The AI Code Slop: Risk or Opportunity?
In a Nutshell
AI-generated code "slop" risks production fragility and unsolved attention management in engineering, but massive software demand outstrips supply amid exploding AI lab revenues (e.g., $1B-$10B in ~1 year vs. historical 3-20+ years) and 100x token cost drops, signaling huge opportunity as tech claims 15-30% of GDP by 2035. SaaS "apocalypse" hype overstates threats to durable enterprise vendors like Samsara, while small technical teams project bespoke vibe-coding onto non-technical giants unwilling to build internally. Founders must adapt fast in this internet-like era of leadership flips: build defensive bundles across workflows, preschedule exit discussions, and rethink durability amid 1-2 year cycles compressing decades of progress.
These notes were generated by AI and may contain inaccuracies.
Anxiety arises from generating enormous amounts of code without anyone reading it, unknown quality, lack of deep codebase understanding, and increased fragility. This is the slop problem but in production codebases: vibe coding slop for lazy engineers. Broader problem is managing human attention to engineering. New companies could solve ticketing systems at risk, with ideas like testing, smart review, agents, formal verification. Open season around this big problem.
Markets melting down about end of software. Discussing if SaaS is dying or projecting five-person startup behavior onto Fortune 100. Covering real trends: incredible revenue growth, collapsing token costs, faster vendor turnover. Hype vs. opportunity. Changing bottlenecks in building software companies, parallels to internet and cloud eras.
Meta trends people get right, specific companies wrong. Premise: SaaS and proprietary software replaced by AI, vibe code internally, no need to pay for Salesforce. Shortsighted in near term. Durable companies like Samsara (fleet management app, distributed via enterprise sales, in-cab camera sensors, support via agents) won't be vibe coded. Massive market correction with long-run truth, short-run for certain companies.
Sierra and similar shifting from per-seat software to utilization-based customer support agents. Real shift impacting prior per-seat SaaS, not every SaaS company. Short-term overstated; long run unknown.
Portfolio companies with hundreds of millions revenue committed to max token usage, few great people (<50 engineers), scaled to ~100 salespeople. Growing AI natives view: vibe sales not happening.
Fleet managers won't write own apps for giant surface area. Strong market reaction overstated, especially for durable companies.
Assumptions by engineers/builders that everyone wants to make own software. Not everyone does; some want others to do it.
Engineers' labor-focused view: should you build Jira in most orgs? Not best use if focused on product.
Five-person startups vibe coding CRM: before that, used spreadsheets. Fine for limited niche, technical teams, custom/bespoke. Doesn't mean Fortune 100 displaces CRM with internal weekend hack.
Projecting small technical startup behavior onto big enterprises. Internal software building impressive, but aggressive extrapolation early doesn't make sense.
Engineer pissed at $10/seat Jira won't want change management at Bank of America: security, opinions on workflows, maintenance.
Production of code not bottleneck if spec known, interesting but overstates software vendor problem share.
Huge demand for software products (AI included). Software eating the world; AI eating the world. Mark's post thoughtful.
Little engineering supply relative to demand. AI boosts productivity gets sucked up; teams hire engineers. Nature of work shifting.
Engineer mindsets: bespoke craftsmanship (aesthetics, code quality, artisanal); utility for product building; math aspects. Subset less happy in new world, like indie game devs vs. EA workers.
Great bespoke engineers unhappy at larger companies as coding tools accelerate, against their approach. Utility-focused freeing.
Variance in reactions based on utility function for work.
Engineering identity based on difficulty/skill ranking: impressive types less hard for agents. Enjoyability and identity elements.
Applied Intuition founder essay: keep your identity small – wonderful advice for adaptability.
Abundant software generation addresses unsolved problems.
Abundant code generation: code quality unsolved. No one reading generated code leads to unknown quality, no deep understanding, fragility – slop in production.
Ticketing at risk. Broader: managing human attention to engineering unsolved. Ideas: testing, smart review, agents, formal verification.
Agents not making big vendor decisions yet; overstated (partnerships like Cognition/Claude spinning Supabase/AWS instances, always happened).
Long run agents true for agenic commerce, persona understanding.
Noisy moment of overstatement despite pro-AI progress belief: SaaS apocalypse, extrapolating future as here (e.g., BD deals). Multi-book stuff human-generated. Hype month with planted marketing (agents cutting humans from Reddit-like forums).
Exciting behavior, subset marketing.
Things tapping deep emotional human-like reactions for marketing (Mold book).
Demos differ from full software reality; equity research misses structural advantages, competitive moats, distribution battles, code production.
Loosening bottleneck on expensive software production cool: more ideas expressed (embedded views in software like engineering, enterprise sales, productivity e.g. Notion).
Revolution; involved early in coding companies, excited, transformational.
Ignore hype noise. Signal: growth pace of biggest companies underpriced.
Jared's Capital IQ data: years to grow from $1B to $10B revenue.
- ADP: 20+ years
- Adobe: ~20 years
- Salesforce/SAP: 8-9 years
- Microsoft: 7-8 years
- Google/Meta/AWS: 3-5 years
- AI labs (OpenAI/Anthropic): ~1 year
Projections $10B to $100B:
- Microsoft: 27 years
- Google/AWS/Meta: 10+ years
- AI labs: 3-5 years
Fastest time to massive revenue in software history. Internet enables global liquidity, easier distribution, higher GDP. Capabilities create enormous value at scale.
Token pricing collapse for equivalent models.
- GPT-4 level: 21 months, $37 to $0.25/million tokens (150x drop).
- o1 equivalent: Dec 2024 $26 to Nov 2025 $0.30/million tokens (88x drop in 11 months).
Pricing collapse + insane revenue ramp.
Inference: clouds (Base10, Mobile, Fireworks) or large providers; 1000x consumption growth, getting efficient. Revenue grows slower but wild.
Human brain: 12-20 watts (dim light bulb, sleeping monitor). One GPU fan blade.
Efficiency room vs. human brain.
Incumbents committed to vendors, have money to fight back – but reflexivity: if labs/app companies grow to $1B run rate fast, valuations rise, incumbents lack currency to compete.
SF housing rising from lab tenders, employee sales; Nvidia from $100B to trillions.
Tech % GDP US, market caps:
- 2005: Google $100B, Exxon world's most valuable at $100B.
- 2018: Apple first $1T company; tech ~30% S&P.
Top eight tech companies had about 10% of S&P market cap in 2005. Now they have $23 trillion market cap, over 50% of S&P value. They went from 4% of GDP in 2005 to 12% today. Question is what proportion of GDP becomes tech. AI drives this by augmenting services and jobs, converting them to software/tech spend. Assumptions on growth rates lead to tech at 15-20% of GDP or up to 30% by 2035. This implies even bigger market caps for tech companies, a metric for terminal values 10 years out and basis for growth.
If everything gets bigger, $1 billion valuation is no longer late stage. Even now, companies raise at $1B with $2M revenue. Runway for foundational companies is much larger than conventional wisdom. 10-15 years ago, sustainable $5B market cap was hard; companies hit it once every couple years. Biggest market caps were low hundreds of billions. Last 15 years, everything grew 10x: trillion-dollar caps, more $100B companies.
Question: how many more trillion-dollar companies can be supported (2, 3, dozen, 50)? If everything pulls up, how to invest over company lifetime or as founder think about end state? Startup fail rate: could go up as value aggregates into platforms (historical pattern). Platforms forward integrate into key verticals.
- Microsoft integrated Office suite (Excel, PowerPoint, Word) into OS, killed/bought competitors.
- Google forward integrated into vertical searches (travel, local).
Labs will forward integrate into interesting apps on top (already seeing with code). Implications for startups: which verticals durable vs. eaten by labs?
Counter: broader tail effect with more valuable startups due to more market cap, global liquidity from internet. Tail dominates because surface area for technology expands rapidly. But market cap is power law: head and torso aggregate most value (true for Google ad revenue too, despite long tail claims). YC value concentrated in few companies (e.g., 80% in 5). Won't change in AI era. More $100B businesses as surface area grows, head gets even bigger.
Time to market leadership and revenue scale compressing. Handful of companies hit zero to $100M+ run rate faster than SaaS 10 years ago. Valuations grow with that. Some durable, but leadership can flip (e.g., Anthropic, OpenAI). Risk vs. 10 years ago when category leadership at scale felt unassailable.
Internet wave: 1999-2000, ~450 companies public each year, total 1-2,000. Dozen to two dozen still relevant; rest died or bought. 10 years later, assumed unassailables flipped:
- Social: Friendster, MySpace -> Facebook.
- Payments: Braintree, PayPal existed; Stripe won (payments is fragmented oligopoly).
For founders: assess business durability and exit timing. Most companies have 12-month peak value window, then crash. Small handful: never sell. Most: sell at right time.
Tip: Preschedule board meeting 1-2x/year to discuss exits non-emotionally (Horus's advice from Opsware). "Nope, still not time" or "Competitive dynamic shifted, offer beats next 5 years—now's time."
Internet era had dominants that peaked then exited:
- AOL exited to Time Warner at peak.
- Yahoo web's front page.
- Netscape browser, then Internet Explorer.
- eBay market.
SaaS era felt more protected at scale, incremental growth. AI more like internet: does growth compound to control or rethink exits?
80s example: Lotus 1-2-3 first spreadsheet, explosive growth to hundreds of millions revenue. Collapsed into IBM as Microsoft Excel took market. Looked durable killer app, but died.
Cracks appearing now: as labs shift, usage changes, differentiation/defensibility shifts, ask if next 6 months is peak value. Preschedule board meetings for logical discussion.
Mark Cuban sold early product to Yahoo for billions at peak, collared stock to protect as it dropped—one of best financial engineering moves, made him billionaire.
Internet flips from new distribution, performance, interfaces, user behavior, exploration. SaaS was replacement market, niches, cheap distribution. AI: next lab capability jump could reset leaderboard. Surface area shifts (agents vs. IDE, voice) reallocate power.
Defense: Build bundle—multi-product surface area, cross-sell into orgs, become default workflow across 5-10 aspects of vertical. Bundles defensive (also offensive). Avoid singular thing easy to clone/displace.
SaaS "do one thing well" led to being bought, but bad advice. Pre-SaaS (platform wave) companies multi-product/acquisitive. SaaS velocity slow (build out internet). AI velocity high: decade-long cycles now 1-2 years due to scaling laws, reasoning, post-training.
Founders: mindset shift—every 2 years = 10 years, react quickly to changes.
Defenses not dependent on incremental feature growth: platforms, ecosystems, networks, bundles, hardware (e.g., Samsara). Don't over-rotate on last era; be intellectually honest about market position, speed of change, true control points. Lots coming, lots shifting.
Keep No Priors: AI, Machine Learning, Tech, & Startups in your library
Save the videos and channels worth coming back to, and find them again in one place.
