Hello Builders,
This week, we start wrapping up 2025, a year that has been transformative in many aspects: the AI ecosystem was defined by a powerful tension between a necessary, sobering hype correction and the relentless pace of underlying technical progress. While industry leaders like Sam Altman and Mark Zuckerberg openly acknowledge a financial bubble of historic proportions, with trillions of dollars in planned infrastructure spending dwarfing the dot-com era, the fundamental limitations of current LLM-based approaches are becoming clearer. Yann LeCun’s departure from Meta to pursue alternative architectures underscores a growing schism in the path to AGI, just as Chinese open-weight models like Qwen and DeepSeek are outpacing their US counterparts in technology. This market irrationality is paradoxically juxtaposed with concrete, measurable breakthroughs: OpenAI’s new FrontierScience benchmark revealed a staggering improvement in scientific reasoning capabilities over the last two years, proving that even as the hype bubble swells, the technology’s real-world potential to accelerate science is advancing rapidly.

This week’s signal in the noise

• The AI bubble is now openly acknowledged by tech leaders, with planned infrastructure spending reaching an unprecedented $12 trillion scale, dwarfing previous tech booms.

• Yann LeCun’s departure from Meta signals a fundamental schism in AI research, challenging the dominant LLM scaling hypothesis and seeking new paths to AGI.

• Chinese open-weight models are now technologically leading the ecosystem, creating a new competitive and geopolitical landscape for AI development.

• OpenAI’s FrontierScience benchmark demonstrates a dramatic leap in scientific reasoning (from 39% to 92% on GPQA in two years), proving tangible progress amidst the hype.

The Great Hype Correction of 2025

2025 is being framed as the year of the great AI “reckoning,” a necessary market and intellectual correction to the inflated expectations that have been set since 2022. The symbolic peak of this hype cycle was the underwhelming launch of GPT-5 in August, which, after months of grand promises, felt merely incremental. This has led to a broader questioning of the exponential progress narrative, comparing the current state of LLMs to the mature smartphone market, where annual updates bring minor improvements rather than revolutions. This shift demands a strategic recalibration, shifting focus from chasing ever-larger models to finding sustainable, high-value applications for the powerful yet limited tools that exist today. It’s a prompt to reassess technology roadmaps, distinguish between genuine capability and marketing-driven hype, and prepare for a period of disillusionment in which only the most resilient and value-focused applications will thrive.

https://www.technologyreview.com/2025/12/15/1129174/the-great-ai-hype-correction-of-2025/

The Schism in the Path to AGI: Beyond Token Prediction

A foundational debate, crystallized in a public discussion between Meta’s Yann LeCun and DeepMind’s Adam Brown, is challenging the industry’s core architectural assumptions. LeCun argues that the dominant approach of autoregressively predicting discrete tokens is a dead end for achieving true intelligence, as it fails to grasp the continuous, high-dimensional nature of the real world. He highlights the massive inefficiency of LLMs compared to a child’s learning process, advocating for new architectures like JEPA that learn abstract world models. This critique gained significant weight with the news of LeCun’s departure from Meta to form a new company focused on “Advanced Machine Intelligence” (AMI). This development signals a critical juncture for technology strategy, suggesting that diversifying research and development efforts beyond simply scaling existing LLMs is crucial. Betting entirely on the current paradigm may mean ignoring the architectural breakthroughs required for the next level of machine intelligence and real-world interaction.

https://the-decoder.com/the-case-against-predicting-tokens-to-build-agi/

The East is Open: China’s Dominance in Open-Weight Models

The competitive landscape for foundational models has been quietly redrawn, with Chinese companies now leading the open-weight ecosystem. The release of DeepSeek R1 in January 2025 was a watershed moment, technologically challenging and, in some cases, surpassing Western counterparts. Today, Alibaba’s Qwen model family is the most downloaded in the world, and models from labs like DeepSeek and Moonshot AI are setting new performance benchmarks, particularly in math and reasoning. While US companies like Airbnb are beginning to adopt these models for their cost and performance advantages, broader adoption is hampered by compliance and geopolitical concerns. This trend presents a complex strategic challenge: ignoring the superior performance of these models is a competitive risk, but adopting them introduces significant supply chain and data governance questions that must be carefully navigated.

Understanding AI

The best Chinese open-weight models — and the strongest US rivals

DeepSeek’s release of R1 in January shocked the world. It came just four months after OpenAI announced its first reasoning model, o1. The model parameters were released openly. And DeepSeek R1 powered the first consumer-facing chatbot to show the full chain of thought before answering…

10 months ago · 66 likes · 4 comments · Kai Williams

The $12 Trillion Bubble: AI’s Unprecedented Infrastructure Bet

The financial scale of the AI boom has entered a new territory of irrational exuberance, with industry leaders now openly discussing being in a bubble. The numbers are staggering: OpenAI has pledged $500 billion for data centers, with a moonshot goal of building infrastructure that could cost over $12 trillion. Bain estimates the industry needs to generate $2 trillion in annual revenue by 2030 just to justify the current spending wave—more than the combined 2024 revenue of Amazon, Apple, Alphabet, Microsoft, Meta, and Nvidia. Unprofitable startups like OpenAI and Anthropic are projected to burn through $140 billion and $20 billion, respectively, before 2030. This environment, characterized by circular deals and a chase for AGI, creates systemic risk. For decision-makers, this underscores the critical need for rigorous ROI analysis on AI investments and highlights the immense financial risk concentrated in a few foundational model providers, whose potential failure could have cascading effects across the industry.

Link: https://www.technologyreview.com/2025/12/15/1129183/what-even-is-the-ai-bubble/

From 39% to 92%: Measuring Real Progress in Scientific Reasoning

Cutting through the hype cycle, OpenAI has provided concrete evidence of dramatic capability improvements with its new FrontierScience benchmark. On the “Google-Proof” GPQA science benchmark, OpenAI’s models improved from a 39% score (GPT-4 in 2023) to a remarkable 92% with GPT-5.2 in 2025, surpassing the 70% expert baseline. The new, more challenging FrontierScience benchmark, created by PhDs and Olympiad medalists, shows GPT-5.2 achieving 77% on Olympiad-level problems and 25% on open-ended research tasks. This data provides a crucial signal: while true open-ended research remains a challenge, AI's ability to perform high-level, structured scientific reasoning is advancing at an astonishing rate. This indicates that the technology is ready to be deployed to accelerate complex, structured workflows in R&D, engineering, and data analysis, moving beyond simple automation to become a genuine partner in discovery and innovation.

Link: https://openai.com/index/frontierscience/

The Agentic Stack

Hello Builders, issue #5