What Is Jev? The AI Model That Doesn't Generate Text

Credibility score: 72/100 — Mostly Credible. Mixed credibility - some claims are solid, others need verification.

BSmeter analyzed "What Is Jev? The AI Model That Doesn't Generate Text" and rated it 72/100 for credibility (a BS score of 28/100 — mostly credible), on 2026-10-01. Its weakest claim — "Claiming many software decisions are quick 'judgment calls,' not System Two." — scored 45/100 and was flagged as confidence mismatch. 27 claims were checked against the video transcript. Scores are produced by BSmeter's AI analysis of the transcript, not independent human verification.

Of 27 claims analyzed: 0 scored under 40, 3 between 40 and 69, and 24 at 70 or above.

Claims analyzed

Setting the stage for a new AI model, Jev. — No Frame (75/100)

At 0:00

Just an intro, setting the stage for what's to come. No tricks here, yet. 😈

Why this score: The speaker is simply introducing the topic of the video, 'Jev,' without making any claims that require scrutiny at this point. It's a straightforward opening.

Original quote: “All right, I want to talk about Jev.”

Highlighting Jev's unusual characteristic of not generating text. — No Frame (75/100)

At 0:08

Describing a core feature. 'Unusual' is just a descriptor, not a claim. 😈

Why this score: The speaker is describing a key, differentiating feature of the Jev AI model. This is presented as a characteristic, not a claim requiring external validation, and it's framed neutrally.

Original quote: “And it's a bit unusual because it's an AI model that doesn't actually write any text.”

Labeling Jev as a 'system one model' borrowed from Daniel Kahneman. — No Frame (75/100)

At 0:21

Connecting it to Kahneman's 'System One' — a known concept. Just a reference. 😈

Why this score: The speaker is categorizing Jev using terminology ('system one model') borrowed from a well-known figure, Daniel Kahneman. This is a classification and a reference, not a claim that is being framed to mislead. It's an attempt to provide context through an established concept.

Original quote: “Now type safe causes a new type of AI model. It's a system one model, which is a name borrowed from Daniel Kahneman”

Explaining Kahneman's System One thinking — straightforward setup. — No Frame (75/100)

At 0:30

He's just laying out the groundwork from Kahneman's book. No tricks here, just setting the stage. 😈

Why this score: The speaker is accurately describing Daniel Kahneman's concept of System One thinking from 'Thinking, Fast and Slow,' using a simple, universally understood example (2x2=4). This is a foundational concept for his later points, presented without manipulation.

Original quote: “Kahneman describes two modes of thinking, human thinking, that is, and the first of those modes is called system one thinking. And system one thinking is fast and automatic. So what's two times two? You instantly know the answer is four without having to really work it out in your head. That system…”

Defining System Two thinking with a quick teleprompter joke. — No Frame (75/100)

At 1:02

Still just explaining Kahneman, and a little self-aware joke about the teleprompter. I appreciate the honesty. 😈

Why this score: The speaker continues to explain Kahneman's System Two thinking, contrasting it with System One. The example (17x24) clearly illustrates the 'slow and deliberate' nature. The self-deprecating joke about the teleprompter adds a human touch without undermining the core explanation.

Original quote: “The other type of thinking is classified as system two thinking. And that's slow and deliberate. So if I ask you what's 17 times 24? Well, you're probably going to need to work through that one step by step to figure it out. Or just look at the answer on your teleprompter, which is 408. There we…”

Claiming many software decisions are quick 'judgment calls,' not System Two. — Confidence Mismatch (45/100)

At 1:59

He's pivoting from human thinking to 'a lot of decisions in software' with zero data. That's a leap, not a bridge. 💀

Why this score: The speaker makes a broad generalization about 'a lot of decisions inside software' being 'quick judgment calls' rather than System Two thinking. While this might be true in some contexts, it's presented as a definitive statement without any supporting evidence, examples, or specific metrics to back up the 'a lot' claim. It's an assertion made with confidence that isn't yet earned.

Original quote: “But, but, a lot of decisions inside software are not really system to they're just kind of quick judgment calls.”

LLMs lack reliable confidence scores — a problem Jev solves. 😈 — No Frame (75/100)

At 3:42

He's laying out a real limitation of current LLMs. No trick here, just the cold, hard truth of their internal uncertainty. 🔥

Why this score: The speaker accurately points out that while LLMs can generate answers, their self-reported confidence scores often don't reflect true probabilities. This is a known challenge in AI interpretability and reliability, making it a straightforward, factual observation. Jev is presented as a solution to this. It's a clean setup for Jev's utility.

Original quote: “Now that mostly works, but it doesn't come back with a reliable sense of how sure the model is. How confident is a large language model that it got the answer to these questions right? Well, we could ask the model, but whatever it responds with wouldn't necessarily actually match the real…”

Jev 'comes up with probabilities' and takes in 'state' data. 😈 — No Frame (75/100)

At 4:04

He's explaining Jev's core function and input. It's a direct description of the model's design. No smoke and mirrors yet. 😈

Why this score: The speaker is introducing Jev's unique selling proposition: its ability to provide probabilities and its input mechanism ('state'). This is a foundational explanation of the model, not a manipulative claim. It's setting the stage for how Jev differs from traditional LLMs.

Original quote: “But Jev well, give does come up with probabilities. So if we send these questions to Jev, we're going to get a very different looking response. So it takes in two things. It takes in what's called state which is the data the decision is about. So that might include our support email.”

Jev's speed advantage over LLMs due to non-text generation. — No Frame (75/100)

At 5:13

He's laying out the core difference here, no tricks. Jev's design is the whole point. 😈

Why this score: The speaker is explaining a fundamental architectural difference between Jev and traditional LLMs, specifically how Jev's lack of text generation allows for simultaneous answer return, which is a direct, verifiable design choice. This is a straightforward explanation of a technical feature.

Original quote: “Now all three of these questions go in in a single request. And whereas an would write it Json answer out one token at a time, Jev doesn't generate any text. So all three answers, they all come back at once.”

Jev is 'so much faster and cheaper' than LLMs for these questions. — Confidence Mismatch (45/100)

At 5:32

He says 'so much faster and cheaper' like it's a given — but where's the actual comparison? Show me the numbers, mortal. 💀

Why this score: The speaker makes a strong, unqualified claim about Jev being 'so much faster and cheaper' without providing any specific metrics, benchmarks, or comparative data. While the previous point about simultaneous answers suggests a potential for speed, the magnitude of 'so much' is asserted without evidence, leaving it as a confident statement rather than a data-backed one.

Original quote: “And that's a big part of why this is so much faster and cheaper than an LLM on these types of questions. Now we provide questions as input. What is coming back? I said Jev doesn't give you text.”

Jev outputs probabilities, like a 90% chance for a refund request. — No Frame (75/100)

At 5:42

Explaining the output format with a clear example. No smoke and mirrors here, just the facts. 😈

Why this score: The speaker is clearly defining the output format of the Jev model, using a hypothetical but illustrative example (0.9 probability for a refund request). This is a direct explanation of how the model functions and what kind of data it provides, without any manipulative framing.

Original quote: “What it gives you coming back are probabilities as the output. So for refund it's going to be a single number. Let's say it's going to come back with 0.9, meaning there is a 90% chance that yes, this is a refund request.”

Explaining the RLHF process and its side effect of overconfidence. — No Frame (75/100)

At 6:40

He's laying out the process of RLHF and its known flaw. No tricks here, just the ugly truth of how these things learn. 😈

Why this score: The speaker is describing a well-documented characteristic of AI models trained with Reinforcement Learning from Human Feedback (RLHF), where models can develop a tendency to sound confident even when incorrect, as humans often prefer confident-sounding answers. This is a straightforward explanation of a technical aspect.

Original quote: “Now the model can still pick the wrong options or score any of these questions incorrectly of course, but the probability tells us how likely that is. So how do you train a model to give accurate probabilities like this? Well, let me go back to training traditional LLMs for a moment. Most language…”

Explaining RLVR as a form of reinforcement learning with verifiable rewards. — No Frame (75/100)

At 8:31

Just laying out the basics of RLVR. No tricks here, just definitions. 😈

Why this score: The speaker is defining 'Reinforcement Learning with Verifiable Rewards' and giving examples of how it's scored automatically. This is a straightforward explanation of a technical concept.

Original quote: “But that's just one form of reinforcement learning. Another form is RLVR. That's reinforcement learning with verifiable rewards. And the scoring here is automatic. Did the model get the right answer to the math problem. Does the code it wrote pass its unit tests those sort of things.”

Claiming RLVR is why reasoning models are good at math and coding. — No Frame (75/100)

At 8:56

Connecting the dots between RLVR and model performance. Seems like a logical link. 🔥

Why this score: The speaker is making a direct causal link between the verifiable rewards in RLVR and the improved performance of reasoning models in math and coding. This is a reasonable assertion within the context of AI development.

Original quote: “And that's a big reason that reasoning models today have got so good at math and coding.”

Highlighting RLVR's limitation: only checking final answers, leading to slowness and expense. — No Frame (75/100)

At 9:02

Pointing out the inherent trade-offs and limitations of the current method. Fair enough. 💀

Why this score: The speaker is identifying a specific limitation of RLVR: it only rewards the final correct answer, not the confidence or efficiency of the process. This leads to models being slow and expensive, which is a common challenge in AI development. It's a direct observation of a known issue.

Original quote: “But the reward here only checks whether the final answer was right. It doesn't reward the model for knowing how sure it should be. And those models spend a while working through their chain of thought before they come up with an answer, and that can make them a little bit slow and expensive to run.”

Introducing Jev's RLCD training, acknowledging limited public architecture details. — No Frame (75/100)

At 9:18

Introducing a new concept and admitting the lack of public info. Refreshingly honest. 😈

Why this score: The speaker introduces 'RLCD' as the training method for Jev and explicitly states that TypeSafe hasn't published much about Jev's architecture. This transparency about limited information is a 'No Frame' move, as it sets realistic expectations for the audience.

Original quote: “Now Jev is trained with a different form of reinforcement learning. It's called RLCD reinforcement learning for calibrated decisions. Now type safe haven't published much about Jev's architecture, so I stick to what they have said.”

Explaining Jev's reward system: probabilities for options, rewarded when correct. — No Frame (75/100)

At 9:37

Defining how Jev's reward system works. Straightforward explanation. 🔥

Why this score: The speaker is explaining the core mechanism of Jev's reward system: it assigns probabilities to options and is rewarded when those probabilities are accurate. This is a factual description of the model's function, as presented by its developers.

Original quote: “Well, we know that Jev gives a probability for each option, and it's rewarded when those probabilities turn out to be right.”

Explaining 'calibrated' means an 80% probability prediction is 80% correct. — No Frame (75/100)

At 10:20

Defining 'calibrated' with a clear example. Simple, direct, no hidden agenda. 😈

Why this score: The speaker is illustrating what 'calibrated' means in the context of Jev by giving a concrete example: an 80% probability prediction means it's correct 80% of the time. This is a clear and concise explanation of a technical term.

Original quote: “So when we say that there is a 80% probability that 80% probability is going to be correct”

Explaining 'calibration' with a clear, simple example. — No Frame (75/100)

At 10:30

A straightforward explanation of what 'calibration' means in this context. No tricks, just laying out the premise. 😈

Why this score: The speaker defines 'calibration' with a direct, easy-to-understand example. It's a foundational concept for their argument, presented without any rhetorical manipulation.

Original quote: “that 80% probability is going to be correct about 80% of the time, that's calibration.”

Setting up the probability threshold system for Jev. — No Frame (75/100)

At 10:46

Just explaining the system's logic, setting the stage for how Jev makes decisions. It's clean. 😈

Why this score: The speaker is clearly outlining the mechanism by which Jev uses probability scores to inform code, establishing a threshold system. This is a direct explanation of functionality, not a persuasive claim.

Original quote: “So if we have probabilities we can use them to inform code using a threshold like this. So we've got a probability threshold here. One is certainty zero is definitely not kind of thing.”

Jev automatically processes refunds if the score is above 0.9. — No Frame (75/100)

At 11:05

A clear, conditional statement about how the system works. No hidden agenda here. 😈

Why this score: The speaker is detailing a specific operational rule for Jev: if a probability score exceeds 0.9, the system automatically processes it as a legitimate refund request. This is a direct description of the AI's intended behavior.

Original quote: “So if a probability score is above let's say point nine. So maybe our score comes back here somewhere. That means we can treat this support email as a legit refund request. And we can send that straight to the refund queue for processing.”

Jev's system routes uncertain requests to humans and dismisses low-score requests. — No Frame (75/100)

At 11:29

Explaining the full decision tree, including human intervention and dismissal. It's all laid out. 😈

Why this score: The speaker continues to describe Jev's operational logic, detailing how scores between 0.1 and 0.9 trigger human review due to uncertainty, and scores below 0.1 are dismissed as not being refund requests. This is a transparent explanation of the system's design.

Original quote: “If, however, the score comes back anywhere kind of around between here and let's say point one, well, there's some uncertainty here. So the code that received a number like this would know to actually get a human involved. So now we need to bring in a person to actually check that because we're not…”

Jev's current limitations are stated plainly. — No Frame (75/100)

At 12:42

He's laying out the current limitations of Jev without any spin. Rare honesty. 😈

Why this score: The speaker is upfront about what the AI model 'Jev' can and cannot do, specifically mentioning its current inability with math and counting. No attempt to obscure or downplay these facts. It's a straightforward admission of current capabilities.

Original quote: “But there are limits. Jev only takes text in as of today. It's not very good at math”

Acknowledging Jev's vulnerability to hidden instructions. — No Frame (75/100)

At 12:51

Another honest admission about AI vulnerabilities. No smoke and mirrors here. 😈

Why this score: The speaker directly states that Jev, like other AI models, can be tricked by hidden instructions. This is a clear, unvarnished statement about a known vulnerability in AI systems, not an attempt to mislead or frame. It's a simple, factual observation.

Original quote: “And like any AI model, it can be tricked by instructions hidden in the data is reading.”

Suggesting collaboration between Jev and large language models. — No Frame (75/100)

At 13:05

He's just proposing a practical integration. No trickery, just a suggestion. 😈

Why this score: The speaker is proposing a complementary relationship between Jev and large language models, suggesting they can 'work together.' This is a straightforward suggestion for system design, not a manipulative framing technique. It's a practical idea, not a claim to be fact-checked for truth, but for its framing.

Original quote: “But the two can work together.”

Explaining Jev's namesake and the historical paradox. — No Frame (75/100)

At 14:03

Giving the origin story of the name and the paradox. It's just background info. 😈

Why this score: The speaker provides a historical anecdote about William Stanley Jevons and Jevon's paradox, explaining the origin of the AI model's name. This is a factual historical reference and an explanation of nomenclature, not a claim designed to persuade through manipulation. It's a direct, informative statement.

Original quote: “Jev is named after William Stanley Jevons, who was an economist who pointed out in 1865 that as steam engines got more efficient, Britain actually used more coal.”

Hypothesizing Jev's potential uses — pure speculation, no data. — Just Vibes (50/100)

At 14:32

He's throwing out 'who knows?' like it's a question, but it's just a shrug for a future that isn't here yet. 🤷‍♂️

Why this score: The speaker is speculating about potential future applications for Jev, using phrases like 'it can go into places' and 'who knows?'. This isn't a factual claim about current capabilities or confirmed future uses, but rather a hopeful projection. It's pure conjecture, not evidence.

Original quote: “Like with Jev, well, it can go into places where an LLM would be too slow or too expensive today, like in every row of a database or an every line of a log file. Who knows?”

See the full analysis with timestamps →