AI Trends in 2025

Summary

In this video, GKCS! explores current trends and future predictions in the AI space, emphasizing diminishing returns on large language models, rising focus on cost efficiency, and the growing use of company-specific models. The discussion highlights declining hype around AI, challenges such as hallucinations and lack of goal-setting in LLMs, and the critical role of high-quality data. Looking ahead, more engineers will be hired, research will advance new architectures like JETPACK, and marketing will shift toward practical, measurable AI benefits. The video also critiques misleading benchmarks and overhyped claims, celebrating realistic assessments by experts like Yann LeCun and Geoffrey Hinton.

Takeaways

• 🤖 Large language models (LLMs) show diminishing returns as they scale, with higher costs yielding smaller intelligence gains.
• 💰 Cost optimization is becoming a key focus for companies using AI, often prioritizing efficiency over raw intelligence.
• 🏢 Company-specific AI models trained on proprietary data are rising, providing better control, higher data quality, and lower costs.
• 📉 The hype around LLMs has decreased since the peaks of GPT-3.5 and GPT-4, as companies and users gained practical experience.
• 📊 Limited availability of high-quality data is a major bottleneck for training increasingly large and capable AI models.
• 🛠️ Practical AI adoption focuses on integrating domain knowledge and creating reliable applications rather than pursuing AGI.
• 👩‍💻 Increased hiring of AI engineers is expected as companies implement and manage AI systems effectively.
• 🔬 Research into advanced architectures, like Yann LeCun's Joint Embedding Predictive Architecture, aims to improve internal consistency and goal-setting in AI models.
• ⚠️ Some AI benchmark claims, particularly in coding, may be misleading or gamed, casting doubt on reported performance metrics.
• 🌐 Public narratives around AI replacing humans or achieving AGI soon are largely exaggerated, often driven by marketing or social media hype.
• 📈 Future AI focus will likely emphasize cost reduction, domain-specific performance, ease of integration, and realistic capabilities.
• 👏 Responsible voices like Yann LeCun and Geoffrey Hinton are highlighted for providing measured, accurate assessments of AI capabilities.

Q & A

What is the main topic of the video by GKCS?

The video focuses on predictions in the AI space, analyzing trends, market forces, and potential developments in AI technology from 2025 onwards.

What trend has been observed regarding the return on investment (ROI) of large language models?

The ROI from scaling large language models like GPT-3.5 to GPT-5.2 is decreasing, meaning that as models get bigger, they are not becoming proportionally smarter.

Why are companies focusing more on cost optimization than intelligence in AI models?

Because companies are becoming skilled at using AI with domain-specific knowledge, even smaller or less intelligent models can perform well, making cost reduction more important for widespread adoption.

What are company-specific models and why are they becoming popular?

Company-specific models are AI models trained on proprietary data to serve specific organizational needs, offering lower costs, better data quality, and complete control, unlike general-purpose large language models.

How has the hype around large language models changed from 2022 to 2025?

The hype peaked around GPT-3.5 in 2022 but has decreased significantly by 2025 as companies have experimented with AI and realized its limitations, reducing exaggerated expectations.

What limitations of large language models does the video highlight?

Limitations include hallucinations, inability to set goals independently, internal inconsistency in logic, and a lack of reasoning capabilities, which prevent them from achieving AGI.

Which new AI research directions are mentioned in the video?

Research directions include contrastive learning, diffusion models, and Yann LeCun's JetPack Joint Embedding Predictive Architecture, which aim to improve internal consistency, goal-setting, and learning from less data.

What future developments does GKCS predict for the AI space?

Predictions include increased hiring of AI engineers, responsible marketing, broader adoption of company-specific models, incremental research advancements, and cautious expectations regarding AGI.

What criticism does GKCS offer about AI benchmarks and claims?

GKCS suggests that many benchmark claims, especially for code generation competitions, are misleading or rigged, giving an inflated perception of AI capabilities.

Which industry leaders are praised and criticized for their messaging on AI?

Yann LeCun and Geoffrey Hinton are praised for measured and responsible explanations, while Sam Altman, Mark Zuckerberg, and Elon Musk are criticized for hyped or misleading statements that fueled unrealistic fears and expectations.

Why is data considered a key limiting factor for large language models?

Large language models require vast amounts of high-quality data for training, and the scarcity of new data constrains model improvement, even when infrastructure and algorithms are available.

How are companies like Netflix and Swiggy using AI differently from general-purpose models?

They are building their own foundation models tailored to specific tasks, such as recommendation engines or predicting user preferences, which are more efficient and effective than relying on generic large models.