How 2024 proved the GPT wrapper myth wrong and broke in favour of startups

Executive overview

Two years after ChatGPT launched, the consensus was that OpenAI would capture all the value and every startup building on top of LLMs was just a disposable wrapper. That turned out to be completely wrong. The real winners of 2024 were application-layer startups powered by open-source models, multi-model architectures, and agentic AI — not the foundation model companies.

Open source — driven by Meta's Llama — ended the monopoly pricing narrative. Once the best benchmark model was open source, the moat shifted from model access to product, distribution, and execution. YC batches in 2024 hit 10% week-over-week growth on average — something that had previously only been achievable by the top quartile of companies.

The GPT wrapper myth, debunked

• The ChatGPT store launched and went nowhere; the breakout apps — Perplexity, Glean, Harvey, Photoroom — were all built by startups, not OpenAI.
• Enterprise pilots from 2023 converted into real revenue in 2024; YC companies are hitting $1M ARR faster than ever seen.
• The idea that only foundation model companies with massive capital could compete proved false; Opus Clip reached tens of millions in revenue without raising a Series A.
• The number of companies with a real shot at $100M ARR has grown 10x per decade — from ~15 per year to ~1,500 per year today.

Open source changes the competitive landscape

• Meta's Llama weights leaked, then Meta formalised open source releases — initially dismissed as 18 months behind OpenAI.
• By summer 2024, a Llama-based model topped all major benchmark rankings for the first time, shocking the community.
• Once multiple competitive models exist, monopoly pricing collapses; model quality becomes table stakes and product execution dominates.
• Model routers evolved into model orchestration: companies now use fast, cheap models for high-volume parsing and powerful models (e.g. O1) for complex reasoning — seen in Cursor, Camfer, and fraud detection workflows.
• Post-training on top of open-source base models (e.g. Variant teaching aesthetics for SVG icon generation) is emerging as a way to build durable IP without competing on foundation model training.

The rise of agentic AI

• "Agentic" barely registered in 2024's discourse a year prior; it became the dominant framing by year-end.
• The shift: from chat-like interactions to AI systems that complete multi-step tasks, call external tools, and operate autonomously.
• Agent reliability improved through techniques around structured outputs, tool use, and evaluation infrastructure — moving from demo to production at scale.
• Computer use from Claude opened new categories: startups are building agents that take over computer workflows end-to-end.
• YC fall batch companies are deploying agents handling thousands of support tickets per day at large enterprises.

Voice AI: vertical, not winner-take-all

• Voice AI is one of the highest-traction verticals in recent YC batches.
• Customer support alone is not one market — airline, bank, and B2B SaaS each require entirely different workflow logic.
• The structure mirrors horizontal infrastructure vs. vertical applications: there will be Stripe-like voice infrastructure companies and hundreds of valuable vertical voice apps.
• Language learning, teleconferencing, and customer support are among the earliest areas with multiple strong contenders.

AI coding breaks out

• The majority of YC founders now use Cursor or similar AI IDEs — this tipped decisively in summer 2024.
• Devin demonstrated full automation of large programming tasks; Replit popularised agent-driven coding for non-technical users; Anthropic Artifacts enabled PMs to prototype working UIs.
• Standard programming interviews broke down: AI coding tools ace whiteboard-style questions, pushing companies toward output-based evaluation (build a real thing in two hours, using the tools).
• Hiring patterns are shifting: some post-seed founders are delaying specialist hires in favour of engineers who are fluent with the full AI coding stack and can evaluate model output quality.
• Amazon completed a migration of hundreds of thousands of lines of code — a six-month engineering project — in weeks using LLM agents, flagging what internal enterprise tooling could look like when released externally (cf. AWS).

Scale AI: the classic pivot story

• Scale AI applied to YC with a healthcare appointment-booking idea and pivoted during the batch.
• The data-labeling insight came from Alex's experience at Quora with Amazon Mechanical Turk — he knew it was hard to use and built the version he would have wanted.
• Early traction came almost entirely from Cruise, labeling images for autonomous driving.
• Scale caught two distinct waves: the computer vision / self-driving wave in ML, then the RLHF wave from LLMs.
• Now valued as a multi-billion-dollar business and raising $1B — and still following the pattern of young technical founders who out-execute incumbents over time.

Robotics: early but accelerating

• More robotics founders in YC than any prior year; at least one team (Weave Robotics) targeting a ~$65–70K home robot shipped in 2025.
• LLMs are functioning as the "consciousness" layer — reasoning about goals and context — while specialised action models handle physical tasks like folding laundry.
• The open question: can startups win by building AI/software on commodity hardware, or does winning require coupling hardware and software (favouring Tesla-like vertically integrated players)?
• Robotics has not yet had its ChatGPT moment; self-driving in San Francisco (Waymo) is the closest analogue — fully deployed, used daily, but still not widely understood outside the city.

AR/VR: a notable flop

• The Vision Pro and Quest launched to excitement but neither found a killer app beyond using the headset as a large monitor or for watching films.
• The constraint is physics: fitting sufficient compute and optics into a lightweight form factor remains unsolved engineering.
• The chicken-and-egg problem persists: not enough devices for developers to build apps, not enough apps to drive device purchases.
• The Meta Ray-Ban glasses — no display, just audio and voice — have found more daily utility as a voice AI interface connected to ChatGPT or Claude.

Regulation and the broader environment

• California's SB 1047 failed; Biden's AI executive order looks unlikely to survive under the Trump administration.
• The concern — that certain AI capabilities would require registration or become illegal — has receded for now, though the platform risk of foundation model companies with API control remains real.
• In-person work is back: late-stage startups are prioritising office presence; YC's in-person demo day returned with 1,200 investors and a roughly 10:1 investor-to-company ratio.
• San Francisco: new mayor, moderate board majority, cautious optimism — change is slow but directionally right.