20VC · · 1h 14m
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
With Diogo Almeida
Summary
Diogo Almeida, founder of TypeSafe, joins 20VC to argue that reliability of AI intelligence, not compute or cheap tokens, is the real bottleneck for automation, and that data for specific tasks matters more than scaling. He is bullish on Anthropic's model quality, while criticising Anthropic's coworker framing, OpenAI's usage-metric focus, and the benchmark-maxing incentives at second-tier labs such as Meta (Llama 4) and Google. He calls the SaaS-disruption narrative overblown because incumbents like Intuit understand workflows, and he sees Harvey as a thin copilot. He also expects US open-source models to struggle to catch Chinese rivals, argues the threat of Chinese models depends on post-training and can be mitigated, and says AI demand could burst quickly if FOMO fades.
Key takeaways
- Diogo Almeida (bullish on Anthropic): Anthropic's models are the best, even where Chinese models match them on benchmarks.
- Diogo Almeida (bearish on OpenAI): OpenAI optimized for usage metrics, so it won't match reliability-focused decision models soon.
- Diogo Almeida (bearish on Anthropic): Anthropic's AI-coworker interface is a horseless-carriage approach that misses automation's strengths.
- Diogo Almeida (bearish on Harvey AI): Harvey is a thin wrapper; its pivot to a lawyer copilot shows deep automation didn't work.
- Diogo Almeida (bearish on Meta Platforms): Meta's Llama 4 was too overt about benchmark maxing; second-tier labs deceive themselves on performance.
What was said, by company and theme
23 views from 1 speaker. Every quote is checked against the episode's transcript (how).
Anthropic
Across the shows: 113 speakers in other episodes, 165 bullish and 83 bearish calls, past 30 days
Where everyone stands on Anthropic
Diogo Almeida · TypeSafe AIBullish
Anthropic's models are the best, even where Chinese models match them on benchmarks.
In decision-model work, he says the field has no real competition yet. Anthropic models are the best even where Chinese models match them on benchmarks. Benchmarks have proven misleading over the last nine months, and Anthropic models are not on the Pareto curve of most benchmarks yet still rank above competitors in practice. Developers quickly saw that Anthropic's models were clearly best on new tasks, while benchmark-maxed rivals were not.
“Anthropic models are just the best even if the Chinese models match it in benchmarks. And, like, even when you, like, measure benchmark maxing, you still think, like, the anthropic models are better than this...”at 22:56
Diogo Almeida · TypeSafe AIBearish
Anthropic's AI-coworker interface is a horseless-carriage approach that misses automation's strengths.
He says Anthropic is pushing AI as a coworker with the same interface as a human, which he thinks will be the horseless carriage of the AI era. Automation has different strengths than humans, so the better approach is to find narrow tasks and automate them permanently, as in the bottle-cap example. He thinks general intelligence applied across many narrow tasks makes more sense than a coworker metaphor.
“This is the direction Anthropic is really pushing towards, where it's kind of like the same interface as like a human being. And I believe that this is going to be like the horseless carriages of the AI era.”at 33:01
Diogo Almeida · TypeSafe AIBearish
Anthropic attracts mercenary hires who fake culture interviews because it pays more than anyone.
Diogo says Anthropic wants missionaries but also pays more than anyone else, which creates strong incentives for mercenaries to join. An entire market has emerged to prepare candidates for Anthropic's culture interviews, so people fake their culture fit. He says TypeSafe avoided this by paying good but not top amounts, which filters out hired guns.
“Anthropic, I think, is suffering a little bit from their over mercenaries lately because, like, they want missionaries, obviously, but also they pay more than anyone else.”at 1:04:18
OpenAI
Across the shows: 38 speakers in other episodes, 15 bullish and 32 bearish calls, past 30 days
Where everyone stands on OpenAI
Diogo Almeida · TypeSafe AIBearish
OpenAI optimized for usage metrics, so it won't match reliability-focused decision models soon.
In ChatGPT's early days OpenAI optimized for monthly active users, which is a very different metric from the best chatbot experience. That produced sycophancy and lying, and now the same pattern is applied to coding agents. Without a dedicated model and intense focus on reliability, OpenAI cannot build a comparable decision model anytime soon. He hopes they try because it would be good for the world, but says it would be bad for their business.
“Well, it really depends on, like, their business objectives. ... they cared about, like, monthly active users. That's actually a very different metric to optimize for than the best chatbot experience.”at 23:32
Diogo Almeida · TypeSafe AIBearish
OpenAI researchers over-index on algorithms and scaling, neglecting data, so few are relevant hires.
Diogo does not hire OpenAI researchers because, by construction, they love algorithms and scaling and tend to hate data, which is a low-status topic at most labs. He says only about three people at OpenAI think about the right task, and he would not hire them either, since they are obsessed with their own tasks. He sees most of the hireable ML research pool at OpenAI as algorithm-focused, which he thinks wastes money.
“the researchers at OpenAI, by construction, love to do algorithms and then scaling. They tend to hate data. Data is a very low status thing at most labs”at 1:07:25
Harvey AI
Across the shows: 1 speaker in other episodes, 1 bullish and 0 bearish calls, past 30 days
Diogo Almeida · TypeSafe AIBearish
Harvey is a thin wrapper; its pivot to a lawyer copilot shows deep automation didn't work.
Diogo says Harvey is much thinner than it claims. He recalls that when Harvey worked with OpenAI it wanted to autonomously automate larger and larger projects, but that did not work, so it pivoted to a copilot for lawyers. He argues a copilot allows only so much engineering around it, so domain-specific tools and retrieval over law give limited defensibility. Real value needs automating large chunks of work, not double-digit efficiency gains. He hedges that he has not fully checked, since he cannot afford a license.
“They are much thinner than I... That sounds politically correct right now.”at 38:17
Meta Platforms
Across the shows: 66 speakers in other episodes, 61 bullish and 34 bearish calls, past 30 days
Where everyone stands on Meta Platforms
Diogo Almeida · TypeSafe AIBearish
Meta's Llama 4 was too overt about benchmark maxing; second-tier labs deceive themselves on performance.
Diogo says top-tier labs like OpenAI and Anthropic have no need to cheat and are mostly truthful, while second-tier labs are incentivised to lie to themselves because they want to become top tier and face internal pressure to benchmark max. He cites Llama 4 as an example: Meta was too on the nose about benchmark maxing even though there was so much internal incentive to do it, which hurts credibility.
“This is what happened with [Llama] four, where they were actually too on the nose about benchmark maxing even though there was so much internal incentive to do that.”at 44:05
Alphabet
Across the shows: 73 speakers in other episodes, 61 bullish and 48 bearish calls, past 30 days
Where everyone stands on Alphabet
Diogo Almeida · TypeSafe AIBearish
Google is benchmark-maxing like second-tier labs, so its published benchmark results are questionable.
Diogo argues second-tier labs face strong internal pressure to benchmark max and cannot admit it to themselves, unlike top-tier labs that do not need to cheat. He names Google as currently doing the same thing as Meta's Llama 4 team, which he says was too on the nose about benchmark maxing. The claim is brief and based on his view of incentives, not on specific evidence.
“Same thing for Google right now”at 44:05
Enterprise AI adoption & AI software
Where everyone stands on Enterprise AI adoption & AI software
Diogo Almeida · TypeSafe AI
AI's unreliable decisions, not diffusion, block automating accounting and high-stakes enterprise work.
Models are high-variance: they perform like Einstein about 95% of the time and like Mr Bean the other 5%. In expected value that fails, so companies keep failing to automate accounting across successive AI hype cycles. The bottleneck is reliability of intelligence, not diffusion. Because LLMs can't be trusted with decisions that carry stakes for the company, AI is bolted onto B2B SaaS as a side feature that pushes checking onto the user, rather than sitting deep in the logic of apps.
“I would argue almost everyone who's paying attention to AI knows reliability is the absolute largest bottleneck.”at 6:22
Diogo Almeida · TypeSafe AI
Companies' proprietary AI edge comes from software around models, not their own trained models.
Diogo argues companies build their proprietary intelligence through software around the models rather than by training their own, because training is hard. Companies that claim to own their intelligence are defending against the charge that they are light wrappers with no moat. He sees real differentiation as engineering that encodes understanding of real problems and automates them, with the proprietary part living at the code level, not in the choice of underlying database.
“the way companies get their own proprietary intelligence is not from having their own trained models because it's hard to do that, but by having software around them.”at 37:02
Diogo Almeida · TypeSafe AI
Real automation will be a far larger market than the crowded coding-agent assistance race.
Diogo says he does not analyse the coding agent market. He thinks the further away his product stays from it, the more competitors fight there while he can focus on automation. He believes automation, which he describes as thicker than coding-agent assistance for software engineers, will be a very large market. He is unsure whether the market will grow beyond today's size because FOMO moves the needle, and past hype cycles have crashed.
“the further away we stay from the coding agent market, the more that they keep doing their little fight and we can keep focusing on automation.”at 54:54
Frontier AI labs & model race
Where everyone stands on Frontier AI labs & model race
Diogo Almeida · TypeSafe AI
Data and the right task matter more than compute; scaling stopped paying off after GPT-4.5.
Extending Sutton's bitter lesson, he argues the priority order is right task, then data, then compute, then algorithms. Scaling plateaued at GPT-4.5, which was the largest model by token cost and was not commercially successful. Useful data for a new task often doesn't exist until you define the task, as with RLHF preference data, and what makes data good now depends on frontier models. Synthetic data is a tool that helps small models catch up but is of limited use for frontier models. Coding agent gains came when data entered distribution, and recent gains show diminishing returns.
“data matters a lot more than compute. This is very obvious to anyone who's done any applied ML.”at 8:41
Diogo Almeida · TypeSafe AI
AI benchmarks are misleading; labs have incentives to inflate them.
People use benchmarks as band-aids for not understanding intelligence as a product. Benchmarks are a proxy for the underlying capability, but they are incredibly misleading. Labs have perverse incentives to make their models look better internally and externally, so they don't seem outclassed by frontier models. Over the last nine months, benchmarks have proven unreliable as a measure of real-world performance.
“people try to use benchmarks as Band Aids. Benchmarks are kind of a proxy for understanding the intelligence of the underlying thing. ... maybe, like, the last nine months have learned that benchmarks don't mean shit.”at 22:41
AI bubble & tech valuations
Where everyone stands on AI bubble & tech valuations
Diogo Almeida · TypeSafe AI
Neolab spending is wasteful and current AI investment isn't justified on expected value.
He sees the typical neolab model as hiring ex-lab researchers, buying GPUs, and getting acquired, which he calls a waste of resources. Money going into RL is partly good because models get smarter with diminishing returns, but his expected-value calculation says current spending is not justified. He concedes a venture-style 1% chance of a moonshot could still have positive expected value, and says he would be happy to be proven wrong.
“most NeoLabs are incredibly, incredibly wasteful. It feels like their business model is hire a bunch of, you know, ex lab people, get a bunch of GPUs, have fun, and then eventually get acquired.”at 17:41
Diogo Almeida · TypeSafe AI
AI demand could burst quickly once buyers face the real limits of AI; FOMO is inflating spending.
Diogo believes AI has been a bubble for a while, and that part of demand is driven by FOMO, which is fed by people believing they have good results. He argues demand can burst quite quickly when people face the realities of what AI can do. He says many people spending on AI are miscalibrated and overestimate its capabilities, and that there has been much overpromising and underdelivering. He sees the demand shock, not supply, as the trigger.
“I do believe demand can burst quite quickly when people start facing the realities of what AI can do.”at 50:40
AI pricing & unit economics
Where everyone stands on AI pricing & unit economics
Diogo Almeida · TypeSafe AI
AI token usage will grow enormously, dwarfing current LLM token consumption via Jevons paradox.
He expects a mega Jevons paradox in which AI runs in the background everywhere, generating far more usage than all current LLM token consumption combined. This shift would change public perception of AI from slop to something that improves life. He argues that serving these models efficiently rather than on a subsidized playbook can meet this demand, though efficiency research will need to push further.
“I truly believe that we're gonna, like, mega Jevon's paradox... There's 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 [current] LLMs.”at 25:24
Diogo Almeida · TypeSafe AI
Pretrained intelligence per compute dollar is barely moving; coding gains come from narrow in-distribution data.
Diogo argues intelligence is not one dimension. Benchmark dollars-per-score trends are misleading. The general notion of pretrained intelligence per dollar of compute is, to his understanding, barely moving. Gains come from putting data in distribution for specific tasks. Coding has improved a lot in intelligence per dollar because coding was put in distribution, not because base models became more general. He doubts writing has improved in the last nine months, and scaling coding data does not help other capabilities due to diminishing returns.
“I think that there's a general notion of intelligence, which is kind of, like, pretrained intelligence per dollar of compute, which to my understanding is barely moving right now.”at 41:07
Open-source AI models
Where everyone stands on Open-source AI models
Diogo Almeida · TypeSafe AI
Chinese open models are risky vanilla in coding agents, but the risk can be trained or guarded out.
Diogo says he would not use a Chinese model vanilla in his coding agent, so the threat is real if models are plugged in as-is. But the threat depends on the post-training, and in theory you can remove bias and build guardrails to different extents. He argues there are many ways to get security, such as measuring bias, training it out, and classifying whether tool calls are safe or phoning home to China, so the single-model threat model is too narrow.
“I would not use a Chinese model in my coding agent vanilla”at 46:58
Diogo Almeida · TypeSafe AI
American open-source models unlikely to catch Chinese rivals on true intelligence unless China stops trying.
Diogo's guess is that American open-source models will not catch up on true intelligence, unless China stops trying for ethical reasons rather than capability reasons. He says American models hold themselves to a higher ethical and legal baseline, which may make catching up very hard. He notes that if the field plateaus, a slightly smarter model may not matter, so whether the gap matters is hard to forecast and may be a fool's errand if the US cannot be competitive.
“my guess is no, unless China decides to stop trying, not due to capabilities, but due to ethics. But it might not make a difference.”at 48:31
Software & SaaS
Where everyone stands on Software & SaaS
Diogo Almeida · TypeSafe AI
AI won't cause a SaaSpocalypse; incumbent SaaS is best placed due to workflow knowledge.
He says the SaaSpocalypse is not happening. Incumbent SaaS has familiarity with workflows and the scale to fund the engineering that automation requires, which is the valuable part. Incumbents win because they are closest to the problems, not because of distribution. He gives the example of TurboTax at Intuit, arguing no one at OpenAI or Anthropic knows how to redesign that interface, and the labs are living in an ivory tower.
“Number one, SaaSpocalypse is obviously not happening. ... I think that they are best positioned, not at all due to their distribution, but because they are the ones closest to the problems.”at 31:15
AI & jobs
Where everyone stands on AI & jobs
Diogo Almeida · TypeSafe AI
Firms that cut staff expecting AI to replace them are now backtracking, with long-term costs.
Diogo says a lot of companies fired large parts of their workforce because they believed AI was solving their problems, and now they need to backtrack. He argues these companies did not just extract value or overplay their hand, but bet logically on AI capability. He sees long-term ramifications for their future, so the value extraction story is nuanced, and layoffs based on AI expectations are being reversed.
“there's a lot of companies who are like, let's fire a lot of our workforce because AI is solving all of these problems, and they need to, like, backtrack on that.”at 51:42
AI: overall outlook
Where everyone stands on AI: overall outlook
Diogo Almeida · TypeSafe AI
Real automation is near zero across frontier models; reliability gates whether AI budgets grow.
Diogo argues Jevons paradox, where cheaper AI increases spending, only works if projects succeed reliably. Cheaper intelligence alone does not raise budgets. Customers need enough value from projects that finally work before they bet more. He says that on the metric that matters, real automation, the frontier models he names (Jev, Fable and Astra) are all roughly at zero, so there is a huge amount of automation still to be done and reliability is the key bottleneck.
“So this is why reliability of intelligence is so important. And what I like to say is in the grand metrics that truly matter, which is real automation, JEV, Fable, and Astra are all tied at roughly 0%.”at 52:49
Robotics & autonomous vehicles
Where everyone stands on Robotics & autonomous vehicles
Diogo Almeida · TypeSafe AI
Robotics is the most overhyped AI category; it is furthest from the reliability it needs.
Diogo says robotics is not uncool, but when he looks at reliability, the field is very far away. Many prerequisites that need to be nailed first have not been solved, and robotics is the furthest from meeting that bar among AI categories he considers. This judgment rests on his view of the reliability prerequisites rather than on specific data.
“My instincts tell me robotics, not because robotics is not cool, but I just think that if you look at the grand scheme of reliability, they're just very far away from it.”at 1:07:45
Across the shows
What other episodes Chorus covered said about the same companies in the 30 days up to this one.
- Anthropic113 speakers · 165 bullish, 83 bearish calls
- Alphabet73 speakers · 61 bullish, 48 bearish calls
- Meta Platforms66 speakers · 61 bullish, 34 bearish calls
- OpenAI38 speakers · 15 bullish, 32 bearish calls
- Harvey AI1 speaker · 1 bullish, 0 bearish calls
Topics
- AI automation reliability
- AI coworker product strategy
- AI demand shock
- AI disruption of SaaS
- AI moats and proprietary intelligence
- AI reliability and real automation
- AI-driven layoffs
- Chinese open-source model security
- US open-source catch-up
- benchmark credibility
- benchmark reliability
- coding agents vs automation
- hiring culture
- intelligence per token
- legal AI
- model quality vs competitors
- neolab funding and spending
- reliability of decision models
- reliability vs usage metrics
- research culture and data focus
- robotics hype
- scaling vs training data
- startup funding and profitability
- token consumption growth
More from 20VC
- 20VC: Taking on NVIDIA: Fractile CEO on the Future of Chips | What Everyone Gets Wrong About Capex Spending | The Winners and Losers in Neo Clouds with Walter Goodwin · November 14, 2026
- 20VC: Jensen Huang Investigation: What Happens Now | Sequoia Leave London, Close Scout Program and End Evergreen Fund | NASDAQ Loses $2.5TRN in 24 hours | Anthropic targets $1.4TRN IPO valuation · November 12, 2026
- 20VC: Mistral CEO on Why Energy Is Not the Bottleneck to AI | Why OpenAI and Anthropic Will Come for Your Business | The Future of the Inference Market | Are Chinese Open-Source Models Truly Safe? · November 9, 2026
- 20VC: From $8BN to $19BN: Inside WashU's Endowment Fund | SpaceX Made Up 20% of Our AUM | Why Europe and China Has So Much Opportunity | The Reckoning That is Coming for AI Companies with Scott Wilson · November 7, 2026
- 20VC: OpenAI Cuts Valuation to $1.25TRN | Public Market Meltdown: GOOG Loses $500BN in 12 Hours | DoorDash Buys Instacart for $17BN | Kalshi Raises $1BN and Cognition Eyes $80BN Valuation · November 5, 2026
Ask your AI about this episode
Chorus gives Claude, ChatGPT and other assistants every view from this episode and every other show Chorus follows. Ask who agrees, who changed their mind, and what else was said about the same companies.
Chorus reports what speakers said. It isn't investment advice. How Chorus works · About Chorus