Jun 5, 2025 · 1h 2m · mad
Inside the Paper That Changed AI Forever - Cohere CEO Aidan Gomez on 2025 Agents
0:00 / 0:00
▶ Watch on YouTube →
gold bands on the timeline = statements, start to end. Hover to read, click to jump. CC turns on captions
In this episode of The MAD Podcast, host Matt Turck interviews Aidan Gomez, co-founder and CEO of Cohere and co-author of the seminal "Attention Is All You Need" paper. Gomez discusses the technical development of the Transformer architecture, Cohere's strategic focus on enterprise AI over consumer AGI, and the rapid deployment of agentic workflows in production.
How this conversation actually went
Every chapter scored 0–10 on four independent dynamics. Hover any point for the reasoning behind the score. Matt holds 27.7% of the talking time here. How this is scored →
Matt as informed peer 4.0 Guest teaching 3.4 Guest disagreement 0.5 Matt pushing back 1.2
speaking balance: gold is Matt, purple is the guest (3 minute bins)
Sharpest disagreement ▶ 25:34 Rejection of AGI Culture and EA Ideology
Aidan strongly dismisses the AGI ecosystem, describing it as people cosplaying and LARPing a new religion rather than doing practical, impactful work.
Hardest push from Matt ▶ 11:35 Challenging Google's Supposed Transformer MissMatt pushes the common narrative that Google failed to leverage Transformers, forcing Aidan to correct the premise by highlighting immediate deployments in search and translate.
Biggest teaching moment ▶ 37:51 Explaining Why Synthetic Data Outperforms Human DataAidan reframes human data as inherently lazy and explains why synthetic data produces far better empathetic alignment and performance in enterprise models.
Matt holds his own ▶ 14:00 Citing SSMs and Post-Transformer ParadigmsMatt shows deep technical grasp of cutting-edge AI research by naming Yann LeCun's alternative models and State Space Models to challenge Aidan on post-Transformer R&D.
the scores for every segment, with the reasoning behind each
| Chapter | Topic | Matt as informed peer | Guest teaching | Guest disagreement | Matt pushing back | Why |
|---|---|---|---|---|---|---|
| Cold Emailing Google Brain and the Administrative Mishire | 1 | 2 | 0 | 0 | Matt asks a standard open-ended question about how Aidan became a co-author on the seminal paper. Aidan shares the lighthearted anecdote of how he cold emailed Google Brain researchers and got hired as an intern due to an administrative misclassification assuming he was a PhD student. | |
| Developing the Transformer Architecture and Tensor2Tensor | 5 | 3 | 0 | 1 | Matt demonstrates industry context by bringing up FAIR research lab structure and asking if open-ended academic freedom still exists today. Aidan explains how economic realities and compute resource concentration have pushed research toward product-oriented work streams. | |
| The NeurIPS Submission Sprint and Sinusoidal Positional Encoding | 4 | 5 | 2 | 2 | Matt asks why Google failed to jump on the Transformer architecture immediately, prompting Aidan to gently correct the premise by noting Google integrated it into search and translation right away. Aidan clarifies that Google's miss was specifically not leaning early into pure sequence modeling of internet text as OpenAI did. | |
| The Endurance of Transformers and Potential Successor Architectures | 6 | 3 | 0 | 1 | Matt displays strong technical familiarity by citing Yann LeCun's alternative architectures and State Space Models (SSMs). Aidan explains that hardware/infrastructure lock-in creates a high hurdle for successor architectures, though he remains hopeful for new paradigms. | |
| Test-Time Compute, Reasoning Models, and Extending Beyond STEM | 4 | 4 | 0 | 1 | Matt asks Aidan to unpack test-time compute and reasoning paradigms for the audience. Aidan explains why allocating different compute energy depending on task complexity makes intuitive sense and why reasoning models are accessible to train. | |
| Founding Cohere and Experiencing the Shock of Early Generative AI | 2 | 3 | 0 | 0 | Matt listens as Aidan recounts the early days of founding Cohere and the emotional shock of reading early generative text outputs like a synthetic Wikipedia article written by a model. | |
| Rejecting AGI Ideology in Favor of Practical Enterprise Impact | 3 | 4 | 4 | 2 | Matt asks why Cohere chose enterprise over being an AGI lab. Aidan forcefully rejects the AGI culture, calling it LARPing a new religion, and argues that enterprise productivity is far more meaningful than building God or preaching doom. | |
| Transitioning from Academic Researcher to Enterprise CEO | 4 | 2 | 0 | 0 | Matt notes the trend of academic researchers becoming CEOs and sometimes returning to research labs. Aidan shares his personal evolution into the CEO role, jokingly noting he is now more annoying than helpful to Cohere's modeling team. | |
| Cohere's Full-Stack Architecture: Command, Search, Rerank, and North | 5 | 3 | 0 | 1 | Matt prompts Aidan to detail Cohere's full product stack spanning Command, Search/Rerank, and North. Aidan details how foundational retrieval and generation models integrate to power autonomous agent workflows. | |
| Synthetic Data Effectiveness and Stylistic AI Alignment | 5 | 5 | 2 | 3 | Matt challenges Aidan on synthetic data, citing past industry skepticism regarding model degradation. Aidan dismisses those old takes, educating Matt on why synthetic data is superior because humans are lazy responders while synthetic data yields better stylistic alignment. | |
| Industry-Specific Models and Custom Enterprise Fine-Tuning | 5 | 3 | 0 | 2 | Matt presses on whether enterprise customization happens at the base model level or prompt interface. Aidan explains how Cohere fine-tunes dedicated customer models and generates domain-specific synthetic data under strict privacy boundaries. | |
| Multilingual AI Innovation and Cohere Labs | 4 | 3 | 0 | 1 | Matt inquires about Cohere Labs and whether enterprise multimodal demand is real or speculative. Aidan explains that vision is table stakes for enterprise tasks like PDF document OCR and GUI computer control. | |
| Localized AI Expansion with Regional Enterprise Champions | 5 | 3 | 0 | 2 | Matt notes Cohere's partnership strategy with regional champions like Fujitsu and LG. Aidan highlights Cohere's ability to deploy models directly inside customer VPCs or on-premise, contrasting it with API-only providers. | |
| Cohere North and Autonomous Agent Workflows in Finance | 4 | 4 | 0 | 1 | Matt asks for concrete examples of agentic multi-agent workflows. Aidan provides a detailed scenario of automated wealth management portfolio hedging when major world events break, compressing weeks of research into hours. | |
| The Shift from Proof-of-Concept to Production Execution | 5 | 4 | 1 | 2 | Matt asks whether customers get overwhelmed by agentic possibilities and require heavy consulting, cracking a joke about money spent on Accenture. Aidan corrects the notion, noting enterprise sophistication has matured significantly. | |
| Early Adopter Competitive Edge and Current Technical Boundaries | 4 | 4 | 0 | 1 | Matt asks where the current technical boundaries are for AI agents. Aidan explains that sensitive fields require human-in-the-loop oversight and notes that models are not yet capable of discovering novel science independently. | |
| CEO Aidan Gomez on Global Politics and AI's Macroeconomic Impact | 2 | 2 | 0 | 0 | Matt wraps up by asking broad macro questions. Aidan shares his concerns regarding global political fragmentation while expressing optimism about AI boosting global labor productivity. |
Statements from this episode (21)
Google Brain accidentally hired Aidan Gomez as an undergrad intern
“I think I got in through an administrative mistake because my manager thought I was a PhD student.”
Gomez says RNN and LSTM architectures were complicated and ugly
“Previous RNN models, LSTM models, which were quite complicated and kind of in some ways ugly.”
Gomez: Modern Transformers look strikingly similar to the original 2017 architecture
“And so one of the big shocks is how over the past eight years, how little things have changed. Like it, it's really surprising to me. That the Transformers we train today looks so similar to what was back then.”
Gomez: Google quickly integrated Transformer architecture into Search and Translate
“No, they jumped all over it. So it went to production inside of search, inside of translate, like the existing product suite. And so they, to say that they didn't adopt the transformer architecture would not be correct.”
Gomez: Google failed to lean into language modeling early, unlike OpenAI
“To say they didn't lean hard enough into language modeling, like just pure Sequence modeling of text on the internet. That's, I think the accurate statement. That's what OpenAI did early and uniquely well.”
Gomez: Hardware lock-in raises the bar to replace Transformer architectures
“They built so much infrastructure specialized to the transformer. And so it's like we dug ourselves into this. Well like we now have chips that are being optimized explicitly to that architecture. And so to move architecture, it requires so much effort, energy…”
Gomez: Discrete diffusion models will not replace the Transformer
“Now there are these discrete diffusion models, which do diffusion, which has been super popular for, like, image understanding, image generation. It's doing that same process for language models, but I still don't see that replacing the transformer.”
Gomez: Creating AI reasoning models is dramatically cheaper than pre-training
“It's easy to create a reasoning model. It's dramatically cheaper than pre-training. And so it's accessible. And so there's this huge intelligence uplift that comes for really quite little effort.”
Gomez: AI reasoning models will expand into medicine and physical sciences
“We've just scratched the surface at the moment. It's mostly focused on, you know math problems and this sort of thing. There is a whole world of applications that we need to make it work work in medicine, you know, everything from the pure sciences, physics, c…”
Gomez: Early AGI and EA ecosystem felt like LARPing a new religion
“The whole AGI effective altruist this whole ecosystem, it never resonated with me. It felt like cause playing. It felt like people were LARPing a new religion and all of this stuff like create God.”
Gomez: Specialist AI models will emerge for drug creation and material sciences
“There will be specialist models that emerge to help in things like in pharma, right? The creation of new drugs in material sciences for advanced materials.”
Gomez: AI models will probably surpass the best doctors at prescribing drugs
“Is it better than the world's best doctor at prescribing drugs? Probably not. Will it get there? Probably.”
Gomez: Anyone selling AI doom and gloom is wrong
“Anyone who's selling you doom and gloom, I think is wrong.”
Gomez: Synthetic data makes up the majority of Cohere's training data
“Synthetic data is incredibly effective. It's now the majority of the data that we train on for creating something like command A.”
Gomez: Cohere's Command model will soon become multimodal
“Command command a isn't currently multimodal, but you can imagine Very soon it will be.”
Gomez: Multimodal AI capabilities are table stakes for enterprise data
“But there's also like multimodal is Essential for understanding enterprise data, like PDF documents, where there's graphs and this type of thing, or understanding slide decks. A lot of the modalities that enterprises work in are visual. So it's sort of table s…”
Gomez: Existing AI models fail on non-English enterprise documents
“The current technology that exists doesn't serve their needs, especially in the enterprise world. Like maybe you can get them to speak good enough Japanese at chitchat, but for actual enterprise documents, everything breaks down like immediately.”
Gomez: AI agents differ from SaaS by requiring full human organizational context
“I think one thing that's different about agents and AI compared to other SaaS is that usually you're trying to do something that a human is doing in the organization. And To do that, you need the con, you need the same context that human has.”
Gomez: AI agents cut financial research tasks from a month to hours
“So we can take something that used to be a month. And bring it down to, you know, four hours, eight hours.”