AI 2025-2026: Consensus, Conflict, and the Next Frontier of Scaling Laws
Insight
Insight
Chapter 01 // Annual Review
2025 AI Retrospective:
Which trendsexceeded expectations?
"We used to think 2025 was the compute endgame, but it turned out to be just the entry ticket for 'system players'."
Silicon Valley's most dramatic moments
Looking back at the end of 2024, who would have thought Meta's path would be such a roller coaster? From massive layoffs to Llama 4's strategic pivots, Yann LeCun's persistence intertwined with external skepticism to create a grand drama.
"The news about Meta falling behind was so dramatic. After the layoffs, Llama 4, and even Yann LeCun's 'strategic controversy'—I didn't see that coming at all."
"Actually, Meta's edge lies in Zuck's control and cash flow. Even if Llama 4's release underwhelmed, his push on the application side has been extremely aggressive. Preventing a Google monopoly is the industry consensus."
Enterprise AI: From Hype to Reality
The biggest shift in 2025: people stopped blindly chasing the 'most expensive, most powerful' models. In highly regulated sectors (healthcare, finance, defense), vertically deployed small models became the real prize. Budgets are no longer just for testing the waters, but for systemic investment.
Core Insight:
- "It's not that the bigger the model, the better; it's that the safer and more industry-savvy it is, the better."
- "AI-powered industry has shifted from a single narrative to a collective awakening across all sectors."
User Milestone
1 Billion
OpenAI's user base is nearing 1 billion, but the game between profit margins and compute costs has reached a breaking point.
Scaling Law: Not Yet at the Ceiling
Discussions on whether the Scaling Law has failed peaked in 2025. Although some of Google's attempts raised doubts about the direction, the mainstream consensus still leans toward an 'order of magnitude' of remaining potential.
There is a key cognitive shift here: In the past, we thought data meant just scraping everything off the web; now we realize that data cleaning, weighting, and how to make the model understand Domain Knowledge is the real moat.
"It's not just about daisy-chaining a million GPUs. From training to analyzing why things crashed, there's a lot of delicate work involved."
The 2026 Outlook
2026 will be a
belonging tosystem-level players'tough battle
As the 'wow' threshold is pushed infinitely higher, simple API calls can no longer satisfy the market. AI's business logic is making its final, perilous leap from 'toy' to 'productivity tool'.
Editor's Note
The impact of DeepSeek mentioned in this section is not just a technical victory, but a spark where 'cost optimization' met the 'open-source spirit' in 2025. As corporate consensus solidifies, the battle over model architectures has truly entered a fever pitch—next, we’ll talk about the DeepSeek revelation that caught the world off guard.
CHAPTER 02
AI Development Mainline and Paradigm Shift:
Enterprise AI Consensus Takes Shape
"If 2025 was the bottom of the AI 'Trough of Disillusionment,' the consensus now is: stop talking about vague AGI; we want systems that actually drive business."
We just reviewed those 'failed' predictions from 2025, and there's a general sense that 'compute anxiety' is easing. But immediately, the conversation turns to the variable keeping all the big players awake at night—DeepSeek. This isn't just a win for one model; it marks a radical shift in the entire AI industry paradigm.
The DeepSeek Revelation: Efficiency is King
Honestly, the emergence of DeepSeek R1 was like dropping a bombshell into the stagnant air of Silicon Valley. We used to think the Scaling Law was just 'brute force aesthetics'—throwing money, GPUs, and data at it. But DeepSeek told us:Smart architecture and extreme engineering optimization can achieve the same level of intelligence at 1/10th the cost.
This hit the open-source ecosystem directly. The logic has changed: it's no longer about who has the most parameters, but who can provide the most robust System 2 thinking (slow thinking) within a limited Inference Budget.
The Open-Source Shockwave
DeepSeek proved the extreme efficiency of Chinese teams in model distillation and Reinforcement Learning (RL), forcing established open-source players like Llama to accelerate their iterations.
Cost reduction for the same performance
Paradigm Shift
From 'large parameters' to 'long thinking' (Reasoning models).
"Scaling Law isn't dead; it just shifted from'compute for scale',
to'time for intelligence'."
Data Estimation: Model intelligence growth is shifting from the pre-training phase to the inference phase
Neolab and the 2026 Outlook
During our internal discussions at Neolab, we realized that the battle for next-gen models is no longer about 'how many books you feed them,' but about 'how they solve math problems'.
Keyword: System 2 Thinking
That is, the model's ability to perform self-verification, reflection, and multi-path search before outputting. This is the true stability that enterprise-grade AI requires.
Enterprises are reaching a consensus: AI is no longer just a poet writing verses, but a logically rigorous auditor capable of handling SOPs 24/7.
HOST A
“So where do you think Meta's pressure is coming from? Is it DeepSeek's efficiency or OpenAI's lead?”
GUEST B
“Meta's pressure is that it's being squeezed from both ends. On the high end, it's suppressed by OpenAI's closed-source logic, while on the low end, its efficiency is being ambushed by open-source 'grind kings' like DeepSeek. Its rumored acquisition of Manus is actually just trying to buy a ticket into the 'Agent' track.”
Terminology Note: Test-time Compute
Refers to the amount of computation a model spends during the inference phase (i.e., when answering a question). Through Chain of Thought (CoT) and search algorithms, the model can 'think' longer before answering, significantly improving accuracy for complex problems without needing to infinitely increase the model's parameters.
Summarizing this phase: 2025 completed the mindset shift from 'brute force miracles' to 'architectural miracles.' Next, we'll see how these consensus views dismantle Meta's strategy and how OpenAI was forced onto that crisis-ridden path toward an IPO.
NEXT CHAPTER
Meta's Dilemma and OpenAI's IPO Gamble →
Meta'sGamblewith
OpenAI'sNarrow Gate
“Having just talked about the physical limits of Scaling Laws, we need to look at how these Silicon Valley giants are stabbing each other in the back amidst their 'growth anxiety.'”
Meta Acquiring Manus: A Stroke of Genius or a Strategic Face-palm?
Zuck's recent move really stunned me. Meta has been pushing Llama, branding itself as the 'Android of AI,' but then it turns around and acquires the Manus team. This sends an extremely dangerous (yet real) signal: **Having a foundation model alone is no longer enough to protect your territory; you have to get your hands dirty and build Agents yourself.**
“Strategic Maneuvering”
Meta is undergoing a painful transition from 'compute hegemony' to 'application closed-loop.' Open sourcing is to stall competitors; acquisition is to arm itself.
Speaker A
OpenAI is now valued at $150 billion—who can afford that price? They're currently like a massive, elegant unicorn, but one that eats tens of millions of dollars in compute every day without an immediate way to monetize.
Speaker B
That's why Sam Altman is in a rush to pivot. They aren't just fighting Google for search; they're also competing with Apple for on-device AI and even trying to develop their own chips. OpenAI's road to IPO is essentially a survival drama racing against their 'burn rate.'
AI Big Three Power Quadrant
“Anthropic is staging the AI version of 'surrounding the cities from the countryside': they aren't trying to beat OpenAI on traffic, but in terms of coding and B2B depth, they are terrifyingly solid.”
! Why 2026 will be the true Year One of AI applications?
Everyone is complaining that AI applications right now are nothing but chat, chat, and more chat. But I believe this is just the growing pains of the 'infrastructure phase.' Given current hardware cycles and model evolution, 2025 will be the laboratory stage for Agents.
By 2026, when model inference costs drop by two orders of magnitude and 'Computer Use' capabilities are tuned to 99% accuracy, you'll find that you no longer need to 'open' an app.
Deep Insight:
The 2026 explosion won't happen because models get smarter, but because they become 'cheaper' and more 'obedient.'Scaling Laws face skepticism; the model layer begins a cutthroat competition over long-context and multimodality.
Tech giants begin large-scale M&A; moves like Meta's acquisition of Manus will become the norm.
Vertical Agents replace 40% of junior repetitive tasks in legal, medical, and engineering fields.
Vertical Agent
and the Industry 'Last Mile'
From Anthropic's B2B strategy to the explosive growth of 2026, the question everyone cares about most is: where is the 'entry ticket' for average developers and startups in this grand narrative?
“Since foundation models are already so powerful, do vertical-specific Agents still have a moat? Won't they just get 'flattened' by a single update from OpenAI or Anthropic?”
This is a classic 'panic.' But my view is clear:General models provide intelligence, while vertical Agents provide labor.
Your moat isn't the API you're calling; it's your obsessive, almost pathological understanding of industry workflows. LLMs know how to code, but they don't know your company's complex compliance processes; they know legal statutes, but they don't know a judge's specific preferences. This 'grunt work' is the real moat for vertical opportunities.
Precision Medicine / Biopharma
It's not just a simple literature search; it's an Agent that can understand experimental protocols and automatically interface with lab automation equipment. This level of deep integration is something model providers simply don't care about and can't pull off.
Legal and Compliance Audit
"We don't want a chatbot that can recite the legal code; we want a 'digital lawyer' that can spot the one clause violating 2024 regulations across thousands of contracts in a single glance."
Industrial Supply Chain Scheduling
Here, the Agent needs to interface with ERP systems in real-time. Its value lies in the 'execution' following the decision—automatically placing orders, adjusting inventory, and rerouting logistics.
In the next three years, customized Agents for vertical industries will capture the lion's share of the market's value.
“Don't knock yourself out competing on AI's intelligence,
instead, dig deep into the friction of the physical world."
Agentic Workflow: A concept Andrew Ng has been emphasizing lately. It refers to the idea that instead of chasing a perfect one-shot result, we should let AI work like a human—completing complex tasks through a cycle of 'reflection, tool-seeking, planning, execution, and error correction.' This is the core secret for vertical Agents to handle professional-grade work.
Having discussed the survival rules of vertical tracks, our conversation today is coming to an end.
Finally, let's return to the basics: Is the finish line of this AI race the evolution of tools, or the restructuring of production relations?
Next: Conclusion & Final Thoughts →
Related Episodes
E226｜聊聊DeepMind创始人哈萨比斯：一个科学家与失控的AI竞赛
E224 | Deep Dive into Clawdbot: Why is it the First Phenomenal Product of 2026?
E223 | The Year of AI Application Explosion: Discussing Model Evolution and Commercialization
E222 | The Death of Skinny Jeans: Who Actually Defines Fashion Trends?
Your Catchy English Title