The Computomatix Times

A 21-year-old college student in Austin runs an OnlyFans account that cleared $43,000 in his first 30 days.
The girl on the page doesn't exist.
Her name is Maya, on paper. 22 years old. UCF psychology dropout. 1,247 paying subscribers. Her top fan has paid $1,847 in messages and content over the last month. Average revenue per fan: $34. She does not sleep.
There is nothing to film. There is no one to type. Claude Code writes every message. Flux generates every photo. ElevenLabs generates her voice. Maya is four .md files in a folder on a MacBook in Austin.

This is not theoretical, and he is not first.
Aitana López

A pink-haired AI model from Barcelona, brings in up to €10,000 a month for The Clueless agency.
Fortune covered her in 2024. She has done campaigns with Victoria's Secret, Razer, and Olaplex. A 5-million-follower Latin American actor slid into her Instagram DMs trying to ask her out, not knowing she was AI.
Emily Pellegrini

Emily Pellegrini is 21, from Italy, on paper. She was generated in Midjourney by an anonymous creator who told the Daily Mail how he built her:
"I asked ChatGPT what's the average man's dream girl, and it said brown hair and long legs. So I made her exactly how it said."
None of them figured out she was not real. The creator who built her stays anonymous to this day.
Aitana took eighteen months to build. Emily took fourteen-hour workdays for half a year. Maya was built in four weeks.
Maya is four files.
persona.md is who she is. UCF psych dropout, '04 baby, Pisces. Step-dad she hates, fake brother in Tampa. Loves Lana Del Rey and chicken tendies. 1,400 words of biography that never break.

voice.md is how she sounds. Three audio samples cloned in ElevenLabs. Voice notes drop at 11pm her time. Top fan in Berlin gets them at 6am.
flux.md is what she looks like. 47 reference shots, one LoRA, three lighting setups. Bedroom mirror. Bathroom mirror. Kitchen counter at 2am. Every image has a tiny scar on her left wrist that she never explains.
brain.md is what she remembers. JSON, one entry per subscriber:

Claude Code reads all four before every message. It never forgets a name. It never breaks character. It sleeps when the creator sleeps and catches up at 7am with "sorry babe just woke up 🥺".

Four weeks, in order.

Week 1 — persona.md.


1,200 words. Three sections:
backstory:

forbidden topics:

voice rules:

Read aloud. If it sounds like Wikipedia, delete and restart. Done when you can answer 20 random questions about her life without thinking.
Week 2 — flux.md + LoRA.

Pick a base face. Lock 6–8 descriptors:

Generate 47 variations of that face. Fine-tune a LoRA on them. ~$80 on a rented A100.
Lock three seed ranges, one per setup:

Different seeds across setups = different jawline. Lock them or she stops looking like herself.
Done when 10 fresh generations from each setup pass as the same person.
Week 3 — voice.md + ElevenLabs.
Buy 90 seconds of clean audio from a Fiverr voice actress, ~$40. Clone in ElevenLabs Instant Voice. Ten minutes. Add audio rules to voice.md:

Generate 30 test voice notes. Done when none of them sound like a podcast intro.
Week 4 — brain.md + orchestrator.

Claude Code reads persona.md + voice.md + brain.md before every reply. Then a second pass extracts new facts from the user's message:

System prompt:

Hook order: read .md files → reply → extract → append. Run on a cron every 30 seconds polling the inbox. Done when Claude holds 50 messages without contradicting brain.md.

This was not possible eighteen months ago.

Aitana took eighteen months. She was built before any of this shipped.
Emily took six months. Half the stack was missing.
Maya took four weeks. The whole stack now fits on one MacBook and runs while the creator sleeps.
The next Maya is a weekend.

The math.
$43,000 in revenue.

To a 21-year-old who paid zero dollars for talent, zero for filming, zero for editing.
The only labor cost was four weeks of writing four markdown files.
OnlyFans is the wedge, not the product.
The same four files run an Instagram fitness account in São Paulo. A TikTok cooking persona in Seoul. A Twitch streamer who plays Marvel Rivals at 3am Pacific. An X account that posts crypto takes between sponsored DEX shills.
None of them exist either.
The bottleneck is not compute. It is not GPUs. It is taste — knowing which lies a stranger wants to believe.
Anyone can be a folder now.

Bookmark this. The question isn't if you've followed an AI. It's how many.

For 90 days I logged every single Claude Code session: timestamp, prompt, response, token count, model, exit reason.
430 hours of work. 6 million input tokens. $1,340 in API spend.
Then I sat down with the data and asked the only question that mattered: how much of this was actually doing my work, and how much was overhead?
The answer was uncomfortable: 73% of my tokens went to nine invisible patterns that I'd been doing on autopilot.
Not bad prompts, I write decent prompts.
Not big models when I needed small ones, I knew that one already.
Patterns deeper than that. Patterns nobody talks about because they're invisible until you instrument them.
If you're hitting Claude Max usage limits more than once a week, you have at least 4 of these. Probably 7.
Below: each pattern, exactly how much it costs, and the 30-second fix.

Why this matters now
Anthropic admitted the problem in late March 2026: "people are hitting usage limits in Claude Code way faster than expected."
Max 5 subscribers reported quota exhaustion in 19 minutes instead of the expected 5 hours.
The Pro $200/year users said the limit maxed out every Monday and didn't reset until Saturday - "out of 30 days I get to use Claude 12."
Anthropic acknowledged the issue. The peak-hours quota change in late March explained part of it. The rest was a prompt-caching bug — two independent bugs in the cache layer that silently inflated costs by 10-20x for some sessions.
A user reverse-engineered the Claude Code binary to find it (GitHub issue #40524). Downgrading to v2.1.34 helped some people.
Some of it is real. Most of it is you.
I'm not going to tell you to "use Haiku for simple tasks" or "start a new chat every 15 messages."
Below: the patterns that the obvious advice misses.
The methodology
I ran an HTTP proxy between Claude Code and the Anthropic API. Logged every request: full payload, response, token counts (input/output/cache), latency, model. 90 days. 430 hours of active work.
For each request, I categorized the tokens into:
Productive — content that directly informed my actual question
Cache hit (free) — system prompt + CLAUDE.md cached, no marginal cost
Cache miss (paid) — same content, recomputed because cache expired
Conversation history re-read — re-tokenizing previous messages on every turn
Hook injection — pre-pended context from PreToolUse / UserPromptSubmit hooks
Skill loading — skill SKILL.md content loaded into context for invocations
Tool use overhead — JSON schemas, tool definitions, tool result blocks
Extended thinking — <thinking> blocks
CLAUDE.md — project rules loaded every turn

Productive tokens: 27%. The other 73% is the 9 patterns below, ranked by how much they cost me.
Pattern 1. CLAUDE.md bloat (~14% of total tokens)
The pattern: my CLAUDE.md grew to 4,800 tokens over 6 months. Every turn loaded all 4,800 tokens. Every session loaded them again on each new request. Most of the rules were never relevant to the task at hand.
A 5,000-token CLAUDE.md costs you 5,000 tokens before you've typed a word. Every turn. Every session. A constant baseline tax. Multiply by 200 turns per week and it's 1 million tokens a week of your CLAUDE.md alone.
The 30-second fix:

If you're over 1,500 tokens combined, refactor:
Move framework-specific rules to project-level CLAUDE.md (only loads in that project)
Extract repeated patterns into skills (loaded only when invoked)
Delete anything you can't remember writing
Convert "explain why" verbose rules into 3-word imperatives
I cut mine from 4,800 to 900 tokens. Same behavior. 31% reduction in baseline cost, instantly.
Pattern 2. Conversation history re-reads (~13% of total tokens)
The pattern: every follow-up message re-tokenizes the entire conversation history. By message 30 in a chat, each turn is paying for messages 1-29 to be read again. The math: at ~500 tokens per exchange, message 30 costs 30× message 1.
I had sessions with 60+ messages. The last message was costing 60× the first. Tokens spent re-reading old context: catastrophic.

The 30-second fix:
Edit the prior message instead of follow-up. Up-arrow -> edit -> re-send. The bad exchange gets replaced, not stacked.
Hard cap conversations at 20 messages. When you cross 20, ask Claude to summarize what's been done and start a fresh chat with that summary as the first message.
Use /compact instead of /clear when you need continuity. /compact summarizes and restarts. /clear nukes everything.
I went from 60-message sessions to 15-message average. 40% drop in conversation re-read cost.
Pattern 3. Hook injection waste (~11% of total tokens)
The pattern: I had 4 plugins installed. Three of them registered UserPromptSubmit hooks that injected context. Combined: 6,200 tokens of hook injection on every prompt I submitted, before Claude even read what I asked.
These hooks are designed to be helpful. They inject branch names, recent file changes, instinct summaries, memory snippets. Each one is small. Together they're a wall.
The 30-second fix:

Aud… continue on X ↗

Apple is already dead. They just haven't filed the paperwork.
That's not a hot take. It's a structural read on what just happened in the last six months & what Naval Ravikant confirmed on his podcast last week. The most patient investor in tech & one of the sharpest capital allocators of the last 20 years just gave a verdict on the entire software industry: pure software is uninvestable.
If you're a founder reading this, the question isn't whether you believe it. The question is whether you have 18 months to reposition before the market notices.
For context: Naval founded AngelList, was an early investor in Twitter, Uber, Notion & roughly 200 other companies that shaped the last decade of tech. He doesn't post often. When he does, he picks his words like a man who knows they'll be quoted back at him for years. So when he says "pure software is uninvestable" with no qualifier, it's not commentary. It's a call.
Here's what he said & what it means for everyone building right now.
No one can stop Apple's structural death
Apple isn't going bankrupt. Apple won't disappear from your pocket next year. The collapse Naval is describing isn't operational. It's economic.
Apple's entire $3 trillion valuation rests on one thing: premium hardware margins justified by superior software experience. Take that experience away & Apple becomes Samsung with better build quality. That's exactly what's happening.
The interface layer is commoditizing in real time. Within 24 months, most people won't open apps the way they do today. They'll talk to an agent. The agent will generate whatever interface they need on the fly. Apple's curated app store, the human interface guidelines, the design polish, the ecosystem lock-in - all of it becomes irrelevant when the interface itself is generated in real time by an AI that runs on any phone.
Apple's response to this transition? They licensed Gemini from Google. Their own AI bet underdelivered. The company that built its entire identity on owning the experience layer just outsourced the experience layer to its biggest competitor.
This is the Microsoft-after-mobile playbook running in fast-forward.
Microsoft missed mobile because they refused to build a touch-native OS from the ground up. Their dominance in the previous era convinced them the old paradigm would hold. By the time they accepted the new one, Apple had already won the next decade. Microsoft is still worth $3T today, but Microsoft Windows lost the consumer war they could have won.
Apple is making the exact same mistake right now with AI. They're betting their hardware-first identity will carry them through the agent transition. It won't. When the OS commoditizes, Apple's margins compress to commodity hardware levels. That's a structural revenue collapse in their highest-margin segment, the one that funds everything else.
You can hold Apple stock through this. Just don't pretend you're holding a growth company.
The most valuable hardware company in history is about to find out what its hardware is worth without the software moat.
If your moat is software, you have 18 months
Now the harder part if you're a founder.
Naval said pure software is uninvestable. He's right. But he didn't unpack what that means for the tens of thousands of SaaS companies currently sitting on Series A & Series B valuations they raised in a different world.
It means most of them are already dead. They just don't know it yet.
Here's the math. Your SaaS company exists because building your product was hard. You raised capital because technical execution required a team. Your moat, whether you admit it out loud or not, is the difficulty of replicating what you built.
That difficulty just collapsed.
A 2-person team using Claude Code can now replicate 80% of most B2B SaaS products in under 90 days. Not a toy version. A working version. With proper architecture, basic security, room to scale. The remaining 20% - your specific integrations, your enterprise sales motion, your compliance stack - is real. But it's not a moat. It's friction. & friction gets compressed by the next generation of agents shipping every quarter.
Look at what's already happening. Adobe acquired Figma for $20B in 2022 because Figma's product was structurally hard to build. Today, design tools with 70% of Figma's core functionality are being shipped by solo developers in months. Salesforce is the most valuable SaaS company in history. AI-native CRMs that didn't exist 18 months ago are already eating its mid-market. Workday. ServiceNow. Atlassian. Asana. Every one of them is now a candidate for replacement by an AI-native alternative built by a team smaller than their HR department.
The companies that survive this transition won't be the ones with the best software. The software is going to zero. The companies that survive will be the ones that built something the AI cannot copy:
Distribution. Network effects. Data flywheels. Hardware integration. Brand. Community. Regulatory depth. These are … continue on X ↗

Capital allocation drives the economy.
It decides which drugs get developed, which technologies get built, and which ideas survive long enough to matter.
And yet, for something so central, it still runs on a fragile foundation: human judgment.
Even the best investors are limited. They can only track so many opportunities, process so much information, and make so many high-quality decisions. The difference between average and exceptional allocators is enormous—but that edge is locked inside individual minds.
That makes it hard to examine clearly, hard to reproduce consistently, and hard to improve over time.
AI changes the equation entirely. Agents can absorb more information, connect more dots, and evaluate more possibilities with a consistent standard than humans can on their own. What was once left to individual judgement can become an optimizable system. A new hill to climb.
That’s why I’m starting Abundance. We’re starting in public markets, where feedback is fast and unforgiving. Over time, we expect to extend the same system across other asset classes. We don’t intend to sell or license this technology. We plan to use it ourselves. That also means we’ll be much more private than a typical startup.
If this works, the payoff is much larger than better investing. Capital allocation is not the same as creation, but it helps decide what creation gets the chance to exist. Done better, it helps turn scarce resources into more human progress.
Where We Are Today
We’re a small team of former quant researchers, AI researchers, engineers, and investors based in Palo Alto.
Over the past nine months, we’ve been building and running the system with our own capital. Our results have outperformed the benchmarks by a high margin, while maintaining a high Sharpe and low directional market exposure.
We’ve also raised $100 million in seed equity financing from some of the best investors in Silicon Valley, giving us the runway to build without distraction.
We work in person, in the same room, with unusually tight feedback loops and very little friction. The culture is high intensity and high urgency. We demo several times a day, debate constantly, and ship relentlessly. We care much more about whether an idea is right than about who said it first. The team is tight-knit, collaborative, and fun.
We work very close to the frontier of what current models can actually do, and we’re often able to get more out of them than most people expect.
Some Problems We Are Working On:
Token efficiency in self-improving agents
Robustness in long-running agents (20+ hours)
Identifying and sourcing alternative datasets
Handling extremely large amounts of data while staying within context limits
Who We Are Looking For
We’re planning on adding just 2–3 people this year. We’re looking for individuals who combine:
Exceptional technical depth
Strong commercial judgment
Fluency in math and statistics
Clear, precise thinking & communication
A bias toward action
And, a track record of making things happen without being asked to.
If you are interested in these problems and our mission, we’d love to hear from you. Check out the open roles here: abundanceco.com/

I built an AI app that runs completely on device. No servers, API requests, or data leaving the device. Local models are finally getting good. Just look at the recent releases from @Alibaba_Qwen and @googlegemma to see that the gap is shrinking.
x.com/Alibaba_Qwen/status/2046939764428009914
I am excited about what this unlocks for hardware, consumer, and privacy.
Robots and edge devices can't rely on perfect signal for their user experience. A robot working in a farm, a construction site, or even a living room can't just freeze up the moment it loses connection. Smart glasses and similar devices running local models have lower latency and much better UX.
For consumer apps, local LLMs make free tiers economically viable. Growth at all costs becomes very expensive when every user burns GPUs on your dime. Local inference flips that math.
@signulll is clearly facing this now:
x.com/signulll/status/2044097057825124430
The third vector is privacy. I have mixed feelings here. Anecdotally, most of my friends don't care where their data goes. My guess is, in the future, privacy-sensitive fields like law and medicine will require on-device or on-prem models to stay compliant. Users won't care, but regulated professionals will.
The hardware is already here. We're all walking around with phones that can comfortably run 2B-parameter models. So I wanted to see how far I could push that with a real app.
Ambrosia
I built Ambrosia, an offline AI journaling app to track how my skin responds to different diets and products. I picked skin specifically for two reasons: I've always wanted a better way to watch conditions and progress over time, and photos are a natural unit of a journal entry.
Two things I love about the app:
1. AI-generated labels and trend tracking, all local
2. Everything including the images stays on device
The Base Model
I went with Qwen3-VL-2B: multimodal, small enough to run on my iPhone, and already well-optimized for llama.cpp. I used the Q4_K_M quantized version from @huggingface, and ran everything through @RunAnywhereAI because their on-device SDKs are the best I've used.
Finetuning the Model Locally
I wanted to prove that I could take a base model and finetune it end-to-end locally on my MacBook (M2 Max).
With Codex, I trained a small LoRA adapter on MedQA, then merged and exported it back into the same Q4_K_M GGUF format so that it would work on my iPhone.
I also experimented with autoresearch (@karpathy’s autonomous framework for model training) to maximize performance. I gave Codex the source code and let it sweep configurations. The sweep landed on a lower learning rate (1e-4) as the winner, beating baseline by ~5 points on the micro-run. Scaled up to full validation, it held at 47.88%.

Finetuned Model: huggingface.co/amankishore/qwen3-vl-2b-medqa-gguf
Constraining the Task
Initially I built Ambrosia as a chat first experience. It was bad. I would document a new skin issue, and the model would respond with three long paragraphs about consulting a dermatologist.
I constrained the task. Qwen was much better at generating labels from images and analyzing trends across journal entries. Ambrosia went from a worse medical ChatGPT to a journal with ambient AI trend analysis.
This is why small models are underrated. They should not be compared to general assistants like ChatGPT. They're much better as narrow specialists constrained to a few tasks.
Since this task is disjoint from MedQA, I built a small eval set of journal entries a user might track (acne, texture, redness, etc) and compared the finetuned model against the base. The finetune gave a small but real lift on label quality, while trend analysis was unchanged. For a 2B model doing a task it wasn't trained for, I'll take it.

Edge Intelligence
Ambrosia was a small experiment, but it maps to the shape of the future: take a small base model, aggressively constrain the domain, tune for your specific task, ship it. Repeat.
Chips keep getting faster. Small models keep getting smarter. Eventually these lines will converge. When they do, you’ll have the power of AGI, in the palm of your hand.

We built an API that turns every SEC filing into structured, real-time data.
Dilution ratings across 2,000+ tickers. Active shelf registrations. Cash runway calculations. Bank agreements with ROFR and tail financing clauses. Pump-and-dump risk scores.
The kind of data you'd normally pay $50K+ a year to access, and for some of these fields, data that literally doesn't exist on Bloomberg or any other institutional provider.
Our API is already being used by $1B+ funds, prop desks and investment banks.
And as of today, it's open to retail for the first time.
Here's how we got here, and what you can build with it.
I — Why This Data Doesn't Exist Anywhere Else
Most traders assume that if something matters, Bloomberg has it.
For large-cap equities, that's mostly true.
For small-caps, the space where 90% of retail trading pain comes from dilution, offerings, and pump-and-dumps, the institutional data providers fall apart.
Here's what they're missing:
Float that actually reflects reality. When a company converts debt to shares, the float changes overnight. Most providers don't update for months. We tack it on within 24 hours of the filing. That one field alone influences shelf capacity, offering ability, and downstream dilution risk, and no one else is doing it right.
Right of first refusal and tail financing. When you see H.C. Wainwright underwrite a small-cap offering, there's usually a contract locking the company into them for the next 12–24 months, with tail fees that keep the relationship sticky even if the company switches banks. This data sits inside exhibit agreements buried in filings. Structured. Queryable. Nowhere else.
Accurate cash runway. Most "months of cash remaining" calculations are a quarterly cash divided by a quarterly burn. Ours accounts for recent raises, warrant exercises, and actual operating burn pulled from the most recent 10-Q, updated filing by filing.
Shelf capacity relative to float. A 10M share shelf on a 2M share float is a completely different situation than the same shelf on a 500M share float. We calculate this ratio in real time. Most providers don't even store shelf data in a queryable format.
Pump-and-dump pattern scoring. Per-ticker scores for country, underwriter, float, and scam risk — each derived from structured filing data and paired with social-media evidence of an orchestrated pump-and-dump.
II — What It Took to Build Out This Data
Three years. Sleepless nights. A lot of things that didn't work.
The core problem: SEC filings are text. Thousands of pages of unstructured legal language, filed across dozens of form types, updated constantly. If you want structured data out of them, you either hire a team of analysts to read every filing by hand, or you build a system that can do it reliably at scale.
We chose the second path. Here's the rough shape of what it took:
Monitor every filing that can change capital structure. Not just the obvious ones (10-K, 10-Q, etc). The quiet ones too, with buried warrant exercises, debt conversions in exhibits, prospectus supplements that change shelf capacity mid-flight.
Build parsers for every form type. Each filing type has its own structure, its own language, its own edge cases. What a PIPE looks like in one 8-K exhibit is not what it looks like in another. The parsers have to handle all of it.
Layer AI on top of parsing. AI finds the keywords and phrases that suggest a dilution event, a new agreement, a compliance issue. Then manual verification commits the data. AI gets us 80% of the way; human review catches the edge cases that would otherwise corrupt the dataset.
Iterate constantly. Filing templates change. New deal structures emerge. Companies find new ways to raise capital that didn't exist five years ago. If the system isn't updated in real time, the data decays.
III — What You Can Build With It
The API has a host of endpoints covering dilution ratings, offerings, registrations, Nasdaq compliance, float, ownership, and bank agreements. Here are three things you can build today that would have taken a team of analysts to assemble manually.
1. A Dilution Risk Monitor

A watchlist dashboard that surfaces dilution warning signs across your portfolio in real time.
For each ticker, you get the overall dilution risk rating, active shelf registrations with remaining capacity, Nasdaq compliance deficiencies, and a cash runway calculation that tells you when the company will need to raise. Alert on things like cash dropping below 6 months, a new shelf going effective, or ATM capacity getting large relative to float.
2. Backtest Low-Float Gappers
Use the historical float endpoint to check performance on historical gappers under 1m float. Use the news endpoint 'tags' to see how gappers performed under certain news.
3. A One-Click Due Diligence Report
Take any ticker and generate a full due diligence report in seconds, ownership concentration, float history, reverse split count, ROFR agreements with active banks, upcoming lockup expiration… continue on X ↗