Joelle Pineau on Scaling, Reinforcement Learning and Enterprise AI

Cohere chief AI officer Joelle Pineau talks with Harry Stebbings about reinforcement learning, enterprise deployment, data, security and research. Her predictions and product descriptions reflect her scientific perspective and role at Cohere; the interview does not independently establish future performance, productivity gains or safety.

Pineau reflects on Meta research from 2017 to 2025. Although AI appears to move very quickly, she says it can take years for an idea to mature through the right optimization, compute and data. Her experience is a reminder to distinguish publication from proven deployment.

She has worked on reinforcement learning for over two decades and remains optimistic about learning from rewards. She is less convinced that applying reinforcement learning alone will produce AGI, and identifies learning efficiency as an unresolved problem.

Pineau explains that sequential decisions can compound mistakes, making successful trajectories difficult to find. Testing actions and training policies may require simulators or other environments, which can be costly and difficult to make varied. This is a conceptual explanation rather than a cost estimate for a particular system.

She points to games and mathematical tasks where success can be expressed with a clear reward as areas of progress. Shaping nuanced social behavior is harder to specify mathematically, in her view. Examples such as AlphaGo illustrate one domain, not general competence across tasks.

On inference costs, Pineau describes Cohere’s on-premises model strategy, where clients run models locally and must manage deployment costs. Stebbings asks whether that weakens the provider’s incentive to optimize; she replies that efficient models serve customer needs. This is the company’s stated strategy, not an independent performance comparison.

She says capital planning is difficult because breakthroughs, GPU requirements and returns are uncertain. Researchers and companies must make decisions about data centers, staff and data without the predictability available in more mature industries.

Pineau distinguishes the effects of compute and data from algorithmic advances: she characterizes the former as more incremental and the latter as sometimes nonlinear. She stresses data quality and diversity as well as quantity, and says even a promising published idea may take years to prove at the right scale.

Asked whether scaling laws will continue, she says they have been remarkably robust but do not act alone: algorithmic innovation also matters. She would not bet against further scaling, while acknowledging results do not always match expectations. This is an expert judgment, not a guarantee of a particular trajectory.

She describes algorithmic research as creative and difficult to predict: the space of possible directions is wide, and one may not know whether a choice is right until it is tested. That uncertainty complicates both research and investment decisions.

Pineau says enterprise deployment can provide feedback on what works in productive tasks, complementing academic benchmarks and informing new research. She has a commercial interest in Cohere’s enterprise business; actual customer outcomes are not independently measured in the interview.

Stebbings proposes judging utility by whether AI can replace work performed by the lowest-performing portion of a workforce. Pineau prefers asking whether workers can do much more with AI and says human and machine abilities are complementary. The disagreement concerns proposed yardsticks, not measured replacement or productivity results.

She considers a tenfold productivity increase feasible for many tasks within a few years, citing translation and media examples. Stebbings questions whether that is an even higher bar. Her claim is a forecast and does not establish a tenfold gain across occupations; human specification and verification still take time.

Pineau says well-specified tasks, where a good result can be described precisely, may be easier to automate than ambiguous ones. She expects effects to vary substantially by task. A fast model response does not by itself verify its accuracy or overall economic value.

She acknowledges workers may reasonably fear displacement and says people react differently to rapid change. Her observations about generations and tool use are personal impressions, not a demographic study or a prediction of employment levels.

For enterprises she identifies integration with existing processes and information systems as a major challenge. Cohere emphasizes on-premises deployment, confidentiality and security, she says. Those are company priorities, not proof that a particular deployment is secure or that every customer benefits.

Pineau says agent security introduces risks that are not yet fully understood, alongside known model attack vectors such as prompt injection. She raises impersonation and unauthorized actions as concerns and calls for rigorous standards and testing; she does not claim any defense eliminates all risk.

She notes that disconnecting an agent from the web can reduce some exposure but also limits information access. The appropriate controls depend on the use case. Stebbings asks who should validate agents; Pineau sees roles for public standards and company deployment.

Pineau points to aviation as an example where government standards have contributed to safety and argues AI guardrails should learn from real experience. She is optimistic that standards can reduce uncertainty, while saying policy should not be expected to precede all technological learning. This is a policy view, not a safety certification.

She welcomes models built in different regions as a way to broaden access and diversity, while describing Cohere’s ambition as global rather than exclusively Canadian. She highlights multilingual deployment and language fit in markets such as Japan and Korea; performance claims are not independently benchmarked here.

On research teams, Pineau looks for a combination of vision, technical execution and social cohesion. She argues that assembling only star researchers without complementary skills and a shared goal may be less effective. These are leadership lessons, not a controlled comparison.

Pressed on high-profile talent packages, she says some deeply experienced researchers are valuable and deserve fair compensation, but the whole team need not consist of superstars. Whether multibillion-dollar packages are justified is something time will reveal, she says.

Given a hypothetical $10 billion, she would balance talent and compute and reserve substantial resources for data. She says Cohere is reasonably resourced for its plans, while describing data preparation as increasingly specialized. These are her company assessment and budget priorities, not audited costs.

She explains that difficult enterprise tasks may require domain experts to identify errors, while agent training may require realistic simulated environments. Building simulators and benchmarks can itself be expensive. The point is that data quality and environment design require work beyond simple labeling.

Pineau expects humans to continue guiding and training AI systems, though the nature of the information humans provide will change. Stebbings discloses an investment in a data-services company as they discuss the sector. Neither predicts which individual firms will remain competitive.

Asked whether synthetic data causes model degradation, she says it depends on how it is generated. Recursive language or image generation can lose diversity and degrade; games with well-defined states permit different forms of simulation. These are distinctions about methods, not proof that all synthetic data is safe or harmful.

For code, she suggests transformations and combinations may inject useful diversity, offering a possible route to more synthetic data without the same collapse. She presents this as a hope and research direction, not a guarantee that generated code or training data will be reliable.

Pineau compares today’s uneven code generation with earlier image generation and expects code quality to improve substantially over the next decade. She acknowledges bad code is generated now. An analogy and future forecast cannot replace present verification, testing or security review.

If generated code becomes abundant, she expects selection, intent, critique and verification to remain important human responsibilities. Stebbings questions whether that still counts as a partnership if teams become smaller. They do not establish the eventual number or composition of developer jobs.

She expects interaction to move beyond typing in a box toward voice and possibly gesture or gaze, while treating language as a powerful way to express intent. These are interface predictions, not announced products.

Pineau says she values experimental and theoretical rigor over fixed conviction. She recalls being skeptical that neural networks would remain central to machine learning and says the evidence changed her view. She offers this as an example of updating a research position.

She is skeptical of confidently asserted extreme AI scenarios, including catastrophic or winner-take-all narratives, because she does not think their analysis has sufficient scientific rigor. This is her position, not a finding that severe AI risks are impossible.

Asked whether AI investment resembles a good or bad bubble, she describes high variance with potentially large gains and losses and favors supporting research and risk-taking. This is not a forecast of financial returns or advice suitable for every investor.

Pineau compares evaluations to unit tests: they can provide signals on a particular dimension, but should not be mistaken for the ultimate goal. Enterprise customers care about useful work, she says, so a benchmark score alone is insufficient to establish value or reliability.

She acknowledges universities often have less compute than companies, but says they can explore risky ideas at smaller scale and contribute important research. She values movement of people and knowledge between universities and industry; her examples do not quantify the resource gap.

For early investment she says both technical experience and the ability to assemble teams can matter. She sees promise in healthcare and scientific discovery within five years, but these are research and market predictions, not evidence that any medical system is effective or ready for clinical use.

She is interested in studying populations of interacting AI agents in a sandbox. The idea is a research ambition rather than a demonstrated method for safely deploying autonomous systems.

Discussing children and digital life, Pineau distinguishes social online experiences from individual ones. She describes discussing privacy settings and limiting screens when her children were younger. These are family choices and observations, not universal parenting or clinical advice.

She is concerned about mental health but cautions against assuming a simple causal link between particular platforms or AI and illness. She says people deserve real help and more research. The interview does not establish causation or evaluate treatments.

Pineau says Mark Zuckerberg’s depth of inquiry was a leadership lesson from Meta: understanding the work can make decisions more informed. She favors fair compensation for talent but says someone motivated only by money may not be the right fit. These are personal assessments.

Looking ahead, she is excited by AI for scientific discovery and by smaller, efficient models that can run on one or two GPUs. She cites a reported download example for an older model as evidence of demand for efficiency; the figure is not independently checked.

Pineau argues that closing research and restricting access to ideas would be a mistake for innovation, while acknowledging that some organizations are limiting access. Her prediction that ideas will continue to circulate is a view about the research ecosystem, not a guarantee of openness. The conversation ends with mutual thanks.