Tomasz Tunguz

Student of Startups
Backer of 9 unicorns
Author of http://tomtunguz.com 
Subscribe https://ttunguz.kit.com/ef0a779b9c

Founder

I asked Claude Cowork to read my tools folder. Eleven steps later, it understood how I work.

Over the past year, I built a personal operating system inside Claude Code : scripts to send email, update our CRM, research startups, draft replies. Dozens of small tools wired together. All of it lived in a folder on my laptop, accessible only through the terminal.

Cowork read that folder, parsed each script, & added them to its memory. Now I can do everything I did yesterday, but in a different interface. The capabilities transferred. The container didn’t matter.

My tools don’t belong to the application anymore. They’re portable. In the enterprise, this means laptops given to new employees would have Cowork installed plus a collection of tools specific to each role : the accounting suite, the customer support suite, the executive suite.

The name choice must have been deliberate. Microsoft trained us on copilot for three years : an assistant in the passenger seat, helpful but subordinate. Anthropic chose cowork. You’re working with someone who remembers how you like things done.

We’re entering an era where you just tell the computer what to do. Here’s all my stuff. Here are the five things we need to do today. When we need to see something, a chart, a document, a prototype, an interface will appear on demand.

The current version of Cowork is rough. It’s slow. It crashed twice on startup. It changed the authorization settings for my Claude Code installation. But the promised power is enough to plow through.

Thayer School of Engineering

Dartmouth College

Northern Highlands

Venture Capitalist at Theory

Network Node Ad Targeting

Social network user data advertising

Open profile content identification

Custodian based content identification

Founder of H Tree Capital (and Hoxton Ventures)

USER GENERATED ADVERTISING

Theory Ventures

Student of startups, author of tomtunguz.com, investor, & partner to founders.

General Partner

Redpoint Ventures

Privileged to support the journey of early stage SaaS, data and infrastructure founders. Worked with Looker ($2.6b M&A), MotherDuck, Mysten Labs (SUI), Monte Carlo, Omni, Chorus ($575m to ZI), Expensify (IPO), ThredUp (IPO), Kustomer ($1b sale to META), & others.

Managing Director

Winning with Data

I co-authored Winning with Data with Looker CEO Frank Bien.  Winning with Data: Transform Your Culture, Empower Your People, and Shape the Future explores  the cultural changes big data brings to business, and shows you how to adapt your organization to leverage data to maximum effect.

Co-Author

Google, Inc.

Managing AdSense for Social Networks including the $900M MySpace deal.
Managed AdSense International including product launches in Greece, Israel, Thailand, Indonesia and Bulgaria

Product Manager

Appian Corporation

Architected, developed and deployed systems for the US Government serving DHS

Engineer

Salesforce, Datadog & Epic are building walls. After two decades of flourishing through open APIs & data portability, the software industry’s largest incumbents are locking down.

The PC era began with a black screen & a blinking cursor. This stark prompt was the only gateway, beckoning a user to test the computer’s power.

Today, the interface to AI looks identical : an empty text box & a blinking cursor. It's not enough.

“[Silicon Valley] is the biggest, most volatile petri dish of raw capitalism on the planet.” So what lessons does a successful short-seller living here have? As founder of Crown Capital Management, Scott Fearon has shorted the stocks of over 200 companies.

Jeff Bezos famously said, “Your margin is my opportunity.” Does this maxim apply in software & AI? 

Yes.

Software companies maintain 76% gross margins yet earn almost nothing. Sales, marketing, & research & development consume it all.

In rowing, the coxswain watches everything. How hard you pull on every stroke. Whether you’re in time with the rest of the boat. How your form breaks down when you’re tired. They’re among the most valuable people in a rowing team.

At work, I don’t have a coxswain. So I built a robotic one.

In 1908, 253 American automobile manufacturers competed for the market1. By 1929, just 44 remained. The assembly line didn’t just change how cars were made. It changed who got to make them.

Ford’s Highland Park plant, operational in 1913, slashed the time to build a Model T from 12 hours to 93 minutes. That 90% productivity gain restructured an entire industry.

This week I chatted with an acquaintance who mentioned a board game. I caught half the title & looked for the full title & Amazon link using my AI in Asana.

The system tried with Gemini & failed. The failover to Claude also failed. Rather than continuously iterating with the AI until it worked, I created a Ralph Wiggum loop.

After a decade of success, the modern data stack has entered consolidation. What comes next?

The postmodern data stack is AI.

Talented people get promoted to management. So do talented models. Claude manages code execution. Gemini routes requests across CRM & chat. GPT-5 can coordinate public stock research.

Why now? Tool calling accuracy crossed a threshold.

After a decade of success, the modern data stack has entered consolidation. What comes next?

The postmodern data stack is AI.

The modern data stack created more than $100 billion in market cap with a simple promise : move the data via ETL, transform it inside a cloud data warehouse, put a semantic layer on top of it to unify metric definitions, & analyze it through BI.

This worked brilliantly for structured data, numerical data found within tables. But AI & businesses thrive on unstructured data, such as call transcripts, status reports, web searches, & PDFs.

The semantic layer, especially within large organizations with huge data sets, must now combine structured & unstructured information at scale, critical because all of it feeds AI & powers agents.

Strategic acquisitions prove this shift is already underway. The Datadog acquisition of Metaplane, the Snowflake acquisition of Observe, & the ClickHouse acquisition of Langfuse are the most concrete strategic moves reinforcing this fusion.

All of these initial acquisitions focus on observability & understanding AI systems. First, the volumes of data are enormous & essential to sustaining the breakneck pace of AI innovation.

Second, observability helps inform roadmaps : how are customers using AI? Where are they not successful? How can a business help them grow & expand?

We will see many more acquisitions that lead to a combined postmodern & AI stack, accelerating the consolidation that has already started. In addition, there are many pieces of the AI stack not represented here, such as evaluations & agent orchestration.

Those are next.

“We are only at the beginning phases of AI diffusion & already Microsoft has built an AI business that is larger than some of our biggest franchises.”

CEO Satya Nadella captures Microsoft’s Q2 FY2026 earnings in a sentence. The company beat expectations across revenue ($81.3b, up 17%), earnings per share ($4.14 adjusted vs $3.97 expected), & Azure growth (39%). Yet the stock fell 11% after earnings.

Will designers design first in a world where AI can code software immediately, or just describe the design? Will large enterprises pay for premium observability when AI can migrate & monitor open source competitors?

As Michael Mauboussin writes, there’s information in price. These questions are priced in.

Talented people get promoted to management. So do talented models. Claude manages code execution. Gemini routes requests across CRM & chat. GPT-5 can coordinate public stock research.

Why now? Tool calling accuracy crossed a threshold. Two years ago, GPT-4 succeeded on fewer than 50% of function-calling tasks. Models hallucinated parameters, called wrong endpoints, forgot context mid-conversation. Today, SOTA models exceed 90% accuracy on function-calling benchmarks. Performance of the most recent models, like Gemini 3, is materially better in practice than the benchmarks suggest.

Did we need trillion-parameter models just to make function calls? Surprisingly, yes.

Experiments with small action models, lightweight networks trained only for tool selection, fail in production. They lack world knowledge. Management, it turns out, requires context.

Today, the orchestrator often spawns itself as a subagent (Claude Code spins up another Claude Code). This symmetry won’t last.

The bitter lesson insists ever-larger models should handle everything. But economics push back : distillation & reinforcement fine-tuning produce models 40% smaller & 60% faster while retaining 97% of performance.

Specialized agents from different vendors are emerging. The frontier model becomes the executive, routing requests across specialists. These specialists can be third-party vendors, all vying to be best in their domain.

Constellations of specialists require reliable tool calling. When tool calling works 50% of the time, teams build monoliths, keeping everything inside one model to minimize failure points. When it works 90% of the time, teams route to specialists & compound their capabilities.

The frontier labs will own the orchestration layer. But they can’t own every specialist. Startups that build the best browser-use agent, the best retrieval system, the best BI agent can plug into these constellations & own their niche.

New startup opportunities emerge not from training the largest models, but from training the specialists the executives call first.

“[Silicon Valley] is the biggest, most volatile petri dish of raw capitalism on the planet.” So what lessons does a successful short-seller living here have? As founder of Crown Capital Management, Scott Fearon has shorted the stocks of over 200 companies.

In his book Dead Companies Walking, Fearon distills three decades of meetings with executives into six failure modes :

Despite their differences, they all failed because their leaders made one or more of six common mistakes that I look for :

1. They learned from only the recent past.
2. They relied too heavily on a formula for success.
3. They misread or alienated their customers.
4. They fell victim to a mania.
5. They failed to adapt to tectonic shifts in their industries.
6. They were physically or emotionally removed from their companies’ operations.

The hairpin turn from SaaS to AI amplifies each.

Learning only from the recent past. Software moves in 20-year waves. Mainframe, then client/server, then SaaS. When these waves come, they create bull markets. As they wane, growth flattens & multiples collapse. SaaS multiples have been flat for three years.

Relying too heavily on a formula for success. Sales motions have changed because of dramatic growth rates in product-led growth & multi-million dollar lands. Many of the rules around efficiencies & quotas no longer apply.

Misreading or alienating customers. Customers want to be AI native. They are willing to pay huge sums for education & solutions. Budgets have exploded with 41% of AI spend net new. Selling the same software application that no longer solves the customer’s pain point is a path to churn.

Falling victim to a mania. The AI hype cycle creates pressure to ship half-baked features. Announcing an AI roadmap isn’t the same as delivering value. This distinction will become increasingly stark as long-running agents enter the workforce in 2026.

Failing to adapt to tectonic shifts. If ever this were true, it’s true today.

Being physically or emotionally removed. Teams who don’t use AI daily miss the pace of change. The technology moves too fast for quarterly strategy reviews. If leadership isn’t prompting Claude or GPT every day, they’re already behind.

Cognitive biases are always hard to see in ourselves. A short seller’s mirror is a useful one at this moment in AI.

In 1908, 253 American automobile manufacturers competed for the market1. By 1929, just 44 remained. The assembly line didn’t just change how cars were made. It changed who got to make them.

Ford’s Highland Park plant, operational in 1913, slashed the time to build a Model T from 12 hours to 93 minutes. That 90% productivity gain restructured an entire industry. Manufacturers who couldn’t match Ford’s efficiency faced a simple choice : adapt or exit.

The consolidation was swift. Between 1908 & 1929, 83% of automakers vanished. Some merged. Most failed. The survivors shared a common trait : they adopted Ford’s methods. General Motors, Chrysler & the handful of remaining independents all built assembly lines.

An analogous revolution is happening in software, with important differences. AI coding assistants now reduce development time by 55-81%. The curve is familiar.

Ford took six years to achieve 90% time reduction. AI coding tools reached 81% in five. The slopes are nearly identical.

What happened to auto industry employment? It grew. Massively. In 1910, US auto plants employed 76,000 workers. By 1929, that number reached 471,000. Mass production created mass consumption, & mass consumption demanded more workers.

The real explosion came from second-order effects. By 1929, for every one person building a car, seven others had a job because that car existed.

Dealerships, service stations, repair shops & supply chains employed nearly 4 million people. The industry didn’t just create more manufacturing jobs; it spawned an entirely new economy of “enablers” that was 8x larger than the core manufacturing base.

The software industry will follow a different pattern. In autos, capital intensity consolidated power. It’s the opposite in the world of AI.

AI data centers, the assembly lines of software, enable hundreds of millions of developers to build software as if they had the capabilities of an automobile behemoth. Any developer can access state-of-the-art models with a laptop & a credit card.

Easier software creation means more software. More software means more developers, not fewer. We should expect many thousands of new businesses each year as a result, & a similar explosion in second-order jobs.

For decades, the to-do list has been a catalog of debt, a deceptively thin list of items to do, with icebergs of work hidden beneath the surface.

AI is NVIDIA’s third climb up a steep slope.

First came gaming in the late 2010s.

Then cryptocurrency.

Now artificial intelligence.

In 2012, we cared that we used software. Today we care how we use it.

I asked Claude Cowork to read my tools folder. Eleven steps later, it understood how I work.

taplio

Tomasz Tunguz's Linkedin Analytics

Tomasz Tunguz

Performance ScoreTomasz Tunguz's LinkedIn Performance

Who is engaging with Tomasz

Tomasz Tunguz's Best Posts (last 30 days)

Ready to grow your LinkedIn presence?

taplio

.css-6ytdur{background-image:linear-gradient(284deg, rgba(39, 43, 72, 1) 0%, rgba(33, 47, 183, 1) 100%);color:transparent;-webkit-background-clip:text;background-clip:text;}Tomasz Tunguz's Linkedin Analytics

Tomasz Tunguz

Performance Score.css-goub6s{font-size:var(--chakra-fontSizes-md);font-weight:500;color:var(--chakra-colors-gray-600);display:block;margin-top:var(--chakra-space-1);}Tomasz Tunguz's LinkedIn Performance

Who is engaging with Tomasz

Tomasz Tunguz's Best Posts .css-1emoo84{font-size:var(--chakra-fontSizes-md);font-weight:400;color:var(--chakra-colors-gray-600);}(last 30 days)

Ready to grow your LinkedIn presence?

Tomasz Tunguz's Linkedin Analytics

Performance ScoreTomasz Tunguz's LinkedIn Performance

Tomasz Tunguz's Best Posts (last 30 days)