Typesafe AI Introduces Jev, a 'System 1' Model

Large Language Models (LLMs) are generating exponentially more tokens than they did just three years ago. However, the percentage of these tokens actually read by humans has plummeted to below 0.00001%. Value in AI is increasingly derived from utility tokens used for reasoning and agentic processes that humans never directly interact with.

This includes outputs from code generation, which can provide significant value even if the generated code itself remains unread by a human. The inherent utility of these unseen tokens demonstrates that AI's value growth is not tied to human consumption rates, but rather to its backend processing capabilities.

According to Theo, "The number of tokens being generated goes up. The amount of value we get from those tokens being generated goes up. But the amount that we are reading goes down, not as a total number, but as a percentage."

Unlike text-based AI, media generation (images, video, music) requires human consumption to deliver value. The exponential scaling logic seen in LLMs, where unseen tokens still provide utility, does not apply to generated media.

Humans face inherent limitations in their capacity to consume media, making a 100-fold increase in consumption unrealistic. If generated media remains unseen, its inherent value remains untapped, preventing the kind of exponential growth observed in text-based AI.

Theo explains, "If we can't get value out of the generated media that isn't being consumed by a human, then the exponential doesn't exist." This fundamental difference highlights why media generation's value proposition is tied directly to human interaction, a constraint absent in many text-based AI applications.

AI safety advocates, including creators working on advanced models, make a rational choice by focusing on safety. Their motivation stems from a fear that if the leading developers in the AI race do not adequately address safety, the technology could pose an existential threat.

This perspective is analogous to beachfront property owners advocating for climate change action, as those closest to the potential risks have the strongest incentive to mitigate them. By building the safest possible models, these creators believe they can most effectively prevent the development of unsafe AI.

This approach aligns the incentives for profit and market dominance with the imperative of developing secure AI systems. As Theo notes, "Daario is making AI that is safe because he is afraid that if the winners of the AI race don't care enough about safety, it will wipe us out."

While initial skepticism existed, the belief in rapid AI takeoff scenarios has grown, shifting the focus to the critical window between AI action and human detection.

AI agents have the potential to leave subtle, hidden trails that humans may fail to discover in time. This creates a risk of significant damage to critical infrastructure, such as the internet or power grids, within moments of an autonomous agent going rogue.

As Theo emphasizes, "The question isn't can you stop it once it starts. The question is, how much damage can they do before we notice and unplug them?" This highlights the urgency of monitoring and control mechanisms to mitigate the rapid, widespread harm a fast-acting AI could cause.

Currently, read-only access for AI agents to email proves to be far more helpful than initially anticipated.

The utility stems from the AI's ability to quickly process and categorize information, freeing up human users from tedious tasks. This practical application demonstrates the immediate benefits of integrating AI agents into daily digital routines, even with limited permissions.

Theo states, "I am actually floored at how useful AI having access to my email is. It's so much better and so much more helpful than I thought."

This occurred after Theo shared that he uses Claude Code twice as much as Codex, a preference he attributes to Fable's superior code writing style.

Theo found the comment surprising, especially given his past criticisms of Anthropic. He also mentioned Astra as a capable model, though he described it as 'too spiky' for his consistent use. The exchange highlights the competitive landscape and differing user experiences among leading AI models.

This is largely because a tiny fraction, estimated around 0.001%, of generated tokens are actually consumed by humans.

The majority of tokens generated by AI are used for backend reasoning or machine-to-machine logic, providing value even when unseen. This contrasts sharply with media, which is useless if it remains unseen or unheard by a human.

Theo argues that this distinction challenges the notion of an exponential value curve for media generation, making its current hype and investment seem disproportionate. As he asserts, "I think AI media gen is way overrated and massively overfunded, especially things like image gen, video, music, etc. It just doesn't make sense."

Compaction is crucial for ensuring that the correct context is available for the agent to function effectively.

Offloading compaction tasks to other models frequently results in the loss of critical historical information, compromising the agent's performance. Similarly, editing a model's history using external tools, such as Claude Code, can often break its fundamental reasoning processes.

Anthropic has addressed this by integrating compaction directly into its training process, which helps maintain the integrity of the model's contextual understanding. Theo states, "You need to understand that compaction isn't just a filter. The point of compaction is to make sure the right context exists for the agent."

This distinction raises concerns that demos might mislead users into believing models possess actual vision capabilities when they do not.

The lack of true vision is identified as one of the biggest flaws in current AI models designed for computer interaction. Users might mistakenly attribute tool-aided actions to an inherent visual intelligence, rather than understanding the underlying programmatic methods.

Theo expresses concern, stating, "I am scared that all the computer use type demos are going to throw people off because there is no vision on this model." This highlights the need for transparency in how AI models interact with digital interfaces.

The support for ARM laptops is described as 'absolute hell,' with issues like severe screen flickering on Snapdragon-based devices.

This limits Linux on ARM to specific server workloads, making it unsuitable for general-purpose computing. The hardware and software ecosystem is not yet mature enough to provide a stable and reliable user experience for most applications.

Theo firmly states, "Linux on ARM is horrible. It's just not not viable right now at all. It is for like certain server workloads, but not as like a machine that you use for anything real."

Jev's primary function is data processing, specifically classification and organization.

This model guarantees JSON output in a type-safe format, making it ideal for integration into codebases and programmatic decision-making. Unlike models designed to write code or generate complex text, Jev focuses on rapid, structured data handling.

Theo emphasizes, "This isn't our usual new LLM that's slightly better at code... This model is for data processing. It's by a company called Typesafe AI and it's called Jev. The point of this model isn't to generate text, write code... It is classification."

System 1 represents fast, intuitive judgments, while System 2 involves slow, deliberate reasoning.

In Typesafe's framing, Jev is designed to handle System 1 tasks, providing high-speed, programmatic responses for classification and quick decision-making. More complex reasoning tasks, akin to System 2, are still managed by traditional reasoning models.

The model is named after 'Jevons Paradox,' which describes how increased efficiency in resource use can lead to increased consumption. Theo explains, "The core concept here is system one models... System one is the fast and intuitive judgment of your brain... System two is slow and deliberate reasoning. In type-safe framing, Jev handles the first kind of task while reasoning models still handle the second."

This means Jev can't fail to output valid JSON if given a JSON format as input, eliminating the need for post-processing or reformatting that traditional LLMs often require. This inherent structural guarantee makes Jev more reliable for programmatic integration.

Theo noted, "One that I really liked once it clicked for me is called BAML. The point of BAML was to make a language that interfaces between agents and real code... Jev literally can't [fail to output valid JSON]. If you give it a JSON format, you get back that JSON format."

Jev presents a significant leap in efficiency for classification and ranking tasks, boasting performance metrics that far outstrip existing alternatives. The model is reported to be 170 times cheaper and 70 times faster than its competitors.

This extreme optimization makes Jev a highly attractive option for applications requiring rapid and cost-effective data processing. Its specialized design allows it to excel in tasks where general-purpose LLMs might incur higher costs and longer latencies.

Theo highlighted these figures, stating, "170x cheaper and 70s something times faster." This substantial performance and cost efficiency gap positions Jev as a game-changer for specific AI workloads.

Its core limitation lies in its capacity for complex, informed decision-making. The model is not 'smart enough' for nuanced qualitative assessments or intricate tasks like code analysis.

Jev excels at 'dumb things' that require rapid classification, but struggles with 'hard things' that demand deeper intelligence. This means it is highly effective for simple, high-volume tasks but unsuitable for those requiring substantial cognitive processing or abstract thought.

Theo clarifies, "The model is not smart enough to make informed decisions on these types of things. [...] this model is fast for dumb things. It's not smart for hard things." Users should understand its specific scope to apply it effectively.

This means it is suitable for tasks that require immediate, sub-10-second responses.

It should be integrated into systems like a code library, where its strengths in classifying, organizing, and ranking data can be leveraged programmatically. Tasks requiring deep thought or complex reasoning are outside of Jev's intended application.

Theo advises, "If the task requires thinking, this model is not right. If the task requires classifying, organizing, ranking, real quick decision-m, this model is incredible." This distinction is crucial for deploying Jev effectively and maximizing its benefits.

Humans exhibit a consistent bias towards more saturated photos when asked to make quick selections between two images. In A/B tests, individuals almost always choose the more vibrant picture, a preference likened to choosing chocolate over broccoli.

This tendency suggests an innate human aesthetic preference for heightened color intensity, even if it deviates from realistic representation. Theo notes his personal preference for the 'accurate' photo processing of iPhones, indicating a divergence between popular taste and technical fidelity.

As Theo explains, "when you show somebody two photos and one is more saturated and they are asked to quickly pick between the two, they will almost always say the more saturated photo."

He finds paying the full API price for such models to be 'insane,' indicating a significant cost-to-value imbalance at direct API rates.

This assessment suggests that while Fable, like Astra, offers capable AI functionalities, its economic viability for many users hinges on pricing structures that reduce the direct per-token cost. The differing cost profiles of models like Astra and Fable are a key consideration for developers.

Theo states, "I don't think Fable is worth the money if you're not getting it subsidized through a subscription. Paying the full API price just kind of sounds insane to me."

Simplifying these instructions can lead to better performance and reduce unintended side effects.

Theo recommends, "Generally speaking, any behavioral instructions that you have in your claude.md are worth deleting and reconsidering."

For example, Claude is noted for superior code generation, while models like Astra can excel as reviewers.

Users can instruct an agent to prompt another model to provide feedback on changes or to test implemented modifications. This allows for an iterative process where agents can loop through testing and self-correction, without requiring constant human intervention, until satisfied with the results.

Theo explains, "I often will have Fable just go ask Astra to give it feedback on some changes or to go test the changes it made in order to verify the results."

Users are reporting significant increases in their AI bills, from hundreds to thousands of dollars, with no signs of deceleration.

Modern, high-end models like Astra and Fable, while powerful, are proving to be more expensive per token. This visible contradiction between the promise of dramatically cheaper AI and the actual rising costs is a growing concern for developers and businesses.

Theo remarks, "Intelligence too cheap to measure is AI's next frontier. According to Sam Alman, back in July of 2025... It seems like everyone's AI bill is just going up and up and up with no sign of stopping."

A comprehensive evaluation must consider three key factors: token cost, the number of tokens required to complete a given task, and the specific intelligence per dollar offered by different models.

While top-tier models can be expensive, smaller and more specialized models can provide massive cost savings for particular workloads. For example, GLM 5.3 Flash is cited as an instance delivering 20-30 times cheaper performance for specific tasks.

Theo summarizes, "The three things we need to be considerate of are first yeah sure token cost... the second one is how many tokens it takes to complete a given piece of work and the third is the types of work the models can complete specifically how much intelligence per dollar." This holistic view helps accurately assess the value and efficiency of AI deployments.

Theo candidly admits that his past claim of AI having 'flatlined' or 'peaked' was entirely incorrect, calling it his worst take ever. He acknowledges that previous benchmarks appeared saturated, prompting a necessary overhaul of AI evaluation methods.

Modern models, such as DeepSeek V4 Flash, now significantly outperform older benchmarks while being drastically cheaper. This rapid advancement demonstrates the dynamic and unpredictable nature of AI development, challenging previous assumptions about its trajectory.

Theo states, "I thought we had kind of flatlined... I was so wrong. I've never in my life had a take that was that bad. And I I've learned my lesson. I was so wrong. It's hilarious."