Are we already in the foothills of the AI singularity?

Something strange is happening in AI right now. On paper, benchmarks still show clear limits: today’s models struggle with open-ended reasoning, long-term planning, and learning from experience. But in practice, AI agents are already coding, researching, planning, paying for things, and cutting real-world work from days to minutes.

That tension is the story. The singularity might not start when AI becomes perfect. It might start when imperfect AI becomes powerful enough to speed up everything around it.

Why top AI leaders suddenly say “singularity” out loud

In 2026, some of the people closest to the cutting edge started using language that would have sounded wild just a few years ago.

Demis Hassabis, CEO of Google DeepMind and a key figure behind AlphaFold’s Nobel-winning impact on chemistry, recently said we are “standing in the foothills of the singularity.” He’s known for being cautious, not dramatic—and that makes the statement land harder.

Even more striking, his AGI timeline has shifted. In mid‑2025, he predicted AGI might arrive between 2030 and 2035. Less than a year later, he narrowed that to roughly 2029–2030. That’s a huge acceleration from someone with direct visibility into frontier systems like Gemini (see also everything Google just revealed about Gemini and AI at I/O 2026).

He also framed the impact bluntly: when AGI arrives, he expects it to be like the industrial revolution—multiplied by 10, and happening at 10× the speed.

Other insiders calling 2026 a turning point

Hassabis isn’t alone. Across the ecosystem, leaders are starting to talk as if the singularity is no longer a sci‑fi endpoint, but a phase we might already be entering.

Recent claims and hints include:

• Elon Musk called 2026 “the year of singularity,” saying we’ve already entered it.
• Dario Amodei, CEO of Anthropic, said we don’t know whether current AI models are conscious.
• Patrick Collison, CEO of Stripe, suggested Q1 2026 may be remembered as “the first quarter of the singularity.”
• OpenAI’s VP of research, Aidan Clark, hinted that AGI may have already arrived “in some form.”
• OpenAI president Greg Brockman said the company has “line of sight” to AGI.
• Investor Marc Andreessen claimed AGI was effectively reached about three months ago with the latest frontier models.

These aren’t random Twitter hot takes. They’re coming from people with direct access to internal systems, research roadmaps, and deployment data.

Recursive self-improvement is moving from theory to practice

For decades, the idea of the singularity has been tied to “recursive self-improvement”: AI systems that help build better AI systems, creating a feedback loop of accelerating progress.

We’re not yet at the point where models are fully improving themselves without humans in the loop. But we are seeing what you could call “soft self-improvement”:

• Release cycles are compressing. Big labs used to ship major model upgrades every 6–12 months. Now, significant improvements are landing in weeks.
• AI is automating AI research. Frontier labs are starting to automate large chunks of their own workflows, using swarms of AI agents to run experiments, analyze results, and even write code for new model features.

Hassabis described this personally. With AI coding agents, he’s been “vibe coding” small game prototypes in one or two spare hours at night—projects that would have taken him six months before. That’s not a demo; that’s the head of DeepMind feeling his own capabilities multiply.

AI agents are becoming operational, not just conversational

In 2026, AI agents are shifting from chat interfaces to full-blown operational software. Instead of just answering questions, they’re starting to:

• Plan multi-step tasks
• Call tools and APIs
• Coordinate across systems inside real businesses
• Execute workflows with minimal human supervision

These agents are being embedded into production environments, not just showcased in demos. One big example: AWS has added payment capabilities for autonomous agents, allowing them to complete transactions and take direct action inside enterprise workflows.

We’re also seeing agents move into consumer-facing roles. For instance, Robinhood now allows AI agents to make trading and credit card decisions on behalf of users. This is exactly the kind of “real-world agency” many people once assumed would be tightly regulated or far off in the future.

AI is starting to do real science and math

Some of the most mind-bending changes are happening in science and mathematics, where AI isn’t just summarizing papers—it’s generating new knowledge.

AI proving theorems and solving century-old problems

A system called AxiomRover (also referred to as Axiom Improver) has quietly been publishing mathematical breakthroughs. Since February, eight of its papers have appeared on arXiv, with five already accepted in peer-reviewed journals.

Among other things, it has:

• Proved that 100% of primes are partially regular
• Shown that, under certain conditions, Ramanujan’s tau function misses 100% of primes

These are the kinds of deep, abstract results that used to be the exclusive domain of human mathematicians.

World models for biology

The Chan Zuckerberg Biohub released what they call a “world model” of protein biology. It’s built on:

• ESM-3, a language model trained on 2.88 billion protein sequences
• ESMFold 2 for atomic-level protein structure prediction
• An ESM atlas mapping 6.8 billion proteins

On top of that, multi-agent systems like Co-Scientist and Robin can autonomously:

• Generate hypotheses
• Design experiments
• Analyze data
• Refine research questions

These systems are already showing promise in identifying new drug candidates and targets in biomedical research. Scientists are beginning to use AI not just as a tool, but as a collaborator.

Productivity jumps that don’t look incremental

In business, AI agents are starting to deliver order-of-magnitude gains rather than small optimizations.

SAP’s sustainability-focused AI agents, currently in beta, have reported:

• >50% reduction in packaging
• Huge cuts in compliance review hours
• Scenario simulation time reduced from a full day to about 20 minutes
• Up to 80% reduction in manual GHS classification work
• Over 20% fewer packaging compliance errors

These aren’t tiny efficiency tweaks. They’re the kind of shifts that can reshape entire teams and processes.

There’s also a revealing metric from translation company Translated. Human editors typically need about 1 second per word to review another human’s translation. In 2014, it took ~3.5 seconds per word to fix machine translations. By 2022, that dropped to ~2 seconds. If that curve continues, machine translation could reach human-level editing effort by the end of this decade—or sooner. That’s a concrete, measurable path to “singularity” in at least one domain.

Benchmarks vs. reality: are we at AGI or nowhere close?

Not everyone agrees that we’re near AGI or the singularity, and a lot depends on what you measure.

Yann LeCun, Meta’s chief AI scientist, argues that current systems aren’t truly intelligent. For him, real intelligence means being able to solve new problems without prior training—generalizing from experience, not just pattern-matching from huge datasets.

Oriol Vinyals, co-lead of Google’s Gemini program, takes a middle view. He says today’s models are extremely strong at code and math, and their reasoning keeps getting more general. If someone had shown him these models seven years ago, he says he might have called them AGI. But he also notes that the ability to learn from ongoing experience and produce genuinely new breakthroughs is still missing.

To probe this gap, the ARC Prize Foundation introduced ARC-AGI 3 in March 2026. It’s designed to test interactive, experience-based reasoning: can an agent explore, infer goals, build a world model, and keep learning over time?

Results so far:

• Humans solve 100% of the ARC-AGI 3 environments
• Frontier AI systems, as of March 2026, score under 1%

By this metric, we’re nowhere close to human-level general intelligence. But by others—like coding, math, translation, or agentic workflows—we’re already matching or surpassing humans in narrow but economically critical areas.

The hardware and physics stack is racing ahead

Under the hood, the infrastructure that powers all this is evolving just as fast.

• NVIDIA’s Vera CPU, based on ARM64, has posted the best performance ever seen on ARM, beating top x86-64 chips from Intel and AMD. That’s a big deal for AI workloads that want high performance with better efficiency.
• Envision in Germany reported the first single-molecule spin–photon interface using a triplet ground-state carbene, opening the door to molecular qubits as a serious quantum computing platform.
• CBN Nanotechnologies in Ottawa achieved simultaneous spatial and chemical control over mechanochemical carbon fabrication via an inverted-mode STM—essentially placing atoms on demand.

These advances might sound abstract, but they’re the kind of deep tech that could enable the next generation of AI hardware, quantum systems, and ultra-dense computing—exactly what a self-accelerating AI ecosystem would feed on.

Safety, Mythos, and the policy whiplash

As capabilities accelerate, safety concerns are escalating just as fast.

One major warning sign came from Anthropic’s internal “Mythos” model. According to reports from April 2026, Mythos was deemed too dangerous for public release. Details are scarce, but it’s being treated as an alarming example of how quickly capabilities can outrun current safety and governance frameworks.

Hassabis has been pushing for more urgency on safety. He’s talked about the need for AI executive orders that require testing before new models are released and said safety work itself needs to be “accelerated.” His choice to use provocative language around the singularity was partly to wake up governments, economists, and the public.

Policy responses so far have been messy:

• On May 21, 2026, the U.S. president was expected to sign an executive order creating a voluntary federal review process for frontier AI systems. Hours before the ceremony, it was pulled—apparently over fears it might slow down the U.S. AI industry at a moment when the country believes it has a lead.
• Illinois passed SB 315, requiring frontier labs to publish catastrophic risk plans and undergo third-party AI safety audits—the first law of its kind in the U.S.

The tension is clear: governments want safety and oversight, but they also don’t want to lose the AI race.

Money, deployment, and the enterprise “gap”

One of the clearest signals that we’ve moved past the research-only phase is the money flowing into deployment.

In just a few days in early May 2026, more than $5.5 billion in capital was announced for projects specifically targeting the “deployment gap” in the enterprise sector. The focus is no longer just on building bigger models—it’s on putting them into production at massive scale.

At the same time, other parts of the tech ecosystem are rapidly reorganizing around AI:

• BusPatrol, which installed AI cameras on tens of thousands of U.S. school buses, plans to convert them into automatic license plate readers and share data with law enforcement.
• YouTube is automatically tagging videos with significant AI use.
• Waymo (Alphabet’s self-driving division) is testing models that give autonomous vehicles a kind of “imagination” to handle unpredictable or dangerous scenarios.

These moves show how quickly AI is being woven into physical infrastructure, media, and public safety systems—often faster than regulations can keep up.

AI building better AI tools for science

Google is leaning into the idea of AI-accelerated science with new tools aimed directly at the scientific method. Recent releases include:

• Co-Scientist – an AI collaborator for designing and reasoning about experiments
• AlphaEvolve – tools for exploring and optimizing complex systems
• Empirical research assistants – agents that help with data analysis, experiment planning, and literature review
• NotebookLM – an AI-native notebook environment for working with documents and data

These tools are explicitly designed to help scientists do better science faster—and, in turn, help build better AI. It’s a feedback loop: AI accelerates science, science improves hardware and algorithms, and the cycle repeats.

If you’re interested in how this fits into Google’s broader AI strategy, it’s worth checking out Google’s latest Gemini and AI announcements from I/O 2026.

So… is this actually the singularity?

The classic definition of the singularity is the moment AI surpasses human control and rapidly transforms society in ways we can’t predict—a kind of technological event horizon.

The tricky part is that there’s no clear “ding!” when you cross that line. There’s no single benchmark, no switch that flips from “pre-singularity” to “post-singularity.”

For Hassabis, the term captures something more immediate: the point beyond which meaningful prediction becomes impossible because the transformation is so fast and so deep. In his view, what makes 2026 special isn’t just model performance—it’s the lived reality of agentic AI systems that can plan, act, and deliver multi-step results with minimal human input.

Depending on who you listen to and which metrics you care about, you could argue:

• We’re already in the foothills of the singularity (Hassabis)
• We’ve crossed the threshold (Musk, Andreessen, and some others suggest this)
• We’re still years away, because current systems can’t yet learn from experience like humans (LeCun and other skeptics)

But a few things are hard to deny:

• Release cycles are compressing.
• Capabilities are exploding in key domains like code, math, translation, and scientific discovery.
• Infrastructure—from chips to quantum experiments to nanotech—is advancing at a breakneck pace.
• The people building these systems are using more dramatic language than ever before.

Whether you call this the singularity, the prelude, or just “the acceleration,” the conversation has clearly shifted. The technology is moving faster than almost anyone predicted a year ago—and the rest of society is scrambling to catch up.