Most important take away

Scaling compute is the most reliable but least efficient way to improve AI models; the real leverage is now in data quality and new methods (reasoning, planning, search). The pure model-API business is becoming a low-margin, commoditized race-to-the-bottom, so durable value is accruing at the chip layer below and the application layer above, while companies that become “effective subsidiaries” of their cloud providers are in a structurally dangerous position.

Summary

Actionable insights and patterns from the conversation:

Career and founder lessons

• Resilience through repeated failure: Aidan credits gaming for teaching the “try, fail, try again, fail less” loop and rejecting the one-shot reputation mindset. Treat your career as iterated attempts, not a single bet.
• Be the intern on the important paper: Aidan was the intern on “Attention is All You Need.” Position yourself near research/work that could compound, even in a junior role.
• Conviction beats consensus: OpenAI’s bet on scaling was mocked for years. The biggest career and company wins came from people who held an “irrational” conviction (Ilya on scaling) long before the market agreed.
• Pursue independence deliberately: Cohere has raised ~$1B but explicitly avoids becoming a cloud-provider subsidiary, even when bigger checks are available. Optionality and ownership matter more than raw capital.
• In-person still wins: Despite being pandemic-born and distributed, Cohere centers work in Toronto, London, NYC, and SF offices because the productivity lift from in-person is “unquantifiable.”

Tech patterns and strategy

• Scaling laws still hold, but with diminishing returns per dollar. Doubling compute yields linear intelligence gains — fine if you have unlimited money, otherwise economically unattractive.
• Data quality is the most underrated lever. A single bad example among billions can degrade a model. Most open-source gains have come from better scraping, deduping, upweighting knowledge-rich pages, and synthetic data — not from raw scale.
• The dominant build pattern: prototype on a big general model, then distill into a smaller, focused, verticalized model for production. Expect a world of many models, not three to five.
• Reasoning is the next unlock. Models today must answer instantly; the next generation will think, try, fail, backtrack. The bottleneck is training data showing the reasoning process (the internet only shows conclusions).
• The model layer is being commoditized via price dumping (OpenAI) and free open weights (Meta). If you only sell models via an API, expect near-zero margins short-term. Value accrues to chips below and applications above.
• Chip market is opening up: training is no longer Nvidia-only — TPUs are production-ready and AMD/Trainium are coming. Build for multi-cloud, multi-chip from day one because customers demand optionality.
• Enterprise adoption: the #1 blocker is trust/security around training on customer data. Winning pattern is private deployment inside the customer’s VPC/on-prem — “bring the model to the data.”
• 2024 shift: enterprises moved from POC tourism to urgent production deployment, scared of being caught flat-footed.
• RAG is a step-change for hallucinations and personalization — it grounds answers in auditable sources and lets models reason over private data they were never trained on.
• Agents are the real promise of AI, but whoever builds the best model is structurally advantaged to build the best agent. Pure consumers of someone else’s model are disadvantaged in the agent layer.
• Pricing AI features: two viable strategies — keep price flat to drive expansion (Canva) or charge premium per seat (Microsoft, Salesforce, Notion). Both work; cost of inference is falling fast enough to make either bet reasonable.
• Sticky enterprise software rarely gets displaced; the bigger consumer opportunity is in transformative new experiences (especially voice and social).
• Robotics is the 5–10 year wildcard: foundation models are replacing brittle hard-coded planners, and general-purpose humanoid robotics is close to a tipping point.
• The macro thesis: productivity gains are underhyped. A 5% productivity lift across healthcare or government is civilization-changing — orient products and careers toward genuine productivity, not novelty.

Chapter Summaries

• Origins and gaming: Growing up rural in Ontario with dial-up made Aidan obsessed with making tech faster, pushing him into CS. Gaming taught resilience and the “fail and retry” mindset he sees in successful founders.
• Scaling vs. efficiency: Throwing compute at models works but is the dumbest, most inefficient path. GPT-4-class quality is now achievable in 13B-parameter models thanks to data and method innovations.
• Data and method innovations: Most recent open-source gains come from data quality, scraping, and synthetic data. The next frontier is reasoning/planning models — limited today by lack of training data showing how humans actually work through problems.
• Synthetic data and the model market: Cohere can’t train on enterprise customer data, so it leans hard on synthetic data and focused enterprise use cases. The pure-model API market is heading toward zero margins via price dumping.
• Chips, clouds, and supply chains: Cohere partners for compute, avoids cloud lock-in, and runs on Nvidia, AMD, and TPUs. Chip supply is loosening; the training-chip duopoly is becoming a real multi-vendor market.
• Cohere’s bet and the founding moment: Co-authoring the Transformer paper in 2017 as an intern; ChatGPT was the tipping point that brought the technology to the public. Cohere has raised ~$1B at a ~$5.5B valuation and intentionally avoids becoming a cloud subsidiary.
• Interfaces — chat, voice, GUI: Chat isn’t the universal interface; GUIs still matter. Voice is the magical next interface, with emotion and inflection that’s shocking on first use.
• Diminishing returns and who keeps paying: As models get smarter, average users can’t perceive the gains, but domain experts can — and they’re worth paying for. Costs per flop keep falling, lowering the barrier to last year’s model (which no one wants).
• Consolidation and effective subsidiaries: Adept to Amazon, Inflection to Microsoft — more consolidation is coming. Being a cloud-provider subsidiary distorts incentives and undermines independence.
• OpenAI and the AGI pivot: OpenAI is now effectively a product company. Aidan most admires Ilya’s early conviction on scaling. The AGI mission appears to be taking a backseat to consumer product.
• Enterprise adoption and trust: Security and data privacy are the #1 blockers. Private VPC/on-prem deployment is Cohere’s wedge. Hallucinations are decreasing and RAG is a step-change for grounding and customization.
• Agents and who wins: Hype is justified — agents are the promise of AI. The best agent builder will be whoever controls the underlying model. Consumer apps are more vulnerable to disruption than sticky enterprise software.
• Future bets: Robotics in 5–10 years, voice-first social experiences, and AI as augmentation rather than replacement. Aidan rejects mass-displacement doomerism — humans remain essential, especially in sales and high-stakes accountability.
• Quickfire: Biggest mind-change is the importance of data quality. Easiest round was the first; current scale of fundraising has “broken his brain.” Optimistic on UK tech, less so on continental Europe due to regulation-first culture. In-person work wins. The world is supply-constrained and productivity gains are the most underrated story in AI.