We scan new podcasts and send you the top 5 insights daily.
Rob Ferguson tells founders to burn through Anthropic, OpenAI and AWS subsidies while learning their intelligence needs, then come to Fireworks when inference spend justifies customization. He points to Cursor distilling a coding-only model for better cost per intelligence and speed. That transition is the path to model independence.
For most startups, training a custom foundation model is a waste of capital. The winning strategy is to focus on workflow and proprietary data, building a "headless" product that uses a model router to switch between the cheapest, most effective LLMs for any given task.
The first step for an AI startup is to prove value using the best off-the-shelf models, even if they are expensive. Investing in custom models and post-training is a form of optimization that should only happen after product-market fit is established and there is a clear user signal to optimize for.
For AI startups, large cloud credit grants (e.g., $2M from OpenAI) are no longer just for deferring server costs post-PMF. They are now used directly in the R&D phase to train models, effectively acting as a new form of non-dilutive capital that accelerates early product development.
Frontier model providers like OpenAI are not incentivized to reduce your token costs, creating lock-in. Independent agent labs (e.g., Cognition) act like brokers, routing tasks to the most efficient models, which is a more sustainable and affordable long-term strategy for building your company.
Don't use the most powerful and expensive AI model for every task. Use cheaper, faster models like Anthropic's Haiku for high-volume, simple jobs and reserve powerful models like Opus for complex reasoning. This strategy can reduce costs by over 99%, turning a potential $150 task into a $1.50 one.
A successful strategy for AI startups is to initially leverage state-of-the-art foundation models to acquire users and data. Once sufficient high-quality, domain-specific data is collected, they can train their own specialized models to drastically cut costs and latency.
Startups can manage high initial compute costs by using expensive proprietary models like GPT-4 temporarily. The long-term strategy is to use these models only until more efficient, on-device open-source alternatives become powerful enough for their specific use case, which is estimated to be within 1-2 years.
Building custom AI model evaluations is a late-stage concern for creating defensibility. For pre-seed startups, it's a distraction. The only goal is achieving product-market fit using the cheapest, most accessible models, whether that's OpenAI, Anthropic, or open-source alternatives.
While large enterprises must constrain AI model usage to control costs, startups should embrace 'token-maxxing.' By giving developers unfettered access to the most powerful models, startups gain a crucial productivity and talent-attraction advantage over larger, more bureaucratic competitors.
For early-stage AI startups, the primary goal is finding a repeatable, valuable use case. Aggressively use all available cloud and API credits for this exploration phase. Cost optimization is a secondary problem to solve only after product-market fit is established.