Silicon Valley’s $2 Trillion Gamble: DeepSeek Tops Nature—Is Meta the Biggest Loser of 2025?

In 2025, the global AI landscape shifted at breathtaking speed.

The mystique around Artificial General Intelligence (AGI) began to fade, while Artificial Super Intelligence (ASI) stepped into the spotlight as the next, more consequential frontier.

Anthropic executive Jack Clark issued a stark warning: seismic change is coming—and AI could split the world into two parallel realities.

This moment isn’t the result of a single breakthrough. It’s the product of a long, compounding cycle where technology, capital, jobs, and everyday life are being reshaped at the same time.

Model capabilities surged, yet the “AGI” label remains contested. Research throughout 2025 highlighted major gains in reasoning, multimodal understanding, and agentic behavior—even as debate continued about whether these systems truly qualify as general intelligence.

The investment wave supercharged infrastructure expansion. Global AI funding continued to accelerate, with generative AI attracting $33.9 billion and major tech companies pushing capital expenditures toward $400 billion. Alongside the excitement came renewed anxiety about bubbles—and growing scrutiny around energy consumption.

Workplace transformation also picked up speed, bringing both promise and pressure. AI tools increasingly feel less like an advantage and more like a baseline expectation, especially in competitive job markets.

Meanwhile, AI spread deeper into daily life—but without fully overturning it. Agents and robots entered manufacturing, healthcare, and operations, boosting efficiency in visible ways, yet many people still felt the change was incremental rather than revolutionary.

One message kept resurfacing across the industry: AGI is not the finish line—superintelligence is where the real story begins.

And if the true AI race starts after AGI, then 2025 may be remembered as the year the starting gun became audible.

One year of AI, a thousand years for humanity

Until now, every form of intelligence in nature has been biological—carbon-based, evolved, and embodied.

But the modern large language model may represent something unprecedented: a new category of intelligence that humans have created rather than discovered.

In his 2025 year-end reflections, Andrej Karpathy put it bluntly: this was the first year he—and, in his view, the broader industry—began to internalize the “shape” of LLM intelligence in a more intuitive way.

Across reasoning, multimodal processing, and agents, progress was undeniable. Systems like OpenAI’s o3 series and Google’s Gemini 3 signaled a step-change in capability. The limitations were still real in practical deployment, but the sense of “AGI on the horizon” became a shared consensus in many corners of the field.

On a growing set of tasks, leading models began surpassing human benchmarks.

Stanford’s 2025 AI Index Report noted that AI exceeded human baselines on seven evaluations spanning:

Image classification
Visual reasoning
Medium-level reading comprehension
English language understanding
Multi-task language understanding
Competition-level mathematics
PhD-level scientific questions

The most stubborn gap remained multimodal understanding and reasoning—tasks requiring cross-format, cross-discipline inference across images, charts, diagrams, and text.

Even that gap narrowed quickly.

The MMMU benchmark, designed to test interdisciplinary reasoning with university-level knowledge, became a key reference point. It stands out for four traits: broad coverage (six disciplines, thirty subjects, and roughly 11,500 questions), highly diverse image types, interleaved text-and-image reasoning, and expert-level perception grounded in deep domain knowledge.

Results climbed rapidly:

By late 2023, Google Gemini scored 59.4%.
In 2024, OpenAI’s o1 reached 78.2%.
In 2025, Gemini 3 Pro hit 89.8% on the tougher MMMU-Pro.

At the same time, release cycles became relentless. Frontier labs shipped meaningful updates every 8–12 weeks.

OpenAI’s o3 series, including o3-mini, gained attention for a “think before you answer” style of reasoning—often spending up to 10× more tokens to raise the density of intelligence, while also raising costs.

Google’s Gemini 3 was widely positioned as a multimodal high point, handling text, images, video, and audio with deeper cross-modal reasoning.

Online communities tracked the openness debate closely. Early in the year, Reddit discussions around accessible frontier models spiked.

DeepSeek-R1 and its open distillations dominated attention, though users pointed out a key trade-off: locally runnable releases were distilled models (8B or 32B), not the full 671B system, with performance often described as closer to GPT-3.5 than state-of-the-art frontier levels.

The deeper fascination was with DeepSeek’s open strategy—especially amid reports of dramatically improved training efficiency.

Researchers later claimed they reproduced DeepSeek-R1-Zero–style reinforcement learning training on a 3B-parameter model for under $30, intensifying the conversation about how far cost barriers could realistically fall.

Benchmarks for “general intelligence” were also evolving. On ARC-AGI-1, top performance pushed toward nearly 90%, and on ARC-AGI-2, AI exceeded average human scores.

Still, skepticism remained. Yann LeCun continued to argue that autoregressive LLMs have structural limits and need richer sensory grounding.

Broadly, though, 2025 marked a clear transition: AI shifted from “chatbots” to “agents”—systems that can plan, execute, and coordinate tasks with increasing autonomy.

The AGI final is 2–3 years away—depending on who you ask

If previous years were about scaling models up, 2025 looked more like a push to bring models down to earth.

Competition intensified around coding, reasoning, multimodality, long context, and enterprise reliability. The largest players sprinted to claim territory—fast.

At the same time, the conversation about AI’s future grew bigger and more concrete. Tech leaders increasingly spoke not only about AGI, but about ASI as the ultimate destination.

AGI is commonly framed as AI that can match human intelligence across a wide range of tasks. ASI goes further—systems that exceed human capability.

Key moments reflected that shift:

In June, Mark Zuckerberg formed Meta’s superintelligence lab, targeting “personal superintelligence.”
In September, Sam Altman suggested society should prepare for ASI potentially arriving before 2030.
Anthropic’s CEO argued that by 2027, AI could surpass humans in “almost every domain.”
Elon Musk went even further, predicting AI could exceed the smartest humans as soon as next year.

The implication was clear: no major player wants to miss this wave.

Zuckerberg has said he would rather risk misallocating hundreds of billions than fall behind in the superintelligence era.

After reaching a net worth of $632 billion, Musk reportedly told xAI employees that if the company can survive the next two to three years, it could emerge as an AI winner.

Some leaders, like Databricks’ CEO, suggested AGI is already here. Others, like DeepMind co-founder Demis Hassabis, remained more cautious—placing AGI in a “five to ten year” window.

Timelines differed, but one shared belief stood out: progress is accelerating—and compounding.

The pace made that hard to deny.

Within a single year, OpenAI shipped more than 30 major products and updates:

Early year: efficient models and agents (e.g., Operator, o3-mini)
Mid-year: multimodal and agent tools (e.g., Sora 2, AgentKit), plus open-weight models and GPT-5
End of year: optimized specialist performance (e.g., GPT-5.2 series) and creative tooling (e.g., ChatGPT Images)

Google, Anthropic, and xAI each had their own standout moments.

What felt like magic in January became ordinary by December.

China’s open-source surge—and DeepSeek’s breakout year

Open-source AI in 2025 wasn’t just active—it was electrified.

Around LLaMA, DeepSeek, Mistral, and broader model stacks, a dense ecosystem emerged: fine-tuning frameworks, inference acceleration, and simplified local deployment pipelines that lowered the barrier to entry month after month.

China’s open-source momentum surged—and in this narrative, LLaMA was increasingly framed as losing its center-stage dominance.

DeepSeek became the year’s biggest dark horse.

DeepSeek-R1 was described as the first large model to pass peer review and land on the cover of Nature, while founder Liang Wenfeng was named among Nature’s “10 people who mattered” of the year.

Meanwhile, other architectures saw their hype cool.

Mamba, after an early wave of attention, gradually receded outside research contexts. Discussions argued that Transformers have become deeply optimized across hardware and software ecosystems, making it economically difficult to justify retraining massive models on less-proven architectures—especially when real-world results are comparable or weaker.

The maturity of Transformer tooling imposes high switching costs. Critics also noted limitations in fixed state memory approaches, including difficulty selectively retrieving skipped tokens.

In computer vision, debate continued over whether Vision Transformers have fully replaced CNNs.

Community discussion suggested Transformers increasingly dominate many large-scale tasks, but CNNs and hybrid models remain competitive in smaller datasets, medical imaging, and specialized domains. Some pointed to ConvNeXt as a strong alternative, emphasizing that dataset quality can matter more than architecture choice—and noting practical drawbacks like higher memory demand and challenges with variable image resolution.

Preparing for what comes next

Over the past year, journalist Lee Chong Ming spoke with more than 50 technology leaders about AI—ranging from trillion-dollar executives to young founders betting their careers on what’s next.

Across boardrooms, summits, and podcast conversations, three themes appeared again and again.

1) Use AI—or risk being replaced by someone who does

NVIDIA CEO Jensen Huang repeated a message he’s emphasized multiple times:

Every job will be affected—and immediately. You won’t be replaced by AI, but you may be replaced by someone who uses AI better than you.

Many leaders echoed this, noting that younger employees may have an edge simply because they’ve already normalized AI tools as part of daily work.

Sam Altman, speaking in August on the YouTube program Huge Conversations, said that while some roles will inevitably disappear, recent graduates may be best positioned to adapt.

If he were 22 and just graduating today, he said, he would feel like “the luckiest generation in history.”

He also emphasized a more difficult question: how older workers will adapt as AI reshapes the structure of work.

Stanford professor Fei-Fei Li echoed a similar idea—arguing that the ability to master new tools can matter more than degrees.

At her startup, World Labs, she would not hire engineers who refuse to use AI tools.

That shift is no longer theoretical. It’s already visible in everyday workflows.

2) Soft skills become more valuable in the AI era

Another strong consensus: as AI automates tasks, human skills matter more—not less.

In May, Salesforce Chief Futures Officer Peter Schwartz told media that empathy and collaboration are becoming the most important skills, not programming.

If parents ask what their kids should learn, his answer was simple: learn to work well with other people.

LinkedIn’s Asia-Pacific chief economist Chua Pei Ying made a similar observation in July: communication and collaboration are becoming increasingly important for both senior employees and new graduates.

As AI compresses teams and automates work, the “human layer” becomes a differentiator.

3) Humans must remain at the center of AI

Many leaders repeatedly stressed that human agency must remain central.

Microsoft AI chief Mustafa Suleyman argued that superintelligence must support—not suppress—human autonomy.

In November, he said teams are trying to build a “humanitarian superintelligence,” warning that systems smarter than humans may be difficult to control or align with human interests.

Anthropic CEO Dario Amodei emphasized misuse risks, arguing that as advanced AI lowers barriers in knowledge work, risks and rewards scale together.

Anthropic’s “Responsible Scaling Policy,” he said, focuses on three major risk areas: AI autonomy and chemical, biological, radiological, and nuclear threats—domains where misuse could endanger millions.

Geoffrey Hinton, often called the “godfather of AI,” warned again in August that once systems surpass human intelligence, protecting humanity becomes the central challenge.

The goal, as he framed it, is ensuring that when AI becomes stronger and smarter than us, it still “cares about humans.”

From hype to reality—and a reminder of what matters

In 2025, AI didn’t “overturn everything.”

But it did lay foundations that are hard to ignore.

Karpathy described AI as crossing the threshold of “English as programming.” The implication: more people than ever can instruct machines in natural language—and get meaningful results.

The next phase won’t just be about smarter models. It will be about building a human-centered superintelligence—systems that elevate people instead of sidelining them.

Whether you build AI for a living or simply live alongside it, one takeaway stands out: learning to use AI tools is becoming essential.

This year turned AI from a spectacle into a practical force—and left a clear message behind.

Technology is a tool. Wisdom remains human.