By playing, you agree to uListen's Terms and Privacy Policy and YouTube's Terms of Service.
The Twenty Minute VCJev Creator on Will US Open Models Catch Chinese | The Cost of Intelligence & Tokens | Diogo Almeida
Summaries, chapters, transcripts, and FAQs on this page are made by uListen, not YouTube.
Great conversations run long. Their best ideas don't have to.
uListen is free on iPhone, iPad, Android, Chrome, and Claude.
See howCHAPTERS
Jump to publisher timestampsChapter summaries
TypeSafe & Jev’s core mission: making AI trustworthy enough for real automation
Diogo Almeida frames TypeSafe’s product Jev as an attempt to close the gap between models that seem brilliant in demos and models that can reliably run real business processes. He argues software hasn’t fundamentally changed because models still can’t be trusted to make high-stakes decisions without humans constantly checking them.
- •Jev is positioned around reliability for automation, not just being a specialized model
- •Diogo contrasts human-in-the-loop assistance with running software "in the background" reliably
- •He points to repeated failures in automating whole jobs (e.g., accounting, drive-throughs) as evidence of a reliability gap
- •He argues most SaaS looks the same as a decade ago because models can’t be trusted for big decisions
Reliability of intelligence: “Einstein 95% / Mr. Bean 5%” and why variance blocks adoption
Diogo defines reliability as low variance in decision quality, not uptime. He argues models’ occasional catastrophic mistakes destroy expected value, so companies rationally avoid full automation.
- •Reliability is about judgment consistency, not system availability
- •Models can be extremely capable yet still unusable for simpler high-stakes workflows due to variance
- •The "Einstein vs. Mr. Bean" framing: occasional low-quality outputs are unacceptable for automation
- •He argues businesses need "nines" of intelligent reliability for background operation
“AI’s bitterest lesson”: right task first, then data, then compute, then algorithms
Diogo critiques the earlier belief that scaling compute alone would deliver the next leaps. He argues progress depends primarily on optimizing the right objective, which determines what data exists and how training should proceed.
- •He says the compute-scaling narrative plateaued and stopped paying off the way people expected
- •He proposes a hierarchy: right task → data → compute → algorithms
- •He claims prior North Stars (RLHF: pleasing humans; RLVR: scoring well on benchmarks) don’t map to reliable automation
- •He notes models can be superficially "reliable" at being polite, but not reliable in decision-making
What makes training data valuable: task definition, bootstrapping, and “on-policy” difficulty
Diogo explains that valuable data often doesn’t exist until a task is defined, as with RLHF preference data. He emphasizes that what counts as “good” data depends on whether it actually improves state-of-the-art models, requiring tight feedback loops with strong models.
- •New tasks require inventing new data; RLHF required creating preference comparisons from scratch
- •He argues RLVR’s value is not just verification but task structure and private reasoning leading to better results
- •Good data must be at the right difficulty to push frontier models forward
- •He’s skeptical corporate data lakes contain the right shaped data without major task-aligned transformation
Synthetic data: tool, not idol—and only useful when tightly aimed at the right objective
Diogo resists blanket claims about synthetic data, calling it a technique whose value depends on usage. He says research suggests synthetic data often helps smaller models catch up, and he stresses his team’s focus is wielding it toward reliability of intelligence.
- •He rejects a one-size-fits-all view of synthetic data
- •He cites a common research view: synthetic data often helps small models but not frontier models
- •He says TypeSafe uses synthetic data but in a different way than many others
- •He reiterates the key is "laser focus" on the right task: reliability of intelligence
Skepticism about “Neo labs” and AI spending: waste, incentives, and the risk of a hype-driven market
Diogo argues many new AI labs are wasteful, built around hiring ex-lab talent, buying GPUs, and aiming for acquisition. While he finds some RL investment exciting, he says his expected-value view is that current spending levels aren’t justified.
- •He says many new labs feel resource-wasteful and acquisition-oriented
- •He distinguishes between cool technical progress and whether he’d personally invest
- •He expects a hype cycle and says the field hasn’t reached a "reality" phase yet
- •He states he’d be happy to be proven wrong but currently believes spending is not justified
What people get wrong about Jev: it’s not speed or cost—it’s reliability, and benchmarks miss it
Diogo says the market misunderstands Jev by focusing on cost, speed, and benchmark numbers. He claims Jev’s differentiator is reliability on the axes that matter for automation, and that chasing benchmark-maxing copycats misses the point.
- •He says Jev’s advantage is reliability rather than cost/speed
- •He argues many competitors are "benchmark maxing" and burning goodwill
- •He describes an internal focus on why performance is not 97% and what’s missing (prompting vs model changes)
- •He says developers care less about benchmark scores and more about whether the model is best for new real tasks
Why big providers can’t “just copy” Jev: different org incentives and what they optimize for
Diogo argues leading model providers often optimize for user growth metrics rather than decision reliability. He claims that if a provider genuinely prioritized reliability on decision-making, competition would be more serious—but he believes many don’t operate that way today.
- •He claims large labs often optimize for metrics like usage rather than robust decision reliability
- •He ties some behaviors (like sycophancy) to what systems were optimized for
- •He argues reliability requires dedicated focus and training on a different North Star
- •He says it currently feels like there’s little real competition in his category, though he expects that to change
Decision models everywhere: “mega Jevons paradox,” efficiency, and a new “software researcher” role
Diogo predicts massive growth in decision-model usage that could surpass traditional LLM token consumption. He argues efficiency improvements and close collaboration with developers are key, and he advocates for a new role that blends software engineering with boundary-pushing research.
- •He forecasts a "mega" Jevons paradox where lower cost and higher reliability drive far more usage
- •He claims TypeSafe serves models efficiently and sees more efficiency headroom
- •He describes a desired role of "software researcher" to push software boundaries with AI
- •He proposes AI could remove the "can’t automate" category, expanding what software can do
Why enterprise software hasn’t changed: rational distrust, AI bolted-on, and shifting incentives
Diogo says the shallow integration of AI into enterprise software is logical because models can’t be trusted for high-stakes decisions. He argues companies offload verification onto users by attaching AI side features rather than embedding AI into core logic.
- •He says lack of deep enterprise AI adoption is not a distribution problem
- •He argues unreliable models can’t be trusted for company-stakes decisions
- •He claims AI often ends up "attached to the side" so the user bears the checking burden
- •He calls the current state tragic but consistent with today’s reliability limits
Incumbent SaaS vs AI labs: workflow proximity matters more than distribution
Diogo rejects the idea of a SaaS collapse and argues incumbents may be well-positioned because they understand real workflows and can invest in engineering. He suggests labs are too removed from product details to redesign interfaces and processes effectively.
- •He says a "SaaSpocalypse" isn’t happening
- •He argues winners are those who care most and understand workflows, not necessarily those with distribution
- •He claims SaaS companies have advantages in engineering and familiarity with real tasks
- •He contrasts workflow knowledge (e.g., improving product flows) with labs’ "ivory tower" mindset
Disagreeing with “every company should own its own intelligence”: automation over AI coworkers
Diogo pushes back on the idea that each company should train its own specialized intelligence, critiquing the framing of AI as a “coworker.” He argues automation should lean into machine strengths through narrow, durable systems rather than humanoid-style interfaces.
- •He compares AI coworkers to "horseless carriage" thinking—carrying forward old interfaces
- •He argues automation works best as narrow systems that solve tasks reliably and forever
- •He says the "own your intelligence" claim often serves a moat narrative for startups
- •He believes many claims rely on assumptions like continual learning and narrowness that don’t match reality
Specialized/vertical models and “thin wrappers”: why reliability and engineering depth determine value
Diogo acknowledges specialized models can be cheaper and more narrow, but says benchmarks are misleading and incentives distort claims. Discussing Harvey, he argues copilots have limits, and he wants deeper automation that runs for weeks of work in the background rather than modest productivity gains.
- •He calls benchmarks misleading and notes incentives to claim model superiority
- •He frames specialized model efforts as micro-optimizations during diminishing returns periods
- •He says companies will build proprietary intelligence more through code and engineering than training bespoke models
- •On Harvey, he says they shifted toward a copilot approach, which has limits compared to deep automation
“Intelligence per token” and the public benchmark problem: plateauing, distribution shifts, and misinformation
Diogo argues that measuring progress via benchmark-per-dollar is misleading and that improvements are uneven across domains. He claims real gains come from putting tasks more "in distribution" with data, while many other capabilities haven’t improved much.
- •He argues benchmark-per-dollar trends are "complete bullshit" for measuring real progress
- •He distinguishes general pre-trained intelligence from task-specific improvements via data distribution
- •He claims coding improved mainly because it’s been pushed in distribution, not because general intelligence advanced dramatically
- •He describes a public ecosystem of misinformation where good intelligence measures aren’t shared
Open source vs Chinese models, security threat models, export controls, and planning for an AI bubble burst
Diogo says benchmarks obscure reality and that second-tier labs are especially incentivized to benchmark-max. On geopolitics, he differentiates threat models (vanilla use vs retraining/guardrails), predicts American open models likely won’t catch up on "true intelligence" due to ethics, and shifts to his view that AI demand could shock downward if expectations reset.
- •He says top labs don’t need to cheat, while second-tier labs may benchmark-max under pressure
- •He discusses concerns about Chinese models, backdoors, and safer use via retraining/guardrails and tool-call safety checks
- •He predicts American open-source models likely won’t catch up in true intelligence, though it may not matter if capability plateaus
- •He argues AI may already be in a bubble and a demand shock could happen as people confront overpromises
TypeSafe’s raise, choosing investors, hiring philosophy, and what’s over/underhyped
Diogo describes a chaotic, inbound fundraising process and says he prioritized allies over valuation-maxing. He explains hiring challenges in building a specialized data team, rejects extreme compensation "mercenary" dynamics, calls robotics overhyped, and highlights "software research" and new interfaces as underhyped.
- •He says the round was inbound-heavy, deck-light, and provided signals about real adoption
- •He says they optimized for allies rather than valuation or maximum dollars
- •He describes the hardest hiring area as a specialized data-building team for a new task
- •He rejects $20M-style comp dynamics, prefers non-mercenary culture, calls robotics overhyped and software-interface innovation underhyped
Publisher timestamps
- 0:00Intro
- 1:43Why AI Reliability Is the Biggest Bottleneck to Automation
- 4:51Why Data Matters More Than Compute
- 7:30Why Scaling Compute Alone Stopped Working
- 9:05What Actually Makes AI Training Data Valuable
- 14:00Is Synthetic Data Overhyped?
- 15:06Why Diogo Thinks Most Neo Labs Are “Incredibly Wasteful”
- 17:18What Everyone Gets Wrong About Jev
- 19:46Why OpenAI Can’t Simply Copy Jev
- 20:28Why Benchmarks Don’t Tell You Which Models Are Best
- 23:02Why Decision Models Could Dwarf LLM Token Usage
- 28:12Why Enterprise Software Has Barely Changed Despite AI
- 29:32Do Incumbent SaaS Companies Have the Advantage in AI?
- 30:57Why Diogo Disagrees That Every Company Should Own Its Own Intelligence
- 33:34Are Specialized Enterprise Models Actually Worth It?
- 37:00Is Harvey Too Thin a Wrapper on Frontier Models?
- 39:44Is Intelligence Per Token Really Improving?
- 42:19Why AI Benchmarks Are “Total Bullshit”
- 45:44Will US Open-Source Models Catch Chinese Models?
- 48:34Why Diogo Thinks the Answer Is Probably “No”
- 49:52Why Diogo Thinks AI Is Already in a Bubble
- 53:16How Jevons Paradox Could Create Massive AI Demand
- 57:06Inside TypeSafe’s $750M Raise at a $7.5BN Valuation
- 1:00:49Why Diogo Chose Andreessen and Sequoia
- 1:04:43Why TypeSafe Refuses to Pay $20M AI Compensation Packages
- 1:09:22Why Diogo Doesn’t Hire OpenAI Researchers
- 1:10:50Why Robotics Is the Most Overhyped AI Category
- 1:11:17The Most Underhyped Opportunity in AI