Lessons from Bob McGrew

Bob McGrew is a former Chief Research Officer at OpenAI and an early leader at Palantir. — Y Combinator — AI Agents and AGI.

Infographic for "Lessons from Bob McGrew".

On the Future of AI and AGI

Interview ideas:

On progress seen from inside: From inside a frontier lab, a long pretraining cycle can look slower than the rapid release cadence outsiders see; McGrew contrasted those perspectives in 2024. — Unsupervised Learning Interview.
On human intent: McGrew suggested that even when AI produces an impressive artifact, people still need to decide whether it matches the result they actually want. — Unsupervised Learning Interview.
On the hard work of scaling: McGrew argued that reasoning may be the last fundamental ingredient needed for human-level AI, while stressing that scaling it remains a difficult systems, hardware, optimization, and data problem. — Unsupervised Learning Interview.
On changing progress: Closing a 2024 interview, McGrew expected AI progress to continue without slowing, but to change form. — Unsupervised Learning Interview.
On the year of reasoning: In the 2025 Sequoia interview, McGrew predicted that year would be defined by advances in reasoning, with considerable room for improvement in compute and algorithms. — Sequoia Training Data Interview.
On the absence of a near-term wall: McGrew expected capabilities to continue improving, while warning that the early easy gains from a new reasoning technique would eventually become harder. — Sequoia Training Data Interview.
On reasoning and agents: McGrew argued that longer coherent reasoning could make agents more reliable at planning and taking actions on a person's behalf. — Y Combinator — AI Agents and AGI.
On scientific progress: McGrew posed a hypothetical question: if AI can propose scientific experiments, why might the pace of real-world science rise much less dramatically? He emphasized the gap between generating ideas and carrying out physical experiments. — Y Combinator — AI Agents and AGI.
On possible AI-era roles: McGrew speculated about AI amplifying individual creators and people who manage AI-heavy teams, while acknowledging that future work remains uncertain and may still involve other humans. — Y Combinator — AI Agents and AGI.

Learnings:

On a shift in AI progress: In late 2024, McGrew expected continued AI progress to depend not only on larger pretraining runs but also on reasoning, new product forms, and reliability. — Unsupervised Learning Interview.
On agency: McGrew guessed that choosing worthwhile questions and projects could become a scarce human capability as AI intelligence becomes more abundant. — Unsupervised Learning Interview.
On stages of AI capability: McGrew discussed progressive AI capability levels—reasoning, then more reliable agents and scientific exploration—rather than treating every advance as a single all-or-nothing AGI event. — Y Combinator — AI Agents and AGI.
On test-time compute: McGrew described reasoning models as spending more compute on a longer, coherent chain of work, which can help with tasks that require sustained problem-solving. — Unsupervised Learning Interview.
On robotics as a forecast: In early 2025, McGrew predicted a possible ChatGPT-like inflection for robotics within several years, while noting that building and scaling physical robots is harder than scaling software. — Y Combinator — AI Agents and AGI.
On slower-than-expected adoption: In 2025, McGrew observed that AI's effect on ordinary work and productivity had arrived more slowly than many expected in 2018, even though technical capabilities had advanced. — Y Combinator — AI Agents and AGI.
On human and AI work: McGrew expected some work to combine AI execution with human direction, but said the jobs of future generations are hard to predict. — Y Combinator — AI Agents and AGI.
On scaling pretraining: McGrew identified scaled pretraining as one of the core advances behind modern language models; he also said further pretraining gains face diminishing returns and large compute requirements. — Sequoia Training Data Interview.
On pretraining alongside reasoning: Rather than declaring pretraining finished, McGrew said new architectures, efficiency and context improvements can still require it even as reasoning becomes a major focus. — Sequoia Training Data Interview.

On Startups and Building Products

Interview ideas:

On starting with the best model: McGrew advised AI founders to prove a frontier-model use case with users first, then explore distillation or cheaper models after finding value. — Y Combinator — AI Agents and AGI.
On startup time: For an early AI startup, McGrew argued that speed to a working product and user feedback can matter more initially than model-cost optimization. — Y Combinator — AI Agents and AGI.
On the FDE role: McGrew described a forward-deployed engineer as a technical person embedded with a customer to bridge the gap between what a product does and what the customer needs. — Y Combinator — FDE Playbook.
On unscalable work at scale: McGrew characterized the forward-deployed model as repeatedly doing close, hands-on customer work across market segments, rather than abandoning discovery after the first customer. — Y Combinator — FDE Playbook.
On Palantir's first customers: McGrew explained that Palantir could not learn intelligence analysts' needs through ordinary interviews, so it used demos and customer feedback to discover the real workflow. — Y Combinator — FDE Playbook.
On when to deploy engineers: McGrew said the embedded-engineer approach made sense for Palantir because different customers needed different adaptations; if a product scales cleanly without it, founders should prefer that simpler route. — Y Combinator — FDE Playbook.
On contract size: McGrew said FDE businesses tend toward larger, more flexible contracts as they take on more important customer problems, unlike simple repeatable SaaS pricing. — Y Combinator — FDE Playbook.
On bearing early execution risk: For a startup selling an uncertain enterprise outcome, McGrew suggested sometimes taking early execution risk and asking to be paid when the project works or expands. — Y Combinator — FDE Playbook.
On executive priorities: McGrew warned that embedded deployments need to tackle a top leadership priority; otherwise the organization may lack the energy and authority to push through implementation. — Y Combinator — FDE Playbook.
On the gravel road: McGrew likened an FDE's customer-specific solution to a gravel road; the core product team then has to generalize it into a reusable highway for future customers. — Y Combinator — FDE Playbook.

Learnings:

On discovery from within: McGrew contrasted sales conversations from outside the customer with engineers solving problems inside its workflow, where further valuable use cases can become visible. — Y Combinator — FDE Playbook.
On avoiding a consulting trap: McGrew acknowledged that embedded engineering can become consulting. The test is whether field discovery yields a more repeatable product, lower delivery cost relative to value, and improving margins. — Y Combinator — FDE Playbook.
On AI-agent product discovery: McGrew argued that AI agents create unfamiliar enterprise product categories, making embedded discovery useful where the workflow and desired outcome are not yet known; he cautioned against using FDEs where simpler scaling works. — Y Combinator — FDE Playbook.
On selling outcomes: McGrew said the FDE model sells a solved customer problem rather than merely installing software, which changes how teams choose and price projects. — Y Combinator — FDE Playbook.

On Management and People

Interview ideas:

On caring as a manager: McGrew said managers need to genuinely care about the people they manage, especially when difficult requests arise. — Sequoia Training Data Interview.
On difficult asks: McGrew argued that a manager can sometimes help a colleague through a hard but necessary choice when that colleague trusts the manager to act in their interests. — Sequoia Training Data Interview.
On talented people's weaknesses: McGrew observed that exceptional researchers can have consequential blind spots they do not see themselves, making candid managerial feedback important. — Sequoia Training Data Interview.
On career-altering mistakes: McGrew distinguished ordinary bad days from major career mistakes; in his experience, the latter can come from avoiding a difficult but necessary choice. — Sequoia Training Data Interview.
On humane role changes: McGrew said that, after giving someone a fair chance, telling them candidly that a role is not working can serve both the person and the organization. — Sequoia Training Data Interview.
On loyalty: McGrew described loyalty—demonstrating care for a colleague's interests—as what enables the difficult parts of management. — Sequoia Training Data Interview.

Learnings:

On preventing avoidable failure: McGrew said trusted managers can sometimes help highly capable people confront frightening choices before those choices become career-altering mistakes. — Sequoia Training Data Interview.

Learn more:

On the Future of AI and AGI

On Startups and Building Products

On Management and People

Read next

Explore the surrounding system

Get the weekly briefing.