The R1 Edition (Plus: Let's Put Those LLMs To Good Use)
Published 2 months ago • 4 min read
Of course the biggest event in AI this week was the $593 billion wipe-out of NVIDIA stock that happened on Monday 30th of January. So this week's edition is going to be the DeepSeek edition.
"One hour here is seven years on earth" - Interstellar, Christopher Nolan
In case you somehow missed the news, last week DeepSeek, a Chinese AI startup, released their R1 Large Language Model. Because it had been trained on much fewer GPUs than the frontier models of US AI labs, the stock market panicked.
If anything, this 17% drop in NVIDIA market cap shows us that a lot of tech investors had a very simple investment thesis: "bigger (GPU clusters) == better models". If you remember, just last week OpenAI unveiled Project Stargate, a $500 billion investment in AI data centers.
The New Kid on the Block
The main innovation compared to LLMs like GPT-4o and Claude 3.5 Sonnet is that for R1, DeepSeek used reinforcement learning algorithms to teach their R1 how to reason (edit: this is probably also how OpenAI's o1 was trained).
Reinforcement learning, for those of who don't know it, is giving a model a policy - a goal - it should try to optimise for through trial and error. In this case, the goal these models are given is to become better at reasoning.
By contrast, the reasoning capabilities of previous generation models was the result of a mix of something called "supervised fine-tuning" (training the LLM on a human-curated dataset) and some kind of emergence.
I wrote a more technical report on how LLMs learn here:
Besides this "pure reinforcement learning approach" to teaching the models, the folks at DeepSeek also used a number of techniques to make do with the few GPUs they could get their hands on due to export restriction: Mixture of Experts, hardware optimisations, and knowledge distillation among them.
It turns out the US export embargoes on GPUs to China forced Chinese AI researchers to become creative - which is great, because besides bringing down the cost of using LLMs a lot (R1 is 27x cheaper than OpenAI's o1!), they also open sourced their model. That means you are free to download and use it commercially!
Dario Amodei, CEO of Anthropic (makers of Claude), has an interesting take on this. In his recent blog post, he argues that DeepSeek R1 isn’t some unexpected miracle - it’s right on schedule. Think of it like Moore’s Law for AI: rapid progress that’s simultaneously shocking and utterly predictable.
What This Means for Your Business
The real story here isn’t just about DeepSeek R1 - it’s about what this continuous progression means for your business:
Democratization of AI: Open-source models like DeepSeek R1 are making enterprise-grade AI accessible and affordable for businesses of all sizes. You no longer need Google’s budget to implement powerful AI solutions.
Competitive Advantage: While others are still wondering if they should adopt AI, you can start implementing it today. DeepSeek R1’s capabilities will power agentic AI solutions for content creation, code generation, and data analysis mean so you can:
Automate customer service operations
Generate marketing content at scale
Streamline software development
Process and analyze large documents efficiently
Market Signals: While DeepSeek R1’s release didn’t cause a direct stock market surge, it’s part of a larger trend driving investment in AI infrastructure and solutions. Companies providing AI services or leveraging AI effectively are seeing increased investor interest.
Let's take a look at one of those agentic AI LLM use cases in our case study.
AI Agents: Your New Lead Generation Team
Remember spending countless hours hunting for prospects online? Those days are numbered. Today, AI agents can handle 90% of lead enrichment tasks, transforming how businesses discover and qualify potential customers. Here’s why this matters for your business, and how early adopters are already reaping the benefits.
The Problem: Your Team is Drowning in Manual Prospecting
If your sales team is like most, they’re spending up to 40% of their time on non-selling activities like prospect research and data entry (HubSpot State of Sales Report). That’s two full days per week spent browsing LinkedIn, company websites, and industry directories instead of building relationships and closing deals.
Enter the AI Agents: Your 24/7 Prospecting Team
Tools like Claygent are leading the revolution in automated prospecting.
These AI agents work tirelessly to:
Discover prospects matching your ideal customer profile across multiple online sources
Automatically enrich profiles with verified contact information
Gather company details, social profiles, and personalization insights
Deliver thousands of qualified leads in minutes, not hours
The Numbers Don’t Lie: AI’s Impact on Sales Operations
According to McKinsey research, AI-powered sales tools can:
Reduce sales operations costs by 30-50%
Increase lead conversion rates by up to 50%
Dramatically improve the quality of leads through data-driven targeting
Why 2025 Will Be the Tipping Point
Forrester Research predicts that AI will fundamentally reshape B2B sales. The 90% AI adoption projection for lead enrichment by 2025 isn’t just ambitious – it’s already happening.
Here’s why:
Competitive Pressure: Early adopters are already seeing dramatic improvements in efficiency and conversion rates
Technology Maturation: AI tools are becoming more sophisticated and user-friendly
Cost Benefits: The ROI of AI-powered prospecting is becoming impossible to ignore
Talent Expectations: Top sales professionals increasingly expect AI tools to support their work
What This Means for Your Business
The message is clear: businesses that don’t adapt to AI-powered prospecting risk falling behind. But there’s good news – getting started is easier than you might think.
Action Steps for Forward-Thinking Leaders:
Start small: Test AI prospecting tools like Claygent with a subset of your sales team
Measure the impact: Track time saved and lead quality improvements
Scale gradually: Use the proven ROI to justify broader implementation
Stay compliant: Ensure your AI tools align with data privacy regulations
The Bottom Line
The days of manually browsing for prospects are numbered. AI agents are not just tools – they’re your future competitive advantage. The question isn’t whether to adopt AI-powered prospecting, but how quickly you can implement it to stay ahead of the curve.
This week in AI
The AI Arms Race Heats Up: DeepSeek’s R1 model isn’t an anomaly, but rather a sign of how quickly AI is advancing. As Dario Amodei points out, it’s a predictable step in the rapid evolution of AI. This means businesses need to stay agile and prepared for continuous change in the AI landscape.
Mistral Small 3 can understand and respond to complex tasks just as well as much larger models, but runs faster and uses less computing power. This means more people can now use powerful AI on their personal computers without needing expensive specialized hardware.
Gemini 2.0 Flash is now the default model in the Gemini app for all users. This stable release, which started rolling out on January 30th, 2025, is available on both web and mobile versions of the app. The model promises twice the speed of its predecessor while delivering stronger performance for everyday tasks like brainstorming, learning, and writing.
