AI got its Nobel moment — twice in one week. Meanwhile, the model wars entered a new phase, OpenAI launched a search engine, and investors poured more money into AI than the GDP of most countries.
In this edition:
Two Nobel Prizes go to AI pioneers
OpenAI teaches models to think before answering
ChatGPT becomes a search engine (and what that means for being found)
The quarter's biggest model releases and funding rounds
Google's quantum chip does something useful
Two Nobel Prizes. One Week. AI Arrives in Stockholm.
For decades, AI researchers watched other fields collect Nobel medals. In October 2024, the committee awarded two prizes to AI pioneers in the same week. It was the clearest signal yet that artificial intelligence has crossed from engineering into fundamental science.
Need to know:
On October 8, Geoffrey Hinton and John Hopfield won the Physics prize for foundational work on artificial neural networks. Hopfield built associative memory networks. Hinton developed backpropagation. Every model you've used this year traces back to their research.
The next day, Demis Hassabis and John Jumper took the Chemistry prize for AlphaFold, the system that cracked protein structure prediction after fifty years of failed attempts. Drug development, disease research, and synthetic biology all got faster overnight.
The double award forced a question the scientific establishment had been avoiding: is AI a tool, or is it a science? The Nobel committee's answer was unambiguous. AI has moved from "technology that helps scientists" to "discovery of universal truths."
Hinton himself has spent two years warning about AI's risks. He resigned from Google in 2023 specifically to speak freely about the dangers. Winning a Nobel for the thing you're publicly terrified of is a particular kind of irony.
The signal matters more than the debate. AI isn't adjacent to science anymore. It's inside it. And the institutions that grant legitimacy just said so, in the loudest way they know how.
Breakthrough
OpenAI Teaches Models to Think
OpenAI released o1-preview on September 12. Instead of generating answers word by word, o1 thinks first, producing an internal chain of reasoning before responding.
Need to know:
Previous models answered immediately. o1 works through problems step by step, like a person solving an exam question. The result is dramatically better performance on maths, science, and coding tasks.
o1 matched PhD-level performance on physics, biology, and chemistry benchmarks where GPT-4o scored like an undergraduate. On competitive programming, it jumped from the 11th percentile to the 89th.
By December, OpenAI released the full o1 with 34% fewer errors than the preview, then teased o3, which scored high enough on the ARC-AGI benchmark to restart arguments about what "general intelligence" actually means.
"Think longer, answer better" is a new scaling axis. Instead of just making models bigger, you can make them slower and smarter. That changes both the economics and the ceiling of what's possible.
Product
ChatGPT Becomes a Search Engine
OpenAI launched ChatGPT Search on October 31, turning the world's most popular AI chatbot into a direct competitor to Google. For anyone thinking about how businesses get found online, this is the story that matters most from Q4.
Need to know:
Instead of blue links, ChatGPT now pulls real-time information from the web and delivers answers with source citations inside the conversation. Users are asking it the kinds of questions they used to type into Google: definitions, comparisons, how-to guides, product recommendations.
ChatGPT is already the largest AI traffic referrer, sending more visits to websites than Reddit or LinkedIn. The volume is still small (roughly 0.1% of total web traffic) but the growth curve is steep.
Nearly 60% of ChatGPT's source citations come from URLs that don't rank in the top 20 organic Google results. Well-structured content with clear headings, statistics, and quotations gets cited 30-40% more often.
The rules for getting found are changing. Most businesses haven't started playing the new game.
The Roundup: The Model Wars Accelerate
Google ships Gemini 2.0. Google released its most capable model family in December, led by Gemini 2.0 Flash. Alongside it came Project Astra (a universal AI assistant prototype), Project Mariner (an agent that takes actions in Chrome), and Jules, an AI coding agent. Google is clearly betting that models alone aren't enough — agents that do things are the next battleground. Build
DeepSeek V3 crashes the party. The Chinese lab released DeepSeek V3 in December with a Mixture-of-Experts architecture that matched or beat leading proprietary models on key benchmarks — at a fraction of the training cost. Open-source AI just got a lot more competitive, and the assumption that frontier models require frontier budgets took a hit. Breakthrough
Anthropic upgrades Claude. October brought an upgraded Claude 3.5 Sonnet and the new Claude 3.5 Haiku. Sonnet's upgrade was significant enough that it stayed competitive with o1-preview on reasoning tasks. Anthropic also signed a deal with Palantir and AWS to bring Claude to U.S. intelligence and defence agencies — a clear signal that the "safety-first" lab is playing to win government contracts. Product / Industry
Sora finally ships. After ten months of anticipation, OpenAI released Sora Turbo on December 9. Users can generate 1080p videos up to 20 seconds long. The physics are still wonky and complex actions fall apart, but the floor for AI video just rose dramatically. Plus subscribers get 50 videos per month included. Product
Meta makes Llama leaner. Meta released Llama updates in October that are four times faster and 56% smaller than their predecessors. The goal: sophisticated AI running on smartphones. The open-source model race is now a size and efficiency race, not just a capability race. Build
The Roundup: Follow the Money
OpenAI raises $6.6 billion. The round valued OpenAI at $157 billion, making it the most valuable private tech company on Earth. The money came as OpenAI continued its transition from non-profit research lab to something that looks a lot more like a tech giant. Money
xAI doubles its valuation. Elon Musk's AI company raised $6 billion in Series C funding in November, reaching a $50 billion valuation — doubling in six months. xAI has spent aggressively on compute, building one of the world's largest GPU clusters in Memphis. Money
Databricks joins the $10B club. The data and AI platform raised $10 billion at a $62 billion valuation, one of the largest private funding rounds in tech history. Enterprise AI infrastructure is now as richly valued as the model companies themselves. Money
AI funding crosses $100 billion for 2024. Total AI venture funding hit $100 billion for the year, up 80% from 2023. Nearly one-third of all global venture capital went to AI companies. In Q4 alone, 62% of North American startup funding went to AI. The concentration is extreme — mega-rounds above $100 million accounted for over 80% of all AI funding. Money
First international AI treaty signed. The European Commission signed CETS No. 225 — the first legally binding international AI treaty. It requires AI systems to respect human dignity, transparency, and accountability throughout their lifecycle. Canada launched its AI Safety Institute. Japan signalled regulations coming for generative AI. The regulatory machinery is warming up globally. Trust & Power
Quick Bytes
Google's Willow quantum chip completed a benchmark calculation in under five minutes that would take a classical supercomputer 10 septillion years.
Perplexity cemented itself as the third major AI search interface, hitting 100 million monthly users by the end of Q4.
The EU AI Act entered into force on August 1, with prohibited practices and AI literacy obligations kicking in from February 2025.
Microsoft unveiled Copilot Actions at Ignite 2024, automating meeting summaries and presentation drafts across Microsoft 365.
Google released Veo 2 and Imagen 3 in December, making AI-generated video and images a multi-vendor race heading into 2025.
Google's AI Overviews now cite an average of 13 sources per response, up from about 7 earlier in the year. The surface area for AI-driven discovery is expanding fast.