If 2024 had a theme, it would be "evolution through iteration." The AI landscape didn't just grow—it matured, pivoted, and occasionally surprised us. Here's what defined the year.
Small Models, Big Ambitions
2024 taught us that smaller doesn't mean weaker. Small language models (SLMs) have started taking over tasks that once needed their bulkier, power-hungry predecessors. Take Llama 3.3 70B—it holds its own against the older Llama 3.2 405B, proving you can pack performance without the bloat.
Why does this matter? Because smaller models are perfect for local, on-device intelligence and agent swarms. Imagine a fleet of hyper-focused agents, each crushing specific tasks with precision. This isn't just an optimization—it's a whole new way of thinking about intelligence.
Agentic Workflows Take Center Stage
Speaking of agents, they've become the poster child of modern AI engineering. Forget the days of trying to cram every task into one massive pipeline. Now we build modular, orchestrated flows—one agent handles PII masking, another takes care of prompt injection detection, and yet another dives into language understanding.
Each agent knows its job and does it well, with handoffs that feel almost… elegant. This isn't just about scalability; it's about adaptability. It's like the microservices movement met foundation models, and the result is game-changing.
RAG: Not Quite Dead, but Less Glamorous
Remember when Retrieval-Augmented Generation (RAG) was the hot new thing? Yeah, it's still here, but the sparkle has dimmed. The rise of large-context-window models like Phi3 (with a whopping 128k tokens!) has made RAG less essential for grounding.
Now, instead of building elaborate RAG pipelines, you just toss a bigger context into your prompt. Sure, this makes your prompt cache heavier, but who's counting bytes when you can simplify architecture? Sometimes progress means making yesterday's hacks obsolete.
Code Generation: No Longer Just a Party Trick
Code generation with LLMs hit a new level this year. Models like OpenAI's O1-mini and Anthropic's Claude Sonnet 3.5 turned coding from a solo sport into a relay, where AI takes the first lap. Python and Java benefitted the most—unsurprisingly—but even smaller, niche languages started getting love.
Open-source tools are still crawling while the proprietary ones are sprinting. If you're a developer, though, this was the year you stopped wondering if AI could help and started thinking about how much you should rely on it.
Anthropic's Claude 3.5: Quietly Taking Over
Anthropic's Claude 3.5 wasn't flashy, but it was impactful. Its handling of artifact management (basically, organizing your mess) won it a dedicated fanbase. And with Amazon's backing, Claude has become the Swiss Army knife of AI tools.
It's hard not to root for a model that gets things done without needing to scream about it. Sometimes, the quiet ones really are the most impressive.
Everyday AI: Convenient, Yet Complicated
Anthropic and OpenAI both made strides in integrating AI into daily computing, but the journey hasn't been smooth. Privacy concerns are holding back mainstream adoption, while Microsoft's Recall feature feels like the kid at the party trying too hard.
Still, there's no doubt that we're heading toward an era where computers respond to us the way we've always wanted them to—if we can figure out how to trust them with our data.
Multi-Modal Models: Beyond Words
Vision-language models like Colpali have turned multi-modal AI from a niche curiosity into a serious field. Text, audio, images, video—you name it, these models handle it. Open-source contributions keep pushing the envelope here, and it's refreshing to see innovation that feels genuinely collaborative.
The pace of progress has been dizzying, but it's also thrilling. This is where things start feeling futuristic.
Google's Year of Redemption
Google finally showed up to the LLM party in style. The Gemini models didn't just perform—they dominated, topping benchmarks like MMLU. And yes, Geoff Hinton and Demis Hassabis won Nobel Prizes, sparking debates about whether AI contributions belong in the same category as, say, curing diseases.
But let's be real—Google earned its moment. Gemini's neutral tone and smooth apps prove that patience pays off. If anything, they've reminded us that the tortoise can still win the race.
NVidia: The Kingpin of AI
If 2024 had a corporate MVP, it was NVidia. From GPUs to AI-specialized hardware, they've solidified their place as the supplier of choice for the AI gold rush. Their product roadmap looks more like a takeover plan, and honestly, it's working.
The question is, can they maintain this momentum when the market inevitably shifts? For now, they're the Goliath—and everyone else is just trying to keep up.
Ethics and Regulation: Stuck in Neutral
AI ethics and fairness debates took a backseat this year, and it's hard not to feel a little uneasy about it. The models are more complex, the stakes are higher, and the voices pushing for accountability are quieter.
Meanwhile, Europe's strict regulations created enough friction for Meta to pull back its offerings. The tug-of-war between innovation and oversight is far from over, and 2024 felt like a year where the scales tipped in favor of big tech.
The LLM Arms Race Intensifies
OpenAI's O1 launch stole the spotlight with its clever implementation of chain-of-thought reasoning, but it wasn't the only headline. Alibaba's Qwen series quickly followed, proving that the open-source community can move fast when it wants to. Even Amazon got in on the action with Nova, finally giving itself a seat at the table.
If this year taught us anything, it's that the competition is fierce—and consumers are the real winners.
Apple's AI Ecosystem: Slowly, Surely
Apple's AI strategy felt like a classic Apple move: quiet, deliberate, and deeply integrated. Their new models run smoothly on Apple Silicon, even if they're not leading the pack in raw power.
What's impressive is how seamlessly they blend hardware, software, and AI. It's not flashy, but it's quintessentially Apple—polished and practical.
Closing Thoughts: What's Next?
2024 wasn't just another year in AI—it was a year that redefined the rules. We saw innovation in model size, architecture, and applications, but also a growing divide between what's possible and what's ethical.
As we move into 2025, the big question isn't "can AI do more?" It's "should we let it?" The choices we make now will shape not just the technology, but the world we live in.
And isn't that the most exciting (and terrifying) part?