TypeSafe's Jev Hits 1 Trillion Tokens a Day as Founder Rejects Benchmarks, Refusals, and RLHF in Favor of a New AI Category
Launch-week interview reveals the technical philosophy and business bets behind "system one" AI models, aimed at machines rather than chatbots
TypeSafe, the AI startup founded by former OpenAI researcher Diogo, used the launch of its new model line, Jev, to introduce what it calls a new category of artificial intelligence entirely: "system one models," designed to be consumed by code rather than humans. In an extended interview conducted during launch week, Diogo laid out a set of unusually blunt positions on benchmarks, safety alignment, fine-tuning, and the current state of frontier AI research that investors tracking the space will want to understand, regardless of whether they ultimately back the company.
A New Model Category, Built for Machines Not Chatbots
The core thesis of TypeSafe is that the entire industry has been optimizing large language models for the wrong consumer. Pre-trained models were built for autocomplete, RLHF-tuned chatbots like ChatGPT and Claude were built to satisfy human raters, and reasoning models optimized with RLVR were built to solve benchmarkable problems. Jev, Diogo argues, is the first model built explicitly so that "code is the consumer." He described it as "machine-native, large, programmable," and said the naming — a reference to Jevons paradox — reflects the company's singular focus on intelligence per dollar. "Jev will be the name of models that will be on the frontier of intelligence per dollar," he said, adding that reliability, cost, calibration and speed are all in tension and that TypeSafe has chosen to go "all out" on the intelligence-per-dollar axis specifically.
The company frames this as solving what Diogo calls the central paradox of the current AI moment: models capable of tackling millennium-prize-level math problems still cannot automate basic, repetitive knowledge work. "We have this supercharged engine of automation that just does not have the right plugs and stuff to plug into all of this economically valuable work," he said, arguing that the gap is not a capability problem but a design and interface problem.
Adoption Numbers Ahead of the Company's Own Expectations
Perhaps the most concrete data point of the interview: Jev has already surpassed 1 trillion tokens processed per day, a figure Diogo emphasized reflects genuine machine-to-machine usage running continuously, "even at night," rather than one-off human testing. The launch video itself has drawn 38 million views, which Diogo's co-host noted puts it ahead of Nvidia's widely watched Jensen Huang keynote moments (74 million lifetime) on a per-week basis and ahead of other major recent tech launches such as Fable's demo (57 million). That said, Diogo was candid that the company was unprepared for the reception. He said that prior to launch, "more than half the people we had play with it just did not get it," and that the non-technical side of the company feared they were "selling a vitamin and not a painkiller." TypeSafe had almost no revenue before this week. He also pushed back on the idea that the launch reflected marketing savvy: "We don't have a marketer, also hiring," he said, adding that waitlist signups are a vanity metric the company has learned to largely ignore — the real signal, he said, is how aggressively users ask for higher rate limits once they experience the product, since that indicates genuine dependency rather than curiosity.
No Public Benchmarks, By Design
TypeSafe has taken an unusually adversarial stance toward the industry's standard credibility signal: public benchmarks. Diogo said the company is "extremely anti-public benchmark" and only "medium" in favor of private benchmarks used as internal proxies, arguing that public benchmarks are trivially gamed even by labs that don't intend to game them. "Back in the old days, every lab had a team to collect data that looks like MMLU to make it look better, which is just benchmarking with extra steps," he said. Instead, TypeSafe is betting that trust in model intelligence will be established through what he called "vibes" — real-world reliability demonstrated inside actual workflows — rather than leaderboard placement. He acknowledged the company maintains internal evaluations but said guarding against gaming those evaluations is "one of the most important things" the team enforces internally, given how easily incentives corrupt measurement.
Refusals Are "A Type Error," Not a Safety Feature
One of the more provocative positions Diogo staked out is his rejection of safety-alignment-style refusals within an API product, which he distinguished sharply from consumer chat products like ChatGPT or Claude, where he says such guardrails make sense. In an API context, he argued, a refusal breaks software non-deterministically and unpredictably. "If you're chatting with a bot and a refusal happens, that's annoying — but you can work with it," he said. "If this is a dependency running in the background, what happens if that refuses? That is straight up insanity." He was careful to note this is not a blanket objection to safety as a concept, but a belief that capability alignment (making models follow developer instructions precisely) and safety alignment (making models follow someone else's preferences, such as the platform's) are fundamentally different problems that should not be conflated at the model layer. On the sensitive question of models being used in warfare, he said he personally would prefer the technology not be used to hurt people but declined to build that preference into the model itself, arguing doing so "fractures" the model's general intelligence.
A New Post-Training Paradigm the Company Calls RLCD
Diogo, who worked on the RLHF-based InstructGPT effort at OpenAI — the predecessor to ChatGPT's instruction-following behavior — argued that the industry has only successfully established two or three genuinely new training "north stars" in the history of large language models: RLHF (optimizing for human preference and instruction-following) and RLVR (optimizing for benchmark-verifiable reasoning). TypeSafe is proposing a third, which it internally calls RLCD, oriented around programmatic reliability rather than pleasing a human rater or solving a verifiable puzzle. Notably, the company has not yet published a paper describing the technique. Diogo also used the interview to explain why he believes RLHF-tuned chat models are structurally prone to sycophancy, overconfidence, and hallucination — describing this as a consequence of "mode dropping," where models become overconfident to avoid being penalized for hedging, distorting their underlying probability distributions. He connected this to Yann LeCun's well-known critique that autoregressive models are doomed to compound errors over long sequences, calling LeCun "among the most accurate" thinkers in AI but arguing the mechanism is calibration collapse from RLHF rather than an inherent flaw in token-by-token generation.
Skepticism Toward the Industry's "Pace the Frontier" Safety Consensus
Asked about growing alignment among frontier labs around slowing AI capability progress for safety reasons, Diogo offered a pointed rebuttal, arguing the framing assumes every lab must keep scaling RLVR to remain competitive — an assumption he rejects for TypeSafe's own model family. "I obviously don't think I need to do more RLVR on our models," he said. "I think zero is the optimal amount for our shape." He characterized the industry's pacing debate as "a bit of a sleight of hand," suggesting that labs granting models broad, unconstrained autonomy during training to maximize capability are also the ones later citing the resulting risk as justification for coordinated slowdowns, without acknowledging that alternative model designs exist. He separately noted, without elaborating on specifics, that some of these safety conversations are shaped by political considerations tied to upcoming elections rather than purely technical risk assessments.
Coding Agents Could Be Reshaped by Multi-Model Architectures
Diogo flagged the coding agent market — currently led by Anthropic's Claude Code and OpenAI's Codex — as an area where TypeSafe's approach could prove disruptive. He noted that both leading products are architected around a single foundation model, which made sense in an environment where all providers offered roughly comparable capability shapes. He argued that many independent open coding agent projects are now integrating Jev and racing to find differentiated use cases that a single-model architecture cannot easily replicate. Separately, Diogo discussed a forthcoming internal document — expanding on a piece he wrote titled "KV Cache Rules Everything Around Me" — arguing that today's coding agents are architecturally constrained by dependence on a single model's key-value cache, which locks agents into one model, discourages proper software decomposition, and complicates sub-agent task delegation, context compaction, and multi-agent state sharing. He said freeing agents from that constraint could unlock materially different, cheaper, and more modular agent designs.
Product Roadmap: No Determinism, No Fine-Tuning (For Now), Possible Model Tiers
On specific product design choices, Diogo confirmed Jev does not support deterministic outputs (same input guaranteeing same output), arguing that "robustness" — similar inputs producing similar outputs — is the more important property, and that determinism would come at a direct cost to the company's core intelligence-per-dollar optimization. He left the door open to offering deterministic variants if customer demand justified the tradeoff. Fine-tuning is also not currently offered, and Diogo was explicit about why, noting that OpenAI, Anthropic, and Google have all rolled back fine-tuning features after finding them to be more of a liability than a benefit: "It could just be a foot gun," he said, though he did not rule out revisiting fine-tuning or offering multiple model sizes in the future as the company's understanding of demand matures. He also disclosed that TypeSafe deliberately avoids training on customer data, even though it could likely obtain rights to do so, because it wants to avoid overfitting the model to present-day use cases at the expense of future, harder-to-imagine applications.
On business use cases, Diogo pointed to "dark data" — large enterprise datasets that were previously too expensive to run through LLMs — alongside coding agents as the company's two largest near-term revenue drivers, with real-time applications (particularly e-commerce, where latency directly affects conversion) and enterprise observability and verification workloads as secondary categories. He was candid that multi-hop reasoning remains a weak point relative to single-hop tasks, and that some demoed capabilities, including full computer and voice control, are not yet reliable enough for production and emerged somewhat unexpectedly from developers rather than the company's own roadmap.
From OpenAI's RLHF Team to a Two-Year Stealth Build
Diogo said the idea for TypeSafe crystallized while working on InstructGPT at OpenAI, when he became convinced that instruction-following chat products, while commercially successful, were a narrow and ultimately limiting application of the underlying technology — one that produced value mostly in categories like AI-generated marketing copy. He said he raised the idea internally at OpenAI, including directly with Sam Altman, but ultimately concluded a startup could move faster than continuing internally. He also disclosed that the earliest version of InstructGPT was trained using an unpublished, custom algorithm he built himself because standard proximal policy optimization was too slow to clean the underlying preference data — a detail not widely known publicly. On the broader wave of new AI labs founded by former frontier researchers, Diogo was unusually harsh, saying "most Neo-labs are crap" and that many lack a clear technical north star, arguing that pure research pedigree without a defined useful task rarely creates value.