I’ve spent the last decade building and investing in AI companies. As COO & CRO at Weights & Biases, I watched the ML tooling market go from niche to default infrastructure. Now, as a GP at B Capital leading our AI investing practice, I’m seeing the same pattern play out one layer up the stack.

The short version: we’re moving past the “AI agent” phase into something more interesting and more valuable. I’m calling it the AI co-worker.

Three eras in two years

The progression is simple:

AI Tools (2023): “Help me write this.” ChatGPT, basic copilots. Humans do the work, AI assists on demand. No context, no memory, no action.

AI Agents (2024-25): “Do this for me.” Cursor, Claude Code, Codex CLI, customer support bots. AI executes defined tasks end-to-end. Can use tools, take actions, complete workflows. But stateless, no organizational memory.

AI Co-Workers (2025+): “Own this with me.” Persistent memory, learns on the job, plans and prioritizes autonomously. A colleague that grows with you.

The difference between an agent and a co-worker isn’t just branding. An agent completes a task and forgets. A co-worker remembers what you discussed yesterday, understands your team’s conventions, knows your codebase, and gets better over time. That’s a fundamentally different product category.

Why now

Four technical curves crossed in 2024-25 and are still accelerating. The data from April 2026 is staggering compared to even six months ago.

Reasoning keeps compounding. SWE-bench Verified went from 2% in early 2024 to 81% by January 2026, and Claude Mythos Preview just posted 93.9%. But the benchmark itself is arguably saturated. OpenAI flagged contamination concerns and stopped reporting Verified scores, recommending SWE-bench Pro instead, where top scores sit around 46-57%. The more meaningful metric: METR’s time horizon analysis shows frontier models now reliably complete tasks that take human experts ~5 hours, with the capability doubling time accelerating to 4.3 months. These aren’t just coding tasks anymore. Claude Code writes 135,000 GitHub commits per day. 4% of all public commits on GitHub are now authored by AI, with projections of 20%+ by year-end.

From protocols to harnesses. MCP was the story of 2025. The story of 2026 is the full stack built on top of it: CLIs (Claude Code, Codex CLI, Cursor agent), SDKs (Claude Agent SDK, Codex SDK), and multi-agent orchestration (Agent Teams, subagent architectures with dedicated context windows per task). Codex CLI has 67,000+ GitHub stars. Claude Code’s Agent Teams feature lets multiple AI instances collaborate in parallel, each with its own context window. Both Anthropic and OpenAI now treat coding agents as their primary growth vector, and they’re converging on similar primitives: CLAUDE.md, AGENTS.md, hooks, plan mode, background execution.

Memory and context at scale. Opus 4.6 runs a 1M token context window in beta. GPT-5.4 reportedly pushes to 2M. But the real shift is persistent memory across sessions and organizational context (code, docs, tickets, CRM) that makes generic AI tools enterprise-ready. This is still the weakest link and where infrastructure startups have the most room.

Costs keep falling, but usage grows faster. Inference costs have dropped roughly 1,000x over three years at equivalent performance levels. GPT-4 quality now runs at $0.06/M tokens from budget providers. Epoch AI found price declines ranging from 9x to 900x per year depending on the benchmark. But total inference spend keeps rising because agents burn 3-4x more tokens than simple chat, and agentic workflows run continuously. The economics work for always-on AI teammates, and enterprises are proving it: Anthropic just hit $30B ARR in April 2026, up from $1B fifteen months ago. Claude Code alone generates $2.5B+ in run-rate revenue.

These curves aren’t slowing down. They’re compounding.

Where the value sits

The demand signal is unmistakable. $252 billion in corporate AI investment in 2024. Enterprise AI spend per company averaging $4.5-7M and expected to grow to $11.6M. Eight of the Fortune 10 are Claude customers. One in five businesses on Ramp now pay for Anthropic, up from one in twenty-five a year ago.

But here’s the number that matters most: only 5% of AI pilots achieve measurable P&L impact. 95% fail to deliver ROI. The bottleneck is not demand. It’s the infrastructure to deploy and scale.

This is where I focus. Three application categories and three infrastructure categories.

Applications: Software engineering ($370B addressable), sales & GTM ($245B), and finance/CFO office ($215B). In engineering, basic code generation is commoditized. The new frontier is enterprise context and verifiable domains (formal proofs, security, AI research). In sales, the market is flooded with AI SDRs, but winners will own the system of record and close the loop on what converts. In finance, high-volume rules-based workflows are ideal for AI co-workers that replace analysts, not just augment them.

Infrastructure: Agentic memory & context (the gap between model memory and organizational memory), orchestration & multi-agent coordination (agents don’t collaborate well yet, with each other or with humans), and production observability (when agents run 24/7, ops teams need visibility into what’s working). The middle layer of the stack is under-invested. These are the missing pieces blocking enterprise adoption.

What makes a co-worker defensible

I look for five things:

1. Team. AI technical depth plus domain expertise. Founder-market fit matters more than early traction in this market.

2. Data moat. As models commoditize, unique high-quality data with continuous feedback loops becomes the primary differentiator.

3. Workflow embedding. Deep integration into daily work creates switching costs. If the product is indispensable to someone’s Tuesday, it’s defensible.

4. Progressive defensibility. Technical moats alone don’t last in AI. You need a plan to layer defenses over time through data accumulation, network effects, and customer lock-in.

5. Economics that work. GTM and pricing that support AI unit economics for both the company and the customer. This is harder than it sounds when your COGS is inference.

The risks worth naming

Hyperscaler competition. OpenAI, Anthropic, and Google are all building agent platforms. But history shows best-of-breed wins in enterprise. AWS didn’t kill Datadog, Snowflake, or MongoDB.

Rapid commoditization. What’s differentiated today may be table stakes in 12 months. AI moats erode faster than traditional software. This is why I filter for data flywheels and workflow depth over features.

Enterprise adoption could be slower than expected. Security, compliance, change management. The 95% pilot failure rate could persist. But this is exactly why infrastructure matters. The failure rate is the problem worth solving.

Where I’m putting capital

I’ve been building this thesis with real investments: Perplexity (which just launched Computer, a multi-model agentic system that orchestrates 19 frontier models to execute end-to-end workflows, turning a search company into a general-purpose AI co-worker), Goodfire (interpretability and trust layer for AI systems), Code Metal (verifiable AI code translation for mission-critical industries like defense and automotive, where “probably correct” doesn’t cut it), Axiom (AI mathematician for formal proofs), and Unblocked (organizational memory for engineering teams). Over $250M deployed across AI co-worker and infrastructure investments platform-wide.

The acceleration is visible in real time. When Anthropic launched Cowork in January 2026 (four engineers built it in ten days, with most code written by Claude Code itself), global SaaS stocks lost roughly $2 trillion in market cap. Investors recognized that agentic AI tools were coming for traditional enterprise software. That’s not hype. That’s the market pricing in the shift I’ve been investing around.

The pattern I keep seeing: the companies that win aren’t the ones with the best model. They’re the ones that best understand the job their user is trying to do, and then build a system that gets better at that job every day.

That’s what a co-worker does. And that’s where the next wave of enterprise value gets created.

I’m actively looking at Seed through Series C companies in AI co-worker applications and enabling infrastructure. If you’re building in this space, reach out: yanda@b.capital

If you’re a technical founder and your board is telling you it’s time to hire a professional CRO, read this before you make that call.

I’m not going to argue that you should never hire one. But I am going to tell you what happened when a company I was involved with hired two of them, what I learned when I took the job myself despite having no business running a sales org, and why the conventional playbook fails so often at technical companies.

Chapter 1: The Founder CRO

When Lukas Biewald founded Weights & Biases, he wasn’t a sales guy, but he wasn’t new to building companies. He’d previously founded and led Figure Eight (formerly CrowdFlower), and before that had been at Yahoo and Powerset. What set Lukas apart was something harder to put on a resume: he seemed put on earth to improve the life of AI engineers. That deep curiosity about their problems and how to solve them shaped everything about how he built W&B’s early team.

The first sales hires reflected that. It was a small group, and several of them had followed Lukas from Figure Eight into the next adventure. They spent most of their time in consultative relationships with customers. They weren’t running traditional sales cycles. They were doing customer development, sitting with AI researchers, understanding workflows, figuring out where experiment tracking fit into real AI pipelines. They’d built their careers at companies like Domino, Figure Eight, and Scale, where you couldn’t survive without technical depth. They could talk to an AI engineer about hyperparameter sweeps and mean it.

Revenue was growing well, with the kind of customer relationships that compound. Users trusted the team, expanded their usage, and referred colleagues. The foundation was solid.

Then W&B raised a round, and the board did what boards do: suggested it was time to bring in a professional sales leader to “scale the GTM motion.” The founder’s scrappy team had gotten the company this far, the thinking went, but now it was time to professionalize.

SaaStr puts the failure rate for first VP Sales hires at SaaS startups at 70%. Average tenure: about 7 months. For technical and AI-focused products, I think it’s higher. But every board thinks their hire will be in the other 30%.

Chapter 2: The Copy-Paste CRO

The first professional CRO came from a well-known DevOps company, where he’d been a regional VP of Sales. He’d never run an entire company’s go-to-market, and had no prior experience selling to AI practitioners, but the DevOps background looked close enough on paper.

I should own my part here: I was on the W&B board at the time, having led their Series A for Coatue in 2019. I agreed we should make the hire. His resume was impressive. At least I owned fixing the problem I helped create in a way most investors never consider, but more on that below.

His approach was to apply his previous company’s playbook at W&B. Not adapt it, apply it: the same org structure, the same hiring profile, the same sales methodology. To wit, he put in place geography-based rep territories, the same way you’d carve up a DevOps platform sale. We ended up with a rep trying to sell AI experiment tracking tooling in Alabama and Georgia, wondering why she couldn’t make quota. He scaled the team fast (from about 5 to ~20 in seven months) but optimized for speed over fit. The hires skewed toward traditional enterprise sellers rather than the technically fluent reps the market required.

Here’s what made this hard to catch: the top-line GTM metrics looked decent. ARR was growing. If you only looked at the dashboard, you’d think the hire was working. But when we dug in, virtually all of the growth was being driven by those early salespeople Lukas had brought on. The new hires were producing almost nothing. The professional sales machine was a mirage. The founder’s technical sellers were carrying the entire number.

Nine months later, he was gone.

Chapter 3: The Unlikely CRO

When I joined W&B as COO in June 2021, I made Lukas a promise: I’d work on whatever kept him up at night. A few months later, he told me that thing was now sales. I thought he was joking.

Nothing in my background says “you should run a sales org.” I’m a product guy. I wrote code at Google and Microsoft. I founded four venture-backed companies. I was a GP at Coatue Ventures leading AI investments. I’d spent years around data science and AI teams but had never carried a quota, never managed a pipeline, never run a QBR.

But a promise is a promise. Lukas saw some assets I missed: I’d been on the W&B board since the Series A. I knew the product & the customers. I’d spent years hanging out with AI engineers, caring about their problems. I understood why someone would choose W&B over MLflow or a homegrown solution, not because I’d read a battlecard but because I’d lived in the ecosystem.

When I took over as CRO, I didn’t try to invent a new sales methodology. I went back to what had been working before. I let go of most of the team the first CRO had brought on and scaled back down to Lukas’ original hires: the technical, consultative sellers who’d been carrying the number all along. Then I used their profile as the template to build the team back up: people who’d built their careers at AI and data companies, who could hold a real conversation with a practitioner, who were curious about the problems rather than just the quota.

One of the first things we changed was territory design. Instead of geography, we segmented by vertical, focusing on the industries we knew trained and fine-tuned models: tech, healthcare, automotive, financial services, government. We gathered data on the number of AI engineers at thousands of companies and did our best to give each rep a territory with a similarly-sized pool of practitioners to target (contract sizes were seat-based then, so this mattered).

We executed quarter on quarter. I ran data science concurrently with sales, which sounds odd but was a structural advantage: I could instrument our pipeline like I’d instrument a product, build dashboards that showed where deals were stalling, and test interventions with the same rigor I’d apply to an experiment. I got my hands dirty across the funnel: helped write SDR sequences, sat in on first pitches, negotiated down-to-the-wire renewals. Not because I needed to close every deal, but because participating across the funnel is the only way to understand where it breaks.

I also couldn’t have done any of this alone. One of the most important things I did was assemble a sales leadership team that filled in my skill gaps. What started as a tiger team to right the ship after the first CRO became a permanent group I relied on for advice, pushback on strategy, and the kind of late-night work sessions where we’d be categorizing prospects for territories at 11pm. They were willing to do whatever was necessary for the company to succeed. They know who they are, and I couldn’t have done it without them.

Chapter 4: The Delegator CRO

Despite our results, after about three quarters, the board pressure returned. The feeling was that the company needed a “career CRO” to take it to the next level, someone who’d scaled enterprise sales orgs before and could build the machine for the long term. So we decided to try again.

This time we ran a more thorough search and hired someone with a stronger profile: years of enterprise sales leadership at data and analytics companies, prior CRO experience, a career spanning institutional finance and global software sales. He was professional, organized, and had a track record.

He wanted to run the team the way a CRO at a larger-scale company would: build the org, build the machine, manage from above. He wasn't interested in getting personally involved in specific deals or spending time with specific customers. His career had been selling to business buyers, from institutional equities to enterprise analytics. He'd never sold to AI practitioners, and at a company where technical credibility is table stakes, that mismatch showed up in every customer call and team meeting.

The GTM metrics got worse: growth and net new ARR decelerated, churn went up, and a few of our best salespeople left because they wanted a leader with deeper technical fluency. On the retention side, the lack of executive hands-on involvement with top customers led to surprise churns and contractions.

After about a year, we parted ways. Lukas knew who had fixed the broken engine once before: “if at first you succeed, try, try again”. I took the CRO role back for the second time, mid-2023.

This rebuild was harder than the first one. Morale was low. The numbers had been tough for several quarters and the team felt like they’d underperformed, even though much of that was a leadership problem, not a talent problem. We needed to jazz the org back to life, not just restructure it. Some of the people who’d left were the best sellers we’d had. I made it my mission to bring them back, and succeeded in many cases. We rebuilt both the sales and CS motions to be what they’d been before: more customer-oriented, more technical, more interested in outcomes than activity metrics.

One thing we did to fix the churn problem was assign executive champions across our leadership team to our biggest and most discriminating customers. This wasn’t a CRM field or a quarterly check-in. It meant our leaders were personally accountable for those relationships. It helped us uncover issues customers had been complaining about for a year, allowed us to focus our engineering efforts on addressing them quickly, and got the relationships back on track. The best part: those feature requests ended up being useful for our whole customer base. Just like in the early days of W&B, the early adopters were tastemakers in how to build in AI. Listening to them made the product better for everyone.

We got back to hitting targets within one quarter. The days and nights of that quarter definitely blur together, but we got there. When I stepped away in mid-2024, the person who replaced me was also an operator and former CEO who’d built AI companies himself, not a central-casting sales leader. The pattern held.

W&B was eventually acquired by CoreWeave. The GTM organization we built was a big part of why.

What I’d Tell You

If you’re the technical founder sitting across from your board right now, here’s what I’d tell you.

Trust your instincts on hiring. Lukas built a better sales team on instinct than two professional CROs built on purpose. He hired people who were curious, technical, and consultative. He didn’t know all the sales jargon for what he was doing, but he knew what kind of person could have a real conversation with an AI researcher. He also trusted that I’d be good at running sales, even when I thought that was nuts. If you find yourself thinking “this salesperson seems impressive but I wouldn’t want them talking to my users unsupervised,” trust that feeling. And if you see someone unconventional who deeply understands your customers, trust that too.

Get your hands dirty before you hire a leader. You need to participate in enough deals, from first outbound email to signed contract, that you understand the real objections (not the polite ones), the buying process (not the org chart version), and the competitive dynamics (not the Gartner quadrant version). Write outbound copy, sit in on pitches, negotiate pricing. That experience is irreplaceable, and it’s the only way to build a sales methodology that reflects your market.

Document the pattern, not the pitch. Most sales playbooks are decks full of talk tracks, and most of them are useless. What you need to write down is the buying pattern: who initiates, who blocks, what triggers urgency, what information they need at each stage. When you hand this to a hire, you’re giving them a map, not a script. Maps scale. Scripts don’t.

Your first sales hire should be a full-cycle AE, not a VP. Hire someone hungry who doesn’t mind doing their own prospecting, demos, and negotiations. Bringing on a VP Sales when you have zero reps is like hiring a general with no army. The VP comes when you have 4-5 AEs producing consistently and the job shifts from selling to managing.

Hire technical sellers, not “enterprise reps.” The best hires at W&B were people who had sales experience but had built their careers selling technical products to technical buyers at companies like Domino, Figure Eight, and Scale. They understood the ecosystem and could hold real conversations with practitioners while knowing how to run a deal process and close. The sweet spot is salespeople with technical depth. They’re not cheap (they know what that combination of skills is worth), but the ROI is better than paying less for enterprise reps who can’t earn trust with your buyers. Be prepared to defend these hires and their compensation to other leaders in the company, including your CFO. Non-traditional sellers often come with non-traditional comp expectations, and you’ll need to make the case that the premium is worth it. But winning often requires innovation in the playbook, and you have to follow that conviction even when others aren’t doing it the same way. Every company succeeds for its own reasons.

Know when to hire above yourself. There’s a revenue threshold (roughly $15-25M ARR in my experience) where the job changes from “how do we win this deal” to “how do we build a machine that works without me in the room.” That’s when a professional sales leader starts to make sense. But not before. And when you do hire, make sure they can go deep with your buyers. The lesson from W&B isn’t “never hire a CRO.” It’s “never hire a CRO who can’t hold their own in a 30-minute technical conversation with your most sophisticated customer.”

Sell like an engineer, not like a salesperson. Technical buyers have finely-tuned BS detectors. They want to see something real, have an honest conversation about where the product is strong and where it’s not, and make their own decision. Some of the best sales conversations I had at W&B were the ones where I said “honestly, we’re not great at that yet, but here’s our roadmap and here’s why we’ll get there.” Counterintuitive, but that kind of candor builds more trust than any pitch deck. When you fly out to a customer to personally apologize for issues and walk through the roadmap for fixing them, that carries more weight than any QBR slide ever will.

Why 2026 Is Different

AI is making the founder-as-CRO path more viable, not less. The parts of sales that technical founders hate (and are bad at) are the parts AI handles well: CRM hygiene, follow-up sequences, call prep, meeting summaries, competitive intel, lead scoring. A technical founder in 2026 with the right AI tooling can run an effective sales operation without the overhead that used to require a full GTM team. You focus on what only you can do: having a real technical conversation about a real problem with a real buyer. AI handles the scaffolding.

This is part of why we invest in what we call “AI coworkers” at B Capital: persistent, learning AI systems that integrate into workflows and get better over time. Sales is one of the clearest early use cases. If you’re building in this space, I’d love to talk.

When This Doesn’t Work

If you’re selling to non-technical buyers (CFOs, CHROs, procurement), the founder advantage is smaller because domain credibility matters less than relationship credibility. If your product is horizontal, the buyer set is too diverse for one person to deeply understand. And if you hate talking to customers, forcing yourself through it will produce worse results than hiring someone energized by it.

This is also about the first sales leader. At scale, you need professional sales management. The question isn’t whether to eventually hire a CRO: it’s when and who.

The Signal I Look For

As an investor, the signal I care about most in early-stage AI companies isn’t the pitch deck or the TAM slide. It’s whether the founder can get on a call with a skeptical VP of Engineering and, within 20 minutes, have that person saying “when can we start?”

That ability isn’t taught in any sales methodology. It comes from building the thing yourself and knowing why it matters. The founders who have it build companies that grow efficiently. The ones who outsource it too early spend their Series A cycling through sales leaders while their competitors close deals.

The conventional wisdom says technical founders need a “professional” to run sales. The data says those professionals fail 70% of the time. I watched it happen twice at the same company before doing it myself, and the company did just fine.

Time the conventional wisdom got an update.

We wrote this article for B Capital News & Insights.

Last weekend, the Citrini Research “2028 Global Intelligence Crisis” memo went viral, racking up roughly 16 million views after Michael Burry amplified it. IBM dropped 13%, and we saw broad weakness across software, payments, and delivery stocks. The market panicked.

The memo paints a vivid picture: AI replaces white-collar labor faster than the economy can absorb, consumer demand collapses, unemployment spikes past 10%, and the S&P draws down nearly 40%. They call it “Ghost GDP,” where output shows up in profits but doesn’t circulate because displaced workers have lost their income.

It’s a clean left-tail story, and it’s wrong as a base case, but it’s directionally correct on the structural shift underneath, which is exactly where the investment opportunity sits.

What the Doomers Get Right

Strip away the compressed timeline and the stacked worst-case assumptions, and several of Citrini’s structural observations hold up.

White-collar work is the near-term fault line. A large share of knowledge work is read, write, decide, and coordinate. That’s exactly where current models are strongest. McKinsey estimates 60-70% of employee time is automatable by AI. The pressure will show up first in back office, operations, finance, support, sales enablement, and parts of legal and compliance. These aren’t theoretical targets; these are the functions where we see enterprise buyers already pulling budget.

AI is moving from tool to coworker. The real shift isn’t chatbots getting smarter, it’s AI gaining persistent memory, learning on the job, and planning autonomously. We went from “help me write this” (early ChatGPT, basic copilots) to “do this for me” (Codex, Claude Code, support bots) and are now entering “own this with me.” That last phase changes org design, not just task execution.

The distributional tension is real. If gains accrue primarily to capital while labor lags, demand weakens and politics get volatile. Even without collapse, wage, tax, and benefit debates will intensify. That affects regulation and procurement behavior. Investors who ignore this are underpricing political risk.

Take rates will face pressure. Agents routing around software is overstated near term, but the direction is correct. As discovery, evaluation, and execution become automated, friction-based pricing power erodes. The winners will be workflow-embedded products and infrastructure providers.

Where It Falls Apart

The Citrini scenario only works if every aggressive assumption resolves in the same direction simultaneously within 24 months. That’s not analysis, that’s a horror story presented as a base case.

Enterprise adoption doesn’t move at 𝕏 speed. Capability may be exponential, but deployment is not. Data restructuring, compliance, procurement cycles, and retraining are multi-year arcs. Only 5% of AI pilots currently achieve measurable P&L impact, per MIT research. 32% stall after pilot. METR’s own randomized controlled trial found experienced developers were 19% slower with AI tools, even as benchmarks showed superhuman coding performance, a stark reminder that lab capability and production value are different things. The bottleneck is not demand: 92% of Fortune 500 already use ChatGPT, 82% of executives plan AI agent integration within three years, and $252 billion went into corporate AI investment in 2024 alone. The bottleneck is the infrastructure to deploy and scale AI coworkers in production.

“Ghost GDP” confuses distribution with destruction. Labor savings don’t vanish; they reallocate through lower prices, capex, profits, dividends, and tax revenue. The issue is distribution and timing, not whether gains circulate. This is an important distinction for investors: the value gets created; the question is where it accrues.

Policy is treated as inert. Automatic stabilizers, monetary easing, fiscal transfers, mortgage forbearance, and credit restructurings historically interrupt demand spirals. The memo assumes none of these mechanisms activate. That’s not how economies function under stress.

Business execution is understated. Code is cheap, but trust, compliance, distribution, and operational execution are not. Agents may pressure take rates, but they don’t eliminate institutional infrastructure in two years.

The Narrative Shock Creates Real Opportunity

The market is conflating a long-term structural shift with a 24-month crisis scenario. That dislocation creates two types of opportunity: mispriced growth exposures in public markets and private-market picks-and-shovels that win regardless of macro path. The latter is where we’re focused.

1. AI Co-Worker Applications: Replacing Hours, Not Just Tasks

The highest-conviction opportunity is AI coworkers that land in the enterprise with measurable ROI inside 90 days. The investment filter is simple: can it replace measurable hours with compliance, auditability, and a clear feedback loop?

Software Engineering (~$370B addressable). Claude Code and Codex have commoditized code generation. The frontier has moved to enterprise context and verifiable domains where the AI can prove its work is correct: formal proofs, math/physics, security analysis, AI research itself. The moat here is organizational context, not raw coding ability.

Sales & GTM (~$245B addressable). The market is crowded with AI SDRs, and most of them are dead on arrival. The winners own the system of record, learn from outcomes, and close the loop on what converts. Data rights are the moat; if you can’t observe what works and improve autonomously, you’re a feature, not a company.

Finance & CFO Office (~$215B addressable). High-volume operational workflows with clear accountability: AR/AP, collections, procurement ops, FP&A, compliance reporting. These processes are rules-based but manual-intensive, making them ideal for AI coworkers. The companies we’re most excited about are the ones replacing FP&A analysts, not just augmenting them, where “better” is quantifiable and feedback is continuous.

When AI works in the enterprise, the returns are significant: $3.70 ROI per dollar on average, with top performers hitting $10.30. Gartner projects 33% of enterprise software will include agentic AI by 2028, with 15% of daily decisions made by agents.

2. The Infrastructure Layer: Where Durable Value Accrues

The application layer gets the headlines. The infrastructure layer gets the margins. This is where we believe the market is most under-invested.

Agentic Memory & Context. Models have memory, they don’t have your memory. The gap is organizational context: docs, code, tickets, CRM data, team terminology, approval workflows. This is what separates a stateless chatbot from a colleague. The defensibility compounds because memory improves as context accumulates, and switching costs grow over time as the AI learns your organization. Think of it as the unlock from agent to colleague.

Orchestration & Multi-Agent Coordination. Agents don’t yet collaborate well, with each other or with humans. The missing layer includes coordination protocols, escalation paths, and seamless handoffs between AI and human coworkers. The network effects here are powerful: value increases as more agents and humans use the same coordination layer. This is Slack for the human-AI workforce.

Production Observability. When agents run 24/7, ops teams need real-time visibility into what’s working and what’s not. Most existing tools focus on dev and debug, but the real pain is keeping agents reliable at scale. The first company to nail production-first observability for agentic systems owns the category. This is the Datadog opportunity for the agentic era.

Agentic Security. OpenClaw made this viscerally obvious (more on that below), but the principle applies broadly: agents with persistent access to enterprise systems represent a fundamentally new attack surface. Least-privilege access, skill sandboxing, action-level anomaly detection, and identity management for non-human actors. This category barely existed a year ago, and it will be table stakes for enterprise deployment within two years.

3. Why Now: Four Curves Crossed Simultaneously

This thesis isn’t speculative. It’s grounded in four technical inflection points that converged in 2024-25, and the data keeps accelerating.

Reasoning quality is on an exponential curve, and it’s steepening. METR’s time horizon research, which measures the length of tasks AI agents can reliably complete autonomously, shows capability doubling every ~7 months over six years. Their updated TH1.1 methodology (January 2026) suggests recent progress is actually 20% faster than the historical trend, with post-2023 doubling at 131 days. Claude Opus 4.6 now clocks a 50%-time-horizon of roughly 14.5 hours, meaning it can autonomously complete tasks that would take a skilled human half a working day. If this trend continues for 2-4 more years, we’re looking at agents that can reliably execute week-long projects. MIT Technology Review called METR’s time horizon plot “the most misunderstood graph in AI.” The misunderstanding cuts both ways: doomers extrapolate it to imminent catastrophe, skeptics dismiss it as benchmark gaming. The investment-relevant reading is that autonomous task capability is compounding on a steep, consistent curve, and the gap between what agents can do on benchmarks and what they actually do in production is precisely the market we’re investing into.

That gap is real, by the way: METR’s own developer productivity RCT (July 2025) found that experienced open-source developers were 19% slower when using AI coding tools, despite believing they were 20% faster. Algorithmic benchmarks overstate real-world performance because they can’t capture code quality, context understanding, and integration complexity. This is the deployment gap, and this is the opportunity.

Tool use is standardizing, but not the way anyone expected. MCP was supposed to be the universal connector between AI agents and external services, and it has real traction: 17,000+ servers on MCP.so, OAuth-native authentication, and adoption by OpenAI, Google, and Microsoft. But MCP isn’t the whole story anymore. Skills (reusable prompt-and-script bundles that encode domain knowledge into agent behavior) have exploded: 96,000+ on SkillsMP, 5,700+ on ClawHub, all built on the SKILL.md standard that emerged from coding agents like Claude Code and Codex CLI. Meanwhile, CLIs are emerging as a surprisingly effective third pattern. Agents have been trained to be exceptionally good at using command-line tools, and CLIs handle authentication, structured output, and composability through patterns that have been battle-tested for decades. Karpathy called it publicly: build for agents by exposing functionality via CLI, publishing task-specific skills, and shipping MCP servers. The investment implication is that the “tool use” layer is not a single protocol but an ecosystem of complementary patterns. Skills encode knowledge, MCP provides authenticated access, and CLIs offer execution efficiency. The companies building the orchestration, discovery, and security layers across all three will own the integration tier of the agentic stack.

Context windows expanded to 2M+ tokens. Persistent memory across sessions is now possible. AI can remember what it learned yesterday.

Inference costs collapsed 200x. GPT-4 equivalent capability now costs $0.40 per million tokens versus $20 in 2022. DeepSeek pushed this even further, running 90% cheaper than Western providers.

These aren’t independent trends. They’re compounding. AI coworkers are now technically feasible, economically viable, and enterprise-ready. The question is no longer if but how fast enterprises can deploy them.

4. The OpenClaw Moment: What 157K Stars in 60 Days Tells Us About Where AI Is Headed

If you want a single case study that encapsulates the entire AI coworker opportunity and its risks, look at OpenClaw.

OpenClaw (formerly Moltbot, formerly Clawdbot) started as a weekend side project by developer Peter Steinberger: a personal AI assistant that runs locally on your machine and connects to your messaging apps, email, calendar, and file systems to act autonomously on your behalf. Not a chatbot, not a copilot, but an agent that does things for you across the tools you already use.

The reception was extraordinary. OpenClaw hit 100,000 GitHub stars faster than Linux, Kubernetes, or any project in GitHub history. It crossed 157,000 stars within 60 days. On January 30, 2026 alone, it gained 34,168 stars in 48 hours. The project spawned Moltbook, an AI-only social network where only agents could post, which hit 1.5 million registered agents in five days and drew coverage from Fortune, CNBC, and TechCrunch. Y Combinator’s podcast team showed up in lobster costumes. “Claw” became Silicon Valley slang for locally-hosted AI agents.

This is demand signal, not hype signal. People don’t want another chatbot; they want AI that manages their inbox, controls their schedule, organizes their files, and executes multi-step workflows while they do something else. OpenClaw’s value proposition was blunt: “AI that actually does things, not just talks.” That resonated so strongly that OpenAI took notice. On February 14, Steinberger announced he was joining OpenAI, a move widely interpreted as OpenAI’s play to acquire agentic AI talent after their $3B bid for Windsurf (Codeium) fell through. The project transitioned to an independent open-source foundation under MIT license, mirroring the governance model of Linux and Kubernetes.

The enthusiasm is the bullish signal. Now here’s the cautionary one.

Within weeks of going viral, SecurityScorecard found over 40,000 exposed OpenClaw instances on the public internet, 63% of them vulnerable to remote code execution. Researchers discovered 400+ malicious “skills” on ClawHub (OpenClaw’s marketplace) distributing infostealers, remote access trojans, and backdoors disguised as legitimate automation tools. A critical one-click RCE vulnerability (CVE-2026-25253) meant attackers could compromise systems through a single link without the user installing anything. Skills execute with full agent and system permissions, with no sandboxing and no least-privilege access. Users were following YouTube tutorials that never mentioned security, deploying agents on cloud servers with authentication set to “none.”

The most alarming development: Hudson Rock documented the first observed case of an infostealer harvesting an entire AI agent configuration, not just browser passwords but the complete identity, permissions, and API keys of a personal AI agent. That’s a new attack surface that didn’t exist twelve months ago, and infostealers are now targeting AI personas as high-value assets.

This matters for our thesis on three levels.

First, the demand is real and it’s massive. 157K stars in 60 days, OpenAI acquiring the creator, and “claw” entering the tech lexicon as a verb are not indicators of a fad. Consumers and developers are telling us, loudly, that they want persistent AI agents with real autonomy over their digital lives. The enterprise version of this same demand is the AI coworker.

Second, agentic security is not a feature request, it’s a category. When agents have persistent access to email, calendars, financial accounts, and code repositories, the blast radius of a single compromise is an order of magnitude larger than a stolen password. Enterprise buyers will not deploy AI coworkers at scale without permissions frameworks, action-level anomaly detection, skill sandboxing, and audit trails. Every CISO who reads the OpenClaw postmortems becomes a buyer for agentic security tooling.

Third, OpenClaw draws a bright line between consumer-grade agent experiments and enterprise-grade AI coworkers. The difference is infrastructure: identity, access control, policy enforcement, observability, and governance. OpenClaw shipped the agent but not the infrastructure underneath it. That infrastructure layer is exactly where we’re investing.

5. The PE and Credit Warning

The Citrini memo should be a warning sign for PE-backed SaaS with weak differentiation and friction-based economics. If your portfolio company’s pricing power depends on being embedded in a workflow that an AI agent can route around, your margins are on borrowed time.

The selective opportunity in PE and credit: buy-and-build modernization plays where AI compresses COGS and SG&A, and distressed recurring revenue assets deeply embedded in workflows that agents will need to run through, not around. Avoid aspirational ARR quality and unsecured exposure if growth stalls.

6. Services as a Wedge

Implementation friction is real, which makes transformation services investable. Mapping processes, instrumenting data, integrating systems, training organizations, and building governance layers. These capabilities scale if paired with product and repeatable playbooks. Enterprises want outcomes, not tools, and the companies that combine software and delivery to implement coworkers and redesign processes will compound.

What We’re Looking For

Our investment filtering criteria across this thesis:

Team. AI technical depth plus domain expertise. Exceptional founders with strong founder-market fit over traction alone. In a market moving this fast, the team’s ability to navigate rapid shifts matters more than any current metric.

Data Moat Quality. As AI models commoditize, unique high-quality data with continuous feedback loops becomes the primary differentiator. If your data advantage can be replicated by a competitor with a bigger API budget, it’s not a moat.

Workflow Embedding Depth. Deep integration into daily workflows creates switching costs. Products that are indispensable to users’ daily work are defensible; products that sit on top of workflows are features waiting to be absorbed.

Progressive Defensibility. Technical moats alone don’t last in AI, and they may not even exist anymore. We’re looking for clear plans to layer in defenses over time: data accumulation, network effects, regulatory positioning, distribution lock-in.

Economic Value & Business Model Innovation. GTM and pricing that support AI economics for both the company and its customers. Sustainable unit economics are not optional.

We’re actively seeking Seed to Series C investments in AI coworker applications across engineering, sales, finance, and support, as well as enabling infrastructure in memory, orchestration, observability, and agentic security. Check sizes range from $5M to $50M.

Bottom Line

The Citrini piece is a useful stress test, not a prediction. The market appears to be treating a long-term structural shift as an imminent crisis, and that creates dislocation for investors willing to look past the noise.

The structural shift is real. METR’s data shows capability doubling every 4-7 months with no sign of deceleration. AI coworkers will change how enterprises operate, how work gets organized, and where value accrues. But it will follow enterprise timelines, not Twitter timelines.

If anything, the panic reinforces the thesis: AI coworkers and the enabling infrastructure are where durable value gets built. The 95% pilot failure rate isn’t evidence that AI doesn’t work, it’s the problem we’re investing to solve. OpenClaw showed us what happens when powerful agents ship without enterprise-grade infrastructure underneath them. The companies that bridge the gap from pilot to production, that give enterprises memory, orchestration, observability, security, and compliance, will capture outsized value in the decade ahead.

That’s where we’re putting capital, and that’s where we think you should be looking too.

I wrote this article for B Capital News & Insights

The Model Isn’t the Moat Anymore

For two years, “AI progress” meant “the model got better.” That era is ending.

The evidence is stark: according to MIT’s State of AI in Business 2025 report, only 5% of enterprise GenAI pilots achieve measurable P&L impact. S&P Global found that 42% of companies abandoned most AI initiatives in 2025, up from 17% in 2024. The average organization scrapped 46% of AI proof-of-concepts before reaching production.

These aren’t bad models. They’re bad environments.

Model capability is still improving, but for most enterprises it is no longer the limiting constraint. What matters now is everything around the model: integration, governance, distribution, measurement and the ability to learn in production without breaking trust.

One constraint is consistently underweighted in almost every AI strategy deck I see: organizational fit.

If AI is going to deliver durable value, it must function less like a tool and more like a coworker. One that collaborates with humans, operates inside real team workflows and carries context over time.

The winners won’t be the teams with the “smartest” model. They’ll be the teams with the best environment to deploy, trust and continuously improve AI.

The Coworker vs. Tool Distinction

This isn’t semantics. The difference between AI-as-tool and AI-as-coworker determines whether value compounds or collapses.

A tool waits to be invoked. It processes inputs and returns outputs. It has no memory of your organization, no understanding of how your team actually works, no awareness of who should approve what. Every session starts from zero.

A coworker maintains context across interactions. It knows your domain, your team’s terminology and your approval workflows. It can be delegated to, supervised and held accountable. It gets better at its job over time because it learns from outcomes, not just prompts.

The MIT data validates this distinction. Their research found that vendor-built solutions succeed at a 67% rate, while internal builds fail at a 67% rate. Why? Vendors who win are the ones building coworker-like systems with deep workflow integration, not generic tools bolted onto existing processes.

Consider what a new human hire experiences: onboarding, permissions, a manager, feedback loops, access to institutional knowledge, clear escalation paths. We don’t hand them a keyboard and expect productivity on day one. Yet that’s exactly how most enterprises deploy AI.

The Shift: From Capability Race to Execution Advantage

In practice, this shift shows up when AI performs well in pilots but fails to survive first contact with real workflows. Buyers are no longer asking “Does it ace a benchmark?” They’re asking a different class of questions altogether:

• Integration: Can it plug into my workflows without rewriting my org chart?

• Governance: Can it touch sensitive data without creating security, privacy or compliance blowback?

• Accountability: Who is responsible when it’s wrong?

• Measurement: Can we evaluate it in production and improve it safely?

• Scale: Can we roll it out to thousands of users without adoption collapsing?

• Collaboration: Can it work with my team like a competent new hire, or does it just generate text?

Those are not model questions. Those are execution questions.

And execution compounds. Deployment creates feedback. Feedback enables improvement. Improvement drives adoption. Adoption earns deeper integration. That loop becomes the moat.

McKinsey’s 2025 AI survey confirms this pattern: organizations reporting “significant” financial returns are twice as likely to have redesigned end-to-end workflows before selecting models. The execution advantage emerges less from initial capability and more from the ability to learn safely in production over time.

5 Ingredients of AI Execution Advantage

By “environment,” I mean the structural conditions that let AI compound in production. These conditions determine whether improvement accumulates or stalls after deployment.

1. Integration Surface How quickly AI can ship into real workflows.

Value compounds fastest when AI lives inside the system of record, removes steps, reduces cycle time and tightens feedback loops. The MIT research shows that ROI is lowest in sales and marketing pilots, where most GenAI budgets are concentrated, and highest in back-office automation where integration is deepest.

The integration question isn’t “can we connect via API?” It’s “can we embed deeply enough to observe outcomes and improve?”

2. Data Rights and Governance What the system can legally and operationally learn from in production.

If you can’t observe outcomes, you can’t improve. If you can’t improve, you don’t compound. Companies that solve this and can learn from production without violating governance will outperform those that can’t.

3. Distribution and Procurement How deployments become default, not optional.

AI doesn’t win by demos. It wins by rollout. PwC’s 2025 survey found that 79% of organizations have adopted AI agents at some level, but only 35% report broad adoption, and 68% say half or fewer employees interact with agents in their daily work. The gap between “we have AI” and “AI is how we work” is primarily a distribution problem.

4. Production Learning Loop Evaluation, monitoring and improvement without breaking trust.

Real-world evaluation tied to business KPIs. Monitoring for drift and failure modes. Human routing for uncertainty. Continuous improvement with governance guardrails. Gartner predicts that 30% of GenAI projects will be abandoned after proof-of-concept by the end of 2025. Not because the technology failed, but because organizations couldn’t build the infrastructure to improve safely in production.

Organizational Fit AI must function as a coworker, not a tool.

This is the missing pillar that most AI strategies ignore entirely. Enterprises are networks of roles, permissions, incentives and handoffs. “Agentic” only works when AI behaves like a well-scoped teammate: collaboration mechanics inside existing workflows, identity and least-privilege permissions, durable memory and context and on-the-job learning that operates without violating governance.

When I evaluate AI companies, I ask: “Would you hire this system as a junior employee?” If the answer requires caveats about supervision, permissions and trust boundaries, you’ve identified the product work that matters.

Where AI Value Compounds Fastest

Value concentrates where execution environments support compounding

Instrumented digital workflows where shipping is fast and telemetry is rich. Software development, customer support, back-office operations. Anywhere outcomes can be observed quickly and iteration is cheap.

High-volume operational workflows with clear accountability and measurable outcomes. Claims processing, compliance review, financial operations. Environments where “better” is quantifiable and feedback is continuous.

Physical operations with telemetry and hard KPIs. Manufacturing, logistics, healthcare delivery. Domains where the system of record captures reality and improvement is directly measurable.

Generic assistants without durable data rights, distribution leverage and a compounding learning loop get competed down to commodity margins.

What This Means for Founders

Markets are not just industries. They are execution environments. This favors teams that optimize for compounding environments over early polish or surface-level performance.

Wedge into the system of record. Don’t build alongside the workflow. Become the workflow. The difference between “we integrate with Salesforce” and “we are where deals get done” is the difference between tool and coworker.

Secure data rights early. The legal and operational ability to learn from production is a moat. Companies that negotiate this upfront, while offering clear value exchange, will outperform those who treat it as a Phase 2 problem.

Design for procurement from day one. Audit logs, SSO, role-based access and compliance certifications. These aren’t features; they’re prerequisites for the environments where AI compounds.

Treat evaluation as product. If you can’t show measurable improvement on business KPIs, you can’t justify continued investment.

Build the AI coworker layer. Collaboration, identity, permissions, memory and handoffs. This is the unsexy work that separates pilots from production systems.

Environments that support compounding often look weaker early yet outperform over time. This allows founders to look wrong early and still be right in the long run.

What This Means for Enterprises

Buying AI like ordinary software and expecting it to behave like ordinary software does not work. AI systems improve only when they are treated as production systems with owners, feedback and failure modes.

Establish an AI operating model. Clear owners, defined accountability and incident response. Who is responsible when the AI makes a mistake? If you can’t answer this question, you’re not ready for production.

Tie AI performance to business KPIs. Not accuracy metrics, not user satisfaction scores. Actual business outcomes: revenue, cost, cycle time and error rates.

Reduce fragmentation where learning loops need consistency. Every team using a different AI tool means every team learning in isolation. Consolidation isn’t about cost savings; it’s about compounding.

Treat AI coworker fit as a first-class requirement. When evaluating vendors, ask: “How does this integrate with how my team actually works?” Not how it works in a demo. How it works in your environment, with your permissions, your approval flows and your existing tools.

What This Means for Investors

Model quality is no longer the primary diligence question. Instead, evaluate:

Ownership of the integration surface. Does the company control the system of record, or are they dependent on someone else’s platform?

Durable data rights and credible governance. Can they legally and operationally learn from production? Is their data strategy a moat or a liability?

A scalable distribution path. Can they reach thousands of users without a proportional increase in sales and support costs?

Evidence of a production learning loop. Are they improving from deployment, or shipping static models?

A credible path to AI coworker fit. Can they function inside real enterprise environments with real permissions and real accountability?

We believe the best AI investments right now are companies building execution infrastructure, not model capability alone. The model layer is commoditizing; the execution layer is where durable value will be built.

How This Shows Up in Our Portfolio

This framework has shaped our investing strategy for some time. A few examples:

Perplexity: Enterprise knowledge work is an execution environment problem. Perplexity’s enterprise offering is explicitly about deploying AI into organizational context: collaboration in Spaces, answers from organizational apps and files, enterprise permissioning, auditability and “no training on your data.” This is governance, distribution and coworker-fit working together in production.

Unblocked: A literal AI coworker for engineering teams. Unblocked plugs into the tools engineers already use, connects code, documentation and conversations, supplies shared team context that makes other AI coding tools more effective. Enterprise fit is table stakes: SSO, RBAC, audit logs and security posture designed for production.

Goodfire: If you care about production reliability, you eventually care about controlling behavior, not just prompting it. Goodfire is building interpretability tooling that surfaces failure modes, enables behavior design and supports durable fixes. This maps directly to the production learning loop and governance required for AI systems to improve safely.

Axiom: In domains where correctness is existential, value shifts toward systems that can reason rigorously and be evaluated against hard truth. Axiom’s focus on an AI mathematician is a wedge into verifiability-first reasoning. It’s upstream capability in service of downstream production requirements.

Where Advantage Compounds

For the next decade, the biggest AI outcomes will not come from “the model got better.”

They will likely come from environments where AI can be deployed, trusted, measured and improved continuously inside real workflows. The environments we choose to build in will determine which AI systems endure.

The data is already pointing the way: 95% of pilots fail not because AI doesn’t work, but because organizations haven’t built the necessary working environment. The 5% that succeed share common characteristics: deep workflow integration, clear governance, production learning loops and organizational fit.

Capability is table stakes. Execution advantage is the moat.

The question for founders, enterprises and investors isn’t “which model is best?” It’s “which environments support compounding?”

Build there, and if you’re already building there, I’d love to talk to you.

If you’re looking for the worksheet and want to skip reading the associated post, you can find it here.

Who Is This For?

If you’re reading this, you may already have an offer from an early-stage startup and are looking to understand the value of the equity portion of your comp package. For simplicity, let’s call this prospective employer Early Inc. You may even have multiple offers, including one from a later-stage company (Later Corp) that you’re comparing it with.

Let’s start with three things you should NOT do at this juncture:

1. Discount the value of the equity to zero and focus solely on the base/cash portion of the package. If you choose your startup correctly, the equity will be the most significant portion of your compensation, possibly by orders of magnitude of your cash comp.

2. Choose to work at a later-stage or public company strictly because their equity is easier to value. There are many reasons to choose a specific later-stage company over a specific earlier stage one, but ease of equity calculation shouldn’t be one of them.

3. Multiply the shares you’d be granted by either the price-per-share (PPS) of the most recent round of financing or the most recent 409A valuation. Both of these calculations will provide you with the current value of the equity, essentially the value at which you’re “buying” the equity, but they won’t give you the value at exit, which is when the equity would be “equivalent to cash.” It’s the latter you’re looking to understand.

Estimating the value of startup equity at exit is hard work. An early-stage company faces a long and arduous journey prior to an exit that could take a number of forms, e.g. an acquihire, a medium-sized sale for product and technology, a larger acquisition, or an IPO in addition to a shutdown in the worst case or, on the flip side, the moonshot scenario of becoming a deca-billion-dollar category-defining disruptor.

The best way to account for all of these variables is by performing an optimism-weighted calculation.

Why Does This Matter?

Let me first tell you about a mistake I made many years ago.

In 2005, I had an offer from Google, a tech giant fresh off of its IPO. I also had an offer from Facebook, which was a tiny early-stage startup at the time. The value of Google’s equity in “2005 value” was 10x that of Facebook’s. I joined Google.

When Facebook went public seven years later, the value of both equity offers in “2012 dollars” (when Facebook’s equity was finally “equivalent to cash”) flipped dramatically — Facebook’s equity offer was all of a sudden 50x greater than Google’s.

Now, don’t get me wrong. I had a great time at Google and met many of my closest friends there, but I wish I had known to perform my own optimism-weighted calculation for both companies before making my final choice.

I want to help you avoid this mistake.

How Does It Work?

First, you’ll need to get some data from the company:

1. The number of shares they are granting you (this should be in the offer)

2. The latest valuation

3. The latest price per share (PPS)

If they are unwilling to share this data, don’t join! It’s a massive red flag for a company to withhold any important materials you might need to make an informed decision about the next decade of your life.

Second, you need to forecast a series of scenarios that the company could ultimately wind up in. At the very least, I would propose sketching out the following:

• A shutdown whereby the company ends up being worth $0

• A small acquihire

• A modest sale to a competitor

• A meaningful sale to a large private company or a public one

• An IPO

• Becoming a category-defining public company

Next to each scenario, you should estimate the probability of the outcome as well as the amount the company would be worth.

One thing you absolutely do NOT want to do is ask the founders or hiring managers to provide these forecasts. You can and should do your own due diligence regarding the company’s thought process on the grander vision, the opportunities that lie ahead, the pitfalls, etc. But you need to fill in the outcomes and probabilities yourself because you want your optimism reflected back, not theirs.

There are three questions I like to ask the company representatives to help me think through the possible outcomes and their probabilities:

1. In your wildest dreams what does this company look like? And what needs to go right for that outcome to be realized?

2. The company hasn’t achieved your wildest dreams. What happened and what does it look like instead?

3. Tell me about the worst-case scenario, i.e. the company failed. What happened and why?

You should end up with a table that looks like this, but with your own scenarios and numbers:

Next Steps:

1. For each scenario, compute the price per share for that scenario using the following formula:

2. Compute the optimism-probability weighted average PPS

3. Multiply that PPS by the number of shares you are being offered

This calculation is incredibly informative as it’s a reflection of your own view of the future potential of Early Inc. The more excited you are about it being a big category-defining company, the more you will see that reflected back in the optimism-weighted value of your shares and vice versa.

To make this interactive, I’ve set up a worksheet that you can copy and play around with here.

What Are Reasonable Probabilities?

What is a reasonable expectation for the probability of a company shutting down or going public? Base rates can inform your judgment. For the cohort of startups founded since 1995, the histogram of the probability of outcomes looks like this:

That said, if you’re relying exclusively on the base rates above, you’re essentially telling yourself that Early Inc is an “average company,” and that could end up dramatically underestimating the outlier outcomes that you’re so carefully attempting to factor-in. So you should use these base rates as a sanity check rather than a replacement for your own probability projections.

Advanced: Additional Dilution

In an effort to increase the accuracy of your final projections, albeit slightly, you could also estimate how much additional dilution the company will incur from future rounds of financing. I wouldn’t worry too much about getting this 100% correct as the effects from additional dilution will likely be subsumed by the margin of error of your probability estimations anyway. But the general adage here is that the bigger the exit valuation (as a multiple of today’s valuation), the more rounds of financing (and dilution) the company is likely to incur along the way.

To incorporate this, you’d have to make a small tweak to the PPS calculation in the algorithm above:

This variable is reflected in the linked worksheet. If you want to ignore additional dilution, just set it to 0% for every scenario.

Comparing Offers

The optimism-weighted approach is also incredibly useful if you want to compare two companies against each other. Fill in the values for each entity and see how the weighted average differs. This approach affords the ability to do an “apples-to-apples” comparison between an early-stage opportunity with a later-stage one, which the more commonly used approach of multiplying the granted shares by the latest valuation PPS does not.

For illustration purposes, let’s look at two hypothetical scenarios:

1. Versus another Early Stage Offer

First, let’s compare Early Inc to OtherEarly Corp. To showcase the importance of the optimism-weighted probabilities, I’ve made the values equal across the board for both companies except for the P(Outcome) columns (Columns C and K in the worksheet). Notice how this impacts the end result: Early Inc’s optimism-weighted value for your shares is more than 40% greater than for OtherEarly Corp. Also compare the (useless) value today to the optimism-weighted average: the latter is 120x greater.

2. Versus a Later Stage Offer

Now, let’s compare Early Inc to a later-stage company (Later Corp) with a current valuation that’s 20x greater. In this case, Later Corp’s valuation today is $200M to Early Inc’s $10M, and the today value of Later Corp’s equity grant is 5x that of Early Inc’s.

The probability of worse outcomes for Later Corp is also lower, given that the mortality rate of startups tends to decrease as they raise more money. Nevertheless, since we have a more optimistic view for Early Inc over the long run, the optimism-weighted average reveals that Early Inc shares are worth more than 3x those of Later Corp.

Conclusion

Is the optimism-weighted value the correct one? The only way to find out is to work at Early Inc for the next few years, or decades. If your own optimism for the company suggests that you should take the offer, you hopefully won’t regret it, even if your calculations are a little off.

I’ve been geeking out on Deep Learning lately, taking Andrew Ng’s awesome Deep Learning specialization on Coursera and my friend Lukas Biewald’s awesome ML class. I wanted to build my own Deep Learning desktop so I can train models much faster than on my Mac laptop (or even than on an AWS Deep Learning AMI). With Lukas’ help & tutelage, we made it happen.

In case you’re interested in doing the same, here’s the box we built. Most of the time was spent configuring the software properly. To save you from some of the pain, I’ve tried to excruciatingly detail the steps I took. Most of what I did was Google-based debugging. There are great resources on the Internet on how to remove obstacles. Wherever possible, I tried to link to the original source material I learned from.

Getting the parts

The whole setup revolves around an NVIDIA GPU, the workhorse that drives the machine learning magic. I opted for a high-end 1080Ti which is ~1/2 of the total cost of the buildout below. The speed at which it crunches through models has been totally worth it.

• GPU: EVGA GEForce GTX 1080Ti 11GB. I only got one, though the motherboard supports 2

• Motherboard: ROG STRIX Z370 GAMING (Wi-Fi AC)

• CPU: 8th Gen Intel Core i7–8700 3.2GHz 6C/12T LGA-1151 12MB Cache. Make sure it’s 8th Gen to match the Motherboard

• RAM: Corsair Vengeance LPX 16GB (2x8GB) DDR4 DRAM 3200MHz

• Hard Drive: WD Blue 3D NAND SAA SSD M.2 2280 500GB SSD. Make sure it’s M.2 so it fits in the motherboard

• Case: Fractal Design Mini-C. Mostly make sure it’s Micro-ATX so it fits the motherboard

• Power Supply: Seasonic Focus 750 Gold (SSR-750FM). Make sure it has enough Wattage to drive the machine with the hungry GPUs

Other things you’ll need (which I assume you already have):

• A Monitor, Keyboard, and Mouse. If not, I recommend getting an HDMI-compatible monitor and USB Keyboard and USB Mouse. Any decent ones should do here.

• An Ethernet cable

• A decent-sized (64GB) USB stick you can use to make a bootable Ubuntu

The whole thing, at the Central Computers physical store in San Francisco cost a little over $2200. On the plus side, they installed the CPU, RAM, and HD onto the motherboard, significantly simplifying the installation. It would probably costs a bit less if ordered online (sales taxes) but the peace of mind was worth it.

Setting up the Hardware

Given that the nice folks at Central Computers already took care of the CPU, RAM, HD, all we had to do was:

• Install the motherboard into the case. This was relatively straightforward just a few screws

• Install the power supply into the case. Similarly straightforward once we realized we could remove the case sides

• Install the GPU unit into the motherboard

• Connect the power supply to the fans, motherboard, GPU, etc. This is where it’s really important to read the motherboard manual and follow the instructions precisely. This will save you a lot of heartache.

Overall the hardware bit was the easy part.

Setting up the Software

This turned out to be quite a journey with 4x re-installs of various versions of Ubuntu, wacky tethering to get network, etc. I’ll spare you the story and tell you what worked.

Prepping the OS

Ubuntu 16.04.4 LTS. I needed to install it on a bootable USB stick. I have a Mac, so used these great instructions.

Getting internet on the box

Originally, Ubuntu 16.04 won’t recognize the Realtek WiFi chip on the motherboard (we fix this later), so you’ll need to connect the desktop to wired internet using the aforementioned ethernet cable. If you’re not close to a wired ethernet port (I wasn’t), it’s quite fortunate my Mac let me share my wifi over wired Ethernet. This was a great save.

Installing the OS

Boot with the USB stick in and follow the instructions. If all has gone well, you can connect to network and fetch the updates during the install. When it’s done, reboot.

Messing with the video drivers, part 1

Once the OS was installed, the NVIDIA drivers on Ubuntu 16 started causing trouble. In particular I saw an error that said something like: dev/sda1: clean, 552599/6111232 files, 7119295/24414464 blocks. Fortunately there’s a fix.

Press CTRL-ALT-F1. This will get you into TTY mode. Then uninstall any old NVIDIA drivers:

sudo apt-get remove nvidia-*
sudo apt-get autoremove

Switching to text & setting up SSH

At this stage, I caveat that I never really got lightdm (Ubuntu’s GUI) working. I didn’t really care because the goal was to run a headless machine that I can SSH into from my Mac. So I:

• Switched to text/console mode

• Set up SSH by following the Outside Access section of this great post

Removing Nouveau

Before installing the NVIDIA drivers, you have to remove the open source Nouveau drivers that Ubuntu installs by default. Fortunately this post details how to do it. Follow Steps 1–3 but ignore the NVIDIA run step as we’ll install the NVIDIA drivers another way (via CUDA below)

Machine Learning Software Stack

I got this machine working with

• CUDA 9.0

• CUDNN 7.0.5.11

• Tensorflow 1.7

Note the latest release as of this writing was CUDA 9.1 and CUDNN 7.1.1 but I couldn’t get Tensorflow to work with these yet.

Installing CUDA 9.0

1. Follow NVIDIA’s Pre-Installation Actions to the letter

2. Download the Base Installer and follow the installation instructions. One important change in step 4, run instead sudo apt-get install cuda-9-0 to make sure you install version 9.0 and not the latest version (was 9.1 as of this reading)

3. Follow NVIDIA’s Post-Installation Actions to the letter. I performed the Recommended Actions (7.2) and the Optional Actions (7.3) as well.

Installing CUDNN 7.0.5.11

1. Get yourself an NVIDIA Developer account if you don’t already have one. It’s free

2. Download CUDNN. Make sure you get the one that says Download cuDNN v7.0.5 (Dec 5, 2017), for CUDA 9.0 You’ll need 3 files: (1) cuDNN v7.0.5 Runtime Library for Ubuntu16.04 (Deb), (2) cuDNN v7.0.5 Developer Library for Ubuntu16.04 (Deb), (3) cuDNN v7.0.5 Code Samples and User Guide for Ubuntu16.04 (Deb)

3. Get those files onto your Deep Learning Box. As a Hack, Dropbox links work with wget so I downloaded the three files on my Mac, moved them to Dropbox, then used wget via SSH to grab them onto the Deep Learning box

4. Install CUDNN by following Step 9 & 10 of this doc. Make sure you use the filenames you downloaded.

Install Tensorflow & Keras

You can now install one of the pre-built Tensorflow libraries. Follow the Installing with native pip instructions from Tensorflow. I built for python3. Make sure to install tensorflow-gpu to take advantage of your fancy GPU. Make sure to test it works.

You can also install Keras by running sudo pip3 install keras

At this point congratulate yourself. You have a (mostly) working Deep Learning box.

Enabling the WiFi chip

If you’re satisfied with a wired Ethernet connection and don’t need to get the WiFi chip working, you can safely skip this section. As I was tethering off my Mac, I was hungry to get the onboard Realtek WiFi chip working.

The following worked to make the OS recognize the chipset:

sudo apt update
sudo apt install git
git clone https://github.com/rtlwifi-linux/rtlwifi-next
cd rtlwifi-next
make
sudo make install
sudo modprobe rtl8822be

Connecting to Wifi

Once the chipset was recognized, I needed to connect to my home WiFi network. This set of instructions was great for that. My wifi adapter was wlp4s0. Yours may vary.

Fixing WiFi sleep reconnection errors

I noticed that my Wifi would randomly disconnect. That was no fun. I followed these instructions, noting that my wifi chipset was rtl8822be so that’s the name you have to use for the rest of the instructions.

Advanced: Install Tensorflow from Source

If it bothers you that Tensorflow’s pre-built libraries doesn’t take advantage of CPU optimizations like FMA and AVX2, you can build Tensorflow from source to get all the optimization goodies. To do this:

1. Grab Tensorflow sources. I used version r1.7, git checkout r1.7

2. Follow Step 11 from this doc, but make sure to include the CPU optimizations you need in the build script. They’re of the form --copt=-m<opt so for me, to build with AVX2 and FMA, my build script looked like bazel build --config=opt --copt=-mavx2 --=-mfma --config=cuda --incompatible_load_argument_is_label=false //tensorflow/tools/pip_package:build_pip_package

Make sure to first remove the Tensorflow you installed from prebuild libraries sudo pip3 uninstall tensorflow-gpu and test that Tensorflow works after you’ve build it.

Conclusion

Building your own Deep Learning box can be frustrating during the process, but the fun of seeing models blast through it once you have it running will be well worth it. If you run into bugs, the Internet likely has an answer for every nasty OS, CUDA, CUDNN, etc question. Keep plowing towards Deep Learning goodness.

There are many great guides to build great Deep Learning boxes, including Lukas’ O’Reilly write-up from last year. This is my own. Feel free to follow your own sherpa or make your own path.

Thank you to Lukas who was a great guide:

I recently read the Lessons of History by Will and Ariel Durant. The book is essentially a self-authored Cliff’s Notes of their 11 tome The Story of Civilization and it’s incredible. In the spirit of summary², I’ve tried to jot down some of the more meaningful passages for me in bullets below and wanted to share them.

Note: I didn’t put quotes around anything because none of the thoughts below are mine. You can assume they are either direct quotes or close paraphrases. I really recommend you read the book: it’s only about ~100 pages… or listen to the audiobook which has interviews with the Durants at the end of each chapter.

History Itself

• Knowledge of past events is incomplete, inaccurate, clouded by missing evidence & biased historians.

• Most history is guessing, and the rest is prejudice.

• Conclusions from the past to the future are made more hazardous by acceleration of change.

Geology

• Human history is a brief spot in space.

• Geology changes every day: sea encroaches on land, land on sea. Mountains rise/fall, valleys become deserts, etc..

• Climate no longer controls us but still limits us. A tornado can ruin in an hour the city that took a century to build.

• If temperature changes by 20%: thriving zone→ lethargy.

• Civilizations grew along water. It’s the life of organisms & towns and cheap transport for trade.

• In 1492 oceans trumped seas. Then airplane changed the map of civilization again as trade routes don’t have to follow the rivers anymore. Coastal cities will lose some luster. Landlocked countries will gain power.

Biology

• History is a fragment of biology. Laws of biology apply to history.

• First lesson is life is competition. Life is peaceful when food abounds, violent when mouths outrun the food.

• Cooperation is real but a tool of competition. We cooperate in our group to compete better vs. other groups. True in families, communities, clubs, parties, “race”, nations, etc

• Competing entities have same traits as competing individuals: acquisitiveness, pugnacity, partisanship, pride

• War is a nation’s way of eating

• Second lesson is life is selection. Some organisms fail & some succeed. Some are better equipped than others to meet tests of survival.

• We are born unfree and unequal to Nature: subject to physical and psychological heredity, diversely endowed in health, strength, mental capacity & character. Nature loves difference.

• Inequality grows with the complexity of civilization. Economic development specializes functions & differentiates abilities. Makes men unequally valuable to their group.

• Freedom & equality are sworn enemies. When one prevails the other dies. Utopias of equality are doomed.

• Third lesson is life must breed. Nature has no use for organisms that can’t reproduce. If there is too much supply of humans, 3 agents for restoring peace: famine, pestilence, war. Lower birth rates on other hand reduce economic & political power of the group within society.

Race

• Role of race seems rather preliminary than creative (despite past historians claims to the contrary)

• It is not the race that makes the civilization, it is the civilization that makes the people: circumstances geographical, economic, and political create a culture, and the culture creates a human type.

• “Racial” antipathies have roots in ethnic origin, are predominantly generated differences of acquired culture: language, dress, habits, morals, or religion. There is no cure for such antipathies except a broadened education

Character

• History shows little alteration in the conduct of humankind. Means change, motives don’t.

• Rich & poor have the same impulses. Nothing more common than rebels in power adopting the methods that they condemned when they were still rebels.

• Evolution has been social not biological.

• The hero has a place in history in forming unstereotyped responses to new & unpredictable events.

• Conservative who resists change is as important as radical who proposes it. New ideas should be heard for the sake of the few that are good, but also need to go through the mill of objection & opposition. They have to survive the crucible.

Morals

• Moral codes differ because they adjust to the situation. Morals in hunter era were very different from agricultural era: greed, cruelty, polygamy.

• Man’s sins are the relics of his rise rather than the stigmata of his fall

• Agriculture regime required new virtues: industriousness > bravery, regularity > violence, peace > war, children = economic assets. Strong familial authority.

• Industrial revolution changed it again: children = individuals not economic assets, marriage delayed, authority of parents not as important, education > religion

Religion

• Indispensable to a happy land/age

• It has kept the poor from murdering the rich. Inequality dooms many to poverty or defeat: supernatural hope staves off despair. Destroy the hope → class war

• Religion has many lives & a habit of resurrection. It prevails when laws are feeble and morals must bear the burden of maintaining social order. Fades with the power of law.

• There is no significant example in history, before our time, of a society successfully maintaining moral life without the aid of religion.

• As long as there is poverty there will be gods.

Economics

• The men who can manage men manage the men who can manage only things, and the men who can manage money manage all.

• History is inflationary, and money is the last thing a wise man will hoard

• Economic system must rely on some form of profit motive. Substitutes (slavery, police supervision, ideological enthusiasm) are too unproductive & expensive and/or transient

• Men are judged by their ability to produce — except in war, when they are ranked according to their ability to destroy.

• Majority of abilities is gathered in a minority of men. Concentration of wealth is a natural result, recurs in history. Rate of concentration varies with economic freedom permitted by morals & laws.

• Concentration may reach a point where strength of number in the many poor rivals the ability in the few rich. Critical juncture. Solved either by legislation redistributing wealth or revolution redistributing poverty. Then power/wealth concentrates again. This is the eternal cycle.

Socialism

• The struggle of socialism against capitalism is part of the historic rhythm in the concentration and dispersion of wealth.

• Capitalist fulfilled a creative function: savings→productive capital, financed mechanization, created flow of goods from producer→consumer.

• History has had many socialist interludes: Sumeria, Egypt under Ptolemies, Rome under Docletian, China under Szuma Ch’ien and many others.

• Over time the two are getting closer. The fear of capitalism has compelled socialism to widen freedom, and the fear of socialism has compelled capitalism to increase equality.

Government

• Men love freedom. Freedom of individuals requires some regulation of conduct. Hence first condition of freedom is its limitation. Make it absolute and it dies in chaos.

• Power naturally converges to the center as it’s ineffective when divided & spread.

• Monarchy is the most natural form of government, judged by prevalence in history. Democracies have been hectic interludes. Monarchy has a middling record. When hereditary, it leads to stupidity, nepotism, and extravagance.

• Most governments have been oligarchies: aristocracies (chosen by birth), theocracies (by religion), democracies (by wealth). It’s unnatural for a majority to rule because can’t be organized for united & specific action.

• Majority can do no more than periodically throw out the minority and set up another.

• History doesn’t justify revolutions. Effects would have in most cases come with economic developments. To break sharply with the past is to court the madness that may follow the shock of sudden blows or mutilations. Violent revolutions don’t redistribute wealth: they destroy it.

• Democracy is the most difficult of all forms of government, since it requires the widest spread of intelligence, and we forgot to make ourselves intelligent when we made ourselves sovereign.

• Democracy has embedded class conflict. Every advance in the complexity of the economy puts an added premium upon superior ability, and intensifies the concentration of wealth, responsibility, and political power.

• All deductions having been made, democracy has done less harm, and more good, than any other form of government.

• If equality of educational opportunity can be established, democracy will be real and justified. Though men cannot be equal, their access to education and opportunity can be made more nearly equal. The rights of man are not rights to office and power, but the rights of entry into every avenue that may nourish and test a man’s fitness for office and power.

War

• War is one of the constants of history

• Cause of war is the same as cause of competition between individuals. The state has our instincts but not our restraints.

• States will unite in basic co-operation only when they are in common attacked from without.

• We may make contact with ambitious species on other planets or stars; soon thereafter there will be interplanetary war. Then, and only then, will we of this earth be one.

Growth & Decay

• History is littered with ruins of civilizations. Life has no inherent claim to eternity , whether in individuals or in states.

• History repeats itself, but only in outline & in the large.

• Civilizations don’t quite die, only their frame is gone & habitat changed & spread. Homer has more readers now than in his own day. Selective survival of creative minds is the most real and beneficial of immortalities

Is Progress Real?

• What does progress mean? If increase in happiness, then no.

• Define as increasing control of the environment by life. On a long-range view: table of mortality rates indicates their is progress.

• We idolize past civilizations but forget their high rates of infant mortality, short lifespan, great disease, etc.

• If education is the transmission of civilization, we are unquestionably progressing. Our finest contemporary achievement is our unprecedented expenditure of wealth and toil in the provision of higher education for all.

• If a man is fortunate he will, before he dies, gather up as much as he can of his civilized heritage and transmit it to his children. And to his final breath he will be grateful for this inexhaustible legacy, knowing that it is our nourishing mother and our lasting life.

If you want an even more concise summary, I encourage you to check out Ben McCarthy’s great tweestorm.

On Tuesday, I was honored to lead two roundtable discussions for the Influencer Series on the topic of Fulfilling the Promise of Industrial IoT. Thank you to Ravi Belani and Michelle Gonzalez for inviting me to speak, for Nima Badiey and Deb Noller for each moderating one of the discussions, and for USVP for hosting us.

I was asked to lead with a controversial point of view. Having recently read Homo Deus by Yuval Harari, and strongly ascribing to his description of humanity’s potential future, I labeled myself a humanist, then asked the room which jobs will be left for us humans?

In his book, Harari suggests that it would be significantly harder for AI to replace a Paleolithic hunter-gatherer than it would to replace a modern human. The reason being that the hunter-gatherer needed to do many things well: hunt, gather nuts & seeds, know which plants were edible, tan hides, build structures, etc… The modern human has professionalized themselves to do one thing very well: manipulate Excel spreadsheets or write Go backend code — while the rest of her needs are fulfilled by other deeply professionalized humans. Doing many things well is hard for AI, whereas doing one thing really well is a machine’s home turf.

One might suggest that we are no closer to creating conscious machines, but Harari suggests this is no consolation. Though intelligence & consciousness have historically been strongly intertwined, this is no longer the case: intelligence is decoupling from consciousness. For example, a Waymo self-driving car does not carry out thoughtful conversations with its passengers, contemplate the beauty of the starlit sky, nor get frustrated at the driver who cut it off in traffic — but it is fast becoming a better driver than you & I. It turns out that, given the choice between intelligence and consciousness, society— economically, militarily, politically, etc… — picks intelligence over consciousness every time. What will happen when machines are more intelligent than us (even if less conscious)?

I suggested that the room — filled with IIOT experts from big companies, VCs, startups, and thought leaders — might think the first jobs to go would be the industrial jobs. After all, the rage — and thesis for many VCs and entrepreneurs— is to usher the lights-out factory through a combination of IIOT, robotics, and AI. I however believe many industrial jobs are safer than those of most knowledge workers. With some exceptions, industrial workers still have to navigate difficult terrain, manipulate heavy tools, and make on-the-fly “one-off” decisions that don’t rely on a wealth of data to pattern-match from. As such, they are more akin to the hunter-gatherer who needed to be proficient at many things. Knowledge workers on the other hand have largely become data processing engines for emails, Slack messages, spreadsheets, radiology reports, etc… You may not even need to build robots to replace knowledge workers because they rarely move around to do work — computers will suffice.

Having framed my argument, I opened up the conversation to the broader group by asking:

• What jobs do you think machines will do “last”?

• What will happen to un-augmented humans in the job market & how do we address this on a societal scale?

Both roundtable conversations were heated, touching on topics ranging from:

• Will machines be able to emulate and/or replace human intuition and emotions — and does it matter if they do or don’t?

• Can we teach machines to treat humans nicely when the time comes, and how is our treatment of animals and pets a good or bad model for how super-intelligent beings might treat humans in the future?

• Are knowledge workers really more at risk?

• Are minds shaped by certain professions more or less “plastic” (eg. re-trainable) than others?

• Is Universal Basic Income (UBI) necessary, and will it be enough? In particular, UBI provides a means for survival, but does not provide a sense of purpose — can humans live without purpose?

• What about human happiness? Does it matter, especially if it can be manipulated chemically with prescription or illegal drugs?

Some of the professions suggested as being “last to go” included:

• Care/empathy givers — less about providing of the technical medical care itself, but more the delivery mechanism for empathy & bedside manner.

• Human organizers — whatever groups humanity congregates in, these groups will need to be organized & motivated, and humans may best be suited for this task.

• Archeologists, NGO members, etc. other jobs that care specifically about the human experience, vs. the economic or military outcomes.

• Humans in the service of training/improving machine learning systems — teachers, but for the AI.

The general consensus in the room was that super-intelligent systems are coming. For industrial & knowledge workers alike, one way to improve our chances in the future market for a purpose-driven life is to embrace that change and “merge” with machines to make ourselves hybrid super-intelligent beings.

While the future of this merger may sound like science fiction today — cybernetic limbs, neural lace, brain-enhancing chemicals, etc… — the first steps of this merger are already happening. Many of us in the room already consider ourselves cyborgs given how we use our smartphones as an extension of our brains. It’s been a long time since I’ve had to remember a birthday, a phone number, driving directions, my task or shopping list.

Mediated by today’s mobile devices, workers — both desk-less and desk-bound — can have access to a vast repository of knowledge and instructions. They have the means to expand their own cognitive abilities to continue to keep pace with the advancing tide of technology. Working in conjunction with smart data on a mobile device today may be the best step to prepare us for a future where hybrid human/machine intelligence is the norm. It may sound strange, but I think it is considerably more appealing than becoming economically and socially irrelevant.

Clearly this is a passion of mine. At Parsable, the company I co-founded, we are helping industrial workers globally do their jobs right every time by augmenting them (today) with a mobile collaboration & workflow platform they access via their mobile device. Their ability to collaborate in real-time and have immediate access to all of the best-practice standard work instructions and operating procedures gives them a leg up in an increasingly complex and technologically advanced world. As technology progresses, our software may end up getting delivered via wearable devices, or even by neural lace. Irrespective of delivery mechanism, the goal of making industrial workers safer & more productive, and thus help secure their place in the future economy, remains our north star.

I’m a humanist, and there’s hope for us humans.

I thought I’d learned what I needed to know about negotiation. I went to a prestigious business school and took their negotiation class, learning all about Getting to Yes, BATNA, and other fancy acronyms. I’d also bargained my fair share in both work and personal life.

Yet, I felt the tools I had at my disposal were meant for some alternate reality where people behave like dispassionate, rational robots: doing math in their heads to get to logical negotiation outcomes. The negotiations I’d been in were instead with passionate, irrational (including myself) humans: sometimes getting angry or sad, often making decisions that didn’t “make any sense” (to me). I was pretty sure the outcomes we were getting to were subpar, both for me and for them: a lot of splitting the difference, mostly to make the negotiations — which felt uncomfortable for all parties — stop.

Note, when I say negotiation, I’m speaking broadly: from negotiating with my fiancée on who should walk the dog tonight, to negotiating with an teammate on why this feature needs to be built urgently, to negotiating with a customer who’s called me angry about something, ... Each negotiation more emotional than the next, yet with a methodology that taught that emotions didn’t matter in negotiations. I was stumped.

Then, I discovered Never Split the Difference by Chris Voss. The book exposed me to a new way of negotiating. It replaced the toolkit meant for negotiating with rational robots with a toolkit meant for negotiating with humans. It built a foundation for negotiation on the basis of understanding the other party through empathy and active listening skills. It taught me that labeling — not ignoring — emotions is the key to a successful negotiated outcome. It exposed the value of open-ended calibrated questions, and forever banished yes/no-answer questions from my repertoire. It taught me polite ways to say no and mean it, without offending the other party. Most importantly, it brought a framework that lets me deeply learn (& yearn) to understand what the other party needs, wants, and desires — and work with them to achieve an outcome where I get my goals met, without ever splitting the difference again.

And it has worked wonders. Since reading this book, I have:

• Forged a better relationship with my fiancée by actively listening to her before jointly finding solutions

• Negotiated successful resolutions to emotionally charged topics with coworkers & friends

• Brought angry customers — who felt we had failed them — back from the brink to trusting us again

• Forged a better relationship with my business partners by understanding how they value time, silence, relationships, surprises, etc…

• Gotten discounts on things that I didn’t think could be discounted, just by using my name

• Gotten to the front of the waiting line at busy restaurants

• Said no to bad deals, because no deal is better than a bad one

• and the list goes on.

I warn you that this book is the start of a rabbit hole you might want to keep digging down. I’ve recommended it to anyone who will listen (cf. this blog post), personally bought it 29 times as a gift for friends & coworkers alike, taken an online class taught by the author’s son (a brilliant negotiator in his own right), etc…

Negotiation, in the broadest sense as described above, is something I want to become an expert in, because I now understand that every conversation is a negotiation. This is the most useful skill you can learn and apply.

It all started with this book. Are you too busy to read it?

Never Split the Difference: Negotiating As If Your Life Depended On It
Never Split the Difference: Negotiating As If Your Life Depended On It - Kindle edition by Chris Voss, Tahl Raz…www.amazon.com

It was a great honor to recently come back to my alma mater, Rice University, to give the opening keynote at the Rice Alliance for Technology & Entrepreneurship’s 14th annual Energy & Clean Technology Venture Forum.

My first experience with entrepreneurship dates back to a presentation I gave in May 2000 at the Rice Alliance — then their 6th venture forum— on a startup idea that two classmates and I had developed for the monitoring & control of industrial process flow lines. It was called EJP Technologies, though you’re unlikely to find it on the web.

After work at Microsoft and Google, and four startups later — not counting the ill-fated EJP Technologies — it was fun to come full circle and sit before the next batch of young entrepreneurs. In the 16 years since I graduated from Rice, its commitment to entrepreneurship has only increased, as has the quality of the presenters and startups.

The talk was not videotaped, but I’ve put a link to the slide deck here for those interested.

Thank you Rice Alliance for giving me this opportunity, and for admitting me into your ranks 20 years ago.

For many organizations — large/small, traditional/techy — culture is oft defined by a list of words describing shared values. Many include words you might expect, like customer obsessed (Microsoft), empathy (Slack), results-oriented (Salesforce). Others sound more unique like draw the owl (Twilio), you can make money without doing evil (Google), servant’s heart (Southwest Airlines)

At Parsable, we believe culture serves two purposes:

1. It allows teammates — especially as we grow — to have a shared understanding of what it means to be a good Parsable citizen. It’s our Constitution: how we should behave every day in the service of our mission.

2. It guides us to hire folks who share in these common values and screen out folks who are bad fits. It’s our Immigration Policy.

In that light, I found the words above which are meant to describe culture lacking, because:

1. How do I know when I close my eyes and imagine empathy that I’m imagining the same thing as my co-workers? How can we have a shared understanding of culture if it’s described by words for which we don’t have a shared understanding? We don’t want a vague Constitution.

2. How can I screen a candidate for draw the owl or customer obsessed? What questions do I ask that candidate? What does a good answer look like? A bad answer? We don’t want a vague Immigration Policy.

3. Where do these culture words come from in the first place? Are they a reflection of the values of the founders/execs, decided by a committee, a grab bag of words that sound good? We want guides that reflect the cultural aspirations of the company: the best selves we want to continue to become.

I was looking for a better, more authentic way to describe & implement our culture. A few months ago, Christina Kelly at First Round (FRC is one of our awesome set of investors) recommended I read Hiring for Attitude, by Mark Murphy. The book details a simple and powerful algorithm to discover, describe & enforce culture. I encourage you to read it, particularly if you’ve founded or are running a company.

Simple and powerful, but not easy: it took ~6 weeks and significant effort to discover & describe our culture. The book suggests hiring a consultancy to help, but as a founder, I thought it important to drive this journey myself.

The process looked like this:

• I interviewed 15 team members (we were about 40 at the time) for 1–2 hours each, sampled randomly across the entire company (Engineering, Product, Design, Sales, Customer Operations, Execs) and across our three offices (San Francisco, Austin, Vancouver). I asked them the following 4 questions:

1. Please list 3-to-5 attitudes or personality traits that you think describe the most successful people at our Company. When describing each trait, describe it in a such a way where two strangers could observe the behavior.

2. Please list 3-to-5 attitudes or personality traits that you think describe the least successful people at our Company. When describing each trait, describe it in a such a way where two strangers could observe the behavior.

3. Think of someone in the organization who truly represents our culture. Could you tell me about a time he or she did something that exemplifies having the right attitude? It could be something big or small, but it should be something that made an impression on you. Also tell me why it made an impression.

4. Without naming names, think of someone who works (or worked) in the organization who did not represent our culture. Could you tell me about a time this person did something that exemplifies having the wrong attitude? It could be something big or small, but it should be something that made an impression on you. Also tell me why it made an impression.

• I collected ~500 stories from these answers, and grouped them into themes which formed the basis of our cultural pillars (we discovered we have 7 of them). It’s important to note that my teammates did most of the work here: by generating the stories. Grouping them just involved finding similarities among the stories, both negative & positive. My dining room table looked like this for about a month.

  

• The stories above were also used to generate Word Pictures for each cultural pillar. You turn stories into Word Pictures by removing names and turning them into “When ___ , I ____ .” statements. The principle is that empathy can mean different things to different people, but when teammates read “When talking to someone, I ask open-ended questions and seek to understand their perspective before adding my inputs.” it’s pretty likely they will imagine the same scene = Word → Picture. To read more about the power of behavioral specificity and Word Pictures, I encourage you to read the book. It’s really worth it.

• Finally, the themes above were used to generate interview questions for each cultural pillar. These interview questions have a very specific format, also described in the book.

• Interview questions without an answer guide — to discern good vs. bad answers — are not very useful. To create our Answer Guide, I surveyed the entire company with these same questions (10 of them in total) and collected hundreds of answers.

• I mapped these answers into Positive Signals (congruence with the cultural pillar) and Warning Signs (sign of lack of congruence). These real answers serve as our Answer Guide when comparing answers given by our interview candidates to the same questions.

• Because we love Star Wars at Parsable, we ascribed a different Star Wars character as the “patron totem” for each cultural pillar.

During this process, two things of note happened:

1. Teammates opened up and told me amazing stories of their peers going above & beyond: coming through for a teammate in time of need, making personal sacrifices for the good of the team & company, etc. Often, these were silent heroes, so I hadn’t known of many of these feats until I performed the interviews. That alone made running this process myself worth it. We recognized these Parsable culture heroes with LEGO Star Wars characters for the specific cultural value they exemplified with their actions.

2. Some of our answers to our own interview questions (mine too) ended up in the Warning Signs section for some cultural pillar. A team & company is a work in constant progress, and this exemplifies the aspirational nature of culture: we are not all good at all things. Rather than get demoralized, the team has rallied and seen this as an opportunity to improve. It helps validate why seeing opportunity in difficult situations is one of our cultural pillars.

This is only a milestone in Parsable’s long journey. If you’re curious to live & shape our culture, consider joining us. As reward for introspective and action-oriented work, many LEGO Star Wars characters may be in your future.