The 7.5 Billion AI Without Chat

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? Oho, that smells lucky. Hilton's unmatched hospitality can keep up with any superstition. Even a marching bandwink 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. This episode is brought to you by PayPal. Holiday Rush? Skip the mall traffic and shop online with PayPal. Shop your favorite online stores like eBay, Fanatics, Walmart, Sephora, Target and TikTok shop. And choose to pay now or pay later at checkout. Plus pay in four with no interest and no late fees. Shop now with PayPal. Pay in four subject to approval. Learn more at PayPal.com slash pay in four. PayPal Incorporated NMLS 910457.

1:01When 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 andd.com slash podcast. That's indeed.com slash podcast. Terms and conditions apply. Need a hiring hero? This is a job for Indeed Sponsored Jobs. Type-safe AI just raised $870 million at a $7.5 billion valuation for an artificial intelligence model that specifically avoids talking to human beings. Yeah, the model is called Jeff. It was launched literally weeks ago by a team of heavyweights. And somehow almost 30% of the Fortune 500 is already using it to run their backend software. Which brings up the central question for you as you look at this sector. If the biggest tech companies in the world are already fast following with exact copies of this architecture,

2:03does simply being the first to cut out the chat interface entirely just to buy a $7.5 billion price tag? Well, you look at the core metric they released to justify that. They stated they accidentally served trillions of tokens a day within those first few weeks of launch. That volume of processing is the mechanism they used to capture 29.4% of the Fortune 500. We should probably look at the people driving this to understand where that volume came from. The founders are Diego Almeida, Sasha Shang and Eric Spock Gaffney. Almeida came out of OpenAI. He's one of the co-inventors of chat GPT, GPT-4, RLHF, and Instruct GPT. Right. So you have the exact engineering team that built the conversational paradigm we currently live in. They built the systems that essentially taught artificial intelligence how to speak to us. RLHF being reinforcement learning from human feedback, the process that actually made chat bots polite and helpful. Exactly. Because before RLHF, a language model was essentially just an aggressive auto-complete engine.

3:07If you gave it half a sentence, it would just guess the next word based on internet data. Which meant it could easily start spitting out random Wikipedia articles or highly toxic forum posts. Right. It didn't know how to behave. Yeah. So RLHF changed that. They hired humans to sit there and grade the model's responses. They gave it a thumbs up when it acted like a helpful assistant and a thumbs down when it went off the rails. That process trained the model to prioritize a very specific conversational tone. And Instruct GPT was the architecture that layered on top of that, allowing the model to follow a user's instruction step by step, rather than just predicting text. So these three founders literally wrote the playbook on making AI chatty. And their new company is built entirely on the premise that making AI chatty is a severe limitation for enterprise software. They're building the next iteration by actively removing the conversational layer they invented.

4:07I do want to push back on the word accidentally when we talk about their launch volume. They claim they accidentally served trillions of tokens a day. If you work in enterprise software, you know that capturing a third of the Fortune 500 is a grueling process. Oh, absolutely. It requires endless sales meetings, security reviews, compliance checks. You do not accidentally integrate your software into the core back end of 30% of the largest companies on earth in a matter of weeks. If you're not subscribed yet, take a second and hit follow on whatever podcast after you're using. It helps us keep making this. We appreciate you being here. But getting back to that adoption rate, doing that so fast means a critical pre-existing bottleneck just suddenly broke open. Yeah, these companies were already trying to automate their back end systems failing and desperately waiting for a solution that actually worked. A trillion tokens a day is a volume that is just difficult to contextualize because a token is roughly like a piece of a word.

5:09When you use a traditional language model to read a document and generate a summary, you might use a few thousand tokens. Right. So to serve trillions of tokens daily, these large corporations had to have plugged Jeff into their systems and instantly routed thousands of parallel continuous core operations through it every single second. Which breaks the known limits of enterprise software adoption curves entirely. Normally, a large corporation tests a new tool on a non-critical workflow for six months. They put it in a sandbox environment. But with Jeff, they bypassed the testing phase and went straight to production. They did that because the model solves a very specific, painful friction point. And they hit that adoption metric specifically because of what the model does not do. It doesn't generate paragraphs of text. Exactly. Jeff eliminates text output entirely to reduce token consumption. It is built strictly for software. It takes in real world data and returns fast, structured decisions that code acts on directly.

6:09Because the industry has been relying on frontier models for basic back end tasks. Developers were taking these heavy conversational models and trying to force them to read data and return a simple answer. The documentation for Jeff says it replaces those heavy models with composable, semantically capable primitives. I mean, that phrase is a bit dense, but think of a primitive, like a basic building block in software, a single Lego brick, a logic gate. It takes an input, applies a rule, and gives a binary output. Right. Semantically capable just means this specific logic gate can read and understand human language or messy data. But it still only outputs a simple, structured decision. Using a standard, large language model for a back end software task is like... College football is back. So, Hilton called in 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.

7:12Even a marching bandwink 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. To approval, learn more at paypal.com slash payin' for paypal incorporated NMLS 910457. 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 andd.com slash podcast.

8:17That's indeed.com slash podcast, terms and conditions apply. Need a hiring hero? This is a job for indeed sponsor jobs. Hiring a famous novelist to flip a light switch. You don't need a beautifully crafted sentence. Explain to you that the room is now bright. You just need a binary decision. Jev is the switch flipper. Imagine you're running the back-end operations for a global logistics company. You have millions of shipping manifests coming in every day from different vendors. And they're all formatted differently. Some are PDFs, some are emails, some are spreadsheets. And you just need to extract the destination, the weight, and the customs code. Right. But if you use a conversational model for that, the model receives the manifest and starts writing a polite response. It says, certainly I can help you with that. The destination is Chicago. The weight is two tons. Which is useless to a computer. Your back-end routing software does not know how to read. Certainly I can help you with that. It just needs the data points formatted in this specific JSON file so it can dispatch the truck. So developers were spending all their time writing extra code just to strip away the conversational pleasantries so their systems wouldn't crash.

9:23And every extra word generated by a conversational model costs money and adds latency. Let's do the auto-regressive text generation. Traditional models generate responses one token at a time. It's like watching someone type on a keyboard one letter a second. Yeah. If the model decides to add pleasantries to its output, your software pipeline has to wait for those extra words to finish generating before it can execute the next step. And Jeff strips all of that away. It gives you the raw semantic capability of a language model. The ability to read the messy manifest, constraints strictly to data formats that machines can read instantly. There is no typing effect. It just outputs the structured data. But removing the text generation completely limits the model's consumer appeal. Regular users want a chat interface. If a normal person typed a prompt into Jeff, they would just get a raw data file back. Which is completely unreadable to anyone outside of a development environment. Right. But consumer appeal is a trap. The tech industry became obsessed with building models that talk to people because it demos well on the stage.

10:30You can show it to a journalist and they write an article about it. But the actual money in the global economy is in silent invisible back end automation. Because a consumer might generate a few hundred tokens a day asking a chatbot for a recipe or help with a homework assignment. A single fortune 500 company will generate a billion tokens an hour just categorizing their internal database entries. Exactly. Developers inside these massive companies realize they could replace their slow expensive frontier models with Jeff. They swap the API key and their code. And suddenly they're back in software ran ten times faster and cost a fraction of the price. The friction was entirely gone. But that immediate adoption creates a completely different problem for types of AI. Dropping a silent model that makes millions of structured decisions a second instantly tests the physical limits of their server racks. Yeah, the bottleneck shifted overnight. The model itself is no longer the limiting factor. The physical infrastructure running the model is when you drop the text generation.

11:32You actually increase the physical hardware stress. A conversational model streams text slowly to a single user, which is relatively easy on the network. Jeff is handling millions of discrete concurrent API calls per second from enterprise back end systems, which tests the network input and output limits. And the server elasticity far more than generating a paragraph. Turning this into actual business value requires monitoring latency, resource utilization and reliability. It just an unprecedented volume. The engineering demands have moved entirely to GPU optimization, efficient inference, scalable infrastructure using Kubernetes and LLM ops. Kubernetes is critical here. If you're managing physical servers for this kind of traffic, you cannot do it manually. Think of Kubernetes like a restaurant manager. If a massive bus of tourists pulls up to the restaurant, the manager instantly pulls out extra tables and shares from the back to accommodate them. And when they leave the manager folds the tables up and puts them away so they aren't taking up space. Right. In this case, the tables are server containers.

12:33When a Fortune 500 company runs a batch process of 10 million database queries, Kubernetes spins up thousands of individual server containers to handle the traffic spike and shuts them down when it drops. If you don't have that infrastructure scaled perfectly, your latency spikes. A structured decision that arrives three seconds late is useless to a live software pipeline. And LLM ops is the other half of that equation. It stands for large language model operations. It covers observability and monitoring. Because if a server rack goes down in a data center in Virginia, your traffic has to route to a facility in Ohio in milliseconds or a major corporations core workflow fails. This connects directly to that $870 million series A funding. You don't raise almost a billion dollars weeks after launch just to hire a few more software engineers. That capital is specifically for scaling product deployment and boosting physical computing capacity. They need to buy heavy iron. They need actual data centers, cooling systems and massive clusters of GPUs just to keep the lights on under the weight of those trillions of tokens.

13:37The center of gravity and artificial intelligence is shifting from the brain to the pipes. For the last few years, the focus was entirely on the intellectual breakthrough of making a computer understand language. Now strong model serving is just as important as the model itself. It changes the focus of development from purely intellectual breakthroughs to heavy industrial scale engineering. You're essentially building a digital utility company. You're building power lines. And building power lines requires capital from people who understand enterprise utility. The investors backing this reflect that shift. The series A was led by Andreessen Horowitz and Martin Cassado is joining the board. Sequoia capital participated to along with DCVC and DCVC actually led the initial $40 million seed round and was advised by Wilson Sonsini. Yeah, and DCVC has a stated philosophy for the specific investment. They phrase it as building prod, not God. That phrase is a direct philosophical rejection of companies trying to build artificial general intelligence.

14:38AGI being the idea of an omnipotent artificial intelligence that can do anything a human can do. College football is back. So Hilton called in 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 bandwink 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. This episode is brought to you by PayPal. Holiday rush? Skip the mall traffic and shop online with PayPal. Shop your favorite online stores like eBay, Fanatics, Walmart, Sephora, Target and TikTok shop and choose to pay now or pay later at checkout. Plus paying for with no interest and no late fees. Shop now with PayPal. Pay in for sub to approval. Learn more at PayPal.com slash Pay in for.

15:40PayPal Incorporated NMLS 910457. 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. And potentially more. Type Safe AI is rejecting that sci-fi narrative completely. They are focused on production, efficiency and immediate utility. I do wonder if capping your ambition at production rather than artificial general intelligence eventually limits your valuation. If you aim to build an omnipotent AI, the theoretical financial ceiling is infinite. You could theoretically replace the entire human workforce. But if you aim to build a very efficient logic gate, you are capping your market to the current enterprise software budget.

16:46But I strongly disagree that it's a bad thing. Utility is what actually pays the bills in the real economy. The company's chasing artificial general intelligence are burning billions of dollars on research projects with no clear path to profitability. They're waiting for a magical breakthrough that might be a decade away. But Jeva is generating cash flow right now. Investors are desperate for sustainable revenue models. They are tired of funding research labs that just produce interesting demos. They value to practically brand new company at $7.5 billion because it integrates seamlessly into the budgets companies already have. By focusing entirely on production, they made themselves indispensable to the businesses that actually power the global economy. They prove there is massive money, impure, unglamorous utility. But the moment you prove that, the giants instantly wake up. Within weeks of Jeva's launch, open AI, AWS, CloudFlare and various open source projects, fast followed with their own non-text decision models. This brings up a critical industry debate.

17:46We have to look at whether Jeva is a truly unique, defensible model created by an interesting lab. Or if Type-Safe AI was just the first to market with a clever idea on how to use existing technology. You have to look at the brutal reality of the current tech ecosystem. If AWS or CloudFlare bake a non-text decision model directly into the server's companies already use, Type-Safe AI's head start might mean nothing. AWS already hosts the infrastructure for a massive portion of the internet. A fortune 500 company is already paying AWS for their servers, their storage, their security. If AWS puts a switch in their dashboard that gives developers the exact same structured decision capability, it is hard to see how Type-Safe AI retains those clients. Because how do you compete with AWS on distribution? You're asking a company to set up a new vendor relationship route their data outside of their existing AWS environment and pay a separate bill, all for a feature that AWS is offering natively. Type-Safe AI's execution quality and pace are their only real mode.

18:48They have the $870 million in capital and they have the engineering talent of Elmeda, Shang, and Gaffney. Their investors have specifically noted that their speed of deployment is completely anomalous. They're betting that they can innovate on the infrastructure and the LLM ops faster than AWS can clone their features. They have to offer better observability, lower latency and more efficient GPU optimization than Amazon or CloudFlare. Which changes the narrative from a clear victory to a high-stake survival race. They prove the market exists. They secure the funding and they capture the initial user base. Now they are in a dead sprint against the most well-resourced infrastructure companies on Earth to maintain their position. It forces them to constantly stay ahead on the unglamorous parts of the technology. They can't rest on the algorithm itself. If they fail at the infrastructure level, if their latency drops by a few milliseconds, the model's intelligence will not save them. The client will just flip the switch back to AWS. Exactly. The future of artificial intelligence in the enterprise isn't going to be chatty.

19:50It's going to be composed of incredibly fast, cheap, invisible logic gates, processing trillions of tokens in the background to keep software running. And with AWS and OpenAI instantly cloning this approach, we really have to wonder if Typesafe AI just successfully built a $7.5 billion company. Or if they just handed the blueprint for the next phase of the internet to their biggest competitors. If you're not subscribed yet, take a second, hit follow and 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. New and exclusive holiday decor just launched online at the Home Depot. Check out the viral grand Duchess collection with Smart Holiday Lights you can control from your phone. Easily customize the Twinkle across trees, wreaths and garlands to match your holiday vibe. Create stunning light effects, then switch it up with the top of your fingers. Even the man in red will be impressed.

20:52Shop the holiday decor you want online only at Home Depot.com. Close your eyes, exhale, fill your body relax, let go of whatever you're carrying today. Well, I'm letting go of the worry that I wouldn't get my new contacts in time for this class. I got them delivered free from 1-800 contacts. Oh my gosh, they're so fast. And breathe. Oh, sorry. I almost couldn't breathe when I saw the discount they gave me on my first order. Oh, sorry. Namaste. Visit 1-800-contacts.com today to save on your first order. 1-800-contacts. Your heart can tell you a lot about your health. Apple Watch Series 12 measures your heart rate every 5 seconds with the most accurate heart rate sensing and awareable. So your vital zap, now with heart rate variability, can tell you when something is off. And your readiness score can let you know when to rest and when to push. Here are the story in every heartbeat with Apple Watch Series 12. The features described are for wellness purposes only and not for medical use. iPhone 11 or later required based on Apple conducted study of heart rate accuracy August 2026.

21:55Visit Apple.com slash Apple Watch Series 12.