Teahose.
SIGN IN
NEW HERE — WHAT TEAHOSE DOES
We read the entire AI & tech firehose — so you don't have to.
PODPodcastsAll-In, No Priors, Acquired…
NEWNewslettersStratechery, Newcomer…
PAPPapersPhysical AI research
PHProduct Huntdaily launches
VCInvestor ScoutSequoia, a16z, Benchmark…
CLAUDE DISTILLS →
7 reads, 30 sec each — free, 6 AM ET.
+ a live graph of the companies, people & themes underneath.
HOME/20VC/20VC: Jev Creator on Why Data Is…
POD
// EPISODE
20VC

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

DATE November 16, 2026SOURCE 20VCPARTICIPANTS DIOGO ALMEIDA, HARRY STEBBINGS
// KEY TAKEAWAYS6 ITEMS
  1. 01Reliability of intelligence, not capability, is AI's binding constraint
  2. 02"AI's Bitterest Lesson": the right task > data > compute > algorithms
  3. 03Compute scaling hit a wall at GPT-4.5
  4. 04A "mega Jevons Paradox" in tokens, driven by automation that actually works
  5. 05AI as a "fourth logic gate" and a coming Cambrian explosion of software
  6. 06AI "coworkers" are the horseless carriage of the AI era

1. Key Themes

Reliability of intelligence, not capability, is AI's binding constraint

Almeida's core thesis is that frontier models are already smart enough, but their intelligence is "jagged." Companies can't hand off real automation because the expected-value math fails. He distinguishes this from uptime and throughput: he means consistency of judgment.

"You get Einstein like maybe 95% of the time and Mr. Bean the other 5%. And the problem is that in expected value it just doesn't work out. Therefore companies quite wisely learn that you can't actually automate stuff." [00:00:00]

He frames the diagnosis as a training-objective problem, not a distribution problem: "It's not a diffusion problem in the slightest." [00:07:01] The product ambition for Jev is "run in the background software" with "nines of intelligent reliability." [00:07:40]

"AI's Bitterest Lesson": the right task > data > compute > algorithms

Almeida reworks Sutton's Bitter Lesson into a priority ordering. Compute-scaling was a "weird outlier" era; the durable truth is that data and, above all, the choice of training task determine outcomes.

"In order of importance, there's doing the right task, then data, then compute, then algorithms." [00:12:22]

He argues that pre-Jev, the only two "right tasks" in LLM land were RLHF ("please humans") and RLVR ("score high on benchmarks"), and neither produces software you can trust in the background: "We just happened to optimize for the human in the loop." [00:09:35] He also claims the data for a new task doesn't exist until you define the task: "Bootstrapping a new task is hard because by construction the data doesn't exist." [00:13:22]

Compute scaling hit a wall at GPT-4.5

His evidence that brute-force scaling plateaued is that the largest model ever served (by token price) wasn't commercially successful.

"GPT 4.5 is still the largest model ever by token cost. And it was not commercially successful." [00:11:10]

He adds that the right data for training isn't sitting in corporate data lakes ("it's probably not shaped the right way" [00:14:39]) and that good data quality is a function of how good the frontier models are, so data vendors need cutting-edge model access. [00:14:15]

A "mega Jevons Paradox" in tokens, driven by automation that actually works

Almeida predicts that reliable, cheap, efficient models will unlock vastly more usage than chat/LLM consumption.

"I truly believe that we are going to like mega Jevons Paradox. It is not going to be a subtle Jevons... There is going to be like tons and tons and tons of usage. And I think that it will completely dwarf at least the total token consumption of string LLMs." [00:25:01]

He explains the mechanism: cheaper, more reliable models make previously failing projects succeed, which raises budgets rather than lowering them. "You're like, hey this actually works. Let's keep betting until the budget matches what it was before and even increases beyond that." [00:53:20] And on current automation share: "In the grand metrics that truly matter, which is real automation, Jev, Fable, and Astra are all tied at roughly 0%." [00:53:20]

AI as a "fourth logic gate" and a coming Cambrian explosion of software

He argues that AI has transformed how software gets written but not the software itself, and that this is the biggest underappreciated opportunity.

"AI is so clearly awesome yet software itself hasn't changed. Writing software has very much changed. But the software is the same and that's tragic." [00:00:27]

His analogy is the blue LED, which enabled screens and efficient white light, not just a 50% bigger LED market. "My hope is that AI ends up becoming like a fourth fundamental logic gate... it should expand what we can do by like hundreds of thousands of X." [00:29:06] He wants to create a new role, the "software researcher," to push software boundaries, and names this as his most underhyped category. [00:27:12], [01:08:55]

AI "coworkers" are the horseless carriage of the AI era

Pushing back on the "own your own intelligence" thesis (Fireworks, Modal), he argues the right form of automation looks nothing like a human colleague.

"I believe that the AI coworkers are a form of that [horseless carriages]. It just makes sense that if you're doing automation, automation will have a different strength profile than a human would be and you'd want to lean into the strength of automation." [00:33:10]

His alternative: find extremely narrow tasks ("twist bottle cap on bottle") and automate them forever, as in the Industrial Revolution. Proprietary company intelligence, he says, lives "at the code level," not in bespoke trained models. [00:37:32]

Benchmarks are distorted by incentives, tiered by lab position

He gives a structural theory of benchmark honesty based on incentives.

"The top tier labs are incentivized to be truthful... The second tier labs are incentivized to lie to themselves. They want to think that they can become a first tier lab... This is what happened with Llama 4 where they were actually too on the nose about benchmark max." [00:43:42]

He argues the people hurt most are those who don't have time to play with models, and that despite Chinese models matching on benchmarks, "Anthropic models are just the best." [00:00:00]

Enterprise software's "AI on the side" state is logical, not a failure

Because LLMs are unreliable, companies put AI only where stakes fall on the user, not the company.

"You can always make them make decisions with stakes to the user. Because companies tend to not give a shit about their users... What you end up having is like the same B2B SaaS with AI attached to the side." [00:29:56]

On who wins, he says incumbents are best positioned "not at all due to their distribution but because they are the ones closest to the problems." [00:31:42]

Hype is a weaker moat than assumed; reputation compounds with real users

Almeida updated his belief about hype: after copycat launches, real customers tried them and rejected them.

"It matters much less than I thought... When we talk to real people everyone is like yeah we tried them they suck. No one cares. We can tell already that you are the best and you care the most." [01:06:24]

Talent: missionaries over mercenaries, and the hardest hire is data

He is deliberately not paying lab-level packages and thinks it filters for the right people.

"We got real clarity that we do not want mercenaries... Anthropic I think is suffering a little bit from their over mercenaries lately... There's an entire market now to prepare people for Anthropic culture interviews." [01:04:10]

The role he can't fill is data: "We have such a hard time hiring for it because no one knows how to get into it." [01:03:24]


2. Contrarian Perspectives

Narrow, company-owned models are likely the wrong long-term bet

Against the prevailing view (voiced by Harry citing Fireworks and Modal) that owning proprietary intelligence is the holy grail, Almeida says general models plus engineering win.

"I would think that the labs' approach of AI coworker but general intelligence makes a lot more sense to me because pretty much any company if you gave them infinite interns that cost like a thousand X less, you could probably find a lot of productive use for them." [00:34:08]

He argues narrow models are a sign of plateau-era micro-optimization, and that "the way companies get their own proprietary intelligence is not from having their own trained models" but from code-level engineering. [00:36:32]

Frontier labs may not truly care about reliability

When Harry pushed that labs must want enterprise revenue, Almeida countered that incentives point elsewhere (MAU for ChatGPT, coding-agent usage now), and that this creates a moat gap.

"I think people have a hard time understanding intelligence as a product which is why people try to use benchmarks as band-aids." [00:22:41]

He says his competitors-in-spirit are "trying to benchmark max, get hype, show numbers that are not relevant like speed and cost" when "the real thing they need to do is hunker up in the bunker and make the models really really smart." [00:20:36]

AI has been a bubble, and a demand shock (not supply) is the risk

He argues that some demand is FOMO built on miscalibrated expectations of what AI can do.

"I believe that AI has been a bubble for a while. So I would like to keep planning for an eventuality of bubble burst." [00:00:27] Elaborating: "I think that's the kind of thing that can cause a demand shock. I think that there's just been so much overpromise under deliver in the AI space." [00:50:38]

He also notes companies that fired staff on AI promises "need to backtrack." [00:51:47]

Robotics is the most overhyped category; "neolabs" are wasteful

Almeida says robotics is far from the reliability prerequisites ("there's like tons of prereqs that still haven't been nailed" [01:08:27]) and rejects the neolab label: "most neolabs are incredibly incredibly wasteful... hire a bunch of ex-lab people, get a bunch of GPUs, have fun, and then eventually get acquired." [00:00:00] He also thinks the American open-source push is largely "pure bullshit and optics" because the labs aren't aiming for the top tier. [00:44:41]

Don't hire from OpenAI

He says researchers there are selected for loving algorithms and scaling, while data is "a very low status thing at most labs," and the few who think about the right task wouldn't want his.

"By construction all that's left of the ML research that could be hireable are the ones obsessed with algorithms. That to me is the kind of thing to do when you want to throw money in the furnace." [01:08:05]


3. Companies Identified

TypeSafe / Jev

Maker of Jev, a model focused on reliable decision-making for automation. Harry says it is "used by a third of the Fortune 500," valued at $7.5B, with a round led by Andreessen's Martin Casado and Sequoia participating. Harry cites Sequoia's Pat Grady saying it went "from zero to 100 million in revenue in just seven days." [00:00:55] Almeida says it is profitable and cheap to serve: "We are not doing the standard subsidized playbook. We actually can serve these things incredibly efficiently." [00:26:15]

Anthropic

AI lab. Almeida credits it with the best models despite non-leading benchmarks. "Anthropic models are just the best even if the Chinese models match it in benchmarks." [00:00:00] He also says it is pushing the "AI coworker" interface, which he considers a horseless-carriage pattern. [00:33:10]

Andreessen Horowitz

Lead investor in Jev, chosen for philosophical alignment. "Andreessen... they care so so so much." [00:59:56]

Sequoia

Investor in Jev. Almeida says he picked them for one partner: "I didn't want their money. I wanted Sonia Wang." [01:00:26]

OpenAI

Frontier lab; mentioned as a competitor that moved quickly on decision-model products (per Harry) and as a source of research talent he avoids hiring. Almeida credits it with top-tier honesty about benchmarks: "I believe that OpenAI and Anthropic are mostly truthful." [00:43:42] He also notes GPT-4.5 as evidence scaling plateaued. [00:11:10]

Intuit (TurboTax)

Used as an example of an incumbent best positioned to redesign workflows with AI: "If we are in TurboTax, where could we change the interface to make it smoother? And there's not a chance anyone at OpenAI or Anthropic knows how to do that." [00:31:42]

Fireworks

Inference/model-customization platform (guest Lin on a prior episode, per Harry). Referenced as arguing for owning your own intelligence, which Almeida disputes. [00:32:45]

Modal

Infrastructure company (founder Eric on a prior episode, per Harry). Referenced for the same "own your intelligence" view and for talent packages up to $20M. [00:32:45], [01:04:00]

Harvey

Legal AI company. Harry defends it as vertically deep; Almeida says it is "much thinner than I would want," had a dream of automating larger projects autonomously that "didn't end up working," and pivoted to a "copilot for lawyers." [00:38:32]

Ramp

Cited by Harry as a Fireworks customer building custom models, and as a Vanta customer. [00:34:41]

Cognition

Coding-agent company; Harry says it is on pace for a $1.6B revenue business and may have "got some wrong" building its own models. [00:34:41], [00:54:26]

Claude Code, Codex, Cursor, Factory

Coding-agent competitors Harry lists. Almeida says he doesn't analyze this market but notes Claude Code changed the metric labs optimize for ("the MAU equivalent of coding agents"). [00:24:07]

Surge, Mercor (data vendors)

Harry asked whether data-provider businesses are good; Harry believes the TAM extends beyond LLM labs to enterprises. Almeida deflected and said how much labs will spend on data once they must be sustainable businesses is hard to ballpark. [00:15:08]

Meta (Llama 4)

Almeida's example of benchmark-maxing too visibly. [00:43:42]

Google, DeepSeek and Chinese labs

Almeida says Google and some Chinese labs aim for the top and so don't explicitly benchmark max, while "a lot of the people fighting for consolation prizes are just doing completely insane levels of benchmark maxing." [00:44:12] He notes DeepSeek 4.1 Flash was not trained the way a hardware-ban test implied. [00:49:03]

Reflection, Poolside, Thinking Machines

American open-source / lab contenders Harry names regarding the push for an American open-source leader. [00:45:49]

Blitzy, Flex, Vanta

Episode sponsors. Blitzy: autonomous enterprise software development with knowledge-graph code understanding; Flex: finance platform ("Flex.one"); Vanta: agentic compliance trusted by "more than 16,000 companies like Ramp, Harvey, and Writer." [01:13:09]


4. People Identified

Diogo Almeida

Founder of TypeSafe, maker of Jev. Known for deep RLHF experience ("when we made RLHF" [00:12:51]) and a contrarian, technologist-idealist stance. His stated operating principle: "Make great decisions slowly." [01:01:27]

Harry Stebbings

Host of 20VC; interviewer.

Martin Casado (Andreessen Horowitz)

Led the round. Almeida says he and Ben Horowitz felt "like old friends talking about from our hearts," and credits Casado with coining "Sassapalooza." [00:31:16], [00:59:56]

Ben Horowitz

Andreessen co-founder; Almeida recorded a podcast with him and Casado. [00:59:56]

Sonia Wang (Sequoia)

Almeida chose Sequoia for her: "She is not just brilliant but a very unique type of brilliant... I would get a different very interesting perspective that should be part of our strategic cards." [01:00:58]

Pat Grady (Sequoia)

Cited by Harry for the "zero to 100 million in revenue in just seven days" claim. [00:00:55]

Richard Sutton

Author of "The Bitter Lesson" (compute over algorithms), which Almeida revises into "AI's Bitterest Lesson." [00:08:08]

Dwarkesh Patel

Almeida recalls a Twitter post showing data was many times more effective than algorithms. [00:11:38]

Will DePue

Almeida attributes to him the result on chips/hardware used in Chinese model training ("that was Will DePue's result"). [00:49:03]

Aditya Ramesh

Almeida praises "Aditya" for inventing text-to-image at OpenAI as an example of a researcher obsessed with a particular task: "He invented an entire field." [01:08:05]

Eric (Modal) and Lin (Fireworks)

Prior 20VC guests; Eric reportedly fights $20M packages for grads. [01:04:00]

Sonali (Sequoia)

Harry cites her question on why reliability is the biggest obstacle to automation. [00:05:55]


5. Operating Insights

Raise for allies, not for valuation

Almeida turned down higher-priced demand because he didn't need the capital and wanted aligned partners for a fight against giants.

"We don't need to valuation max... What we need is allies to fight the largest companies in the world. Let's find the best people to be part of the team." [00:57:38]

Tactic: when inbound is overwhelming, optimize term sheets for strategic thinkers rather than price. He ran the round with no deck, asking VCs for market signals instead. [00:56:28]

Pay "good, not insane" to filter out hired guns

Refusing lab-level comp is a culture filter, not just a budget decision.

"We want to pay people like good amounts that is not the highest thing ever... This is the kind of thing that takes care of people and also filters out the hired guns." [01:04:38]

Hire for technophilia, and build an internal lab for data

His one cross-role trait: love of technology, so lunchtime conversations stay at the frontier. [01:05:20] Operationally, he built an internal "data demon" team ("the best lab in the world but only internal to a company") and treats data/task design as the core craft rather than an input. [01:02:56]

Treat a model failure as a diagnosis problem: prompt, education, or model

For a developer tool, a 97% accuracy gap is investigated across three levers.

"Why is it 97% accurate? What was the missing piece? And to be clear it is not always a model thing. It is very oftentimes a prompting thing... We need to figure out is there some nuance in this language that could help people make more robust prompts, or do we need to make the model smarter?" [00:21:05]

Under-promise, over-deliver and stay off the coding-agent turf

He sees strategic value in letting competitors fight in coding agents while Jev focuses on automation. [00:54:49] Combined with his view that benchmark hype fades with real users, the tactic is to let real customer evaluation, not announcements, be the moat. [01:06:24]


6. Overlooked Insights

Jev's real defensibility is a data-creation capability, and "good data" is a moving target

Buried in the discussion of data is a structural moat: because quality data depends on frontier-model access and defining a new task requires creating data that doesn't exist, the advantage compounds for whoever defines the task first.

"What even good data means is now a function of how good the models are. This is why people who make data need access to the cutting edge models. Because what actually makes the data good is whether or not it improves the cutting edge models." [00:14:15]

Paired with "The thing is that pre-Jev, the two right tasks that existed in LLM land were RLHF... and RLVR," [00:08:37] it implies that sustaining a lead means continually inventing new task definitions and data, which is exactly the capability he struggles to hire for.

Use cases for the "decision model" include securing other AI, a quiet demand driver

In passing, Almeida notes a big current use: classifying whether other agents' tool calls are safe.

"That is actually a huge use case for us right now which is Jev classifying if the tool calls are safe or are they phoning home to China or whatever else." [00:48:05]

This reframes the Chinese open-source threat: a reliable classifier layer can make cheaper, riskier models usable, positioning reliability-focused models as security infrastructure for the broader (including open-weight) model ecosystem, a market that scales with every agent deployed.