The Twenty Minute VCCohere Founder, Nick Frosst: How To Compete with OpenAI & Anthropic, and Sam Altman’s AI Disservice
CHAPTERS
- 0:00 – 0:48
Cold open: Nick Frosst says AGI rhetoric is a disservice (Sam Altman critique)
Nick opens with a forceful critique of claims that AGI is near, arguing that repeated dramatic predictions have been wrong and academically irresponsible. He frames “existential threat” messaging as harmful to public understanding of the tech and to productive policy discussions.
- •Claims that public talk about “how close AGI is” has been misleading
- •Calls out high-profile “existential threat” messaging to world leaders
- •Argues sensational predictions distract from real, tractable risks
- •Sets the tone: pro-technology but anti-hype
- 0:48 – 2:14
Working with Geoff Hinton: playful intuition over rigid math
Nick reflects on being Geoff Hinton’s first hire at Google Brain and what that taught him about research. He emphasizes Hinton’s creative, analogy-driven approach and the importance of intuition and curiosity in pushing breakthroughs.
- •Research as playful exploration rather than purely formal derivations
- •Use of physical analogies to reason about optimizers and loss functions
- •Curiosity-driven “what would happen if…” mindset
- •Surprise at how much intuition drives great research
- 2:14 – 3:20
Did Google miss the ChatGPT moment despite inventing transformers?
Nick discusses the irony that transformers were created at Google, yet rapid commercialization and scaling happened elsewhere. He points to organizational and incentive dynamics, noting that many transformer authors left to keep building on the idea.
- •Transformer architecture originated at Google (2017)
- •Google didn’t scale/commercialize transformers quickly internally
- •DeepMind still has strong talent and output despite consolidation
- •Talent migration: key transformer contributors left to keep advancing it
- 3:20 – 4:34
What Cohere is: a foundation model company optimized for enterprise tool use
Nick defines Cohere as a foundation model company with a singular focus on enterprise deployments rather than consumer chat. He explains their goal: models that reliably use business tools/APIs and company data to complete real workplace tasks.
- •Cohere is one of <20 global companies training frontier LLMs
- •Differentiation: enterprise-first, not consumer engagement
- •Focus on tool use, API calling, and secure access to business data
- •Goal is workplace augmentation and production reliability
- 4:34 – 5:41
Training for enterprise: synthetic companies, synthetic emails, real constraints
Nick explains how enterprise focus changes training priorities and data. Cohere generates synthetic enterprise environments (fake orgs, emails, internal APIs) to teach models how to operate in realistic work contexts—while still relying on real data foundations.
- •Enterprise models aren’t trained to maximize “conversation engagement”
- •Synthetic data simulates business workflows and tool ecosystems
- •Real-world data is still necessary to seed high-quality synthesis
- •In-house annotators still produce non-synthetic, high-quality data
- 5:41 – 7:16
What’s the real bottleneck: data quality beats algorithms (and compute isn’t everything)
The conversation turns to constraints across compute, algorithms, and data. Nick argues core architectures haven’t changed much, and that usefulness hinges on acquiring and generating high-quality data rather than expecting algorithmic magic.
- •Transformers remain the dominant architecture nearly a decade on
- •Training pipeline evolved (base model + SFT/RLHF + other RL methods)
- •Algorithms aren’t the main limiter for utility improvements
- •High-quality real data and derived synthetic data are key bottlenecks
- 7:16 – 11:22
Scaling laws skepticism: GPT-5, plateaus, and what “AGI” even means
Nick challenges the notion that simply scaling compute leads to linear product or capability gains, using GPT-5 vs GPT-4 sentiment as a prompt. He offers a practical definition of AGI (“treat a computer like a person”) and argues current LLMs won’t get us there—while still being transformative for enterprise automation.
- •Compute alone can yield diminishing or messy product outcomes
- •Product experience matters (e.g., auto-selection vs speed)
- •AGI definition proposed: a computer you treat like a person
- •LLMs are powerful for enterprise workflows (e.g., expenses) but not AGI engines
- 11:22 – 18:13
Models vs apps: specialization spectrum and why benchmarks miss enterprise value
Nick explains that model and application layers are intertwined: the best product often needs a model trained for that interface and workflow. He then critiques common eval culture, arguing many benchmarks track “training on the benchmark” more than real business usefulness.
- •Value capture is tied to training models for specific product interfaces
- •Market trend: general foundation + domain refinement (not one model per micro-task)
- •Benchmarks change constantly (LM1B → HellaSwag → math/ARC-style tasks)
- •Enterprise customers care about production/ROI, not leaderboard hype
- 18:13 – 20:08
How fast are models really changing? Chips, training times, and transformer sameness
They compare rapid model release cycles with slower hardware evolution. Nick notes training now takes months and, despite frequent releases and incremental training improvements, the underlying transformer paradigm remains largely unchanged.
- •Model iteration feels fast, but training cycles are still long (months)
- •Chip progress and deployment constraints operate on different timescales
- •Training methods evolve, but core “next-token” transformer remains constant
- •Industry dichotomy: constant releases vs fundamental architectural stability
- 20:08 – 23:44
The talent market: huge pay, retention realities, and spending $5M for impact
The discussion shifts to the AI talent war and headline-grabbing compensation. Nick is skeptical of exaggerated stories but agrees some individuals create enormous value; he frames compensation as only one factor among purpose, stability, and alignment.
- •Some compensation headlines may be overstated or unstable
- •Top researchers can legitimately create outsized value
- •Cohere actively thinks about making the company a great place to work
- •Would pay $5M if value justifies it; equity already rewards many contributors
- 23:44 – 33:00
AGI hype vs real societal risk: jobs, augmentation, and income inequality policy
Nick argues the most damaging hype is AGI/existential-risk framing because it crowds out concrete conversations about labor shifts, resilience, and inequality. He expects major workforce changes akin to past industrial revolutions, and says outcomes depend heavily on policy choices.
- •AGI/existential-risk rhetoric is confusing and often unhelpful
- •LLMs will change work; many tasks will be automated or reduced
- •Models don’t create independent scientific breakthroughs (yet) in Nick’s view
- •Income inequality impact depends on labor policy, not inevitability
- 33:00 – 36:02
Open vs closed models: Cohere’s “middle ground” and avoiding competitor obsession
Nick outlines Cohere’s stance: release weights for non-commercial use while monetizing commercial deployments. He argues founders should avoid both extremes—obsessing over competitors or ignoring them entirely—and stay grounded in customer value.
- •Cohere releases weights for research/non-commercial use; commercial requires a relationship
- •Positioning: credibility and transparency without giving away the business
- •Predicts/notes some “open” players may trend more closed over time
- •Founder focus: avoid hype-driven micro-comparisons; prioritize customer outcomes
- 36:02 – 38:43
The future of prompting: less “prompt craft,” more literacy about model limits
Nick predicts prompting will remain as an interaction mode, but “prompt engineering” as a specialized skill will fade as models better match human intent. The durable skill becomes understanding how models are trained, where they fail, and rejecting “digital god” thinking.
- •Prompting stays, but the need to “trick” models declines
- •Early prompting hacks reflected web-text training artifacts (e.g., “In summary:”)
- •Key competency: understanding model training, behavior, and emergent limits
- •Building/using LLMs requires non-magical, mechanistic thinking
- 38:43 – 1:05:12
Building Cohere as an enterprise company: fundraising, efficiency, competition, and sovereignty
Nick discusses Cohere’s $600M fundraising context, emphasizing increased investor maturity and concrete customer deployments. He details a strategy built on efficiency (models that fit on two GPUs), forward-deployed engineering for production, and a view that sovereign models are infrastructure shaped by geopolitics.
- •Fundraising shifted from “what is this?” to customer-specific proof (RBC, Fujitsu, LG)
- •Efficiency strategy: enterprise deployability; models designed to fit on two GPUs
- •Forward-deployed engineers help enterprises reach production; not a crutch for bad tech
- •Sovereignty: countries want models as infrastructure; Canada positioning can be an asset
- 1:05:12 – 1:14:38
Quickfire finale: regulation pitfalls, China, 2026 predictions, tools, and being contrarian
In rapid Q&A, Nick returns to his criticism of AGI fear-messaging and warns against benchmark-driven regulation. He downplays fears of China “beating” the West soon, predicts near-term agentic workplace automation (like automated expenses), shares his favorite tools, and reflects on curiosity/contrarianism as both strength and weakness.
- •Regulation risk: choosing gameable benchmarks as proxies for “AGI” and freezing progress
- •China: strong models, but not clearly surpassing top Western ones (in his view)
- •2026 prediction: reliable enterprise agents completing multi-step tasks (e.g., expense filing)
- •Workflow tools: North internally; Cursor for coding; cost sensitivity at company scale
- •Self-assessment: curiosity + contrarianism drives breakthroughs but can cause misses
