Dr. Arjun Jain on Fast Code AI: Why the $100 Million AI Engineers Know Scaling Laws Are Dead
A Yann LeCun protégé explains how the industry is pivoting from the "Software TAM" to the "Salary TAM" through outcome-based pricing and small language models
In December 2024, at the NeurIPS conference, Ilya Sutskever made a confession that sent ripples through Silicon Valley. The former OpenAI chief scientist admitted what insiders had suspected for months: we have one internet, and we’ve used it all. For Dr. Arjun Jain, founder of Bengaluru-based Fast Code AI, this wasn’t news. It was vindication.
“I think scaling laws have already stagnated,” Arjun told me during our recent conversation. “Even if you had a lot more data, I don’t think those scaling laws are holding, and it’s becoming asymptotic in terms of their performance.”
The numbers back him up. Doubling a model from 1 trillion to 2 trillion parameters now yields only 2-5% improvement, a dramatic slowdown from the exponential gains of the 2013-2020 era. This stagnation explains why AI engineers command salaries approaching $100 million. When you’re spending billions on compute clusters, the cost of a mistake dwarfs any individual salary.
Check out the video of the conversation here or read on for insights.
The Engineer Who Saw the Wall Coming
Arjun’s skepticism about brute-force scaling comes from lived experience across AI’s evolution. After leaving Yahoo in 2007 (”I think I’m a very lousy employee”), he pursued a PhD at Germany’s Max Planck Institute working on what we’d now call generative AI, a decade before ChatGPT. His viral “Movie Reshape” research, which made actors look taller or more muscular in video, caught Hollywood’s attention and led to six months at Weta Digital solving motion capture problems for Snowy the dog in Spielberg’s “The Adventures of Tintin.”
Then came the AlexNet moment. As a postdoc at NYU (2013-2015) working under Turing Award winner Yann LeCun, Arjun witnessed the deep learning explosion firsthand. His labmates included Wojciech Zaremba, who went on to co-found OpenAI.
AlexNet trained a neural network on this ImageNet data on GPUs and it beat the older algorithms, which were more of these handcrafted algorithms, by a big margin. And this is when we knew that this is going to take over completely.
But before founding Fast Code AI, Arjun performed what might be the most relevant engineering feat for today’s cost-conscious AI landscape: at Mercedes-Benz, he compressed a 6-gigabyte neural network down to 300 kilobytes, a 20,000x reduction. The model had to run on a 2-watt FPGA chip embedded in a car’s rearview mirror, monitoring driver attention in real-time. This compression expertise, now deployed in production Mercedes vehicles, represents exactly what enterprises need as inference costs spiral.
We had to fit it into like a small two-watt Xilinx FPGA with 110 DSPs, and eventually we had like 300 kilobytes of weight.
This work taught him that intelligence isn’t about size but about efficiency.
Selling Digital Labor, Not Software Seats
Today, Fast Code AI’s 45-person team is attacking a problem most enterprise software vendors are afraid to confront: their business model is obsolete.
Traditional SaaS companies target the “Software TAM,” the 1-5% of enterprise revenue allocated to productivity tools. They charge per seat, per month. But AI agents are designed to reduce headcount. When an agent makes a team 50% more efficient, the client needs 50% fewer seats. The vendor cannibalizes its own revenue.
Fast Code AI targets something far larger: the “Salary TAM,” the 30-70% of enterprise revenue spent on human labor. They do this through outcome-based pricing, taking 5-10% of the value their AI agents create.
If we save X for you, give us 0.1X or even like 0.05X. So I think companies in their transformation journey, all they want to do is reach this ROI.
Consider their procurement agent. Large enterprises typically ignore contracts under $10,000 because the negotiation cost exceeds potential savings. Fast Code’s agent handles this “tail spend” automatically, reading RFPs, emailing vendors, comparing quotes, and negotiating terms based on historical data. If the agent saves a client $500,000 annually on previously ignored contracts, Fast Code takes $25,000 to $50,000. For the CFO, it’s found money.
The AI agents market is projected to grow from $7.63 billion in 2025 to $182.97 billion by 2033, a compound annual growth rate of 49.6%. Fast Code’s bet is that this growth will be captured not by selling software seats, but by selling digital labor.
The Self-Play Advantage
What separates Fast Code from competitors is their reinforcement learning pipeline. Rather than simply prompting GPT-4, they build simulation environments that mirror the client’s workflows. For a procurement use case, they create digital twins of vendors, departments, and the procurement software itself. The AI agent then engages in “self-play,” negotiating thousands of times against simulated vendors before ever sending a real email.
We create these different personas, these different vendors... And then these LLMs do something to the software, it goes and clicks a few buttons, and we essentially rig them up with prompts, with personas, where LLMs are essentially then replicating what humans would be doing.
The approach mirrors how AlphaGo mastered the game of Go, applied to enterprise workflows. Crucially, Fast Code uses small language models of 7 billion parameters, fine-tuned on client data rather than relying on massive foundation models. These SLMs run in the client’s private cloud, addressing data sovereignty concerns while dramatically reducing inference costs compared to calling OpenAI’s API millions of times.
The industry’s response to the scaling plateau has been “test-time compute,” making models think longer rather than bigger. This reasoning underlies OpenAI’s o1 model, which generates hidden chains of thought before answering. Fast Code applies this concept by allowing their agents to simulate outcomes in RL environments before taking action, turning test-time compute into a practical deployment strategy.
Building After the Plateau
When I asked Arjun about advice for founders navigating this landscape, his answer was characteristically grounded: start with empathy, identify friction in boring back-office processes where competition is light, and iterate fast.
No idea comes out fully formed. You will pivot many times and your idea will evolve and it’ll become kind of real after a while. Starting is the most important thing. Once you take the first four steps, you will get to see what is on the next four.
His former advisor Yann LeCun calls current LLMs “stochastic parrots,” sophisticated text predictors lacking true understanding of physics, causality, or the ability to learn from single examples the way humans do. Arjun agrees. True intelligence requires stateful memory and grounded understanding of physical reality. We’re not there yet, and we won’t brute-force our way there with bigger models.
As the AI industry grapples with the end of easy scaling, Fast Code AI represents a different path: smaller models, deeper domain expertise, and business models aligned with value creation rather than seat counts. While Silicon Valley chases AGI with ever-larger clusters, Arjun is building the digital workforce enterprises will actually pay for today.
Listen now!
Other ways to listen:
Your Feedback matters
As always, I’d love to hear your thoughts! Whether it’s about this episode or ideas you’ve been playing around with, shoot me an email at ad@thepodium.in. Your feedback keeps these conversations going, and I’m always up for chatting about your startup ideas too.
Until next time,
Your Host,
Akshay Datt