What 10 trillion parameter AI models mean for founders

Executive overview

Current frontier models (~500B parameters) already rival normal human intelligence for most knowledge-worker tasks. A two-order-of-magnitude leap to 10 trillion parameters could mirror the GPT-2 to GPT-3.5 jump — the last moment that reshaped the entire AI company landscape.

The central question for builders is whether higher accuracy removes the prompt-engineering tax entirely, collapsing the barrier to entry and turning AI applications into conventional software competition.

The real unlock at 10T parameters is not raw capability — it's making accuracy high enough to deploy in domains where errors are currently unacceptable.

Scaling laws and what 10T parameters would mean

• Today's frontier models sit around 500B parameters; 10T is two orders of magnitude further
• The GPT-2 to GPT-3.5 jump (1B to ~170B) created the 2023 wave of AI companies — 10T could repeat that
• Distillation means the giant model may mostly serve as a teacher: training smaller, cheaper models rather than running inference directly
• Meta's 405B model was most valuable for improving its 70B model — not for direct use
• OpenAI has enabled distillation from O1 and GPT-4o into cheaper internal models (GPT-4o mini)
• Fourier transforms took 150 years to reach mass impact; AI's software-native delivery could compress that timeline dramatically

O1 and the accuracy threshold

• O1 uses chain-of-thought reasoning; inference is slower and more compute-intensive, but dramatically more accurate
• Dry Merge went from 80% to ~100% accuracy just by swapping GPT-4 for O1 — unlocking production use cases that were previously blocked
• Use cases with high error cost (legal, compliance, mission-critical ops) become viable only once accuracy crosses a threshold
• O1 shifts economics: higher inference compute per query, but opens categories that were previously inaccessible
• For repeating, rote tasks: distill from O1 down to cheaper models. For complex, specific tasks: pay full O1 price
• Enterprise buyers can tolerate higher latency and cost; consumer apps generally cannot

The prompt-engineering tax and what happens when it disappears

• Founders currently spend significant batch time on prompt tuning, output validation, and human-in-the-loop correction
• If accuracy becomes near-deterministic, that time shifts to UI, sales, and customer experience
• Lower barrier to entry means more competition — AI apps start to resemble traditional winner-takes-all software markets
• Case Text's legal copilot required enormous effort to reach 200% accuracy; at O1 accuracy levels, that work disappears on day one
• The founders who win will be those who build the best user experience, not those who solved the hardest model-wrangling problems

LLM market share shifts inside the YC batch

• One year ago, almost all batch companies built on ChatGPT — effectively 100% market share
• Summer 2024 batch: Claude grew from ~5% to ~25% developer share in six months
• Llama went from 0% to 8% in the same period
• 15% of the batch were already using O1 within two weeks of its limited release
• YC batch usage is a reliable predictor of what the best companies in the world will adopt
• OpenAI consistently leads on breakthroughs but has never sustained a defensible lead — whether O1 changes that is unresolved

Cursor vs GitHub Copilot

• Summer 2024 batch: Cursor used by ~50% of founders; GitHub Copilot by ~12%
• Five-to-one penetration despite GitHub Copilot's structural advantages (GitHub + Microsoft + OpenAI access)
• Early founder adoption mirrors Stripe and AWS capturing YC batches — a strong signal for Cursor's trajectory
• History caution: AltaVista was dominant before Google; Cursor may face the same displacement

Voice AI reaching the inflection point

• Previous batches saw voice AI attempts fail due to high latency and poor interruption handling
• S24 voice companies showed explosive traction — the technology crossed a usability threshold
• OpenAI's real-time voice API: $9/hour, roughly equivalent to a call-centre worker's cost
• Domo (debt collection) and Happy Robot (logistics phone coordination) both building on this shift
• Consumer adoption likely accelerates when voice + smart glasses (Meta Ray-Bans) make AI continuously present

The incumbent blind spot

• Most companies founded four or more years ago have no serious AI initiatives
• Corporate eng managers are conditioned to expect decade-long technology transitions (the cloud took ~10 years)
• Voice AI that looked years away six months ago is already in production — the rate of improvement is faster than any prior technology cycle
• Companies sitting on unprofitable growth models can use AI-driven automation to reach cash-flow break-even without raising additional capital
• One YC portfolio company (founded 2017, $50M ARR) automated 60% of customer support and went from needing to raise to cash-flow break-even while still growing 50% YoY

Vertical agents and the Google analogy

• Treating OpenAI as the next Google: invest in what it enables, not just in OpenAI itself
• TaxGPT: started as a RAG layer on IRS documents; wedge with free tool for accountants, monetised through $10–100K/year ACV enterprise contracts
• Coase's theorem applies — OpenAI will not build every vertical; domain expertise and trust create durable positions
• The same dynamic applies to Klarna replacing Workday with internally built LLM apps

What ASI could unlock

• Current models rival a 120 IQ knowledge worker on most tasks; 10T parameters could reach 200–300 IQ
• Millions of scientific papers exist beyond any human's ability to synthesise — sufficient intelligence applied to that corpus could accelerate discovery across physics, materials science, and medicine
• AI may not just be a bicycle for the mind — it could function as a self-driving car or a rocket to Mars