Quantum Computing in the NISQ era and beyond

OverviewBalancedalloy voice

Picture the moment we're in as a kind of base camp. We've climbed high enough to see the terrain, but not high enough to plant a flag on the summit. John Preskill calls this the Noisy Intermediate-Scale Quantum era, or NISQ, and it's defined less by a magic number of qubits than by a tradeoff we can't yet beat. We have dozens of controllable quantum bits, but every operation introduces a little noise. That means you can run only so many steps before the signal fades into static. In rough terms, devices with around 50 to 100 qubits can already push past brute-force classical simulation, but when you try to run deep circuits, the computation decays. You get a practical ceiling around a thousand two-qubit gates and circuit depths under about a hundred time steps before errors overwhelm the result. That might sound discouraging after a decade of glossy demos and big corporate investments. Preskill's advice is to reframe the goal. He sees the near term as a period of discovery, not disruption. Fault-tolerant quantum computers—the kind that correct their own errors as they run—are coming on a timeline measured in decades. Not months. Not even a handful of years. So the point of NISQ hardware isn't to take Wall Street by storm. It's to light the path to machines that eventually could. Why hang in there at all? Because of entanglement. Preskill calls it the entanglement frontier, and it's the key resource that makes quantum information different from anything in a laptop. In a quantum device, information is stored in correlations across many parts at once. Look at the parts one by one and you see gibberish; measure them together and a global pattern snaps into focus. This is why physicists get excited: deeply entangled states let you do things that are hard to even write down classically, let alone compute. There are three solid reasons to expect quantum devices to outperform classical ones sometimes. First, we already know algorithms—Shor's factoring is the famous example—that outpace the best-known classical methods in principle. Second, complexity theory points to scenarios where sampling the output of a quantum circuit—essentially drawing from a weird, highly correlated probability distribution—should be hard for a classical computer to mimic efficiently. Third, despite heroic efforts, no general classical simulation method has emerged that scales to arbitrary quantum circuits. Put together, the case for quantum advantage is strong, but not universal. Preskill is clear: worst-case NP-hard problems, like an adversarial traveling salesman instance, aren't expected to vanish under quantum speedups. And for many real tasks, any advantage may be modest or depend on specific structure. So where does that leave us? In the sweet spot that Richard Feynman pointed to decades ago. If nature is quantum, then simulating quantum dynamics is the natural job for a quantum computer. Preskill leans into that logic, echoing a theme from Laughlin and Pines: strongly correlated matter—complex molecules, exotic phases, high-temperature superconductors—pushes classical methods to the edge. Entanglement isn't a bug to tame; it's the engine that lets a quantum simulator explore those systems faithfully. If you're going to bet on a NISQ payoff, this is the place to bet it: quantum chemistry, materials, and many-body physics where classical surrogates falter. The hitch, again, is the hardware. The best two-qubit gate errors today cluster around the tenth-of-a-percent range and often higher. Readout—the act of measuring a qubit—still makes mistakes, especially in superconducting platforms where error probabilities hover around one percent. And even when your gates are decent, speed and connectivity matter. Superconducting circuits can be about a thousand times faster than trapped ions, but speed alone doesn't erase cumulative error if each step still wobbles. Who you can talk to on the chip, and how reliably you can fabricate those connections, matters as much as raw qubit count. Here's a simple way Preskill frames the depth problem. Imagine a circuit with G gates. If the chance of an error per gate isn't below about one over G, then the odds of a catastrophic mistake somewhere during the run are uncomfortably high. That's the intuition behind the thousand-gate wall you hear cited. Current error rates and crosstalk mean you hit that wall quickly, and practical demonstrations tend to stay shallow—dozens of time steps rather than hundreds—just to keep the signal alive. You can feel the constraint: it's not that you lack ideas, it's that the medium blurs them as you try to draw finer detail. This is why error mitigation and error correction are not the same story. In the short run, you can mitigate: calibrate better, cancel some noise, design circuits that are more forgiving. Those tricks help, and they'll improve. But true scalability needs fault tolerance—encoding a logical qubit into many physical ones and repairing errors as they happen. The cost is steep. To run meaningful algorithms with thousands of protected qubits, you likely need millions of physical qubits. That's the "quantum chasm" Preskill talks about, the yawning gap between the few hundred qubits we can manage now and the army we'll need to compute reliably at scale. Meanwhile, there's more than one way to do quantum. Analog simulators—think cold atoms in optical lattices or Rydberg arrays—don't execute gate sequences. They evolve according to a Hamiltonian that resembles the model you care about. Over the past 15 years, this approach has been a vibrant playground, letting physicists watch quantum dynamics in regimes that confound classical numerics. The price is control. Your lab system only approximates your target model, and you mostly trust features that are universal—robust to small tweaks or imperfections. Quantum annealers sit in yet another corner. D-Wave's two thousand Q, with roughly two thousand qubits, is a flagship example. It tackles optimization by slowly transforming one energy landscape into another. That's a neat idea, and it's quantum, but Preskill is blunt: so far, there's no convincing theoretical argument or experimental evidence that annealers outperform the best classical tools on the same problems. Part of the nuance is something called stoquasticity—when the underlying math has only non-negative off-diagonal terms, classical simulations often cope well. Move to non-stoquastic designs and you might change the game, but that's still a research frontier. Gate-based, digital quantum computers are the universal ones—the machines that can, in principle, simulate any quantum system and run any quantum algorithm. They're also the ones inching toward fault tolerance. In the NISQ moment, the wins come from staying within the noise budget. That's why hybrid quantum-and-classical methods have become the emblematic play. With the Variational Quantum Eigensolver, or VQE, you use a quantum chip to prepare a flexible family of states, then a classical optimizer nudges the parameters to lower the energy and approximate a molecule's ground state. With the Quantum Approximate Optimization Algorithm, or QAOA, you alternate two kinds of simple operations to steer toward good solutions of a combinatorial problem, again tuning parameters in a classical loop. These methods are deliberately shallow. They're built to survive in a noisy world, and sometimes they borrow insights from tensor networks—mathematical tools for capturing limited entanglement—to stretch their reach. But even for hybrids, you have to ask the question that keeps Preskill cautious: compared to what? Any claim of quantum advantage needs to be benchmarked against the best classical algorithms, not last year's code. History has already served a few reminders. A much-hyped quantum recommendation algorithm promised exponential speedups under certain assumptions, only to be met by a "quantum-inspired" classical method that eroded the claim. The celebrated Harrow-Hassidim-Lloyd algorithm for solving linear systems can be exponentially faster in theory, but it leans on sparse, well-conditioned matrices and a quantum random access memory to serve inputs—pieces that look out of reach for NISQ devices. And there's a subtle trap here: when you modify a circuit to be more noise-resilient, you sometimes make it easier for a clever classical algorithm to mimic the output distribution. Good engineering can, ironically, undermine your claim of irreducible quantumness. So validation is not just a formality; it's the game. If you're sampling from a quantum circuit to demonstrate complexity-theoretic hardness, can you show that a classical simulator can't match your statistics within noise? If you're doing chemistry with VQE, can you establish that your energy estimates beat the best classical approximations at the same cost? This is where NISQ work earns its keep: by being clear-eyed about baselines and by publishing the full context—hardware noise, calibration drifts, classical post-processing—so the community can tell what really happened. Amid the caveats, the progress is real and measurable. Google and IBM have both run devices in the 50 to 70 qubit range—think of Google's 72-qubit chip and IBM's 50-qubit machine—as testbeds for exactly the kind of shallow, entanglement-rich circuits that define NISQ. The lesson from those campaigns is sobering and encouraging at once. With careful calibration, smart circuit design, and lots of statistics, you can explore regimes that would have been unthinkable a decade ago. But every extra qubit and every extra layer raises the stakes for gate fidelity, crosstalk, and readout accuracy. Connectivity and fabrication yield become the quiet bottlenecks that decide how many of your nominal qubits are actually useful. If you zoom out, a pattern comes into focus. Analog simulators will continue to illuminate universal features of many-body physics, especially where you can lean on robustness to small errors. Digital, gate-based devices will keep pressing against the noise ceiling, using variational and other hybrid methods to extract insight from shallow circuits. And over time, if gate fidelities improve and architectures mature, the brutal overhead of error correction drops to a manageable multiplier, and universal, fault-tolerant machines will outshine analog approaches with precision and scale. Preskill's throughline is simple and sturdy: the entanglement frontier justifies the work now and points to what comes next. Use NISQ devices as scientific instruments. Stress-test algorithms like QAOA and VQE where they make physical sense. Invest in better gates, smarter layouts, and sharper validation. And keep expectations realistic about near-term commercial impact. Where does that leave us? In a decade defined more by discovery than by disruption. We'll map the contours of what shallow, noisy circuits can do, and in doing so, we'll train the next generation of hardware and algorithms. Some milestones will be splashy, many will be incremental, and the most important ones might look, from the outside, like engineering footnotes. That's okay. The climb to fault tolerance is steep, but the route is visible: higher fidelity, better architectures, disciplined benchmarking, and a steady focus on the physics that truly needs quantum.

We have dozens of controllable quantum bits, but every operation introduces a little noise. That means you can run only so many steps before the signal fades into static. In rough terms, devices with around 50 to 100 qubits can already push past brute-force classical simulation, but when you try to run deep circuits, the computation decays.

You get a practical ceiling around a thousand two-qubit gates and circuit depths under about a hundred time steps before errors overwhelm the result.

That might sound discouraging after a decade of glossy demos and big corporate investments. Preskill's advice is to reframe the goal. He sees the near term as a period of discovery, not disruption.

Fault-tolerant quantum computers—the kind that correct their own errors as they run—are coming on a timeline measured in decades. Not months. Not even a handful of years.

So the point of NISQ hardware isn't to take Wall Street by storm. It's to light the path to machines that eventually could.

Why hang in there at all? Because of entanglement. Preskill calls it the entanglement frontier, and it's the key resource that makes quantum information different from anything in a laptop.

In a quantum device, information is stored in correlations across many parts at once. Look at the parts one by one and you see gibberish; measure them together and a global pattern snaps into focus. This is why physicists get excited: deeply entangled states let you do things that are hard to even write down classically, let alone compute.

There are three solid reasons to expect quantum devices to outperform classical ones sometimes. First, we already know algorithms—Shor's factoring is the famous example—that outpace the best-known classical methods in principle. Second, complexity theory points to scenarios where sampling the output of a quantum circuit—essentially drawing from a weird, highly correlated probability distribution—should be hard for a classical computer to mimic efficiently.

Third, despite heroic efforts, no general classical simulation method has emerged that scales to arbitrary quantum circuits. Put together, the case for quantum advantage is strong, but not universal. Preskill is clear: worst-case NP-hard problems, like an adversarial traveling salesman instance, aren't expected to vanish under quantum speedups.

And for many real tasks, any advantage may be modest or depend on specific structure.

So where does that leave us? In the sweet spot that Richard Feynman pointed to decades ago. If nature is quantum, then simulating quantum dynamics is the natural job for a quantum computer.

Preskill leans into that logic, echoing a theme from Laughlin and Pines: strongly correlated matter—complex molecules, exotic phases, high-temperature superconductors—pushes classical methods to the edge. Entanglement isn't a bug to tame; it's the engine that lets a quantum simulator explore those systems faithfully. If you're going to bet on a NISQ payoff, this is the place to bet it: quantum chemistry, materials, and many-body physics where classical surrogates falter.

The hitch, again, is the hardware. The best two-qubit gate errors today cluster around the tenth-of-a-percent range and often higher. Readout—the act of measuring a qubit—still makes mistakes, especially in superconducting platforms where error probabilities hover around one percent.

And even when your gates are decent, speed and connectivity matter. Superconducting circuits can be about a thousand times faster than trapped ions, but speed alone doesn't erase cumulative error if each step still wobbles. Who you can talk to on the chip, and how reliably you can fabricate those connections, matters as much as raw qubit count.

Here's a simple way Preskill frames the depth problem. Imagine a circuit with G gates. If the chance of an error per gate isn't below about one over G, then the odds of a catastrophic mistake somewhere during the run are uncomfortably high.

That's the intuition behind the thousand-gate wall you hear cited. Current error rates and crosstalk mean you hit that wall quickly, and practical demonstrations tend to stay shallow—dozens of time steps rather than hundreds—just to keep the signal alive. You can feel the constraint: it's not that you lack ideas, it's that the medium blurs them as you try to draw finer detail.

This is why error mitigation and error correction are not the same story. In the short run, you can mitigate: calibrate better, cancel some noise, design circuits that are more forgiving. Those tricks help, and they'll improve.

But true scalability needs fault tolerance—encoding a logical qubit into many physical ones and repairing errors as they happen. The cost is steep. To run meaningful algorithms with thousands of protected qubits, you likely need millions of physical qubits.

That's the "quantum chasm" Preskill talks about, the yawning gap between the few hundred qubits we can manage now and the army we'll need to compute reliably at scale.

Meanwhile, there's more than one way to do quantum. Analog simulators—think cold atoms in optical lattices or Rydberg arrays—don't execute gate sequences. They evolve according to a Hamiltonian that resembles the model you care about.

Over the past 15 years, this approach has been a vibrant playground, letting physicists watch quantum dynamics in regimes that confound classical numerics. The price is control. Your lab system only approximates your target model, and you mostly trust features that are universal—robust to small tweaks or imperfections.

Quantum annealers sit in yet another corner. D-Wave's two thousand Q, with roughly two thousand qubits, is a flagship example. It tackles optimization by slowly transforming one energy landscape into another.

That's a neat idea, and it's quantum, but Preskill is blunt: so far, there's no convincing theoretical argument or experimental evidence that annealers outperform the best classical tools on the same problems. Part of the nuance is something called stoquasticity—when the underlying math has only non-negative off-diagonal terms, classical simulations often cope well. Move to non-stoquastic designs and you might change the game, but that's still a research frontier.

Gate-based, digital quantum computers are the universal ones—the machines that can, in principle, simulate any quantum system and run any quantum algorithm. They're also the ones inching toward fault tolerance. In the NISQ moment, the wins come from staying within the noise budget.

That's why hybrid quantum-and-classical methods have become the emblematic play. With the Variational Quantum Eigensolver, or VQE, you use a quantum chip to prepare a flexible family of states, then a classical optimizer nudges the parameters to lower the energy and approximate a molecule's ground state. With the Quantum Approximate Optimization Algorithm, or QAOA, you alternate two kinds of simple operations to steer toward good solutions of a combinatorial problem, again tuning parameters in a classical loop.

These methods are deliberately shallow. They're built to survive in a noisy world, and sometimes they borrow insights from tensor networks—mathematical tools for capturing limited entanglement—to stretch their reach.

But even for hybrids, you have to ask the question that keeps Preskill cautious: compared to what? Any claim of quantum advantage needs to be benchmarked against the best classical algorithms, not last year's code. History has already served a few reminders.

A much-hyped quantum recommendation algorithm promised exponential speedups under certain assumptions, only to be met by a "quantum-inspired" classical method that eroded the claim. The celebrated Harrow-Hassidim-Lloyd algorithm for solving linear systems can be exponentially faster in theory, but it leans on sparse, well-conditioned matrices and a quantum random access memory to serve inputs—pieces that look out of reach for NISQ devices. And there's a subtle trap here: when you modify a circuit to be more noise-resilient, you sometimes make it easier for a clever classical algorithm to mimic the output distribution.

Good engineering can, ironically, undermine your claim of irreducible quantumness.

So validation is not just a formality; it's the game. If you're sampling from a quantum circuit to demonstrate complexity-theoretic hardness, can you show that a classical simulator can't match your statistics within noise? If you're doing chemistry with VQE, can you establish that your energy estimates beat the best classical approximations at the same cost?

This is where NISQ work earns its keep: by being clear-eyed about baselines and by publishing the full context—hardware noise, calibration drifts, classical post-processing—so the community can tell what really happened.

Amid the caveats, the progress is real and measurable. Google and IBM have both run devices in the 50 to 70 qubit range—think of Google's 72-qubit chip and IBM's 50-qubit machine—as testbeds for exactly the kind of shallow, entanglement-rich circuits that define NISQ. The lesson from those campaigns is sobering and encouraging at once.

With careful calibration, smart circuit design, and lots of statistics, you can explore regimes that would have been unthinkable a decade ago. But every extra qubit and every extra layer raises the stakes for gate fidelity, crosstalk, and readout accuracy. Connectivity and fabrication yield become the quiet bottlenecks that decide how many of your nominal qubits are actually useful.

If you zoom out, a pattern comes into focus. Analog simulators will continue to illuminate universal features of many-body physics, especially where you can lean on robustness to small errors. Digital, gate-based devices will keep pressing against the noise ceiling, using variational and other hybrid methods to extract insight from shallow circuits.

And over time, if gate fidelities improve and architectures mature, the brutal overhead of error correction drops to a manageable multiplier, and universal, fault-tolerant machines will outshine analog approaches with precision and scale.

Preskill's throughline is simple and sturdy: the entanglement frontier justifies the work now and points to what comes next. Use NISQ devices as scientific instruments. Stress-test algorithms like QAOA and VQE where they make physical sense.

Invest in better gates, smarter layouts, and sharper validation. And keep expectations realistic about near-term commercial impact.

Where does that leave us? In a decade defined more by discovery than by disruption. We'll map the contours of what shallow, noisy circuits can do, and in doing so, we'll train the next generation of hardware and algorithms.

Some milestones will be splashy, many will be incremental, and the most important ones might look, from the outside, like engineering footnotes. That's okay. The climb to fault tolerance is steep, but the route is visible: higher fidelity, better architectures, disciplined benchmarking, and a steady focus on the physics that truly needs quantum.

Related lectures