Editorial illustration for AI Labs' Breakthroughs Are "Pushing Us Further Away" From What People Need
AI Labs Missing the Mark, O'Reilly Warns
AI Labs' Breakthroughs Are "Pushing Us Further Away" From What People Need
Tim O'Reilly has been repeating the same line for three decades: create more value than you capture. He said it about open-source software in the 1990s, when he was publishing manuals for Perl and Linux while Microsoft was busy trying to lock the industry into Windows. Now he's saying it about AI, and he's aiming it squarely at the companies building the biggest models.
O'Reilly, who coined the term "Web 2.0" and has spent years advising startups and investors through his firm O'Reilly AlphaTech Ventures, argues that OpenAI, Google, and their peers are repeating Microsoft's old mistake. They're chasing scale and lock-in instead of building systems people can actually shape and control. His fix isn't just open-weight models, the kind Meta releases with Llama. He wants the entire stack opened up, architecture included, so users and developers can participate rather than just consume.
I sat down with O'Reilly to press him on this, and on where he thinks the industry has miscalculated. We didn't agree on everything, including AI's role in making original work. Here's where he started.
The big labs are reading the future wrong. They have told themselves a narrative where having the biggest, best model is the key to the future. Big models like Claude are optimizing for particular use cases, but they aren't necessarily the use cases that people want.
Why this matters
O'Reilly's warning cuts against the industry's own scoreboard. OpenAI and Anthropic measure progress in benchmark scores and parameter counts; O'Reilly is measuring who actually gets to use the thing and what it costs them. Those aren't the same metric, and conflating them is how you end up with frontier models that impress investors but sit unused by the small businesses and workers who could actually benefit from cheaper, dumber, more accessible tools.
His China comparison is the sharper point: diffusion beats supremacy if the goal is real-world adoption rather than leaderboard position. For founders building on top of frontier APIs, this is a hint about where the margin actually is, not at the top of the capability curve, but in the boring work of getting smaller models into hands that don't currently have them. For researchers, it's a reminder that "state of the art" is a lab metric, not a market one.
Locking developers into one hyperscaler's stack looks a lot like the browser wars, and we've seen how that story ends for the company doing the locking.
Related Reading
Market Trends AI drives three initiatives to accelerate global action for planet protection
Market Trends Akamai data shows AI training bots and content‑fetching bots rise since July
Market Trends NVIDIA and Google Cloud let developers scale AI from prototype to production
From Archives Kids Compare Current AI to the First Car or Airplane
From Archives NVIDIA NeMo Switchyard Routes AI Agent Workloads Across LLMs
Common Questions Answered
What is Tim O'Reilly's main criticism of big AI labs like OpenAI and Anthropic?
O'Reilly argues that big labs are optimizing for the wrong use cases and prioritizing model size over practical accessibility and affordability. He contends that frontier models impress investors with benchmark scores but remain unused by small businesses and workers who would benefit from cheaper, more accessible AI tools.
How does O'Reilly's principle of 'create more value than you capture' apply to AI development?
O'Reilly has consistently advocated this philosophy across industries, from open-source software in the 1990s to AI today. He believes AI companies should focus on creating value for users rather than maximizing proprietary model advantages, which would lead to more widespread adoption and benefit.
What metrics does O'Reilly use to measure AI progress differently from the industry standard?
While OpenAI and Anthropic measure progress through benchmark scores and parameter counts, O'Reilly measures progress by who actually gets to use AI tools and at what cost. His alternative metrics prioritize accessibility and affordability over raw model performance and scale.
Why does O'Reilly believe big AI models like Claude are problematic despite their capabilities?
O'Reilly contends that big labs have misread the future by assuming the biggest and best models are the key to success. These large models optimize for particular use cases that don't necessarily align with what people actually need, resulting in powerful tools that remain inaccessible to those who could benefit most.
Further Reading
- The Missing Half of the AI Economy - O'Reilly
- How AI Is Accelerating Scientific Discovery - Stanford HAI
- How AI is Transforming Scientific Discovery While Keeping Humans at the Center - Stanford HAI
- Nine Breakthroughs Made Possible by AI - UC San Diego Today
- Reflections from ICML and GECCO 2025 - Cognizant AI Lab Blog
