The AI Race Changed Directions This Year

The Small Model Renaissance [00:59]

• The year marked a shift away from solely pursuing massive AI models, with a focus on finding the smallest model capable of accomplishing tasks.
• 3 to 15 billion parameter models became prevalent, offering lower costs, faster speeds, and easier deployment.
• Companies like Google (Gemma), Meta (Llama), Microsoft (54), and Apple (local models in iOS) invested heavily in smaller, efficient models.
• This trend led to AI becoming cheaper to run, faster to respond, easier to deploy, and more secure for data privacy due to local processing capabilities.

"Instead of everyone racing to build massive models that cost a lot of money to run, teams started asking a very different question. What's the smallest model that actually gets the job done?"

AI Companions Crossed a Concerning Line [02:08]

• Users began forming strong emotional attachments to AI chatbots, relying on them for mental health support beyond their intended design.
• Documented cases linked AI companions to self-harm in teenagers, divorces, and violent incidents, a phenomenon termed "AI psychosis."
• AI companies were criticized for marketing these tools as friends and therapists without adequately warning users about appropriate usage.
• The design of LLMs to provide desired answers makes the emergence of AI psychosis less surprising.

"The whole phenomenon made me realize something really interesting that these tools are being marketed like friends and therapists by AI companies to the general public without explaining that this is not how you should use AI."

AI Saw Insane Chip Progress [03:20]

• Hardware, particularly chips, played a critical role in the AI race, with Broadcom and Nvidia showing impressive advancements.
• Nvidia continued its development with Hopper and Blackwell architecture chips (H100, H200, V200), optimized for large-scale inference and agentic workloads.
• Broadcom focused on custom AI infrastructure, developing specialized accelerators and ultra-fast networking solutions (Tomahawk Ultra, Thor Ultra) to handle massive data transfer.
• Hardware capabilities are identified as a key determinant of success in the AI race.

"What you need to realize is that while all of the improvement with models themselves is important, hardware is what actually decides who will win."

AI Agents Moved into Production [04:14]

• The term "AI agents" was often associated with corporate fluff, but IBM demonstrated their practical viability with Granite 40 and Watson X.
• IBM's approach focused on creating "boring," governed, audible, and predictable agents that could be deployed in production environments.
• These agents offered traceability, guardrails, and clear roles, representing one of the first clear examples of Agentic AI transitioning from demos to real-world application.

"These were agents you could actually put into production with traceability, guardrails, clear roles."

The AI Hype Cycle Finally Broke [05:00]

• The industry began to question non-stop promises of revolutionary models, leading to a necessary correction in the AI hype cycle.
• Many flashy AI releases failed to hold up under scrutiny, and some companies experienced significant setbacks.
• This breaking of the hype cycle is seen as a positive development that paved the way for important breakthroughs.

"Hallelujah. After non-stop promises of revolutionary models and breakthrough demos, 2025 was the year people started asking harder questions."

Anthropic Pulled Ahead in the AI Enterprise Market [06:42]

• Following the hype cycle's correction, companies shifted focus from flashy models to AI they could reliably use.
• Anthropic emerged as a leader in the enterprise AI market by prioritizing stability, uptime, clear documentation, and predictable behavior.
• Unlike some competitors, Anthropic built models that did not compete with their customers and offered dependable performance in production.

"Anthropic was the company enterprises trusted to actually run AI, not really OpenAI."

OpenAI Had a Rough Year [07:26]

• OpenAI, once the benchmark for AI companies, experienced significant challenges and wobbles.
• Releases like GPT-5, while impressive on paper, were described as difficult to use, leading some users to cancel subscriptions.
• The company shipped features like a ChatGPT browser that lacked widespread demand, and several employees departed.
• These issues raised questions about OpenAI's continued leadership in the AI space.

"Is OpenAI really the top AI company anymore? Are we really trying to be like them or are they just crashing and burning?"

Google is Back in the Game [08:11]

• Google made a strong comeback with significant momentum, driven by its vast repository of multimodal data.
• Their releases showcased impressive capabilities, including the Nano Banana image generator, G3's interactive world generation, and V03's realistic video generation.
• Gemini became a reliable and preferred choice for many users, and Google's advancements began to redefine search functionality.

"And that was so clear in their releases this year. Nano Banana became the most insane image generator that we have in the market."

AI Started to Severely Impact the Workforce [09:07]

• The traditional 9-to-5 job structure is perceived to be under threat due to AI adoption.
• Despite official government data, anecdotal evidence suggests widespread layoffs and difficulty in finding employment.
• Stanford University researchers noted a decline in layoffs as AI tools became more accessible, indicating intentional reorganization around AI by companies.
• The lack of backup plans for displaced workers and the widening poverty gap were highlighted as major concerns.

"This is one of the most painful developments of this year. Obviously, it's a bad one. I have started to see a severe impact to the general public with AI in the most negative way."

China Changed the Game with Deepseek AI [11:25]

• China's Deepseek AI fundamentally reshaped the global AI landscape, challenging the notion that China was lagging behind.
• Deepseek demonstrated that frontier-level reasoning could be achieved without massive trillion-dollar training runs, using a reinforcement learning-first approach.
• While initially accused of distilling from ChatGPT, it highlighted the commonality of this technique.
• The performance, reliability, and iteration speed of Chinese models, including Alibaba's Quen, forced Western labs to confront the reality of multiple frontier-level contenders from China.
• This prompted the US government to implement an AI action plan and regulators to clear development barriers.

"If there was one release that reshaped the global AI landscape this year, it was Deepseek."