''' I told a founder not to use GPT-4 last week. He was shocked. "But Sahin," he said, "isn't bigger always better in AI?"
That’s the myth, isn't it? The tech world is obsessed with the size of language models. How many billions of parameters? How many GPUs did it take to train? We celebrate these massive models like they're the pyramids of our time. But I’m here to tell you, we’re looking at the wrong pyramids.
The real revolution isn’t happening at the top with these behemoths. It’s happening in the trenches, with small language models (SLMs). And it’s about to change everything.
The Billion-Parameter Elephant in the Room
Let’s be real. For 99% of applications, a massive, multi-billion parameter model is overkill. It’s like using a sledgehammer to crack a nut. I’ve seen it dozens of times in my portfolio. A startup gets a ton of funding, and the first thing they do is go for the biggest, most expensive model on the market. They think it’s a silver bullet.
But then reality hits. The inference costs are astronomical. The latency is a killer for user experience. And forget about running it on-device. You’re completely tethered to an API, with all the privacy and reliability headaches that come with it.
I remember one of my early investments, a company building a real-time translation app. They started with a huge model, and the performance was… sluggish. The lag made conversations feel unnatural. We switched to a much smaller, fine-tuned model, and the difference was night and day. The translations were just as accurate for their specific use case, but the speed was incredible. The app felt alive.
That’s the thing about SLMs. They’re not just smaller; they’re smarter. They’re specialized. They’re efficient. They’re the nimble fighter jets to the LLMs’ lumbering bombers. ''' '''
The "Good Enough" Revolution
We have this obsession in tech with perfection. We want the model that scores 99.9% on every benchmark. But for most real-world problems, 95% is not just good enough—it's often better. Why? Because you can get to 95% with a model that's 100x smaller and faster.
Think about it. Do you need a model with a deep understanding of 18th-century poetry to power your customer service chatbot? Probably not. You need a model that understands your product, your customers' common questions, and your company's tone of voice. That’s it.
At RemoteTeam, which was acquired by Gusto, we built an internal tool to help summarize long email threads. We could have used a massive, off-the-shelf model. Instead, we took a much smaller open-source model and fine-tuned it on our own email data. The result? It was faster, cheaper, and frankly, better at understanding our internal jargon and context than any generic model could have been.
This is the power of specialization. It’s a lesson I’ve learned over and over again across my 200+ angel investments. The companies that win aren’t always the ones with the biggest tech stack. They’re the ones that are the most resourceful and focused.
Fine-Tuning is the Superpower
This brings me to my next point: fine-tuning. If LLMs are the raw engine blocks, fine-tuning is the process of turning them into high-performance racing engines for a specific track.
Companies like Scale AI and Hugging Face—both of which I’m proud to have in my portfolio—are at the forefront of this. They’re building the tools that allow anyone to take a base model and adapt it to their unique needs. This is a fundamental shift in how we build with AI.
Instead of being a consumer of a single, monolithic model, you become a creator. You can distill the knowledge of a massive model into a smaller, more efficient one. This process, known as model distillation, is one of the most exciting areas in AI right now. You get the best of both worlds: the power of a large model, with the efficiency of a small one.
I’ve seen teams build incredible things this way. A legal tech startup that fine-tuned a model to review contracts, catching specific clauses that a general model would miss. A healthcare company that built a diagnostic tool trained on a specific set of medical images, achieving near-human accuracy at a fraction of the cost.
This is where the real innovation is happening. Not in the race to a trillion parameters, but in the smart application of models that are the right size for the job. ''' '''
The Magic of On-Device AI
Here’s where it gets really exciting: on-device AI. The ability to run a powerful model directly on your phone, your laptop, or even your car.
This is the holy grail. Why? Two reasons: privacy and experience.
First, privacy. When the model runs on your device, your data stays on your device. It’s not being sent to a server owned by a massive tech company. For applications in healthcare, finance, or even just personal journaling, this is not a feature—it’s a requirement. I’ve seen multi-million dollar enterprise deals fall apart because of data privacy concerns. On-device SLMs solve this problem elegantly.
Second, the user experience is just… better. It’s faster. It works offline. You’re not at the mercy of a spotty internet connection. Think about the apps you love to use. They’re the ones that feel instant and reliable. That’s what on-device AI delivers. It’s that feeling of magic when something just works, seamlessly.
I have an investment in a company building an AI-powered camera app. It does real-time image enhancement. If they had to send every frame to the cloud for processing, it would be a non-starter. The lag would make it unusable. But with a highly optimized SLM running locally, it feels like you have a professional photo editor in your pocket. That’s the power of small.
The Democratization of AI
For me, this is the most important point. The obsession with massive, closed-door models creates a world where only a handful of giant corporations can afford to innovate. That’s not the future I want to build. That’s not the world where the best ideas win.
SLMs, especially when combined with the open-source movement, are a democratizing force. They lower the barrier to entry for everyone. You don’t need a hundred-million-dollar funding round to build a world-class AI product anymore. You need a good idea, a specific problem to solve, and a talented team that knows how to work with these smaller, more accessible models.
This is why I’ve invested in companies like Anthropic, OpenAI, and Hugging Face. They are creating the tools and the foundational models that empower the next generation of builders. They are providing the picks and shovels for the AI gold rush, but I’m telling you, the biggest veins of gold aren’t in the biggest mountains.
We’re seeing a Cambrian explosion of creativity. Small, dedicated teams are building amazing things:
- Tools that help developers write code faster.
- Apps that help people learn new languages.
- Assistants that can summarize your meetings and draft your emails.
Most of these are not running on the biggest model available. They are running on the right model.
Stop Chasing Size, Start Chasing Value
So, the next time you hear about a new model with a staggering number of parameters, ask yourself: who is this really for? Is it for the builders, the innovators, the problem-solvers? Or is it for the marketing department of a tech giant?
My advice to every founder I meet is this: stop chasing the hype. Stop thinking that bigger is better. Instead, get obsessed with your user’s problem. Get obsessed with the data that will help you solve it. Find the smallest, most efficient model that can do the job, and then fine-tune the hell out of it.
That’s how you build a product that people love. That’s how you build a business that lasts. The future of AI isn’t about the size of the model. It’s about the scale of the impact. And for that, you need to think small. '''
The Hidden Cost: AI and the Environment
There's another dirty secret about the race for bigger models that we don't talk about enough: the environmental cost. Training a single, massive language model can have a carbon footprint equivalent to hundreds of transatlantic flights. It consumes enormous amounts of electricity and water for cooling data centers. As an industry, we can't just ignore this.
I was talking to a founder in the climate tech space, and she put it perfectly: "We're trying to solve the world's biggest problems with a technology that is, in itself, creating a huge environmental problem." It's a paradox we have to confront.
This is another area where SLMs shine. They are exponentially cheaper and less energy-intensive to train and run. A smaller model doesn't just save you money on your cloud bill; it reduces your product's carbon footprint. For a generation of users and builders who are increasingly conscious of their environmental impact, this isn't just a nice-to-have. It's a core part of building a responsible and sustainable business.
More Than "Good Enough" - The Joy of a Snappy Product
I want to double-down on my point about the "Good Enough" revolution. It's not about settling for mediocrity. It's about understanding the point of diminishing returns. The difference between a model that's 95% accurate and one that's 98% accurate can be a 1000x increase in computational cost. Is that marginal 3% improvement worth a slow, expensive, and clunky user experience? Almost never.
I remember the early days of MovieLaLa, my second company, which was acquired by Gfycat. We were building a movie discovery app. We experimented with a very complex recommendation engine. It was technically brilliant and could predict with incredible accuracy what movie a user might like. But it was slow. It took a few seconds to generate recommendations. Users hated it. They'd open the app, get impatient, and leave.
We switched to a much simpler, faster algorithm. The recommendations were slightly less personalized, but they appeared instantly. Our user engagement and retention shot through the roof. We learned a critical lesson: in consumer products, speed is a feature. A snappy, responsive experience that delivers good value is always better than a slow, ponderous one that delivers perfect value. SLMs are the key to building those snappy experiences in the age of AI.
A Personal Story of Fine-Tuning
Let me share one more story. A few years ago, I was helping a portfolio company that was building a tool for writers. The idea was to create an AI writing assistant that could adapt to your personal style. They started by using a large, generic model, but the results were bland. The AI's suggestions all sounded the same, like a corporate press release. It was stripping the author's voice away, not enhancing it.
The team was about to give up. I pushed them to try a different approach. We took a smaller, open-source model and had it analyze a large corpus of the user's own writing. We fine-tuned the model not on the entire internet, but on the unique voice of one person. The results were astonishing. The AI started making suggestions that felt personal, insightful, and genuinely helpful. It learned the author's cadence, their favorite words, their sense of humor. It became a true collaborator.
That company is now thriving. They found their niche not by building the biggest model, but by building the most personal one. That's the magic of fine-tuning. It’s the path from generic intelligence to specialized, personal genius.
My Challenge to You
So, I'm throwing down the gauntlet. To the founders, the builders, the engineers, the creators: I challenge you to think small. I challenge you to ignore the hype and focus on the user. I challenge you to find a niche, a specific problem that you can solve 10x better than anyone else, not with a bigger model, but with a smarter one.
The next unicorn in AI won't be the company that builds a 10-trillion parameter model. It will be the company that uses a 1-billion parameter model to solve a billion-dollar problem. It will be a company that values efficiency, privacy, and user experience over brute force.
Build something fast. Build something focused. Build something that people love to use. That's the revolution. And it's just getting started.
Frequently Asked Questions
How can I apply this thinking to my own situation?
Start by identifying the core principle behind the opinion, not the specific example. Then ask yourself: does this principle apply to my context? If yes, test it in a small, low-risk way before going all in.
What experience informs this perspective?
This perspective comes from over a decade of building companies in Silicon Valley, two successful exits (RemoteTeam to Gusto, MovieLaLa to Gfycat), and investing in 200+ startups including Anthropic, OpenAI, and Scale AI. I write about what I've lived.
What's the most common pushback you get on this?
People often push back by citing exceptions or edge cases. And they're usually right that exceptions exist. But building a strategy around exceptions rather than patterns is a losing game for most founders.
Do all experts agree with this view?
No, and that's fine. The best ideas in business are often contrarian. I share my perspective based on my experience and data, but I encourage you to seek out opposing viewpoints and form your own conclusions.