AI product management
In a joint conversation moderated by Sarah Guo, OpenAI CPO Kevin Weil and Anthropic CPO Mike Krieger argue that building AI products requires a genuinely new product-management discipline — centered on writing and iterating evals, designing for stochastic/partially-reliable model behavior, and tight product-research collaboration — because the underlying model capabilities shift unpredictably every couple of months rather than resting on a fixed technology base.
AI product management differs fundamentally from traditional PM because the underlying technology base itself keeps changing unpredictably (Kevin: 'every two months computers can do something computers have never been able to do before in the history of the world'), forcing product decisions to be made before capabilities are even confirmed.
Enterprise AI product work trades the fast aggregate feedback of consumer products for slower but more direct customer feedback, and success depends as much on buyer-side procurement/organizational dynamics as on product quality.
Partial model reliability (e.g. ~60% task success) doesn't preclude shipping a good product — Kevin argues you can design around it via human-in-the-loop correction, editable output, and models that flag their own low confidence, citing GitHub Copilot's early small/imperfect model as proof this pattern already worked.
Writing and grading evals is becoming the central skill of AI product management, to the point that OpenAI's formerly separate 'research PM' and 'product/API PM' roles are merging, and Anthropic tests candidates by having them turn a bad eval into a good one.
AI's non-determinism breaks roughly 25 years of user intuition that identical inputs yield identical outputs, so product design must shift from deterministic bug-fixing toward feedback loops, guardrails, and aggregate-behavior monitoring.
Enterprise AI adoption requires active user education, because unlike consumer power-users, most enterprise employees are encountering chat-based LLMs for the first time and need structured teaching materials and in-product guidance.
New agentic capabilities like Anthropic's Computer Use are still exploratory rather than fully productized, being validated on narrow concrete use cases such as UI testing and repetitive back-office data workflows.
Sophisticated builders (and both companies internally) rely on orchestrating multiple specialized models rather than a single universal model, since a model like o1 excels at reasoning but is slow and not multimodal.
Reasoning models like o1 represent a distinct new scaling axis — inference-time/query-time compute rather than pretraining scale — which the speakers liken to deliberate human reasoning (working through hypotheses) versus the fast, single-pass 'System 1' style of ordinary pretrained-model text completion.
The speakers predict the near future of AI products moves toward proactivity (models that surface relevant information unprompted, based on authorized access to a user's data) and asynchronous interaction (users delegate longer, open-ended tasks and are notified later), superseding instant-chat as the dominant interface.
As models take on more personality and behavioral nuance, 'model behavior' becomes a genuine, separately designed product surface, and users are beginning to form quasi-relational, almost social attachments to specific models.
Co-design/co-research/co-finetune collaboration — Mike describes Anthropic building features like Artifacts and Canvas through tight, iterative collaboration between product/design and research teams rather than research handing off a finished model. Apply: Embed designers/PMs early in research experiments and treat early outputs as learning signals to shape co-fine-tuning, not as guaranteed shippable features.
Three-part product-research capability framework — Mike's framework distinguishing (1) the general 'slope' of intelligence you can watch and plan around, (2) capabilities product deliberately co-invests in and co-fine-tunes with research, and (3) a surprise 'capability front' of new releases (e.g. Computer Use) that product must react to. Apply: Separately track long-run capability trends, planned joint research bets, and unplanned capability drops so roadmap decisions distinguish predictable investment from reactive opportunism.
Designing for partial reliability (the '60% pattern') — A product-design approach for models that are only correct some of the time, assuming human-in-the-loop correction/editing rather than requiring near-perfect output, as GitHub Copilot did with an early imperfect model. Apply: Ship a feature even at ~60% success if the interface lets users quickly review, edit, or reject output rather than blindly trusting it.
Model self-uncertainty signaling — Kevin's proposed technique of having a model recognize its own low confidence and explicitly ask the user for help or clarification instead of guessing. Apply: Train or prompt the model to flag uncertain outputs and request human input, raising combined human+model accuracy above the model's raw success rate.
Eval-writing as a core PM skill — Both speakers describe writing and iterating on evaluations (including exercises like turning a 'crappy eval' into a good one, and OpenAI's internal eval-writing boot camp for PMs) as the emerging central competency of AI product management. Apply: Have every PM practice writing task-specific evals and critiquing weak ones (e.g. via a boot camp or interview exercise) before shipping AI-powered features.
AI writing and grading its own evals — Mike notes Claude is itself good at both authoring and grading evaluations once a customer has defined success criteria. Apply: Ask the model to draft candidate evals and grading rubrics for a task, then use it to score outputs, iterating once a human has defined the success criteria.
'Look at the data' discipline — An approach (attributed to figures like Andrej Karpathy) of manually reading actual model outputs and failure cases rather than trusting small aggregate eval-score deltas, since 'golden answer' grading can itself be flawed. Apply: Before concluding a model regressed or improved from a score change (e.g. 80% vs 78%), read a sample of the actual answers to check whether the eval's grading criteria, not the model, is at fault.
AI-assisted rapid UI prototyping — Using a model like Claude to generate and compare UI options (an 'A/B comparison') before a designer even opens a tool like Figma. Apply: Prompt the model to mock up or compare multiple interface variants to explore many more options faster than traditional design-tool-first workflows allow.
Rolling weekly user research — A continuous research cadence (used at Instagram and now at Anthropic) that brings in new participants every week to test prototypes. Apply: Schedule ongoing weekly sessions with fresh users to surface both interface issues and model-behavior issues in real time rather than doing one-off studies.
Self-referential model grounding — Explicitly giving a model documentation/knowledge about its own product features (e.g. teaching Claude how its own tools work) so it can guide confused users instead of deflecting them elsewhere. Apply: Feed the model accurate, current product documentation about itself so it can answer 'how do I use X' questions and proactively help stuck users, including direct links to docs.
Custom GPTs as enterprise power-user enablement — OpenAI's mechanism letting internal power users build organization-specific GPT configurations (often 'thousands' per company) that make AI usable for less technical colleagues. Apply: Identify internal power users and let them build and share task-specific custom GPTs as an evangelism and adoption mechanism inside an enterprise.
Computer Use for automated UI testing — Using Anthropic's Computer Use capability to verify whether an application does the intended behavior, as an alternative to brittle traditional UI tests that break on unrelated cosmetic changes. Apply: Have the agent interact with the live UI to check outcome-level behavior ('does it do the thing you want') rather than writing fragile selector/pixel-based UI test scripts.
Agentic automation of repetitive data workflows — Exploratory use of agentic models with support and finance teams to automate multi-step, cross-silo data-manipulation tasks (e.g. expense/PR forms) described as 'drudgery.'. Apply: Target agentic automation at high-click-count, repetitive internal workflows first, freeing staff time for higher-value creative work.
Multi-model orchestration — Building workflows that route different sub-tasks to different specialized models rather than relying on one model for everything, since e.g. o1 reasons well but is slow and non-multimodal. Apply: Compose pipelines where one model handles reasoning, another handles speed/multimodal needs, and models can even check or catch each other's outputs.
Test-time/inference-time reasoning scaling (o1 paradigm) — A distinct scaling axis from pretraining: instead of a single fast 'System 1'-style pass, models like o1 spend extra inference-time compute (currently ~30-60 seconds) forming and testing hypotheses before answering, akin to deliberate human problem-solving. Apply: For hard, ambiguous, or multi-step problems, use a reasoning-time-scaled model and accept longer wait times in exchange for more deliberate, higher-quality answers; reserve fast pretrained models for tasks needing an immediate response.
Fine-tuned model cross-checking for reliability — Fine-tuning models for narrow tasks with precise input/output expectations and having one model check another's outputs (e.g. in cybersecurity) to catch inconsistencies and trigger a retry. Apply: Pair a task-specific fine-tuned model with a second model or pass whose job is solely to validate the first model's output before it is trusted or acted upon.
Real-time AI-mediated bidirectional translation — Using a simple system prompt to turn a voice-enabled model (ChatGPT's Advanced Voice Mode) into a live two-way interpreter between two spoken languages. Apply: Instruct the model to translate each language into the other in real time during a live conversation, enabling business or social exchanges across a language barrier.
Personalized interaction-history summarization — A feature (which went viral for Anthropic) where a user asks the model to summarize what it has learned about them from all past interactions, and the model generates a personalized description. Apply: Build a feature that lets users query an aggregated, memory-based profile the model has formed of them from prior conversations, surfacing it back on request.
OpenAI's formerly separate 'research PM' and 'product PM' roles are merging in practice because feature quality now gates on eval-writing and fine-tuning skill rather than product-surface design alone, per Kevin's CSV/code-analysis feature anecdote.
Mike says same-day pilot feedback from two enterprise customers testing an identical model can be diametrically opposed — one says it solved a three-month-old problem, another says it's worse than a competitor's model — because internal evals don't capture the diversity of real customer data and prompting styles.
Mike argues today's models are often 'eval limited' rather than 'intelligence limited': the demonstrated capability ceiling reflects the quality of testing/grading more than a true intelligence limit, and many companies that shipped flashy AI features years ago still can't tell if newer models are actually better for their use case because they never built evals.
GitHub Copilot's original underlying model was reportedly small and imperfect (suggested to be GPT-2-class), yet still commercially valuable — used as evidence that partial-reliability products succeed when designed around easy correction/editing rather than requiring near-perfect output.
Enterprise deal dynamics can be governed entirely by buyer-side procurement rituals disconnected from product quality — Kevin cites a customer requiring 60 days' advance notice before any new launch, joking 'I also would like to know 60 days [in advance].'
Anthropic reportedly had to explicitly teach Claude to explain its own features after user research showed it would otherwise unhelpfully tell confused users to search documentation online themselves; it now gives direct links and proactively helps stuck users.
Early Anthropic Computer Use beta testing produced a widely-cited internal anecdote of the tool ordering Domino's pizza to the office (and over-ordering it), treated as a symbolic 'first AI-ordered delivery' moment that also revealed the tool's early clumsiness.
A viral Anthropic feature letting users ask 'based on everything you know about me... what would you say about me?' reportedly led users to relate to Claude as a quasi-social entity, which the speakers connect directly to why model personality is now treated as its own product design lever (fixed personality vs. customizable per user).
Kevin used ChatGPT's Advanced Voice Mode as a live bidirectional English–Korean interpreter to conduct real business conversations in Korea, framing it as an early realization of a 'Star Trek universal translator in your pocket.'
Both speakers explicitly predict that today's 'magic' model behavior will look like 'garbage' within roughly 12 months given the pace of internal progress they're each witnessing — an acknowledged built-in expiration date on much of their own conversation.
«I've never had such a challenging, interesting, Sleepless product role in my life.»
— 01:21
«here it's like every two months computers can do something computers have never been able to do before in the history of the world»
— 01:56
«I'd argue there are a lot of things that you can actually do when something is 60% right — you just need to really design for it.»
— 10:48
«That's our new... that's the Mendoza Line of AI.»
— 12:12
«There's a very real sense in which models today are not intelligence limited, they're eval limited.»
— 13:13
«Claude is actually good at writing evaluations and also grading them.»
— 14:01
«Nothing beats looking at at data.»
— 16:16
«just tell Claude more about itself.»
— 25:01
«the moment where Domino's shows up at your office and it was ordered entirely by AI is like a very... seminal moment.»
— 27:53
«can we automate the drudgery so you can focus on the creative stuff.»
— 28:57
«imagine what happens if they can think for five hours or five days.»
— 31:44
«Hey ChatGPT I want you to act as a translator when I say something in English I want you to say it in Korean and when you hear something in Korean say it back to me in English.»
— 36:16
«you've got this like Star Trek Universal translator in your pocket.»
— 36:42
«people who are 14 years old are like well I expect the AI to be able to do that and I love that.»
— 37:29
«shipping a product you're shipping like intelligence and intelligence and empathy are like what makes like interpersonal relationships important»
— 39:16
«based on everything you know about me based on all of our past interactions you know what what would you say about me»
— 40:19
«Kevin Mike thank you so much for doing this and giving us a glimpse into the future»
— 40:46
Reception
Many viewers praised the talk as an insightful, refreshingly open cross-company AI product conversation, but a highly-liked comment criticizing one speaker's past crypto promotion and complaints about a distracting on-screen likes counter pull the overall reception down from unambiguously positive to mixed.
A practitioner-level, forward-looking conversation between two sitting AI-company CPOs that offers concrete, source-attributed frameworks (eval-writing as core PM skill, designing for partial reliability, multi-model orchestration) alongside speculative predictions the speakers themselves flag as having a roughly 12-month shelf life.

40:58