co-founder @zommalabs (YC S26) prev at hedge fund, wealth tech, Goldman Sachs. in between 🌁🗽🇸🇬🇰🇷
San Francisco
Joined July 2025
- Tweets182
- Following790
- Followers248
- Likes229
Pinned Tweet
Today we are launching @ZommaLabs
What started as a 2am library project to build a computer use agent -> 100s conversations with hedge funds, payments, fintechs -> getting into @ycombinator -> starting with a billion dollar compliance problem that computer use is uniquely good at = collecting and synthesizing ever increasing info to solve a complex, close bounded high variance human problem
we are only getting started with building better back office for finance using frontier computer use. @CashelFitz @varumugam07 @_theodorayy 🔥
Financial institutions paid $3.8B in AML fines last year while spending $206B on financial crime compliance, with labor its largest cost.
Today we're launching Zomma (@ycombinator S26)
AI operations employees for finance back offices.
We start with compliance: for every cross-border transaction alert, our agent works the case end to end on the same screens your analysts use: gathers the evidence, writes the assessment, drafts the RFI. No APIs. Onboarding is an SOP and a login.
92% on OSWorld. Live compliance-ops pilots with a US RIA and a $10B AUM wealthtech.
If you run compliance or financial-crime ops, or know someone who does, let's talk → zommalabs.com.
Introducing @zommalabs Specialists 😎
First, for Month-end Billing 💸
One customer pain we keep hearing was tedious timesheet and invoice reconciliation that often delays companies getting paid.
BPOs, staffing, or any industry with a large admin overhead.
So we are launching our first specialist that helps you get paid faster with less error.
At Zomma, we are bringing frontier AI research to the real world with computer-use agents that can do end to end work.
Deploying in the wild means:
- Real, meaningful evals, custom built for each company
- Computer-use that can work across ALL systems
- Simple and fast onboarding
If you have a high-volume workflow that was hard to automate or test until now, talk to us 👋
@AnthropicAI just published internal data on how AI changed its own workforce.
What changed:
- Junior share of technical hires: 27% → 9%
- Commits per engineer: 5x in 6 months
- Agents wait on humans ~80% of the time
- Hiring now favors judgment over domain expertise
- Agents: 6x employee hours
Frontier labs are probably the most condensed & extreme test bed of AI-led labor shift given their access to infinite tokens, frontier models and AI-pilledness.
So if this is their present, the market expects this to be others' future. And it's restructuring fast in anticipation.
Hiring signals that worked before (e.g. YOE, tech stack, education) don't work anymore. So what works?
It's hard to describe the intangible yet visceral societal shift we're in.
Original research:
anthropic.com/research/emplo…
jihyun retweeted
Over the last few months at @river_ai_inc we’ve developed what we believe is the best system in the world for building your own internal custom models, driven by 10 years of experience training top models at DeepMind, OpenAI and xAI.
In order to best support our customers, we’ve decided to do something radical: We also deploy our API on-prem on your compute cluster, and if you don’t have compute we will help you find it.
Again and again we’ve seen that custom models designed for your own internal workflows outperform top frontier models on performance, at a fraction of the cost (1 percent or less).
We’re just at the beginning of the custom model era, but its already clear that in the future more than 99% of LLM tokens will be generated on efficient custom models. The economics of it are just too powerful, and it’s far easier than people think once you have the right tools in place.
The reason why small models will dominate is due to the physics of deep learning: A general model that can do everything needs more capacity and training it on the whole distribution of all tasks is more challenging, so the model will be larger and more expensive, and wont perform perfectly on every task. A small model trained on a targeted domain usually outperforms it.
We were able to fine tune an open weight model (Qwen3.5-9B) on proprietary internal data at @coinbase, and it out performed frontier models, at much lower cost and speed, for a specialized task (fraud prevention in this case).
I think we'll see more of this pattern inside companies. The frontier models are amazing generalists. But increasingly you'll see specialist models out perform on more narrow tasks, either working on teams, or by themselves.
We're making it easy to fine tune models internally for our employees. We even have a skill for it in our internal harness that makes it easy to spin these up and start training them:
"/fine-tune-and-serve-llm I want to fine-tune Qwen for payments risk review. help me get started"
I imagine we'll have some backend services that could be replaced entirely by this new architecture.
Introducing Zomma Agent Fleets 🙂↕👇
It's 2026. Why is your customer ticket still stuck for days?
Because most teams still work in a single queue.
With Zomma, you can now dispatch a Fleet of parallel computer-use agents as soon as work comes in.
The biggest hack? Fleet = better outcome with fewer tokens. Each subagent gets just enough context to do the work well, saving tokens and time.
And you can score every subagent against your own eval set.
@ZommaLabs is in early access. If you're building agents for your operations, let's chat 📥
what most people still don't grasp is how many computers we'll need when agents run everything
compute isn't just GPUs for AI training or inference. It's CPUs and memory for your Muse, Instinct, Town, Claude Code, Codex sessions to always be running
this is especially true when enterprise deploys more agents (and it's still VERY early!)
everyday we dynamically scale ~500 VMs up and down for @ZommaLabs computer-use swarms for operations and it will only go up
now that agents can do real work, they just need a place to do it persistently and safely = computers on cloud
.@auchenberg calculated that Muse needs 65,000 CPUs and 75PB of DRAM to serve 100 million users. That's $2.8B in infra for one app.
It made me think. Facebook alone has over 3 billion users. That's nearly 300x scale for one platform. And they have more platforms with similar numbers.
Just do the math. There aren't enough CPUs in the world to run all of this. A worrisome constraint we're heading toward.
why model choice matters for a consumer product:
when using @meetgranola I realized that the model toggle is my "lever" to few shot a task until I'm happy with the output
dropdown >> typing "try again" multiple times to a model
I see a crazy token bill from my cto and he's watching a chess game so I asked him: are you ok
turns out he's casually pushing the frontier of computer-use by being pilled on code mode, making any task more in-distribution for LLMs
so naturally he has to write an X post 😺 @CashelFitz
I made my CUAs build a chess game from scratch, and spawn 2 subagents to play against each other.
Code mode computer-use agent harnesses put development much more in-distribution. This means that the range of tasks the agents can handle is much broader.
The orchestrator not only built a chess game from scratch, it found a public pub/sub server on the internet for the subagents to talk to each other.
Then the frontend was sent to the subagents in the task definition by base64 encoding the repos and telling them to decode and compile on their desktops 🙀
low key these frontier models are a little too creative when sharing code/assets to sandboxed agents 🤔 kinda sus..
we are making banking cool again here in SLC @ransomthoughts @benxiang @Schornack @marcrehberger @Jane_Barratt
today is a pretty special day for a technical woman founder like me. inspiring to see the og ai baddies paving the way ✨
traces are critical for 1) evals 2) auditability 3) training
and figuring out which traces to store is hard
as @ZommaLabs rolls out computer-use swarms for our financial institution customers, we generate and annotate a lot of traces along with the replays and human feedback
we get a lot of questions on buy vs build and ofc anyone can build managed agents, but nailing the LOOP of building evals, curating the traces and improving the model and the harness on top is hard to inhouse, especially given how many workflows each team has to cover
jihyun retweeted
What an insane day in AI. The frontier models just became substantially cheaper, with the Opus 5.5 price cuts, and now with GPT-6 Sol and Luna dropping token prices by 50%.
The rate at which the cost per task (on a like-for-like basis) drops in AI is unlike any other type of technology in history. And every time the cost of AI drops, the use-cases you can deploy agents against dramatically increase. This is Jevons paradox applied to agents.
These improvements will directly lead to broader diffusion of AI in the economy as we can use agents to process all of our data, scan our code for security issues, read through all log data to make decisions, have agent swarms in workflows, and much more. The cost of tokens is directly correlated to these use-cases being opened up at scale.
Please welcome GPT-6 Sol and GPT-6 Luna to the GPT-6 universe.
GPT-6 Sol and Luna build on the advances behind GPT-6 Astra, bringing much of its strengths into faster and more affordable models to support work at scale.
We’ve also made caching and inference more efficient, and we’re passing the savings directly to you: 50% lower API prices for Sol and Luna compared with GPT‑5.6 promotional pricing.
being a founder = chewing glass and loving the taste of blood
a friend called me today asking if he should quit his boring job to be a founder. i asked if he's ready to chew some glass
you've read enough essays to know being a founder is hard
but oh man glass
glass is a hundred of unanswered emails, rejections, deals you thought you had but falls through last minute, not making a decision which actually is just a bad decision, crazy amount of doubt and uncertainty you have to counter. all in a single day :)
and yet for some twisted reason there's nothing else you'd rather be doing
so you start loving the taste of blood instead, as @sparker said. great episode from @patrick_oshag @GabeStengel 🥛🍷
My conversation with Gabe Stengel (@GabeStengel), founder and CEO of Rogo.
For years, Gabe and I have talked about how much of an investor's job AI will eventually do and how he is building Rogo toward that future.
Today, Rogo helps some of the world's largest financial institutions research companies, run diligence and execute M&A.
But Gabe's ambition is much bigger. He is building toward investing superintelligence, where Rogo does much of the work inside investment banks and firms and becomes the venue where they do their deals.
It's a fascinating business and has been so fun watching Gabe build it.
We discuss:
- 10,000 agents searching for one great investment idea
- Which investing skills will still matter
- Why Anthropic/OpenAI won't win finance
- "Chewing glass"
- Why the harness around the models matters so much
- Getting rejected by 40+ investors
- Building an AI native Bloomberg
- Becoming a black hole for talent
Enjoy!
TIMESTAMPS:
0:00 Intro
2:38 Building Rogo
6:12 10,000 AI Agents
12:02 Skills That Still Matter
17:31 Beating OpenAI and Anthropic
28:35 Bloomberg of the AI Era
37:37 Rogo’s Company Brain
44:19 Chewing Glass
53:34 AI-Native Finance
59:21 What Humans Still Do Better
2026 timelines are so warped because yes you rip Salesforce in a day but it still took Meta, Instinct etc 8 months and billions to rip Openclaw
The alumni demo day @ycombinator is truly wild. Surreal to have alumni angels write checks on the spot 🤯
thanks labs for all the computer-use post-training that only make @ZommaLabs better
OpenAI is buying Mac Mini's by the tens of thousands and the interesting part is what this says about where the bottleneck has moved.
The scarce input for this generation of models isn't only GPUs, it's realistic places for agents to practice. Labs already pay heavily for those places. Anthropic leadership has reportedly discussed spending over $1 billion on RL environments in the next year, and vendors like Surge (about $1.2B in revenue from the big labs) and Mercor (valued at $10B) are racing to build task environments (gyms) for them.
Why computer use needs real computers:
Both OpenAI and Anthropic are trying to build models that operate software the way a person does. Look at the screen, click, type, check the result. The training method is reinforcement learning at scale roughly following this process:
1. spin up thousands of real desktop sessions
2. hand the agent tasks like editing and testing code or organizing email
3. score the outcome
4. update the model on what worked
Practitioners describe single runs sustaining this loop for days or weeks, and most of that time goes to the agents working through tasks, not to updating the model weights.
The reason to focus on Mac is that macOS accounts for roughly a fifth of the world's desktops, second only to Windows, and overrepresented among the developers these agents are supposed to help. The only way to get a model that is expert at using a Mac is to have it use a Mac, millions of times. All that practice is showing up on the scoreboard between @OpenAI and @AnthropicAI.
Who's ahead right now in the Computer Use race:
The main eval is OSWorld-Verified and in this test, an agent gets 369 real desktop tasks and passes only if the machine ends up in the right state. Humans scored about 72% when the benchmark launched. The top models are now past that.
- Claude Fable 5 (Anthropic's flagship): 85% (leaderboard.steel.dev/leader…)
- GPT-5.5 (OpenAI's best published on this test): 78.7% (benchlm.ai/benchmarks/osworl…)
NOTE: OpenAI's newest flagship, GPT-5.6 Sol, reports on a different variant: 62.6% on OSWorld 2.0, a newer and harder task set where it's the state of the art (openai.com/index/gpt-5-6/). These are different tests, so the numbers don't compare directly. The honest summary is Anthropic leads the older benchmark, OpenAI leads the newer one.
Why Mac minis and not macOS Virtual Machines in the cloud:
This part has to do with licensing, not physics. Apple's macOS license only permits macOS to run on Apple hardware, so there is no legal way to spin up a macOS VM on an ordinary server fleet the way you would a thousand Linux or Windows machines.
AWS's "Mac instances", which Anthropic uses, are literally racks of physical Mac minis, billed with a 24-hour minimum allocation to comply with the Apple macOS Software License Agreement. So if you want macOS desktops at scale, you buy Macs or you rent real ones. A headless $899 mini drawing 60-120W is the cheapest legal container for a macOS session that exists.
The rent vs buy math:
Buying an M6 Mac mini lists at $899. Renting the closest AWS host, an M4 mini, runs about $1.23 an hour.
So left on 24/7 that's roughly $10,800 a year, or about $6,000 with a 3-year commitment. This means that break even for a single Mini running around the clock, happens in about a month compared to the rental prices.
So Anthropic's route costs 7-12x more per box per year. What it buys: no cash up front, no data center ops for an unusual hardware class, and a fleet that scales down when a run ends.
My guess is it also signals commitment level. You buy what you know you'll use for years and rent what you're still deciding about. Amazon being Anthropic's biggest cloud partner and investor makes AWS-first the natural posture anyway.
What actually happens on these machines:
1. The minis get racked headless and imaged with sandboxed macOS sessions. They are sandboxed because a half trained agent can theoretically click anything, which means practitioners need to keep them off the open web so they don't delete data or buy $5,000 items.
2. Each session runs one agent attempt: screenshot in, action out, for trajectories up to hundreds of steps.
3. An automated verifier checks the end state. Did the code pass its tests, did the email get filed correctly?
4. The trajectories flow back to a GPU cluster where the weights actually update, and the improved model redeploys to the fleet. The Macs mainly generate the training data, and the gradient math almost certainly stays on Nvidia GPUs.
Procurement:
Compared to acquiring GPUs, acquiring Mac Minis is simple. Data center GPU capacity is allocated years in advance. Mac minis sit in retail channels you can buy from this week. The catch is that consumers share those channels.
High memory Mac configs have been out of stock for months and Apple pulled the 512GB Studio from sale at one point, while Mac revenue jumped 29% year over year to $10.3B. Apple even shipped its M6 mini refresh this month instead of the usual fall window.
My read:
GPUs became the commodity layer for training compute, and Mac minis are starting to become part of a commodity layer for something newer, generating agent experience. A $899 desktop is the cheapest legal unit of real macOS practice, and the labs are treating it as a component, not a computer.
what is $1 of token credit worth? the token to dollar rate is opaque but we know the black market exists (i.e. companies selling free tokens for $$$)
this is a 1 sentence summary of why Stripe bought OpenRouter and what seems to be the official financialization of tokens
Stripe’s Will Gaybrick: "We just want to make sure that moving between tokens and dollars is as seamless and safe as moving between dollars and euros. We're at the beginning of this journey, but we think it's going to be a big part of the future of Stripe."
@gaybrick
the latest model observations:
Fable: let me achieve this task in any way possible. common sense is below me. im looking for all the secret tokens in this box to hack my way through
Opus 5: oops i thought i did that my bad
GPT-5.6 Sol: yep ok abandoning my current task to do whatever you tell me now