AI Leaders Debate Open Models, Pricing and the Future of Search

Harry Stebbings introduces a compilation assembled from separate interviews, not a panel that met together. He brings contrasting views on foundation models, pricing, technology incumbents and AI's future into a single edited conversation.

Stability AI co-founder Emad Mostaque predicts that only a handful of companies will train major foundation models in three to five years, naming his own company among the possible participants. This is his 2023 forecast and reflects his company's commercial position.

Intercom co-founder Des Traynor says its tests did not yet show all language models to be equivalent. His team compared conversation quality, hallucinations, trustworthiness and confidence, while recognizing that providers were improving rapidly. At that time, he says, Intercom prioritized quality over cost optimization.

Digits founder Jeff Seibert expects foundation models to become commoditized as open alternatives improve and the cost of running and tuning them falls. His company builds products on top of models, and his view is a prediction rather than a finding that models were already interchangeable.

Mostaque expects rapid changes in model efficiency and argues that national, cultural and personal datasets can help tailor systems to different contexts. He also argues that models are not inherently free of bias. His examples and predictions reflect his own interpretation of technology evolving at the time.

Meta's Yann LeCun argues that effective models need not always be extremely large: pretraining can be followed by fine-tuning, and efficiency improvements may allow some systems to run on local hardware. He speculates that future architectures using objectives and planning could be smaller still.

You.com founder Richard Socher takes a different view, arguing that model size matters for a single system expected to perform many tasks. Runway co-founder Cris Valenzuela says the durable advantage lies less in any particular model than in the people building, learning and adapting quickly. Each has a commercial interest in AI products.

Seibert notes that model and training-data size had correlated with performance, while research was also exploring how to preserve performance in smaller systems. For fine-tuning, he places particular emphasis on the quality of the data rather than size alone.

On open versus closed models, LeCun argues that open development can draw on ideas from a much wider community than any one organization. He points to the work of making smaller Llama models run on personal computers as an example. Meta had its own interest in releasing model weights; openness also varies by license and release.

Contextual AI founder Douwe Kiela says he likes open source but considers it naive to assume open efforts will automatically keep pace. He argues OpenAI had advantages in understanding usage and serving requests at scale. That is his competitive assessment at the time, not a permanent ranking.

Socher agrees OpenAI was ahead, but predicts an open model comparable to an earlier GPT-4 version before the end of the year. He expects researchers and communities to improve open models and make them usable locally. He distinguishes that forecast from the performance of a continually changing proprietary model.

Theory Ventures founder Tomasz Tunguz compares estimated market values of three major cloud infrastructure businesses with a broad set of public application companies. His illustrative calculation puts each group at roughly $2.1 trillion, but the application value is spread across many more companies. He infers that venture investors may find more opportunities in applications if the analogy holds.

Traynor says much of the value in Intercom's then-current AI offering flowed to its model provider, OpenAI. He expects more competition eventually to reduce infrastructure prices, though he anticipates a small number of major providers. In his view, margins follow differentiation wherever it appears in the stack.

Benchmark investor Miles Grimshaw forecasts a move from selling software seats toward selling completed work. He imagines a control center that sets objectives and measures outcomes, with service commitments tied to results rather than uptime alone. He stresses that little of this architecture was visible at the time.

Traynor similarly expects some pricing to reflect work performed, such as resolving customer questions, rather than the number of employees using software. Beyond Work co-founder Christian Lanng predicts consumption pricing and argues that changing demographics and work preferences could push companies to redesign workflows.

Seibert disagrees that AI itself necessarily changes pricing models. He sees AI as an underlying technology: businesses already charging per seat or by consumption may continue using those models and incorporate AI costs accordingly. His position contrasts with the other guests' forecasts.

Lanng criticizes the copilot metaphor, imagining that many separate assistants could create another layer of coordination work. Grimshaw calls copilots an attractive incumbent strategy because existing vendors already control users, distribution and seat-based economics. He thinks startups may have more room to disrupt by offering a different workflow, while acknowledging exceptions.

In a separate excerpt, LeCun imagines highly capable personal assistants becoming a common interface to digital services. He argues that such important infrastructure should be open and imagines community vetting of knowledge in a process resembling Wikipedia. This is a proposal and prediction, not a demonstrated guarantee of accuracy.

Asked about Apple, Traynor predicts major progress in an assistant that can converse and act across devices. He explicitly makes assumptions about Apple's internal work and privacy priorities. Mostaque calls Apple a black box, while Seibert predicts sufficiently capable on-device models could reduce dependence on cloud providers; none claims inside knowledge of Apple's plans.

Tunguz, a former Google employee, expects chat to replace some search use cases and argues Google faces a difficult choice about disrupting its own lucrative business. He sees opportunities for new shopping, travel and other experiences. These are his 2023 views as an investor, not conclusions about subsequent market outcomes.

Traynor describes Google's Bard launch as seeming reactive to ChatGPT and asks whether Google will be willing to disrupt itself. He sketches a hypothetical advertising model involving disclosed sponsored content in answers, and even imagines subsidized devices as a possible distribution strategy. These are proposals, not Google announcements.

Seibert considers Google vulnerable because of its dependence on search and calls for a more aggressive response. Socher, whose You.com competes in search, argues that advertising economics make a sudden change to Google's main experience difficult. Their claims and revenue estimates are opinions and historical estimates, with competitive interests attached.

On Amazon, Mostaque says the company combines engineering capacity with a marketplace approach to models and cites Bedrock. Traynor speculates that Amazon might acquire Anthropic; this is a hypothetical suggestion, not a reported transaction.

LeCun rejects the view that increasingly capable AI must inevitably escape control or cause catastrophe. He expects technology to change jobs and productivity, while acknowledging that societies must address how gains are distributed. His sweeping views about employment and risk are presented as his arguments, not settled consensus.

LeCun hopes AI will expand access to knowledge and creative tools, enabling more people to make text, art, music and video. He acknowledges risks but opposes slowing research in anticipation of inevitable catastrophe. He supports regulation of AI products, particularly those making consequential decisions, while distinguishing that from regulation of research itself.