Skip to content

Ep 97: Jev Creator on Decision Models, Benchmark Fatigue & Why the Labs Missed It

Unsupervised Learning with Jacob Effron

Hosted by Jacob Effron

1h 1m
Transcript

Original podcast

Where to listen

Summary

Written by a model from the transcript; not yet reviewed.

This episode features a conversation between the host Jacob Effron and the guest, the CEO of TypeSafe, discussing decision models, benchmark fatigue, and the future of AI. The guest argues that the current AI industry over-indexes on benchmark optimization and superficial model shapes while ignoring real-world utility, reliability, and developer needs. They explain why machine-native AI and system-one architectures are crucial, and why current coding agents remain constrained by the tyranny of the KV cache. Furthermore, the discussion touches on AI safety, model jaggedness, and the gap between expected capabilities and actual production realities.

  • Developers should push past the passive consumer mindset and actively collaborate on model capabilities and weaknesses. 5:34
  • Treating model reliability as a North Star matters far more than chasing short-term benchmark metrics. 11:38
  • AI engineering design decisions are heavily restricted by append-only constraints of the KV cache. 18:07
  • Coding agents currently yield an estimated 20 to 40 percent speed up when side effects are properly isolated. 56:35

Chapters

  1. 0:00
    Podcast Intro and Host Overview

    Jacob Effron introduces the episode, the guest, and sets the stage for the discussion.

  2. 1:23
    Early Expectations of Jogo and RLCD

    Jogo discusses the initial hype and early confidence in the RLCD approach despite modest early results.

  3. 2:04
    Mission-Driven vs. Product-Driven AI

    Jogo reflects on his personal mission alignment and the tension between AI’s transformative potential and practical usefulness.

  4. 3:10
    Model Transparency and Roadmap Priorities

    Jogo explains how transparency about model limitations informs the roadmap and the balance between reliability and new features.

  5. 6:00
    Benchmark Fatigue and Developer Experience

    Jogo critiques the overemphasis on benchmarks and the impact on developers who must experiment to gauge model quality.

  6. 8:39
    Real-World Feedback vs. Planned Features

    The conversation turns to how much real-world usage informs product decisions compared to internal planning.

  7. 27:14
    KV Cache Tyranny in Coding Agents

    Jogo details the limitations imposed by the KV cache on coding agents and explores potential architectural alternatives.

  8. 40:01
    Five-Year Perspective on AI Progress

    Jogo reflects on how much of Jev’s success relies on recent AI advances and whether similar progress could have been achieved earlier.

  9. 42:42
    Competitive Landscape and Moat Strategy

    The host asks about Jev’s competitive moat and Jogo outlines the challenges of maintaining differentiation in a crowded field.

  10. 44:26
    Funding Strategy and Board Composition

    Discussion of how the company raised capital, avoided subsidies, and built a board of top AI talent.

  11. 45:33
    Quick‑Fire Question: Shifts in AI Thinking

    The host asks what the guest has changed his mind about in the last six weeks, highlighting a shift toward reliability.

  12. 49:03
    Coding Agents: Value and Market Dynamics

    Exploration of the potential productivity gains from coding agents and the uncertain future of their market.

Highlights

  • 2:04

    Low Percentile Mission Drivenstory

    The guest reflects on their time at OpenAI, describing themselves in hindsight as one of the lowest percentile mission-driven employees because the stated mission was intentionally vague.

  • 3:29

    Machine Native AI Callsinsight

    The guest explains that the vast majority of future LLM calls will be made by software for machine consumption rather than human consumption, operating on bits instead of strings.

  • 7:22

    System One Compressioninsight

    The guest argues that AI naturally compresses into system one thinking, making it essential to transparently communicate model weaknesses and jaggedness instead of pretending otherwise.

  • 18:01

    Overpromise and Underdeliverdebate

    The guest expresses concern that the AI industry's focus on new model shapes over actual utility risks burning developer goodwill through overpromising and underdelivering.

  • 31:27

    Tyranny of KV Cachestrategy

    The guest explains how coding agent user experience and advanced routing patterns are severely constrained by the append-only nature of the stateful KV cache.

  • 40:21

    AGI a Year Agostory

    The guest reveals they wrote an internal document at OpenAI arguing that models from the GPT-3.5 era could have met the definition of AGI by performing economically valuable work by volume.

  • 50:57

    Lab Financial Engineeringinsight

    The guest claims that the true greatest strength of major AI labs is financial engineering rather than models or distribution, specifically in how they structure complex compute deals.

Quotes

  • “I think distribution is like a nonsensical answer here, because so many companies have gone all in on that and then gotten burned.”
    Unnamed speaker · 2:56
  • “We have a gigantic job to both push the frontiers of model intelligence reliability, all of these things that we really care about, as well as education to push the limits of software.”
    Unnamed speaker · 22:12
  • “I think that people overuse them for a lot of tasks. And I actually hope the market for that increases too, as people get better at it.”
    Unnamed speaker · 16:26
  • “I think that releasing models that are like kind of more benchmark-optimized than real-world-optimized actually burns that goodwill for people to try things out, which is the only thing that matters.”
    Unnamed speaker · 42:50
  • “I will say that people probably don't believe it, but like, we can run our service profitably, and I hope, I hope that we never have to do the thing of like temporarily subsidizing to get lock-in.”
    Unnamed speaker · 44:26
  • “I think coding agents are massively valuable. It's unclear to me how much of the actual use is truly valuable, like it being like net positive is different than it being worth all of the money put in.”
    Unnamed speaker · 49:03
  • “I think AI safety is not one monolith and there's like a ton of different slices of it. I think that the ASI discussion is actually extremely different from the bio risk discussion, which is very different from like the cyber risk discussion, which is different than like trust and, trust and safety.”
    Unnamed speaker · 56:06
  • Unsupervised Learning 21:21–22:02

About this episode

Jev creator and TypeSafe AI co-founder, Diogo Almeida, joins Jacob Effron to talk about why AI still isn't automating the easy work it's been smart enough to do for years, and why he built a new class of model around reliability. Diogo coauthored OpenAI's InstructGPT paper and spent two years in stealth building Jev, a "system one" model that returns structured decisions instead of text, which launched in September. Diogo and Jacob discuss what the viral launch was really like, why he thought Jev would take a week and it took years, why almost every LLM call should be meant for a machine, and why developers should push model providers harder on jaggedness. They also cover the tyranny of the KV cache in coding agents, why new model shapes are cheap and reliability is the hard part, his estimate that coding agents deliver a 20 to 40% speedup, why he sees financial engineering as the labs' underrated strength, and why he thinks we're two gigantic breakthroughs from true AI agency.

Transcript

0:00
1:01:04

12,531 words · 3 speakers

Speaker names marked are inferred from speaking time and the show’s known hosts — they are not voice-verified.

Topics

  • KV cache
  • Jev
  • coding agents
  • OpenAI
  • Anthropic
  • Jeb
  • GPT-3
  • ASI
  • GPT-3.5
  • dependent typing

Page updated · transcript and summary are machine-generated