AI's biggest bottleneck is unstable switching between Einstein and Mr. Bean
The founder of TypeSafe says enterprises won't hand high-risk decisions to AI not because models aren't smart enough, but because it's Einstein 95% of the time and Mr. Bean 5% of the time—the expected value doesn't work out.
The video won't play here. Listen to the audio instead:
The argument · tap a timestamp to hear it
Insufficient reliability makes the economics of AI automation impossible
Diogo offers an informal but widely-accepted observation: existing large models' problem isn't insufficient overall intelligence—it's excessive variance. You get Einstein 95% of the time and Mr. Bean 5% of the time. This instability kills the expected value calculation, pushing enterprises away from automation entirely. Even if a model could solve thousand-year math problems, it can't handle basic accounting. He emphasizes this isn't about politeness or vulgarity; chat GPT never suddenly spirals into cursing. Rather, it makes flawed judgment calls embedded in text—the exact problem big tech hasn't solved for.
— DiogoGPT-4.5 shows compute scaling has reached its limit
Diogo invokes 'the bitter lesson': doing the right task matters more than data, data more than compute, compute more than algorithms. If priced by inference compute, GPT-4.5 is still the most expensive model ever by token cost—yet commercially it hasn't worked. This breaks the 'scale alone leads to AGI' narrative. He cites Dwarkesh's recent tweet: data influences model quality vastly more than compute. This isn't new to AI; it's something machine learning has always known. The last few years of compute scaling were the exception.
— DiogoBig tech doesn't copy because it doesn't actually care about reliability
Asked why OpenAI doesn't just copy Jev, Diogo is blunt: big tech doesn't realize how little it cares about reliability. Their orgs historically optimized for monthly active users, spawning conversational styles optimized for flattery (sycophantic). Now they optimize around similar metrics for coding agents. His verdict: Anthropic's models are best, even if Chinese models match them on benchmarks—because the whole decision-model (system one) subcategory is benchmarks-gaming, nobody trusts the numbers anyway. That actually gives companies that nail reliability negotiating power.
— DiogoThe moat isn't training your own model, it's the engineering
On companies like Harvey and Cognition that claim proprietary trained models, Diogo sees narrative theater. Without claiming proprietary intelligence, they'd just be thin shells on top of large models. He counters: proprietary intelligence shouldn't come from training your own model, but from engineering and software built around large models. Real value is understanding problems deeply upfront, automating them so humans never repeat the work. He's direct: Harvey aimed for complete project automation early on, didn't achieve it, fell back to 'lawyer's copilot'.
— DiogoTier-one labs tell the truth, tier-two labs lie to themselves
Asked who lies most on benchmarks, Diogo stratifies: tier-one labs (OpenAI, Anthropic) are incentivized to tell the truth—they don't need to cheat to stay on top. Tier-two labs are incentivized to lie to themselves—they want to believe they can become tier-one, internally pressured, too scared to publicly admit benchmark-gaming. (Llama 4 crashed doing it too visibly.) Companies fighting for 'participation trophies'—like US open-source models—engage in naked benchmark-gaming; nobody uses these models anyway, nobody would call them out. The decision-model (system one) field is worse yet: there's no consensus on what good performance even means.
— DiogoThe bubble bursts on the demand side, not the supply side
Diogo admits he's long believed AI has a bubble, and he prepares continuously for its burst. He judges the collapse won't come from GPU scarcity but from demand shock: enterprises making AI purchases badly misestimate actual AI capability—they're systematically too optimistic. Once that over-promise/under-delivery gap gets exposed en masse, demand could collapse fast. He acknowledges reality is messier: some firms bet on AI while cutting staff and got burned; others bet rationally and actually extracted value. Both happen simultaneously.
— DiogoIn fundraising: skip the Deck, refuse the inflated valuation
This funding round was almost entirely inbound. They told VCs directly they had no time to build a Deck; instead the VCs pitched market signals at them. The company deliberately rejected 'valuation maximization' and complex term structures (cock tranches), reasoning: they're already profitable, cash on hand is growing, they don't need much more capital. What they need are battle-hardened allies. Even when someone offered 3x valuation, they declined. The logic was stark: 'That's bubble talk, that's bullshit.'
— DiogoDon't hire OpenAI researchers because they don't do the right work
Diogo explains why the company deliberately doesn't recruit from OpenAI: getting in builds selection bias—a 'drank the Kool-Aid' effect. Plus their researchers instinctively prefer algorithms and scaling, hate data work; data ranks low in most labs. He adds: even if a rare OpenAI person genuinely cared about 'doing the right task,' he wouldn't poach them—those rare people should obviously keep perfecting what they're good at (turning text-to-image into a whole field, for instance). Forcing them into another lane wastes them. The real bottleneck now is hiring the data team. That's harder than salary competition.
— DiogoIn their own words · checked verbatim
you get Einstein like maybe 95% of the time and Mr. Bean the other 5%
Diogo6:10
Anthropic models are just the best, even if the Chinese models match it in benchmarks.
Diogo22:40
ai is so clearly awesome yet software itself hasn't changed
Diogo28:50
benchmarks are total bullshit all benchmarks are bad
Diogo43:19
I believe that AI has been a bubble for a while. So I would like to keep planning for an eventuality of bubble burst.
Diogo49:39
What we need is allies to fight the largest companies in the world.
Diogo57:56
We don't need the money or the valuation.
Diogo58:56
Figures
| Time for Jev to reach $100M | Seven days | 1:00 |
| TypeSafe valuation | $7.5 billion | 1:00 |
| Term sheets received this round | 28 | 56:53 |
| Jev usage among Fortune 500 | One-third | 1:00 |
Glossary
- Jevons paradox
- Efficiency gains lower per-unit costs, raising total consumption
- RLHF
- Training models by scoring outputs based on human preferences
- RLVR
- Training models by scoring outputs based on automatically verifiable results (like unit tests)
How to listen
Founders and investors focused on reliability in enterprise AI deployment, the debate over moats for model companies, and the fundraising and hiring logic of top AI firms.
Opening and closing ad reads (Blitzy, Flex, Vanta) can be skipped.