The Twenty Minute VCCohere's Chief AI Officer, Joelle Pineau: Why Scaling Laws Will Continue & Future of Synthetic Data
CHAPTERS
- 0:00 – 1:16
Scaling laws, uncertainty, and what we still don’t know
Joelle opens with the view that scaling laws have held up surprisingly well, while emphasizing there’s still meaningful uncertainty around vulnerabilities and system behavior. The framing sets up the episode’s central tension: robust empirical gains vs. incomplete understanding of risks and limits.
- •Scaling laws have been “remarkably robust” historically
- •AI systems still have unknown vulnerabilities and failure modes
- •The conversation will balance empirical progress with uncertainty
- 1:16 – 2:22
Meta years: why breakthroughs take time (and why “agents” is a decade)
Joelle reflects on her time at Meta doing fundamental research and how long it can take for ideas to mature. She contrasts the hype-cycle pace with the reality that optimizers, compute, and data often need years to align before an approach works.
- •Research progress often requires years of iteration despite rapid headlines
- •Right optimizer/compute/data combinations unlock results
- •Agents and RL resurgence reflect long-gestating ideas finally finding traction
- 2:22 – 3:53
Reinforcement learning’s comeback—powerful idea, inefficient reality
The discussion turns to why RL is suddenly central again (reasoning models, agents) and whether it’s overhyped. Joelle argues RL is fundamental but warns against assuming vanilla RL leads straight to AGI, noting major inefficiency constraints.
- •RL is a core paradigm (learning via rewards) and will persist
- •Expectations that RL “out of the box” yields AGI are premature
- •Learning efficiency is the key bottleneck for RL at scale
- 3:53 – 6:46
Deep dive: why RL is so hard (compounding errors + costly interaction)
Joelle breaks down RL’s inefficiency: sequential decisions compound mistakes and exploration can resemble finding a needle in a haystack. RL also requires interactive data generation (simulators/environments), which is expensive and limits diversity of training settings.
- •Sequential decision trees make small errors snowball over time
- •Exploration makes RL sample-inefficient compared to supervised learning
- •Effective RL needs interaction/simulation, not just static datasets
- •Environment variety and realistic simulators are costly to build
- 6:46 – 8:32
Training vs inference economics—and Cohere’s on‑prem enterprise bet
Harry asks about the shifting market from training spend to inference spend. Joelle reframes the question through Cohere’s approach: on‑prem deployments move inference cost considerations to customers and create strong incentives for efficient models that fit enterprise constraints.
- •Training/inference cost curves vary widely by deployment model
- •On‑premise enterprise AI changes who bears inference costs
- •Efficiency matters because enterprise adoption depends on practical run-costs
- •Enterprise needs may differ from frontier consumer model priorities
- 8:32 – 11:37
Capital efficiency in AI: unpredictability, step-changes, and where progress comes from
Joelle identifies unpredictability as the core economic challenge: teams, GPUs, and ROI are hard to forecast. She decomposes progress into compute/data (more linear improvements) versus algorithmic innovation (non-linear jumps that can take time to validate).
- •AI investment is hard because breakthroughs and ROI are uncertain
- •Compute and data often deliver more linear gains (with quality/diversity caveats)
- •Algorithms can create discontinuous step-changes (transformers, Adam, reasoning)
- •Great ideas may sit dormant until tried at the right scale/settings
- 11:37 – 14:58
Scaling laws will continue (but they need algorithmic innovation too)
Harry presses on whether scaling laws still apply given today’s emphasis on efficiency. Joelle argues scaling has been consistently reliable and she wouldn’t bet against it, while stressing that scaling alone isn’t enough—algorithmic creativity remains pivotal.
- •Scaling laws have repeatedly surprised skeptics by holding up
- •Scaling doesn’t behave exactly as expected but remains dependable
- •Algorithmic breakthroughs are required alongside scale
- •Algorithm innovation is the most creative—and hardest-to-predict—lever
- 14:58 – 21:51
Enterprise adoption: 10× productivity, integration hurdles, and human change management
Joelle challenges a replacement-focused view of enterprise AI, preferring productivity amplification: can most employees do 10× more with AI? She explains biggest blockers are task ambiguity, workflow/data integration with legacy systems, and the human side of change.
- •Better yardstick than layoffs: broad 10× productivity enablement
- •Biggest gains come from well-specified tasks; ambiguity is hardest
- •Enterprise deployment hinges on integrating with existing processes/systems
- •Adoption requires curiosity and iteration, not “get it right first time”
- 21:51 – 26:12
AI security in the agent era: impersonation risks and the need for standards
Security becomes more complex with agents: beyond LLM prompt attacks, agents can impersonate entities and take actions in the world. Joelle advocates rigorous testing, practical containment (e.g., restricting web access), and a standards/solutions split between governments and companies.
- •Agents introduce new vulnerabilities beyond classic prompt injection
- •Key agent risk: impersonation and unauthorized action-taking
- •Security is a cat-and-mouse dynamic requiring continuous defense
- •Governments can define standards; companies implement scalable solutions
- •Risk can be reduced via isolation (e.g., cutting agents off from the web)
- 26:12 – 28:07
Sovereign models, global deployment, and why multilingual capability matters
The conversation shifts to geopolitics and whether AI winners will be regional. Joelle supports multiple global model builders for diversity and access, positioning Cohere as global while highlighting the practical need for strong multilingual performance across markets.
- •Healthy ecosystem benefits from models built beyond US/China hubs
- •Cohere’s intent is global rather than purely “Canadian”
- •Different regions demand strong local-language performance
- •Multilingual research and internationalization become competitive advantages
- 28:07 – 32:11
Building great AI teams: vision, execution, social glue (not just ‘Galacticos’)
Joelle explains that elite talent matters, but a winning org needs complementary roles: visionaries, execution-focused builders, and team “glue.” She warns that collecting only superstars without focus and cohesion can underperform compared to balanced teams aligned on a North Star.
- •Three critical team ingredients: vision, execution muscle, social glue
- •Over-indexing on one profile (e.g., only superstars) can fail
- •Focus and clarity (North Star) amplifies team effectiveness
- •Top talent is valuable, but composition and collaboration matter more
- 32:11 – 36:38
Why data is getting more expensive—and how the labeling market is evolving
Given a hypothetical $10B, Joelle prioritizes balancing talent and compute, but emphasizes data as increasingly costly. She explains “easy labeling” is largely done; now data work requires domain experts, and synthetic environments/simulators for agents add major cost and creativity demands.
- •Capital allocation requires balance: compute + talent + (often underestimated) data
- •Data costs rise as tasks require specialization and domain expertise
- •Shift from simple labels to complex, business-logic-aligned datasets
- •Agents and RL-like training push demand for simulators and synthetic environments
- 36:38 – 38:42
Synthetic data: when it degrades models, when it helps, and how to preserve diversity
Joelle addresses the risk of model collapse when training on recursively generated synthetic data—especially where diversity is essential (language, open-ended generation). She contrasts closed domains (chess/go) where synthetic generation is well-controlled, and “in-between” domains like coding where diversity can be injected through structured transformations.
- •Synthetic-on-synthetic loops can degrade performance via diversity loss
- •Open-ended domains (language, images) are more vulnerable to distribution collapse
- •Closed worlds (chess/go) support extensive synthetic generation safely
- •Coding may benefit because structure enables diversity injection and transformations
- 38:42 – 45:42
AI coding’s trajectory (like image generation in 2015), and the rise of curation/verification roles
Joelle predicts code generation quality will improve dramatically over a decade, similar to image generation’s leap from 2015 to 2022. As output volume explodes, the key human role shifts toward intent, editorial selection, curation, and verification—reshaping team structures and workflows.
- •Today’s AI coding resembles early, low-quality image generation era
- •Quality is likely to improve significantly over the next ~10 years
- •Future bottleneck becomes selection: choosing the best output from huge volume
- •Curation and verification remain essential human work
- •Team composition changes as creation becomes cheap and oversight becomes central
- 45:42 – 59:02
Investor lens: bubbles, evals as ‘unit tests,’ open vs closed, and what she’d fund
Joelle characterizes the AI moment as high-variance rather than simply “good” or “bad” bubble—rewarding risk tolerance. She defends evals as useful indicators (not ultimate goals), worries about research access disparities but sees universities’ role as idea factories, argues closing AI is a strategic mistake, and highlights scientific discovery/healthcare as exciting investment areas.
- •AI is a high-variance investment environment with big upswings and downswings
- •Benchmarks/evals are helpful diagnostics, but not the goal customers buy
- •Universities still generate top ideas even with less compute; ecosystems benefit from talent flow
- •Closing access harms innovation; ideas will circulate regardless
- •Most exciting opportunities: scientific discovery and healthcare progress within ~5 years
