Diogo Almeida finds AI's blind spot but misprices the liability
By Klaus Berger · Reporting from Frankfurt ·
On the a16z podcast, TypeSafe founder Diogo Almeida rightly mocked software that cannot manage a drive-through, but replacing rigid code with probabilistic guesswork simply shifts the burden of error onto the public.
After capital expenditure, artificial intelligence can solve complex theoretical benchmarks. Yet it still cannot reliably handle an order at a fast-food drive-through. On the latest episode of the a16z podcast, Diogo Almeida, founder of TypeSafe, posed a vital question. Corporate treasurers and public auditors ought to be asking it: where is all the actual automation? OpenAI attempted to automate basic customer support, while enterprise software remained largely unchanged. Almeida conceded that the sector has fallen into an unsustainable cycle of overpromising and underdelivering.
His venture partner host, Martin Casado, was eager to present Almeida’s new product, Jev, as the antidote. TypeSafe proposes embedding probabilistic intelligence directly into program internals rather than using external coding assistants to churn out mediocre boilerplate faster. This marries natural language classification to deterministic state machines. In Casado’s phrase, the goal is to build "prod not god." This transforms corporate software into a resilient engine that acts on developer intent.
The Strongest Case for Probabilistic Logic
The ablest argument for this approach deserves a fair hearing. Conventional software engineering is expensive and brittle. Writing hard-coded conditional rules to parse human reality requires vast teams of engineers. It demands months of maintenance and endless edge-case patches. Businesses could finally automate clerical workflows that currently stall inside human-managed queues. Software components would natively classify intent and output state transitions based on statistical confidence. Casado pointed to an "inverse apocalypse." Established enterprise software providers become dramatically more useful rather than obsolete. This would finally rid the world of rigid multiple-choice forms and bespoke rule engines.
If this worked cleanly, the economic dividend would be immense. Productivity gains have long eluded European and American balance sheets. They might finally show up in output figures rather than mere technology valuations.
The Rulebook Always Penalises Uncertainty
Yet this optimism founders on the most basic principle of ordoliberal economics: the liability principle. Walter Eucken observed a key truth. An economic system remains functional only when those who reap the gains also bear the full cost of failure. Embedding probabilistic classifiers into the core architecture of software does not eliminate risk. It obscures where the risk settles.
When traditional software fails, there is a clear chain of accountability. An engineer wrote a faulty deterministic instruction. An audit log traces the bug, and the vendor or firm absorbs contractual liability. In Almeida’s probabilistic vision, software makes decisions based on confidence levels and statistical thresholds. An automated classification might misroute a municipal welfare disbursement, misinterpret an invoice dispute, or fail an air traffic protocol. Who holds the liability then?
A system that operates on approximate confidence rather than deterministic verification creates moral hazard. Vendors will sell probabilistic efficiency upfront. Meanwhile, ordinary citizens, clients, and municipal administrations absorb the downstream friction of resolving statistical errors. No regulatory framework in Frankfurt, Brussels, or Washington will permit public utilities or critical infrastructure to operate on vague notions of "robustness." An algorithm cannot merely be expected to behave in an "understandable" fashion.
The Historical Record on Silver Bullets
History offers a sobering warning. In the first wave of probabilistic programming, the technology promised to resolve the messy ambiguity of human systems. It was abandoned because commercial accounting and legal compliance demand absolute auditability. You cannot balance a ledger or satisfy a bank regulator with an approximate confidence interval.
Silicon Valley may celebrate replacing deterministic logic with probabilistic classifiers. However, capital markets will eventually demand the same boring discipline that governs any utility. Software that trades deterministic predictability for statistical guesswork does not eliminate enterprise friction. It simply socialises the cost of mistakes onto the public. Automation must answer to a legal rulebook and guarantee institutional liability. Until then, the drive-through will remain stubbornly manual—and rightly so.