Hey - It’s Nico.

Welcome to another Failory edition. This issue takes 5 minutes to read.

If you only have one, here are the 3 most important things:

• Pulley, a startup that built cap table software for startups, shut down — learn why below

• One Founder needs to own GTM

• Jev is a new type of AI that can’t talk — learn why this matters below

This Week In Startups

🔗 Resources

Inside 12 months of AI search experiments

Rise of the Borderless Founder

One Founder needs to own GTM

Growth is a System. Start There.

📰 News

OpenAI releases GPT-6 Sol and Luna

OpenAI tests Sponsored Agents and agent-run campaigns

Meta introduces camera-free AI glasses

Meta’s Muse is outpacing ChatGPT’s early mobile launch

💸 Fundraising

Snorkel AI, a startup that builds training datasets and simulated environments for AI systems, has raised $350 million.

Enveda, a biotech startup that uses AI to discover medicines from plant and microbial molecules, has raised $311 million.

Angle Health, a startup offering AI-enabled health insurance and employee benefits, has raised $200 million.

Ema, a startup developing AI agents for HR, IT, and finance workflows, has raised $77 million.

Fail(St)ory

The Cap Table Challenger

Pulley helped startups keep track of who owned what, model dilution, issue employee equity, and handle the compliance work that followed.

Last week, the company announced it was winding down. Its service will stop on December 8, and customers are being offered an assisted migration to Carta, the much larger rival Pulley spent years trying to displace.

What Was Pulley:

When a startup is created, the founders own all of it. That, however, changes very quickly: Investors put in money, employees receive shares or options, and every new funding round changes how much of the company each person owns.

The record of who owns what is called a cap table. It often begins as a spreadsheet. As the company grows, that spreadsheet becomes harder to update and easier to get wrong.

Carta was already the dominant software for this work. It helped tens of thousands of private companies track ownership and handle the paperwork around fundraising. Pulley entered the same market with a simpler product aimed first at founders and young startups.

A founder could send Pulley the company’s spreadsheet and legal documents. Its team would turn them into a digital cap table showing who owned each part of the company and how those percentages would change after a new investment or employee grant.

Customers could also send documents for signature, issue shares to employees, and give investors and employees their own view of the records. As the startup grew, Pulley sold it valuations for employee shares, tax filings, accounting reports, and help running the equity program.

Pulley combined the software with hands-on service. Its entry plan included concierge onboarding. The company advertised access to cap-table experts seven days a week and in-house valuations completed within days.

This approach attracted plenty of startups. Pulley grew from more than 1,700 customers in 2022 to 4,600 companies by February 2024. The gap remained large: Carta served about 40,000 customers at the time.

The Numbers:

• 💰 Funding: More than $50M

• 🏷️ Reported valuation: $250M in 2022

• 🏢 Customers: 4,600 in February 2024

• 🌱 Startup plan: $1,200/year for 25 stakeholders

• 📈 Growth plan: $3,500/year for 40 stakeholders

• 📅 Service ends: December 8, 2026

Reasons for Failure: 

• Early customers may not have generated enough revenue: Pulley’s entry plan cost $1,200 per year, while qualifying companies could use Carta Launch for free. The model depended on young companies surviving, raising more money, adding employees, and eventually buying valuations, reporting, or managed administration.

• The service promise added a lot of work: Concierge imports, fast responses, expert reviews, and in-house valuations made Pulley easier to trust. They also required people. Bundling this much assistance into relatively modest annual plans probably made operating leverage harder to achieve.

• Carta had a much larger system to grow into: Carta served roughly 40,000 companies by early 2024 and could move customers from a free cap table into valuations, fund administration, and other products. One 2025 finance survey put Carta at 41% of respondents, spreadsheets at 23%, and Pulley at 1%. Pulley won about 400 customers after Carta’s January 2024 trust crisis, but the scale gap remained.

• The funding raised the required outcome: More than $50 million in capital and a reported $250 million valuation meant Pulley needed to become a substantial platform. A smaller, slower-growing business could not justify that valuation, so Pulley had to capture a much larger share of a market dominated by Carta.

Why It Matters: 

• Track what early customers become, because survival, upgrades, support time, and revenue expansion reveal more than logo count.

• Charge separately for concierge help unless each account becomes easier and more profitable to serve over time.

Trend

The Model That Doesn’t Talk

Last week, TypeSafe released Jev and described it as the first of a new class of “System One” models. In its own workflow tests, the company says Jev was up to 193 times faster and 444 times cheaper than an LLM.

Within 24 hours, nearly 13% of paid teams using Vercel’s AI Gateway had tried it, the fastest adoption of any model launched there. The reactions ranged from “the biggest breakthrough since transformers” to “just a hyped classifier.”

I think Jev is genuinely interesting. There’s definitely a lot of hype surrounding it, but this is clearly a launch worth paying attention to.

So today, let’s look at what Jev actually is, what it does differently, where the hype gets ahead of the evidence, and what kinds of products it might make possible.

Why it Matters

• Many AI products use an LLM for decisions that never required an essay: Which support queue owns this ticket? Is this transaction suspicious? Which tool should an agent call?

• Jev handles those small judgments without generating text that the software then has to parse and discard. That makes the calls faster, cheaper, and easier to control.

• Its probabilities may be the most valuable part. A product can act on clear cases, send ambiguous ones to a larger model, and reserve consequential cases for a person.

What Jev Actually Does

Give a chat model a customer email and ask what to do. It generates an answer one token at a time, perhaps wrapped in JSON if you ask nicely.

With Jev, you first define the possible answers. The email becomes the input, and Jev could, for example, choose a queue, score its urgency, or estimate the probability of fraud. It returns only those declared values and the probability distribution behind them.

A regular LLM can return the same JSON, but it is still composing text. Jev’s output space is the answers you supplied. Its job ends after assigning each a probability.

Functionally, Jev is a general-purpose zero-shot classifier. “Zero-shot” means giving it new labels and instructions without training examples. A traditional classifier learns one fixed task and needs retraining when the categories change. Jev accepts a different decision with every request.

Zero-shot classifiers predate Jev. TypeSafe packages the idea into a hosted model with natural-language instructions, typed answers, and useful probabilities. The simplest way to look at it is as a reusable judgment function for software.

What is it Used For?

Jev is useful when software needs to make a small judgment before doing something.

Imagine a support agent with five tools: look up an order, issue a refund, change an account, flag fraud, or ask a human for help. An LLM can decide which tool to call, but that means using a full generative model for what is basically a multiple-choice question.

Jev handles that choice. It reads the customer’s message, returns the most likely tool with a probability, and the agent’s code makes the actual call. Jev does not use the tool itself.

The same idea applies to routing emails, ranking search results, flagging risky transactions, or deciding whether an answer needs human review. Jev handles the small, repeated decisions inside a workflow. An LLM can then focus on the parts that require writing or deeper reasoning.

How Good Is It?

The early independent results are promising.

One test ran Jev and four LLMs through 791 labeled decisions. Jev was roughly as accurate as the small models, ran two to four times faster, and cost five to eight times less than the cheapest ones.

Its median response time was 0.33 seconds, compared with 0.67 seconds for the fastest LLM and 1.17 seconds for the frontier model. In a classification test where each message had one of 77 possible labels, Jev got 78.8% correct, almost identical to two small LLMs at 78.4%.

The most useful result came from combining models. Jev handled decisions above 80% confidence and sent the rest to a frontier model. That setup matched the frontier model’s accuracy at 26% to 28% of the cost and roughly half the average latency.

These tests are still early and small, but the economics are compelling. I think Jev, or models like it, will likely become a normal part of many AI products, handling routine routing, classification, and checks so that larger models can focus on writing and deeper reasoning.

Help Me Improve Failory

That's all for today’s edition.

Cheers,

Nico