Grit

Synthetic Data and the Future of AI | Cohere CEO Aidan Gomez

November 17, 2025

Key Takeaways Copied to clipboard!

  • The core insight of the seminal Transformer paper was its efficiency and suitability for scaling across many GPUs, which proved to be the dominant property for subsequent ML progress. 
  • The pure scaling law for LLMs is showing signs of saturation, evidenced by rumors of smaller next-generation models and a necessary refocusing on better data and training methodologies rather than just brute-force scaling. 
  • Aidan Gomez believes the aggressive fear-mongering and claims of exponential, runaway AI capability by some labs were intellectually dishonest posturing intended to pull up the ladder and disincentivize competition in the AI race. 
  • The EU's current regulatory focus on policing foreign tech companies is seen as protectionism that hinders its ability to build its own competitive technology companies. 
  • True strengthening of Europe will only occur when it organizes to build the next generation of great companies, rather than focusing on protecting itself from others. 
  • Cohere is currently extremely resource-constrained on the people front, hiring across all functions, with a specific callout for ML researchers to reach out. 

Segments

Google’s AI Comeback and Talent
Copied to clipboard!
(00:00:06)
  • Key Takeaway: Google has potentially surpassed OpenAI technologically and possesses significant resources (money, data, talent) to fuel its AI efforts.
  • Summary: Discussion on whether Google missed the AI wave, concluding they have come back strong, potentially surpassing OpenAI with their models, backed by immense resources and talent concentration.
Transformer Paper’s Core Insight
Copied to clipboard!
(00:00:31)
  • Key Takeaway: The core unique insight of the ‘Attention is All You Need’ paper was efficiency, specifically suitability for scaling across many GPUs.
  • Summary: The host asks about the core insight of the consequential paper co-authored by Aidan Gomez. The answer is ‘Efficiency,’ noting its suitability for scaling across GPUs, which led to the dominance of the Transformer architecture.
Introduction to Cohere CEO
Copied to clipboard!
(00:01:21)
  • Key Takeaway: Aiden Gomez, co-founder and CEO of Cohere, is introduced as a Google Brain alum and co-author of the ‘Attention is All You Need’ paper.
  • Summary: The host introduces the show’s premise and welcomes Aiden Gomez, highlighting Cohere’s focus on enterprise LLM infrastructure and its $6.8 billion valuation.
CEO Job as Sales and Travel
Copied to clipboard!
(00:02:04)
  • Key Takeaway: The CEO job at this stage is primarily a sales job, which involves extensive, often grueling, travel.
  • Summary: The conversation shifts to the demands of the CEO role, focusing on the necessity of travel for sales and the host’s strategy of bringing his wife along for dinners to mitigate time away.
Origin of ‘Attention is All You Need’
Copied to clipboard!
(00:04:49)
  • Key Takeaway: Aiden Gomez co-authored the seminal Transformer paper at age 19 as an undergraduate intern.
  • Summary: The host probes the story behind the famous paper, noting Gomez was only 19 and an undergrad intern when it was written with seven other co-authors.
Impact of Open-Sourcing the Transformer
Copied to clipboard!
(00:05:44)
  • Key Takeaway: If Google had kept the Transformer secret, the academic community likely would have developed a similar, highly scalable architecture shortly after.
  • Summary: They discuss whether Google creating the ‘Chat GPT moment’ was dependent on releasing the paper. Gomez suggests the underlying ideas were already in the ether, and someone else would have created something similar within 12-18 months.
The Transformer’s Unique Insight Revisited
Copied to clipboard!
(00:08:25)
  • Key Takeaway: The core insight was building an architecture extremely well-suited for scaling across many GPUs, which proved critical as models scaled exponentially.
  • Summary: Gomez reiterates that the unique insight was efficiency and scalability across GPUs, noting that the project was compressed into about four months in 2017.
Scaling Laws and Diminishing Returns
Copied to clipboard!
(00:12:00)
  • Key Takeaway: There was little conviction that scaling models would yield sustained gains forever; progress is now showing signs of saturation, leading to a refocus on better data/methods over sheer size.
  • Summary: They discuss the surprise that scaling laws held for so long. Gomez notes that recent model iterations suggest diminishing returns, evidenced by rumors that GPT-5 might be smaller, shifting focus to data quality and training methods.
Economic Viability of Scaling
Copied to clipboard!
(00:15:32)
  • Key Takeaway: For consumer chatbots, the massive spend on larger models is entering uneconomic territory as consumers don’t perceive a proportional benefit, though science applications might justify the cost.
  • Summary: The discussion centers on the economics of scaling. While progress is slowing, spend is increasing. For consumers, the ROI isn’t there, but for high-level scientific breakthroughs, massive spending might be justified.
Gomez’s Current Technical Engagement
Copied to clipboard!
(00:17:47)
  • Key Takeaway: Gomez is now primarily a sales/product leader, reading technical papers infrequently, but he still pushes the Cohere team on product-level missing capabilities like learning from experience (memory).
  • Summary: Gomez admits he is ‘cooked’ technically and reads papers rarely. He focuses on product-level gaps, specifically the lack of persistent learning/memory in LLMs, comparing current models to an intern who resets daily.
Critique of AI Hype and Posturing
Copied to clipboard!
(00:28:41)
  • Key Takeaway: Gomez is annoyed by the ‘bluster’ and intellectually dishonest posturing from some labs regarding AGI/ASI timelines, viewing it as a strategy to ‘pull the ladder up’ and scare off competition.
  • Summary: Gomez criticizes the doomsday rhetoric and claims of exponential takeoff, pointing out that multiple models have converged to a similar capability level, suggesting the extreme claims were posturing to discourage participation.
Cohere’s Enterprise Focus and Deployment
Copied to clipboard!
(00:33:04)
  • Key Takeaway: Cohere focuses on enterprise deployment in critical sectors (finance, healthcare) with a unique ability to deploy models on-premise, constrained to fit within two GPUs for efficiency.
  • Summary: Gomez describes Cohere’s mission to integrate LLMs into business systems securely. Their key differentiator is on-premise/air-gapped deployment capability and a strict constraint on model size (max two GPUs).
Enterprise Adoption Shift: POCs to Production
Copied to clipboard!
(00:37:45)
  • Key Takeaway: The enterprise market is shifting from broad, low-appetite Proofs of Concept (POCs) to focused, high-volume production deployment across entire organizations.
  • Summary: Gomez disagrees that enterprise appetite is shrinking; rather, it’s focusing. Companies are moving from testing 30 use cases with a few people to scaling successful bets across tens or hundreds of thousands of employees.
Talent Competition and Cohere’s Philosophy
Copied to clipboard!
(00:43:05)
  • Key Takeaway: The top tier of AI researchers is small (around 150-200), and Cohere avoids competing for ‘mercenary talent’ by focusing on mission-driven individuals seeking generational upside.
  • Summary: They discuss the intense competition for top researchers. Gomez states Cohere won’t match multi-million dollar offers from giants like Meta, preferring talent focused on building a long-term organization.
Gomez’s Metamorphosis from Researcher to CEO
Copied to clipboard!
(00:47:28)
  • Key Takeaway: Gomez finds the shift from research to public-facing CEO uncomfortable and against his nature, but he has learned to manage the shyness to be productive.
  • Summary: The host notes the whiplash of Gomez’s career change. Gomez confirms that faking extroversion is a learned behavior, but he appreciates the platform it provides.
Google’s Strong AI Position
Copied to clipboard!
(00:50:00)
  • Key Takeaway: Google, under Demis Hassabis, has made a strong comeback technologically with Gemini, leveraging its massive resources (money, data, talent) to compete at the highest level.
  • Summary: Gomez expresses pride in Google’s recovery, noting that while they were criticized for missing the initial wave, their models are potentially the best now, fueled by their ‘money printing machine’ and talent concentration.
The Enterprise Opportunity and White-Collar Impact
Copied to clipboard!
(01:00:02)
  • Key Takeaway: The enterprise adoption of current LLMs is still in the basic stages, but the technology is poised to augment white-collar workers—the most supply-constrained segment of the labor market.
  • Summary: Both agree the enterprise is still early, using models for basic tasks. The real impact will be augmenting white-collar jobs (like coding and legal) where demand far outstrips supply.
Future Outlook: Resumed Economic Growth
Copied to clipboard!
(01:02:42)
  • Key Takeaway: Gomez hopes that diffusing AI technology will resume productivity growth across developed economies, reversing recent stagnation that fuels social conflict.
  • Summary: Gomez believes the obvious future effect is increased productivity and new products. He hopes this growth will reverse the dangerous trend of flat GDP per capita seen in places like Canada and Europe.
Critique of European Tech Stagnation
Copied to clipboard!
(01:06:17)
  • Key Takeaway: Europe’s focus on regulation and protectionism, rather than building competitive tech companies, is hindering its economic progress.
  • Summary: Gomez suggests that Europe’s focus on preserving culture and acting as the ‘police’ of other people’s tech (e.g., USB-C mandate) prevents them from building their own competitive hardware and software companies.
EU’s Regulatory Focus
Copied to clipboard!
(01:06:44)
  • Key Takeaway: The EU’s current strategy is perceived as protectionism focused on regulating foreign tech rather than building its own.
  • Summary: The speaker suggests that the EU’s approach is characterized by protectionism, viewing its role as policing other countries’ tech companies through heavy regulation, rather than focusing on developing its own competitive technology firms.
Critique of EU Influence Strategy
Copied to clipboard!
(01:07:12)
  • Key Takeaway: The focus should be on building domestic tech companies, not just regulating external ones.
  • Summary: The speaker argues that the framing for European influence should be centered on building domestic tech companies. They criticize the focus on regulation, citing the USB-C mandate as an example of celebrating regulatory wins over actual hardware or software innovation.
Europe’s Untapped Potential
Copied to clipboard!
(01:07:56)
  • Key Takeaway: Europe possesses the necessary resources (universities, capital) but needs to organize to build the next generation of great companies.
  • Summary: Despite having incredible universities and pools of capital, Europe has failed to organize itself to build leading companies. The speaker, emphasizing their personal connection to Europe, states that true strength will come from building their own firms, not from protecting against others.
Closing Remarks and Future Plans
Copied to clipboard!
(01:08:53)
  • Key Takeaway: The conversation concludes with plans to meet in London and a discussion about Cohere’s hiring needs.
  • Summary: The host and guest wrap up the main discussion, exchange contact information for a future meeting in London, and briefly discuss Cohere’s extensive hiring needs across all functions, especially for ML researchers and sales teams.
Defining ‘Grit’
Copied to clipboard!
(01:10:45)
  • Key Takeaway: The guest defines grit primarily as the ability to withstand pain.
  • Summary: As a final question for the podcast ‘Grit’, the guest is asked what the word means to them. They offer ‘Toughness’ and ultimately define it as ’the ability to withstand pain.’