Chetan Puttagunta is a General Partner at Benchmark and Modest Proposal is a money manager who is one of the most respected thinkers in financial markets. We cover the current scaling challenges of frontier AI models, the incredible proliferation of open source models, and the investing implications across private and public markets.
Principles & Lessons:
1) Pre-training scaling is running into data limits, leading to a shift toward “test-time compute.” Chetan says, “all of the labs have hit some kind of plateauing effect on how we perceive scaling” because “we were going to run out of text data that was generated by human beings.” Now, rather than throwing more compute at ever-larger pre-training runs, labs focus on inference-time reasoning, prompting the model to “look at the problem, come up with a set of potential solutions… and pursue multiple solutions in parallel,” as he explains. This reshapes AI progress from simple “bigger pre-training” to exploring more sophisticated agentic reasoning at inference.
2) A pivot to inference-driven reasoning is more financially efficient and aligns costs with usage. Modest emphasizes that “training… requires $20, 30, 40 billion of CapEx,” and the return on that spend “only comes after you release the model.” In the new reasoning paradigm, “capital expenditures are tied more directly to usage,” because “you’re paying for inference at test time,” which matches real customer demand. This structure might calm concerns about “$50 or $100 billion AI bets,” allowing more incremental and sustainable spending.
3) Meta’s open-source releases are fueling small-team innovation and catching up to frontier models. Chetan points out that “in the last six weeks… we see two to five person teams… match performance, not broadly but in specific use cases, with the frontier models.” He credits open-source LLaMA for standardizing “the entire stack,” letting entrepreneurs “download them, put them on a local machine… and catch up to the frontier.” Modest similarly notes that “Llama’s willingness to remain open is a massive strategic dynamic.”
4) Large incumbents may be less threatened by a single “god-like” model if pre-training scaling stalls. Modest highlights that earlier anxieties focused on a “one or two models to rule them all” scenario if “pre-training scaling” persisted. With synthetic data hitting diminishing returns, the landscape shifts: “You don’t have that doomsday scenario… where there is unstoppable momentum for the biggest labs.” Instead, smaller specialized models and test-time compute hamper the “winner-takes-all” dynamic.
5) The hardware buildout must adapt from giant training clusters to distributed inference. Modest reasons that training previously demanded “50, 100,000 chips utilized… for 9 months,” leaving behind specialized clusters. But if “most scale moves to test-time compute,” networks become “peaky and bursty” across many smaller data centers. As Chetan adds, “You just don’t need as much compute as you did,” so “over capacity might exist on-prem,” allowing enterprises to run AI apps “essentially free.”
6) Enterprise software faces an “innovator’s dilemma” as AI apps unlock faster sales cycles and 10x ROI. Chetan believes that “AI-first solutions” now close deals “in 15 minutes,” leapfrogging incumbents whose systems are “very expensive, built for older architectures,” and can’t quickly re-architect. He cites “immense traction” across areas such as “legal, accounting, circuit board design… it’s not 5% improvement; it’s eliminating entire categories of software and labor.” This creates a distribution unlock for startups.
7) The shift to “reasoning” does not mean progress in AI ends; it steers innovation to new frontiers. Modest clarifies, “People aren’t saying it’s the end of AI progress… it’s full speed ahead,” but rather that “the axis of advancement” has changed from bigger pre-training. Chetan also stresses that labs may still attempt new breakthroughs, so “we should be humble—if they figure out synthetic data or new approaches, we could revert.” Meanwhile, “video and audio data remain wide open” for further leaps.
8) AGI and superintelligence remain ambiguous, but near-term breakthroughs in emergent capabilities seem inevitable. Chetan believes AGI is “very close by,” because “end-to-end tasks like travel booking” will soon be “fully automated.” Modest notes self-play examples, like AlphaGo’s surprising moves or poker bots’ over-betting, showing algorithms exceed human patterns. Both see it as the “boiling frog effect,” where achievements (e.g., passing the Turing test) get normalized. Chetan is “optimistic about continuing leaps,” while Modest sees “recursive self-improvement” as a major unknown – an open question about how AI surpasses its original training bounds.
Transcript