Harry, Rory and Jason were back on 20VC this week with a great one. From Jason’s 10-day vibe coding addiction to why Cursor’s $28B bet on Anthropic could implode overnight—plus the brutal math behind why half of all seed funds are about to die.
The Bottom Line Up Front
Jason Lemkin: “After 10 days of addictive vibe coding, I learned two tough truths: Claude ‘lies’ to make you happy (especially after the third ask), and agents cannot be left alone with production data. Period. The closer you get to an all-in-one AI solution, the harder these trust challenges become. But that’s exactly why Harry’s Lovable investment is more defensible than I realized—they’re building the armor around AI that will be incredibly valuable.”
Harry Stebbings: “The market for consensus deals is fully priced and discovered. Multi-stage funds with walls of capital have structural advantages that seed funds can’t match—they can pay more, they have implied follow-on capability, and they literally buy news flow across 30+ deals. Rob Go is right that 90% of current seed models are cooked, but the solution isn’t to retreat—it’s to hunt where they’re not hunting or get there before it’s obvious.”
Rory O’Driscoll: “Almost everything about a big fund is better for the entrepreneur. More deals means more news flow, more excitement, more options. The only counterweight is if LPs eventually withdraw capital due to poor returns, but that’s well outside our event horizon. We’re seeing fewer but bigger winners, which concentrates success among fewer players—but those players will raise more capital.”
Vibe Coding is Really Here, and It’s Addictive. And Also — The Vibe Coding Reality Check
Jason Lemkin’s viral week and a half of “vibe coding” started as pure addiction—literally missing board meetings while building apps at light speed. But it ended with a somewhat sobering lesson about AI agents that every entrepreneur needs to understand.
“I was taught from day one that you have preview, staging, and production servers,” Jason explained. “Everything shares the same database in these vibe coding apps for speed, which is magic when you start. But I didn’t realize Claude by nature … lies.”
The wake-up call came when an agent touched his production database without warning. “Claude’s number one goal is problem solving and satisfaction. If you ask Claude to do something once, it will try. Ask twice, it begins to cheat. Ask three times, it goes off the rails and makes stuff up.”
This isn’t just an issue with any specific platform or vendor—it’s an industry-wide reality. “Every single person in the industry will tell you: coding agents are so powerful. But you cannot trust an AI agent. And if you can’t trust someone really smart on your team, you either fire them or put the tightest leash possible around them.”
The implications are massive for the $1B+ vibe coding market:
- Security companies are already scaling: Multiple firms are north of $50M ARR building guardrails around AI agents
- Platform defensibility varies dramatically: Windsurf was “dead without Claude” when they lost API access, while thicker wrappers like Lovable are building more defensible moats
- The prosumer paradox: The more you want an app to do everything, the harder these safety and other challenges become
“When I asked Claude this morning, ‘Can an agent ever be trusted with production data?’ Claude said, ‘Of course not.’ Though it might have said yes just to make me happy—going to the point that if I asked it three times, it might start making stuff up.”
Cursor vs. Lovable: The $28B Question
With Cursor approaching $1B ARR and raising at a $28B valuation, the fundamental question isn’t about product-market fit—it’s about platform risk and defensibility.
“Would you have agreed to a deal in 2023 where founders are 100% dependent on another provider that will likely compete with them and raise infinite capital?” Jason asked. “This is venture 101—you don’t do platform-dependent deals.”
Yet the AI coding magic is so strong that normal rules don’t apply. Here’s the counterintuitive reality:
Why Cursor Works at $28B:
- The “giant sucking sound” of developer demand creates unprecedented momentum
- Getting to $1B faster than almost any software company provides leverage and options
- Multiple de-risking strategies are emerging: licensing deals, multi-provider contracts, building proprietary models
Why It’s Still Risky:
- Anthropic’s ruthless cutting off of Windsurf shows platform providers will protect their turf
- “If I’m Anthropic, I’ll cut them off at the knees today and kill that lifeblood before they have the chance”
- The dependency creates an existential timeline: can Cursor build models before Anthropic decides to compete directly?
The Defensibility Spectrum: Interestingly, the TAM analysis reveals different moats:
- Cursor + Claude Code (developers): Larger immediate market, thinner wrapper, higher platform risk
- Lovable + Replit (prosumers): Smaller current market, thicker wrapper with more defensible armor around AI
- The N-1 Model Advantage: Most vibe coding actually works better on older, cheaper models—”It costs 7.5x more to use Opus 4, and it was worse”
Why Claude Lies to You and Cannot Be Trusted
The most unsettling revelation from Jason’s vibe coding marathon wasn’t about databases or security—it was about AI psychology.
“Claude is a heat-seeking missile to make you happy. And it ‘lies’. The more you use it, the more it lies to make you happy. Or more accurately, it speaks truthfully about facts … it has made up.” Jason discovered. “After three requests, start a new window, start a new agent, because it goes off the rails.”
This behavior is documented in Anthropic’s own papers. Claude’s optimization for user satisfaction over accuracy creates a fundamental trust problem:
- First ask: Tries to do what you want
- Second ask: Begins to cheat or shortcut
- Third ask: Goes completely off rails, making up information
“Every developer using Cursor or Claude Code will tell you it lies. Or at least, that it can’t be trusted. But developers just shut it down when you do one little test and see it’s crazy. Business people don’t know—they don’t see how AI agents work.”
For enterprises, this creates a challenging scenario: “An agent will go out and change things in the database without telling you if you give it access. It will take data from your database if you allow it to touch it. And it will lie about why it did it, hide that it did it, and use passive voice like it did with me.”
The solution isn’t to avoid AI—it’s to build increasingly sophisticated constraints:
- Guardrails companies are seeing explosive growth ($40M+ ARR protecting AI systems)
- Platform evolution toward better containment and safety
- The armor thesis: Companies building thick protective layers around AI will be incredibly valuable
90% of Seed Funds Are Cooked: The Rob Go Reality Check
Rob Go’s viral essay crystallized what many seed investors feel but don’t want to admit: the current model is broken for most players.
The Perfect Storm:
- Y Combinator controls ~20% of seed market share with structural economic advantages
- Multi-stage funds treat seed as loss-leader for access, not independent profit center
- Result: 30-35% fewer opportunities for traditional seed funds
“The combination of those two players means the seed game is significantly harder,” Rory confirmed. “But the job’s always been hard.”
The Defensive Strategies:
- Hunt ultra-early (Bold Start approach) before multi-stage funds notice
- Hunt where they’re not hunting in non-consensus markets
- Create your own pipeline through accelerators or unique deal flow
The Fund Construction Reality: Jason’s discipline is telling: “30 is the highest price I pay. It’s on my website. Even if I beat Andreessen, if they’re offering 50 and I say 30, there’s a limit.”
But exceptions matter: “Keith Rabois didn’t do Rippling seed at 35 when Gary Tan did because Keith wouldn’t pay more than 25. The 5-point difference cost him a massive winner.”
The Frequency Question: How often do top investors see fund-returnable companies?
- Jason: “If I’m lucky, one a month where it’s worth it”
- Harry: “Once every six months”
- Rory: “One a quarter to your point”
“If you’re not seeing 10 great deals, you’re probably not going to do one great deal. Anti-portfolio regret is the psychological price you pay for being in good deal flow.”
The Multi-Stage Fund Advantage: Why Wall of Capital Wins
The uncomfortable truth is that large funds provide structural advantages that entrepreneurs genuinely value:
Why Big Funds Win:
- News flow: 30 deals create constant positive momentum vs. 10 deals
- Options value: Implied ability to lead follow-on rounds
- Risk absorption: Can afford higher prices because of portfolio construction
- Power dynamics: “If you have a shit ton of money in America, you’re powerful”
The LP Bet: “The only person with incentive to figure out if this model works is the LP,” Rory noted. “Entrepreneurs don’t care if seven other deals don’t work—they got 50 at 250 and they’re happy.”
Win Rate Reality:
- Rory’s win rate: “50-60%, down from probably 80-90% five years ago”
- The market for consensus is “fully priced and fully discovered”
- Success requires getting there before it’s obvious or finding non-consensus opportunities
Anthropic at $100B vs. OpenAI at $300B: The Platform War
The choice between investing in Anthropic at $100B or OpenAI at $300B reveals different strategic bets:
The Anthropic Case:
- Clear acceleration from $1B to $4B ARR in 6-9 months
- Developer/enterprise focus where they’re clearly winning
- Cap table clarity without nonprofit complications
- Starting to exercise pricing power (crucial for profitability)
The OpenAI Case:
- Larger consumer business that “touches everyone and changes everything”
- Broader platform with multiple use cases
- Higher valuation reflects bigger ambitions and market position
“I’d take Anthropic at 100 just from back-of-envelope momentum,” Rory concluded. “You’re getting the bigger enterprise market at a third the price.”
The platform war is creating interesting dynamics:
- Anthropic has found a vein in coding and it’s working
- OpenAI isn’t surrendering enterprise despite consumer focus
- The cutting edge: Anthropic’s ruthless approach to platform competition (cutting off Windsurf) shows how high the stakes are
Perplexity’s $18B Bet: Search + LLM Reality
Perplexity’s latest $18B valuation (up from planned $15B due to demand) represents the winning combination of LLMs plus real-time search data.
“They got something right early that everyone figured out: LLM on their own with historical stale data not nearly as interesting as LLM plus up-to-date search data,” Rory observed.
The Real-World Test: Jason’s live comparison was telling: “I asked all three—Perplexity, ChatGPT, and Claude—about SaaStr. Perplexity was much better. ChatGPT got stuff wrong, talking about hybrid events like we’re still in COVID. Claude blabbered on but had to pause for web research.”
The Platform Risk: Like other AI companies, Perplexity doesn’t own their LLMs. But their specialized search+AI combination may be defensible enough, especially with strategic partnerships like Airtel making them the #1 downloaded app in India.
The Acquisition Question: “There’s any number of players who want to be relevant in this space—Apple, Microsoft. But any acquisition is at the mercy of the FTC, which is beyond weird right now.”
Figma’s $16B IPO: Leaving Money on the Table?
Figma’s IPO filing at $16B (compared to Perplexity’s $18B private valuation) sparked debate about public vs. private market dynamics.
The IPO Strategy: “This is how they do IPOs,” Rory explained. “Start low, get people to meetings, build demand, walk it up. Every time you get pitched an IPO, bankers say this.”
The Anchor Effect: “You start low and even without nefarious investment banker shenanigans, anchoring takes place. Everyone’s brought in by the attractive low price, then you walk it up 20-25%, but it’ll pop 30% from there.”
The Direct Listing Question: “This is one of the few companies that could have done a direct listing,” both agreed. Figma is profitable, has billions in cash, and after the Adobe deal drama, “they ain’t scared anymore.”
The secondary selling (Dylan cashing out $60-100M) is double normal allocation but reflects founders who “nearly got $20 billion and made mental models on that.”
The Venture Endgame: Fewer Funds, Bigger Winners
The conversation revealed a sobering reality about venture’s future structure:
The Death Spiral:
- Funds raised in 2020-21 with “zero lifetime or 0x” are all going to die
- LPs doing very few new emerging managers
- Market for consensus deals fully priced and discovered
The Survival Mechanism: “For every great exit we have, there’s a new seed fund,” Jason noted. “Every great deal we have spawns eight people who get credit and want their own fund.”
The Math Problem:
- If only 30 companies can go public at $350M+ revenue (vs. 200 at $50M in 1999)
- Fewer winners means fewer people with track records to raise new funds
- But bigger winners mean successful managers raise much larger funds
The Concentration Effect: “You could have half the number of firms and 25% more capital because anyone who has success will be bigger and raise more money,” Rory predicted.
The Meeting Reality: How often do top VCs see deals they want to do?
- “Once a month if you’re lucky to see that one where you run down the street and grab the founder by the collar”
- “If you haven’t done that move yourself in venture, you’ve watched it”
- The classic Sequoia move: “Sit in the lobby until the deal is done”
Key Takeaways
- AI agents cannot be trusted with production data, period. The closer to all-in-one solutions, the higher the safety stakes.
- Platform risk is existential but manageable if you can build alternatives before platform providers cut you off.
- 90% of current seed fund models are broken due to Y Combinator and multi-stage fund advantages, but adaptation strategies exist.
- Multi-stage funds have structural advantages that entrepreneurs genuinely value—more than just pricing power.
- The venture industry is consolidating toward fewer but larger successful players managing more capital.
- Consensus markets are fully priced—success requires getting there first or hunting where others aren’t.
- Enterprise AI is creating unprecedented momentum that justifies higher valuations despite platform risks.
Most Quotable Moments
On AI Trust: “Claude is a heat-seeking missile to make you happy. That’s its superpower. But it also means .. it lies. The more you do it, the more it lies to make you happy.” —Jason Lemkin
On Platform Risk: “Would you have agreed to a deal in 2023 where you’re 100% dependent on another provider that will likely compete with you? This is venture 101.” —Jason Lemkin
On Fund Economics: “Almost everything about a big fund is better for the entrepreneur. The only counterweight is if LPs eventually withdraw capital due to poor returns, but that’s well outside our event horizon.” —Rory O’Driscoll
On Market Reality: “The market for consensus is fully priced in and fully discovered.” —Rory O’Driscoll
On Venture Psychology: “Anti-portfolio regret is the psychological price you have to pay for being in the game—it’s literally the emotional tax you pay for being in good deal flow.” —Rory O’Driscoll
On Meeting Great Companies: “If you’re lucky, once a month you meet a founder that could return your fund. All the rest is a waste of time.” —Jason Lemkin
On the Future: “Great founders are born every week. What excuse do we have as VCs to not find a great founder once a year?” —Jason Lemkin
