Why Jev refuses to write sentences

0:00College football is back. So, Hilton called to me the superstition concierge to make your fan rituals a reality. Need a room to match your lucky number? We got you. Want to make sure our team doesn't wash your lucky jersey? Oh, that smells lucky. Hilton's unmatched hospitality can keep up with any superstition. Even a marching bandwick up call it 555 and 55 seconds. Hit it! When you need a team that will do whatever it takes on game day, it matters where you stay. Hilton, for this day. When you need to build up your team to handle the growing chaos at work, use Indeed Sponsored Jobs. It gives your job post the boost it needs to be seen and helps reach people with the right skills, certifications, and more. Spend less time searching and more time actually interviewing candidates who check all your boxes. Listeners of this show will get a $75 sponsor job credit at ND.com slash podcast. That's ND.com slash podcast, terms and conditions apply. Need a hiring hero? This is a job for Indeed Sponsored Jobs.

1:01Low energy, unfocused, foggy, you might be dehydrated. Gatorade zero powders have been scientifically designed to help improve hydration balance in the body with a clinically proven electrolyte blend and no artificial colors or flavors. Great tasting hydration, no matter where you're going. Shop now online or at retailers nationwide. Gatorade, hydrates better than water. Is it in you? Symptoms noted are signs of mild to moderate dehydration, consult a healthcare professional if symptoms persist. A former open AI researcher named Diego Almeida just built an AI model called Jev for his startup type safe AI. And the financial press is floating valuations in the billions, but the weirdest part isn't the money. It's that Jev explicitly refuses to write a single sentence. I'm right, which goes entirely against the current grain of everything happening in the tech industry right now. Because the entire space is just obsessed with chat bots and generative text. Every product, every interface, and every backend system is being rebuilt

2:01to have this fluid open-ended conversation with you. Yet this specific model operates by just selecting from preset responses and spitting out a strict confidence score. It acts completely counter to the prevailing logic of what an artificial intelligence product is actually supposed to look like today. So why are developers suddenly abandoning conversational language models to use a bot that acts like strict triage nurse? Well, I mean, it really comes down to the underlying economics of what we are actually asking these systems to do in production environments. Would you look at real world applications? Generating free-flowing text is often the exact opposite of what a developer actually needs for a backend system. We have been applying this conversational tool to mechanical sorting problems. And developers are realizing that forcing a conversational AI to do administrative work is computationally expensive and frankly, practically inefficient. Yeah, and you can actually see the friction of that mismatch in the performance data, because polyline swapped

3:03their traditional language models out for Jove in their production environment and found it was 89% faster and 39% cheaper. Oh, wow. And Sam Roam looked at a series of class fire tasks and found an even wider gap. Depending on the workload, moving away from a traditional large language model, made the process up to 400 times cheaper and 200 times faster. Right. And Prithvi.la really crystallized why this gap exists by pointing out a very simple reality of the technology, which is that reading is cheap, but writing is what you pay for. Exactly. We have been forcing language models to write conversational essays. When we really just needed them to sort data, because generating words requires the model to calculate the probability of the next word, token by token over and over again. Every single time it outputs a word, it scans its neural network. It is mathematically exhausting. If you just need a system to verify a condition with a yes or a no, having the model generate a conversational paragraph explaining its reasoning is entirely wasted compute.

4:05You're just burning money on words you don't need. You really have to look at how DotLa tested this concept on Texas government contracts to fully grasp the scale of the waste. He looks specifically at request for proposals. Right, which are incredibly dense bureaucratic documents. Completely. They are filled with legal stipulations, compliance requirements, and specific deadlines. He mapped out 438 complex rules to see how different systems handle the reading and sorting process. He was essentially simulating the kind of grueling administrative work that usually requires a small army of paralegals or compliance officers to sift through line by line. And that is where the math becomes impossible to ignore. And it clarifies exactly why developers are making this shift. Jev cost two cents per document to process. The best open source generative model costs $3.30 per document. Claude Opus cost $4.70. When you scale that up to the thousands of documents of state government processes, the financial implications

5:08are staggering. If you process 100,000 documents using Claude Opus cost nearly half a million dollars. Using Jev costs $2,000. Jev processed 20 times more text than the traditional language models and still came in drastically cheaper. The cost difference isn't a small margin. It is the difference between being able to afford to automate a process and being forced to keep doing it by hand. If you're not subscribed yet, take a second and hit follow on whatever podcast app you're using. It helps us keep making this. We appreciate you being here. It completely redefimes the economic floor of automation. That limits the necessity of generative AI for back-end documents sorting and opens up ultra-cheap automation for legal and bureaucratic review because you are removing the expensive part of the compute process. You can evaluate every single contract a state puts out without incurring the immense overhead of generating text. You are just paying for the system to comprehend the text and output a binary decision.

6:09And think about the everyday administrative bloat inside any large corporation. Think about how many corporate processes are simply reading a document to verify if a specific condition is met. An insurance company processing claims is just checking if the right diagnostic code is present and matches the policy limit. A logistics company is just checking if a supply chain contract includes a forced major clause. Right, you don't need a friendly talkative chatbot to answer those questions. No, you need a switchboard. By eliminating the generation of text, you drop the cost to near zero, which makes entirely new categories of automated administration economically viable. And that exact same principle applies directly to live interactions, not just static paperwork. Matria O'Wogg and Sunkett Bot from Bolna started testing Jeff Unreal phone calls, replacing the traditional conversational language models they have been using to handle live voice interactions with customers. Yeah, Sneha Simanta pointed out that traditional language models are weak at nuanced decision making in those live environments because their output is fundamentally fuzzy.

7:11A chatbot is designed to be creative and conversational, which means it naturally introduces variance. If you ask at the exact same question three times. College football is back. So Hilton called to me the superstition concierge to make your fan rituals a reality. Need a room to match your lucky number? We got you. Want to make sure our team doesn't wash your lucky jersey? Oh, that smells lucky. Hilton's unmatched hospitality can keep up with any superstition. Even a marching band wake up call it 555 and 55 seconds. Hit it! When you need a team that will do whatever it takes on game day, it matters where you stay. Hilton, for this day. When you need to build up your team to handle the growing chaos at work, use Indeed Sponsored Jobs. It gives your job post the boost it needs to be seen and helps reach people with the right skills, certifications, and more. Spend less time searching and more time actually interviewing candidates who check all your boxes. Listeners of this show will get a $75 sponsor job credit at Indeed.com slash podcast.

8:11That's Indeed.com slash podcast, terms and conditions apply. Need a hiring hero? This is a job for Indeed Sponsored Jobs. Low energy, unfocused, foggy, you might be dehydrated. Whether it's hot yoga over 100 degree weather or too much sun, gator light with a specialized blend of five electrolytes was scientifically designed to help move fluids into the body faster for rapid rehydration. Shop now at Retailers Nationwide. Gator light hydrates faster than water. Is it in you? Symptoms noted are signs of mild to moderate dehydration, consult to healthcare professional and symptoms persist. You might get three slightly different responses. That variance is great if you want the AI to write a poem, but it is a massive liability if you are trying to route a customer to the correct billing department. Jeff bypasses that variance by structuring the input and only allowing pre-qualified outputs. Right, because when you rely on a system that generates free text in a customer service setting, you invite hallucinations.

9:14The AI might give you the right answer 90% of the time, but that other 10% of the time, it might invent a completely fictitious return policy or go off on a strange tangent based on a misunderstanding of your prompt. Yeah, by respecting the AI to selecting predefined options rather than generating free text, you get traceable concrete decisions without the random slop that plague so many modern customer service bots. It really forces you to evaluate what kind of user experience you actually want when you interact with the company. If you are calling your internet provider because your router is broken, do you want to chat with an artificial intelligence that pretends to feel sorry for you? Or do you just want the system to instantly and accurately route your problem to the technical support department? A conversational model tries to mimic human empathy. A triage model ignores empathy entirely and just focuses on solving the routing problem perfectly. And the desire to route a problem perfectly is exactly why this architecture becomes mandatory in high stakes environments.

10:16When you move away from a broken internet connection and look at fields where an AI hallucination isn't just an annoyance, but a material risk, the value of a strict, non-conversational model changes entirely. Customer service is one thing, but if you apply this to a field where accuracy is critical, the fuzziness of generative text becomes a severe liability. Which naturally brings us to the ultimate high stakes routing environment, right? Healthcare. Neha Patadia detailed an analogy comparing Jev directly to a human triage nurse. She outlined a scenario involving home blood pressure cuffs. Imagine 10,000 patients sending in daily blood pressure data to a single clinic. A human clinical team cannot review 10,000 readings every single morning. The clear cut normal readings are handled automatically, but the uncertain or abnormal readings are flagged. Jev looks at the data and simply outputs a classification and a score like, does a nurse need to see this? Yes, 87%. The confident score mechanism is the crucial element

11:16in that workflow. It operates fundamentally differently from how something like chat GPT works, because chat GPT was trained via reinforcement learning from human feedback. Human Raiders were shown two answers and told to pick the better one. Humans naturally tend to prefer answers that sound authoritative and certain, so the model learned a bluff. It learned to sound absolutely confident, even when it was entirely wrong. Jev is trained in reverse. It is mathematically rewarded only when it stated confidence matches how often it is actually right. So if Jev says it is 80% confident, it means that out of 100 identical scenarios, it will be correct exactly 80 times. It forces a level of statistical honesty that we just aren't used to seeing from commercial tech tools. The clinical team looking at those blood pressure readings can set the exact threshold for human intervention. They can build a protocol that says anything under a 95% confidence score automatically gets routed to a human doctor. Just think about the societal shift

12:16of an artificial intelligence. Being capable of saying, I am only 40% sure. We have spent the last few years dealing with AI systems that will confidently tell you to put glue on your pizza that it keeps the cheese from sliding off, or systems that will invent legal precedence with absolute certainty, a system that is mathematically calibrated to accurately report its own uncertainty, acts as a reliable filter rather than an unreliable oracle. In healthcare, you don't want a machine guessing a diagnosis to sound smart. You want a machine identifying exactly which patient the human doctor needs to look at first based on mathematical risk. It is a completely utilitarian, unglamorous approach to the technology, and you would assume a highly calibrated silent decision engine isn't exactly the kind of thing that naturally goes viral on social media. Yet, a week prior to launch, type safe AI had 200 followers. The founder, Diogo Almeida, had 1500.

13:17After the launch video dropped, it accumulated 39.5 million views. They secured 150,000 weightless signups in 24 hours, and 13% of paid for cell AI gateway teams started using the model immediately. Yeah, and Nierage Shaw broke down exactly how a niche, non-chatting triage model suddenly took over everyone's social feeds overnight. It was not organic. A group called DOOMERS AI orchestrated 80 to 100 carefully selected engineers, founders, and creators to post their own demos of the tool simultaneously. They perfectly synchronized the post to hit the algorithms at the exact same moment, manufacturing the perception of an absolute consensus. You open your feed and see 50 different engineers and founders you respect all talking about the exact same tool at the exact same minute. Your brain naturally assumes this is the most important breakthrough of the year. It creates this overwhelming sense of inevitability. Following that launch, the company raised $40 million

14:20at a reported valuation of around 200 million with some of the financial press floating that the momentum could eventually push valuations into the billions. It introduces a real tension regarding modern distribution and how tech products actually gained traction today. Obviously, Jeff gave developers something genuinely useful to try, and the product mathematically works, but it asks whether a great product can still win on its own merits, or if orchestrated synchronized coordination is now a mandatory prerequisite for success. You can build the most efficient classifier in the world, saving companies hundreds of thousands of dollars. But if you just post it on a blog and hope for the best, no one will ever see it. The algorithm will completely bury it. Right, it is a calculated manipulation of the algorithmic feed to guarantee visibility, but the success of that synchronized launch also points back to the mindset of the guy who actually built the underlying model. Diego Amada has an incredibly unusual resume for a founder operating at this level.

15:20He really does. He left Google Brain previously, which was arguably one of the most prestigious places to be working on machine learning at the time. Most people in that position immediately jump into another huge tech company or found a startup right away to capitalize on the prestige. Instead, he took a year and a half off to travel, do improv comedy, meditate, and aggressively grind competitive video games. He hit the top 30 in League of Legends and the top two in Hearthstone, which is crazy. But Hearthstone is particularly interesting to look at in the context of what he built later, because it is a digital card game entirely based on probability and complete information and deck mechanics. You are constantly calculating your outs. You are making binary choices based on the statistical likelihood of drawing a specific card. He claimed that reaching the top of the Hearthstone letter was potentially harder than building Jeff. I mean, there is a direct cognitive link between grinding a probability based card game

16:21to the top two in the world and building an AI model entirely focused on mathematical calibration and confidence scores. Both require an absolute obsession with statistical reality. You have to understand exactly what your chances are in any given scenario without lying to yourself. But he noted that the video game dopamine eventually felt empty. The thrill of outsmarting an opponent in a digital card game wears off, because it doesn't leave any lasting impact on the real world. And that realization of emptiness pushed him to transition to open AI a few years ago, back when it was still a relatively obscure research lab focused on fundamental science. He worked on the core training methods that eventually defined the entire industry. He worked on reinforcement learning from human feedback, instruct GPT, and the foundations of chat GPT. He literally went from grinding digital card games in his free time to helping build a foundation of modern artificial intelligence. He realized that working on incredibly complex projects with highly driven people provided a high

17:23that video games simply couldn't match. The intellectual intensity of a small group trying to solve a seemingly impossible problem became the ultimate focus. He directly linked that specific experience to the two core principles he built type safe AI's culture upon. Yeah, and those principles are entirely focused on doing exciting work at the frontier of AI and building a team of smart, driven people who push each other to be better. He essentially built a culture on the specific kind of intellectual intensity he found at open AI before it scaled into an enormous corporate entity. But you have to look past the founder praise and the viral launch numbers to examine the actual math of what the system is doing, because not everyone is buying the hype completely out of the box. Ryan Arman ran a test on the banking 77 benchmark, which is a public customer support data set containing 77 different banking intents, things like a lost card, a fraud charge, or an address. College football is back.

18:23So Hilton called to me the superstition concierge to make your fan rituals a reality. Need a room to match your lucky number? We got you. Want to make sure our team doesn't wash your lucky, Jersey? Oh, that smells lucky. Hilton's unmatched hospitality can keep up with any superstition. Even a marching bandwick up call it 555 and 55 seconds. Hit it! When you need a team that will do whatever it takes on game day, it matters where you stay. Hilton, for this day. When you need to build up your team to handle the growing chaos at work, use Indeed Sponsor Jobs. It gives your job post the boost it needs to be seen and helps reach people with the right skills, certifications, and more. Spend less time searching and more time actually interviewing candidates who check all your boxes. Listeners of this show will get a $75 sponsor job credit at Indeed.com slash podcast. That's Indeed.com slash podcast, terms and conditions apply. Need a hiring hero? This is a job for Indeed Sponsor Jobs. Energy of baja, falta de concentración,

19:24nevlamente al, you might be dehydrated. Gatorade Zero powders, han sido diseñados científicamente para ayudarte a mejorar el balance de hidratación en el cuerpo. Hechos con la mezcla número uno de lecrolitos, sin coloros ni sabores artificiales, shop now online or at retailers nationwide. Gatorade hydrates better than water. Is it in you? Signos de hidratación leve y moderada, consulta tu médico si los índomas persisten. Change. Right. And out of the box, Jeff only hit about 80% accuracy on that benchmark. That is a noticeable drop from the perfection the viral demos were implying when the product launched. I find the mechanics of how they actually fixed that accuracy drop really revealing. To boost that score to 92.4% accuracy, they had to rely on a technique called retrieval augmented few shot classification. For every single user query that comes in, the system has to search a database for the 24 most relevant past examples of that specific issue.

20:25It retrieves those 24 examples, grabs their correct labels, and feeds all of that context into Jevon along with the new query. It is exactly like asking a librarian a question, and before they can answer you, they have to run to the basement, find 24 identical questions from previous years, read all the answers, and then come back to tell you yes or no. The 92.4% accuracy is an measurement of Jev functioning alone. It is a measurement of Jev functioning on top of a heavy, complex retrieval system. And I really struggle with the practicality of that approach for an established business. If a company already has thousands of labeled examples of customer support requests, retrieving 24 of them for every single live request is incredibly clunky. You are sending a gigantic packet of data to the model every single time a customer asks a simple question about their account. It increases latency, and it drastically increases the complexity of the backend infrastructure required to run the system. Our mom contrasted this directly with Oommi's approach to the exact same data set.

21:26They took a completely different path. They trained a small open source in struct model, specifically Quim3-4b, on the banking 77 data set in just 30 minutes. That fine-tuned model hit 93% accuracy without any complex retrieval system attached to it. And crucially, because it is a language model that has been fine-tuned on this specific data, it can actually explain why it made its choice. It isn't just spitting out a binary decision based on a retrieved prompt. The knowledge is baked directly into the weights of the model itself. It actually understands the context. This really points to the actual limits of what GenJ is built for. It is brilliant for brand new categories or situations where you want to avoid the friction of training a model entirely. If you have a constantly shifting set of rules like government procurement guidelines that update every month, you don't want to retrain a neural network every week. You just want a system that can read the new rules and sort the documents. But it might actually be the wrong tool

22:27for established, well-documented enterprise tasks where the categories rarely change. Yeah, if you are a bank dealing with the exact same 77 customer service intents for five years straight, fine-tuning a small model makes much more sense. It is faster, it is self-contained, and it doesn't require a complex database querying mechanism running alongside it. You have to match the tool to the specific structural reality of the problem you are trying to solve. You don't use a wrench to hammer a nail, and you don't use a zero-shot classifier for a deeply established data set. You really have to look at the weird, noisy reality of the current tech market to understand why this model is generating so much attention right now despite those limitations on established data sets. Oh, totally. Because the background atmosphere of the industry is completely chaotic right now, Microsoft is completely overhauling co-pilot to integrate it into every aspect of their software. Meta's muse is topping the charts. Everywhere you look, immense tech companies are pushing generative text into every conceivable interface,

23:31trying to make every digital interaction a conversation. At the exact same time, anthropic is fighting federal black listing battles, and artificial intelligence labs are arguing with the White House over safety protocols. The entire ecosystem is filled with debates about the dangers of models that can generate open-ended, unpredictable content. Which is exactly why a boring, calibrated, non-chatting decision engine is suddenly valued in the hundreds of millions or billions of dollars. When the environment is filled with generative text, hallucinations and constant safety debates, absolute mathematical certainty becomes the most valuable commodity available. A model that refuses to write a sentence cannot write a dangerous sentence. A model that only outputs a predefined classification cannot hallucinate a fake corporate policy. We spent the last few years teaching computers how to talk, pouring billions of dollars into making them sound as human and conversational as possible, only to realize that for most actual business problems, we just needed them to listen carefully and point to the right answer.

24:33It makes you wonder if the future of this technology is highly calibrated, silent decision-making, happening entirely in the background, routing our calls and sorting our documents, without ever saying a single word, how long until we stop realizing we are interacting with artificial intelligence at all? If you're not subscribed yet, take a second and hit follow on whatever app you're using. It helps us keep making this. We appreciate you being here. Also, check out our YouTube channel for more business and tech updates. There's a link in the description.