20VC: Jev Creator on Why Data Is More Important Than Compute | Will US Open-Source Models Catch Chinese | The Cost of Intelligence and Tokens: What No One Sees But Needs to Know with Diogo Almeida
The Twenty Minute VC (20VC): Venture Capital | Startup Funding | The Pitch
Hosted by Harry Stebbings
- 1h 13m
- Transcript
Summary
Written by a model from the transcript; not yet reviewed.
This episode of 20VC features Harry Stebbings in conversation with Diogo Almeida, founder of Typesafe and maker of Jev, discussing the state of artificial intelligence, model reliability, and the economics of data versus compute. Almeida argues that reliability of intelligence is the industry's largest bottleneck, comparing models to getting Einstein ninety-five percent of the time and Mr. Bean the other five percent. They examine why enterprise software has not fundamentally changed despite advanced models, the limitations of current coding agents, and the overemphasis on public benchmarks.
Throughout the discussion, Almeida explains why data matters far more than compute and outlines the concept of doing the right task to achieve true background automation. They debate the competitive landscape of AI models, the reality of the software bubble, and why traditional tech hiring practices and valuation maxing can lead companies astray. The episode offers a candid look at building a high-growth AI startup with profitability and strong technical convictions.
- Reliability of intelligence is the primary bottleneck preventing AI from achieving full end-to-end automation in enterprise software. 6:22
- Data matters significantly more than raw compute when building and scaling advanced machine learning systems. 8:37
- Public benchmarks are heavily gamed and misleading, often failing to reflect true model reliability and performance. 22:48
- Software reliability and deep workflow understanding are more valuable for enterprise moats than thin wrappers over LLMs. 37:49
Chapters
- 0:00Reliability Is the Biggest Bottleneck
The host and guest open the episode by asserting that AI reliability, not compute, is the main hurdle to widespread automation.
- 0:57Introducing Diogo Almeida and Typesafe
Harry introduces Diogo Almeida, founder of Typesafe and creator of Jev, highlighting the company's rapid growth and valuation.
- 5:08What Is Jev and Its Mission
Diogo explains Jev as a non-specialized model focused on reliable intelligence for automation, contrasting it with traditional LLM benchmarks.
- 46:08Chinese Open‑Source Models vs. U.S. Security
Explores the threat posed by Chinese open‑source LLMs and whether U.S. models can keep pace ethically and technically.
- 46:50Threat Models for Coding Agents
Discusses how to guard against bias, backdoors, and security risks when integrating LLMs into coding agents.
- 48:45Export Bans and GPU Restrictions
Examines the impact of U.S. export bans on Chinese GPU tech and how it affects model training and deployment.
- 50:00Neo‑Cloud Bubble and Demand Shock
Analyzes the potential burst of the AI bubble, driven by FOMO and over‑promised capabilities, and its effect on demand.
- 56:17Funding Strategy and VC Lessons
Details the founder’s unconventional fundraising approach, learning from VCs, and prioritizing strategic allies over capital.
- 1:02:57Hiring Data‑Centric Talent
Explores the challenges of recruiting a specialized data team and the importance of aligning hires with the company’s mission.
- 1:05:20Techno‑philia as a Hiring Trait
Highlights the desired passion for technology in hires, emphasizing a non‑sexual, obsessive love for tech.
- 1:09:00Future of AI: Reliability and Governance
Speculates on AI’s role in solving principal‑agent problems, enforcing laws, and uniting society through reliable systems.
Highlights
- 1:12
Zero to 100 Million Revenuedata
Harry Stebbings highlights that Sequoia's Pat Grady reported Jev went from zero to 100 million in revenue in just seven days, backed by a $7.5 billion valuation led by Martin Casado at Andreessen Horowitz alongside Sequoia.
- 6:51
Einstein and Mr Bean Intelligenceinsight
Almeida describes the core unreliability of current AI models as getting Einstein ninety-five percent of the time and Mr. Bean the other five percent. This high variance is why companies cannot safely rely on models for automated accounting or high-stakes decision-making.
- 8:37
AI's Bitterest Lesson on Datainsight
Building on Sutton's bitter lesson that compute beats algorithms, Almeida argues that data is obviously more valuable than compute. The North Star must be defining the right task before scaling infrastructure.
- 19:52
Competitive Advantage of Silencestrategy
Almeida reveals his rule that talking publicly about what Jev is actually doing is a competitive disadvantage because competitors keep missing the mark. Focusing on reliability rather than cost or speed keeps them ahead.
- 43:43
Tiered Labs and Benchmark Maxinginsight
Almeida argues that top-tier labs like OpenAI and Anthropic are mostly truthful because they do not need to cheat, whereas lower-tier labs and open-source contenders engage in intense benchmark maxing for optics.
- 56:42
Raising Capital Without a Pitch Deckstory
Almeida shares how Typesafe's massive funding round happened entirely inbound without a pitch deck because investors and portfolio companies were already experiencing the product's impact firsthand.
- 1:04:46
Filtering Out Mercenaries in Hiringstrategy
Explaining Typesafe's compensation strategy, Almeida notes they deliberately avoid paying the highest market rates to filter out mercenaries and attract missionaries who care about the technology.
Quotes
“Data matters a lot more than compute.”
Unnamed speaker · 0:11 “Reliability is the absolute largest bottleneck.”
Unnamed speaker · 0:00 “If you optimize for the important stuff to be reliable on, which is decision-making and judgment, that is how you get real reliability.”
Unnamed speaker · 10:18 “I think there's a consensus that Chinese open source models are very efficient.”
Harry Stebbings · 46:08 “I would not use a Chinese model in my coding agent vanilla, just to be clear.”
Unnamed speaker · 46:58 “We want to pay people like good amounts. That is not the highest thing ever.”
Unnamed speaker · 1:04:46 “Technophilia. Not non-sexual, just to be clear, non-sexual, OK, relax.”
Unnamed speaker · 1:05:29 “I think that principal-agent problems are abound in the world.”
Unnamed speaker · 1:09:08
Sponsored segments
- Blitzy 1:35–2:43
- Flex 2:38–3:29
- Vanta 3:25–4:29
- Blitzy 1:10:44–1:11:52
- Flex 1:11:48–1:12:39
- Vanta 1:12:35–1:13:37
About this episode
Diogo Almeida is the Co-Founder and CEO of TypeSafe AI, the company behind Jev. Jev is an AI model that lets software make decisions automatically and tells developers how confident it is in each call. https://www.alphaxiv.org/%40diogo-almeida-2?utm_source=chatgpt.comAccording to Sequoia's Pat Grady, it went from zero to $100M in revenue in seven days. TypeSafe's latest round valued the company at $7.5BN, with Martin Casado at Andreessen Horowitz leading and Sequoia also investing. Before TypeSafe, Diogo was a researcher at OpenAI, where he helped invent ChatGPT. AGENDA: 00:00 Why reliability is the biggest bottleneck to AI 10:26 When does doubling compute stop making models smarter? 12:06 If data matters more than compute, which data actually makes models better? 22:00 What stops OpenAI from copying your product the moment it works? 37:53 Are Harvey and Legora just wrappers around frontier models? 40:05 How much smarter is AI actually getting per token? 43:26 Which AI lab makes the most misleading benchmark claims? 48:19 Will American open-source models catch up with China's? 50:01 If demand for GPUs keeps growing, what bursts the AI bubble? 57:39: How we raised $750M in a week with no deck? 1:07:03 Why Typesafe does not hire from OpenAI?
Transcript
14,350 words · 3 speakers
Speaker names marked are inferred from speaking time and the show’s known hosts — they are not voice-verified.
Topics
- Jev
- OpenAI
- Anthropic
- RLVR
- Sequoia
- Harvey
- Andreessen
- RLHF
- Typesafe
- GPT-4.5
Page updated · transcript and summary are machine-generated