The Race To AGI: Where Are We Now In Artificial Intelligence?

Introduction

We’re in the midst of a grand technological experiment. The quest for artificial general intelligence (AGI)—a machine or system with human-level (or beyond) cognitive abilities—is not just science fiction anymore. And so the phrase race to AGI has become shorthand for this high-stakes competition among researchers, companies, and even nations.

In this article I’ll map out where we stand in that race to AGI: what counts as AGI, what progress has been made, what hurdles remain, which players are in the race, what the risks are, and what the near- to medium-term future might look like. My aim: to give you a panoramic, detailed view (yes, nerd mode engaged) of this fascinating, strange, and unsettled domain.

What is AGI? Establishing a Definition

Before we can judge “how far” the race to AGI has come, we need clarity on what we mean by “AGI.”

According to a recent explainer by McKinsey & Company, AGI is a theoretical type of artificial intelligence with capabilities that rival human intelligence across most or all tasks: reasoning, learning, language, perception, adaptation. McKinsey & Company In other words, it’s not the narrow AI of today (which is brilliant in specific tasks) but a machine whose cognitive flexibility is comparable to a skilled human.

Other commentary stresses the distinction between narrow AI (specialized systems) and the desired broad, general capability of AGI. ai-pro.org Some definitions emphasise autonomy: an AGI system could learn new tasks without retraining, shift across domains, and reason robustly about unfamiliar problems.

Because definitions vary, we should keep in mind: the “race to AGI” is somewhat ill-defined. What one organization counts as “AGI” might differ from another.

Why the “Race to AGI” Matters

Why should we care about this race? Several reasons:

Transformative potential — If AGI is achieved, it could reshape economies, science, labour, education, warfare: everything. McKinsey estimates that AGI could “revolutionize every aspect of our lives” once it arrives. McKinsey & Company
Competitive dynamics — Tech companies, governments, and investment funds are treating AGI as a strategic frontier. The race aspect means incentives, resource mobilization, breakthroughs, but also potentially dangerous shortcuts.
Ethical, governance, safety implications — The more advanced AI becomes (towards AGI), the greater the risk of misalignment (machine intentions diverging from human values), misuse, unintended side-effects. Google DeepMind
Scientific insight — Pursuing AGI advances our fundamental understanding of intelligence, cognition, and the interplay between algorithm, data, and computation.

Thus the race is not just hype—though there is plenty of hype—but also a serious intersection of science, engineering, economics and philosophy.

Current State: How Far Along Are We in the Race to AGI?

Let’s survey where things actually stand. Because (spoiler) we’re still not at AGI—but the terrain is shifting.

Advances and Milestones

Modern large-language models (LLMs) like GPT‑4 are extremely capable in their domains: generating text, reasoning about many prompts, coding, translating. Yet these remain narrow tasks. As IBM puts it: they “lack genuine understanding” and cannot adapt to completely novel domains without retraining. ibm.com
Researchers identify that many current systems are “prediction machines” (based on patterns in data) rather than systems with a true world model or reasoning architecture. McKinsey & Company
Some recent academic work points out that scaling up the size of current models may not by itself yield AGI. For example: a 2025 paper argues that model o3 (by OpenAI) achieved high benchmark scores but still did not address the core conceptual gap to AGI. arXiv
On the development side, the company DeepMind (part of Google LLC) published its “Taking a Responsible Path to AGI” blog, outlining how it plans to chart progress, manage safety, and evaluate risk on the path. Google DeepMind
Expert surveys provide some numbers: One compilation shows many experts believe there is a ~50% chance of AGI by 2040-2060 (assuming current trajectories hold) though estimates vary wildly. AIMultiple

Assessment: Speed Check

If I were to give a grade, with some humility: we might say that in the race to AGI we’re still in the “major stride” phase, not yet the “final leg”. We have made significant progress in narrow AI and bridging domains, but the jump to full general intelligence remains elusive.

Key indicators:

Generalisation: Current systems still struggle to transfer knowledge broadly, adapt to new domains without retraining.
Autonomy: Most systems need human-in-the-loop or retraining; true self-learning and self-improvement are rare.
Understanding/Reasoning: The leap from pattern-matching to genuine reasoning (with common sense, abstraction, planning) remains large.
Robustness: Many systems still fail in adversarial or untrained contexts.

Hence, the race to AGI is alive, and acceleration is happening, but the finish line is neither near nor clearly visible.

The Major Hurdles in the Race to AGI

Here are the main obstacles standing between narrow AI and AGI. (Downhill run ahead for the curious.)

1. Conceptual and Theoretical Gaps

We lack a unified theory of intelligence—what exactly does it mean to “understand” or “reason”? Some commentators say the race to AGI is “vibes-based”: meaning we’re scaling models without fully knowing why they work or what the path to AGI is. The Guardian A 2025 paper on embodied intelligence argues that new architectures (embodied learning, sensorimotor feedback loops) may be essential. arXiv
Without a clear roadmap, the race to AGI becomes one of engineering plus guesswork.

2. Computational & Data Requirements

The bigger the model, the more compute, data, energy. Training large models becomes costly and energy-intensive. Scaling up may hit physical limits.
Also, more compute alone may not produce qualitatively different capabilities (see above).

3. Transfer Learning & Generalisation

Narrow AI systems succeed when the tasks align well with their training. But general intelligence demands adaptability: to learn new tasks, apply previous learning to new domains, deal with novelty. That’s hard. Many systems fail outside their training distribution.

4. Autonomy and Embodiment

Humans learn by interacting with the world, having embodied experience, feedback loops, unexpected events. Many AI systems are purely virtual, lacking physical interaction or real-world grounding. Researchers argue this may be a critical missing link on the race to AGI. arXiv
Without embodiment, abstract reasoning may lack the real-world scaffolding that human intelligence uses.

5. Safety, Ethics, Alignment, Governance

Even if the race to AGI succeeds, if the system is misaligned (goals diverge from human values), we risk large scale unintended harm. Safety research, robust governance frameworks and regulation are lagging. Tech Policy Press
An open-letter by the Future of Life Institute recommended pausing giant training runs until safety, alignment and governance are better understood. Wikipedia
So the pace of value-alignment and policy is a bottleneck too.

6. Benchmarking and Evaluation

How will we know when AGI arrives? Which benchmark? Researchers argue that current benchmarks are insufficient because they reflect narrow tasks or distributions of experience. A 2025 academic paper argues that high scores on specific tasks do not imply AGI. arXiv
Better metrics, better definitions, clearer evaluation frameworks are needed.

Who’s in the Race: Players, Strategies & Stakes

Let’s peek behind the scenes: who’s racing, how, and why.

Big Tech & AI Labs

DeepMind/Google: Focused heavily on long-term general intelligence. They’ve published about safety, autonomy, agentic systems.
OpenAI: Aiming toward human-level capability; in its blog and public statements links AGI to future superintelligence.
Anthropic, Microsoft, Meta (Facebook): Also investing heavily.
Governments, national agencies: AI leadership is becoming a strategic issue for national competitiveness, economic growth, defence.

Strategies

Different camps have different tactics:

Scale up: More parameters, more data, more compute.
Architectural innovation: Beyond scaling; new model designs (transformer alternatives, multimodal models, embodied systems).
Embodiment/agency: Agents interacting with environments, robots, feedback loops.
Safety and alignment: Parallel track to build guardrails, value alignment, robust governance.
Hybrid human-machine collaboration: Instead of replacing humans, augmenting them.

High-stakes incentives

Economic gain: early leadership in AGI could lead to massive productivity increases, new industries.
Strategic leadership: dominance in AI implies geopolitical or national advantage.
Reputational prestige: achieving AGI first would be huge.
Risk: The faster the development, the higher the potential for missteps, accidents, misalignment.

What Does “Winning” the Race to AGI Look Like?

What would success look like? Given the ambiguity, let’s sketch possible scenarios (with the caveat: these are working theories).

Scenario A: Human-level AGI arrives

A machine system that reliably performs any intellectual task a human can (or most tasks) — learns autonomously, transfers knowledge, adapts to novel situations, functions in the real world. That would mark a kind of finish line.
In this scenario, the race has a clear winner (company/nation), standards get established, regulation follows, etc.

Scenario B: Narrow AGI / Domain-General but not Fully General

We might see systems that are “AGI-ish”: very broad capability across many domains but still not full human-level flexibility. For example: an AI that handles most digital tasks, design, planning, strategy, but struggles with novel physical embodiment or creative open-ended research.
This would look like a partial win.

Scenario C: No clear winner; incremental progress

The race to AGI might continue for decades with many labs improving capabilities, but none crossing a decisive line. Instead we see many semi-general systems, and humans remain central. The “winner” is ambiguous or shared.

What winning implies

Rapid acceleration of innovation.
New societal models: labour, education, economy.
Big questions of control, alignment, ethics, governance.
Possibly existential risk if mismanaged.

Where the Race to AGI Is Heading: Near Term (Next 5-10 Years)

Given what we know, here’s a plausible roadmap (not prophecy!).

2025-2030

Continued improvements in LLMs, multimodal models (text, image, audio, video, maybe robotics).
More focus on agentic systems: models that plan, act, get feedback, adapt.
Embodied AI research ramps up: robots, simulation environments, interaction loops.
More serious safety and alignment research begins flooding out of labs.
Governments and regulators begin forming frameworks for “capability oversight”.
Some breakthroughs may shift the “vibes” of the race: new architectures, new learning paradigms (e.g., unsupervised, self-supervised, reinforcement plus model-based).
Companies may claim “AGI milestones” — but these will be contested and ambiguous.

2030-2035

If trends accelerate, we may see something approximating general intelligence in practical settings: e.g., systems that outperform humans in many economic tasks, adapt across domains, incorporate world models. Some experts believe this window. AIMultiple
At this point governance, ethics, alignment will become urgent — the race to AGI becomes less speculative, more real.

Beyond 2035

If AGI is achieved, society faces transformation: jobs, creativity, science, power relationships.
If too fast or uncontrolled, risks of misalignment, runaway systems, existential hazards.
Alternatively, we may find the hurdles bigger than expected, and AGI remains decades away.

The Risks & Ethical Dimensions of the Race to AGI

Because yes: this is where the wise-nerd lens comes in. Not all progress is benign, and there are deep philosophical questions.

Misalignment and Autonomous Behaviour

If an AGI system is built, but its goals or values diverge from ours (human values, flourishing), we risk systems that optimise for unintended outcomes. This is sometimes called the “alignment problem”. Google DeepMind
The faster the race to AGI, the higher risk of cutting corners in safety.

Concentration of Power

Who wins the race to AGI may wield enormous power: economically, politically, socially. That concentration raises issues of equity, access, control.

Societal Disruption

If AGI becomes capable of performing most of the tasks humans can do, large scale labour disruptions may follow. Education, purpose, jobs may all shift dramatically.

Existential Risk

While speculative, some philosophers and scientists argue AGI (or beyond, superintelligence) could pose an existential threat to humanity. Wikipedia
Even if low probability, the magnitude of the risk calls for attention.

Ethical and Governance Challenges

Who sets the agenda for AGI development?
How transparent are the labs and governments engaged in the race?
How do we ensure systems respect human dignity, fairness, privacy? A paper from 2023 emphasised the importance of “human-centred design” in AGI development. arXiv
Regulation: Should there be moratoriums? Licensing? International treaties? The Open Letter to “Pause Giant AI Experiments” is one such call. Wikipedia

Reflections: Why the Race to AGI Feels Unsettled

It might seem odd: we’re racing towards something we don’t fully define and don’t know when we’ll reach. Some reflections:

Uncertainty of goal-post: What counts as AGI is fuzzy. Without a clear, agreed benchmark, the race is open-ended.
No clear roadmap: Unlike a moon rocket with fixed parameters, we don’t know the “launch velocity” required for AGI. Some argue we’re using bigger rockets (models) but we may still not know how to break orbit. The Guardian
Engineering vs science tension: Many current efforts are engineering (build big models) rather than foundational science (understand intelligence). The race can sometimes privilege hype and capital over deep understanding.
Narrative, hype, and real progress: As noted by critics, the “race to AGI” narrative can obscure more mundane but critical improvements in narrow AI. Tech Policy Press
Global coordination challenge: The race is international, and alignment (in both senses) is hard: aligning machines with humans, aligning actors with each other, aligning ethics with progress.

What You Should Watch For in the Race to AGI

If you’re curious (which you are, I know), here are signals that matter:

New architectures that go beyond just scaling parameters (e.g., embodied intelligence systems, multimodal agentic systems).
Evidence of genuine transfer learning across distant domains—an AI learned Task A and without retraining handles Task B and Task C well.
Labs publishing safety and alignment frameworks that are concrete, audited, transparent.
Corporate/government policy moves: regulation, oversight, treaty-making around advanced AI systems.
Public demonstrations of systems that adapt, plan, execute, learn new domains with little human intervention.
Benchmarking improvements that shift away from narrow metrics toward more general intelligence metrics (diverse tasks, novel domains).
Ethical incidents, accidents, misalignment cases: as the race accelerates, the probability of such events increases.

Implications for Business, Society & Individuals

The race to AGI is not just for labs and technologists—it affects all of us.

For Business

Companies should start thinking now: what happens if AGI becomes real in the next decade? How will workflows, competition, talent, automation shift? Even before full AGI, narrow-but-very-powerful systems will disrupt industries.

For Society & Policy

Education systems, labour markets, regulation, public discourse all must prepare. We will need frameworks around how advanced AI systems are deployed, how benefits are shared, how risks are mitigated.

For Individuals

One practical takeaway: “learn how to work with AI” seems wiser than “worry about being replaced by AI”. Many experts emphasise augmentation rather than replacement. But being literate about the capabilities and limits of AI (and the race to AGI) is a wise move.

For Philosophers, Ethicists, Citizens

Questions about human identity, value, purpose become more pressing. If machines approach human-level intelligence, what does that say about what it means to be human? How do we maintain human-centric values? The race to AGI forces these questions into the open.

Concluding Thoughts

The race to AGI is not a sprint with a clearly visible finish line—it’s more like a long ultramarathon through foggy terrain, populated with signposts, false summits, and evolving pathways. Yet the stakes are high: the winner (if one emerges) may reshape what it means to live, learn, work, and think.

Today we stand at a pivotal moment: narrow AI is flourishing, architectures are improving, investment is massive—but full AGI remains an open question. The science isn’t settled, the path isn’t clear, yet the momentum is undeniable.

ByJarrett Adwell