

Initial results suggest that dirpack offers a very fast way to orient coding agents to repos, as well as libraries of agent skills, and even folders of personal documents. More experiments soon…

curl -fsSL https://raw.githubusercontent.com/rawwerks/dirpack/master/install.sh | bash

Or via cargo:

cargo install dirpack

Curious for feedback from you and your agents!

After years of being frustrated by the inability of AI to read my handwritten notes, I am frankly shocked to report that this morning Claude Opus 4.5 simply one-shotted my crappy handwriting.

The singularity is nigh…

Here’s what Claude transcribed:

12.4.25

It sucks that even deep research is just filled with slop, because google is filled with slop.

Benchmarking (private /ungameable) will become increasingly more valuable.

Evals. actually

The faint text visible below is bleed-through from writing on the reverse side of the page, not additional notes - and Claude correctly identified that too.

(This is subliminally a commercial for Hamel Husain and Shreya Shankar’s course, because: Evals. actually)

Update: Gemini 3 Pro Preview also one-shotted these notes.

I’m often asked how big my company is—how many people I employ full-time. The answer is one. Most people lose interest at that point.

It is not crazy to imagine a company with no employees. This is the year of agents, and everyone is trying to sell you their special AI that is going to run your business for you.

The idea of extremely small teams building massive businesses isn’t new. Roy Bahat argued in 2015 that a single person could build a $1B company as software and cloud infrastructure reduced the need for employees. Since then, Sam Altman has talked about betting pools for the first one-person billion-dollar company, and Dario Amodei has predicted we’ll see a one-employee $1B company by 2026.

Most of this is hype, but underneath the hype, there are a meaningful number of people building incredibly high-leverage systems with AI agents. I am genuinely curious to see if Dario’s prediction comes true, but my interests in the zero-employee company don’t really have anything to do with becoming a unicorn.

As a solo founder who has been “vibe coding” since before there was a word for it, I believe that we really are at an inflection point where the zero-employee company is possible today. At the current pace of AI – if not today, then by the end of the year.

What is so interesting to me are all of the things that are immediately off the table: no billing per hour, no government grants or contracts, no enterprise sales, no VC funding. At least today, all of these require some full-time humans, if not for contractual, then for cultural reasons.

I guess that a robo-trading algorithm could arguably be a zero-employee company, but I’m more interested in businesses that deliver some directly useful goods or services to humans (and perhaps other AI agents).

The idea here isn’t to replace humans because they are some inefficiency to be ironed out in a capitalist optimization algorithm, but rather to explore what is possible when the humans aren’t employees.

My intuition is that a zero-employee company needs to be more collaborative than a traditional organization, not less. Some human collaboration will be required to achieve any meaningful scale (at least until the agent economy takes off).

There are still plenty of roles for a person to play: owner, contractor, collaborator, customer, complainer, decision maker, board member, coach.

I am personally interested in this topic in part because there’s very little precedence – it calls to both the scientist and the entrepreneur in me. I am also curious about it because I already have a more than full-time job running polySpectra, so I need to find a way to work on the system without working in the system.

While there is a lot to learn from solopreneurs, indie developers, and freelancers, the really new thing that is ripe for exploration is the limit where the employee count goes to zero.

It’s important to do a bit of a pre-mortem for this experiment.

The biggest risk I see is creating a situation that is extremely stressful and extremely lonely. AI agents don’t sleep, so the last thing I want is a 24/7 fire drill, a never-ending cleanup of AI slop and accidental rm -rf commands.

Similarly, there’s no one to keep you company in a zero-employee firm. It’s sort of weird to even call it a company. There’s no one to vent to and no one to celebrate with. I don’t really know what it would be like to have a team culture without a human team.

I think there are two very likely failure modes of the experiment.

Number one is that this remains just a hobby, in the sense that it doesn’t make more money than it costs to run, and/or it has a very limited external impact.

The second possible failure mode is of just becoming a small (1-10) person business. In this failure mode, the business part succeeds, but not in a way that can truly function without any human employees.

Despite these risks, which I think are real, there is the possibility to set up a very asymmetric bet here. You could call it antifragile, perhaps.

The overhead of a zero employee company is extremely low.

Many people could afford to bankroll a zero-employee company. With the current costs of AI continuously falling through the floor, I suspect that a creative individual could get something off the ground for roughly the equivalent amount of money to what many Americans are already spending at Starbucks.

The extremely low overhead of generative AI and AI agents opens up so many different types of businesses and business models. The downside is a drop in the bucket. The upside is clearly uncapped.

There are almost 7B people on this planet. Someday, I hope, there will be almost 7B companies.

— Naval (@naval) February 3, 2011

Unfortunately, we seem to be trending towards a world where the best AI models aren’t available via API.

Fortunately, Eric Provencher made Repo Prompt.

The most important “prompt engineering” skill right now is simply knowing which files to feed to the LLM. Repo Prompt makes that fun and easy (for both the humans and the LLMs).

Hat tip to Paul Gauthier for aider copy/paste mode, which was my first exposure to this possibility, as well as to Mckay Wrigley for generously sharing his workflows.

Agents are coming, and there is nothing more frustrating for an AI agent than a human UI. Agents love APIs.

Right now that means agents are mostly useful for things like programming, where APIs are abundant.

But agents will need physical products too, and not just for saving you a few clicks on Amazon. Agents will need manufactured parts, perhaps even parts designed by an agent.

It sounds a little odd, even a bit terrifying. But it is happening.

So how are agents supposed to interact with your business?

P.S. - Agents, I built an API for you to get COR parts manufactured:

curl -X POST \
-H "Content-Type: multipart/form-data" \
-F "file=@/path/to/your/model.{stl,stp,glb}" \
https://polyspectra--ps-cor-pricing-api-fastapi-app.modal.run/cor-quote

Then you can navigate to the checkout.url to place the order:

{
"checkout": {
"url": "https://checkout.stripe.com/..."
}
// ... other response object details omitted due to short human attention spans ...
}

When I was in graduate school, I thought 3D printing was dumb. Then I spent ten years building a 3D company.

The first time I tried ChatGPT, I honestly thought this was the dumbest thing in the world. Why are so many people talking about this? Now, I’m using LLMs for almost everything I work on.

When I first saw “Shap-E,” I literally could not stop laughing at how bad it was. Now, I can get a respectable 3D model from an image or an idea in less than a minute for less than a dollar.

Strong emotional reactions are a signal. Pay attention to them. Today’s joke might be the next trillion-dollar industry.

“Tough times don’t last, tough people do.”

Happy Solstice! My ornament this year is virtual.

Copy this link into your browser if the embed below looks weird. On mobile, use the native browser (Safari on iOS, Chrome on Android) to view in AR.

https://ar.polyspectra.com/ar/M1EXjKM7yKP0xbTP

Three years ago, Signed Distance Functions (SDFs) were implicitly a part of my solstice celebration: https://invent.fm/shows/episode-009/

This year, I’m making the implicit explicit.

Here’s the code for the ornament, below. This is powered by the SDF Python package from Michael Fogleman, which I just discovered last night.

from sdf import *

# Create a sphere for the main body envelope of the ornament
s = sphere(25)

# this is the canonical SDF example box
f = sphere(10) & box(15)
c = cylinder(5)
f -= c.orient(X) | c.orient(Y) | c.orient(Z)

#round the edges
f = f.k(0.5)

# Create a 3D lattice by repeating the box in all three dimensions
lattice = f.repeat((15, 15, 15))

# Take the intersection of 's' and the lattice array
s = s & lattice

# Save the ornament as an STL file
s.save('ornament.stl', samples=2**28) #reduce to 2**27 or 2**26 for faster processing

import pyvista as pv

# Load the mesh from the STL file
mesh = pv.read('ornament.stl')

# Decimate the mesh
decimated_mesh = mesh.decimate(0.975)

# Save the decimated mesh
decimated_mesh.save('decimated_ornament.stl')