<?xml version="1.0" encoding="UTF-8"?><rss xmlns:dc="http://purl.org/dc/elements/1.1/" xmlns:content="http://purl.org/rss/1.0/modules/content/" xmlns:atom="http://www.w3.org/2005/Atom" version="2.0" xmlns:cc="http://cyber.law.harvard.edu/rss/creativeCommonsRssModule.html">
    <channel>
        <title><![CDATA[Stories by Nishkal Pandya on Medium]]></title>
        <description><![CDATA[Stories by Nishkal Pandya on Medium]]></description>
        <link>https://medium.com/@nishkalspandya?source=rss-3a0d887ba305------2</link>
        <image>
            <url>https://cdn-images-1.medium.com/fit/c/150/150/0*qCSnArddZFGYL3W4</url>
            <title>Stories by Nishkal Pandya on Medium</title>
            <link>https://medium.com/@nishkalspandya?source=rss-3a0d887ba305------2</link>
        </image>
        <generator>Medium</generator>
        <lastBuildDate>Thu, 19 Nov 2026 23:36:07 GMT</lastBuildDate>
        <atom:link href="https://medium.com/@nishkalspandya/feed" rel="self" type="application/rss+xml"/>
        <webMaster><![CDATA[yourfriends@medium.com]]></webMaster>
        <atom:link href="http://medium.superfeedr.com/" rel="hub"/>
        <item>
            <title><![CDATA[Is Quantum Machine Learning Actually Better? I Built Both Models to Find Out.]]></title>
            <link>https://medium.com/@nishkalspandya/is-quantum-machine-learning-actually-better-i-built-both-models-to-find-out-e1af5ae4bb9f?source=rss-3a0d887ba305------2</link>
            <guid isPermaLink="false">https://medium.com/p/e1af5ae4bb9f</guid>
            <category><![CDATA[quantum-ai]]></category>
            <category><![CDATA[ai]]></category>
            <category><![CDATA[quantum-computing]]></category>
            <category><![CDATA[python]]></category>
            <dc:creator><![CDATA[Nishkal Pandya]]></dc:creator>
            <pubDate>Sun, 31 May 2026 02:56:36 GMT</pubDate>
            <atom:updated>2026-05-31T02:56:36.241Z</atom:updated>
            <content:encoded><![CDATA[<p>Replicate 2</p><p>Right now, quantum machine learning is described as the next innovation from classical AI. People say quantum AI has faster computing power, more optimized training, and solves problems that normal, classical systems struggle to achieve.</p><p>However, I decided to test this. I built a classical machine learning model and a quantum machine learning simulation using Qiskit. I then evaluated the models on 2 datasets: the Iris dataset and a randomly generated NumPy dataset to compare time to train and accuracy.</p><p>To test this, I built a classical model and a simulated quantum model.</p><p>For the quantum model, I used the Qiskit library to make a variational quantum classifier. This encodes classical data into quantum states and uses a quantum circuit.</p><p>For the classical model, I used a standard model using logistic regression from Scikit-Learn, which is a machine learning algorithm that classifies data by finding the optimal boundaries between different classes.</p><p>These models are evaluated on the 2 datasets. The first one is the Iris dataset, which is a dataset with 3 species of Iris flowers. This is easy for classical datasets, but difficult to optimize more than the standard it already is. The 2nd model is the randomly generated NumPy dataset, which is not structured and has difficult patterns to recognize.</p><p>Before I talk about the code or the specific results, let me tell you what I noticed. The classical model trained almost instantly in both datasets and was mostly accurate. However, the quantum model needed a lot more training time, almost seconds more than the classical system. For a person who knows almost nothing about the specifics of quantum AI, they would find this surprising due to the models being tested on the same dataset and the promises of quantum AI.</p><p>Once you understand how quantum AI actually works, then these issues make sense.</p><p>Unlike the current classical SVM, the quantum model encodes classical data into quantum states using a ZZFeatureMap. The feature map can encode the data into rotations and entanglements, which are then further sorted using a fidelity quantum kernel.</p><p>This is entirely different than normal machine learning. A traditional system operates in the standard math space, while the quantum kernel is on a multi-dimensional space. This allows the quantum model to represent more complex relationships and patterns than classical models.</p><p>This is shown in the graphs. The classical graph is smooth, and there is a clear separation between 2 classes. In contrast, the quantum ZZ kernel has multiple detailed and fragmented areas. This shows the quantum kernel has more discernible patterns, increasing accuracy.</p><p>Even though this can create high accuracy, it also creates high instability. This has led to a debate between accuracy and stability. Even though the accuracy increased, the decision boundary is much more chaotic and sensitive compared to the classical version. This can also be that the model is looking at local patterns instead of general trends.</p><p>Even though the quantum kernel has strong potential, it incurs significant overhead. The biggest thing is that this code is not executed on a real quantum computer but a simulation using Qiskit. The simulation is very expensive in terms of computing power. Classical computers struggle to simulate quantum systems since the space grows exponentially as qubits and operations are added.</p><p>As a result, every prediction needs more quantum computations. In reality, the quantum model will take less time to train compared to the 10 seconds shown, but it will still take more than a few milliseconds for the Iris dataset.</p><p>This does not prove that the quantum machine is slow. Instead, it shows one of the main bottlenecks of quantum computing research: Most of the experiments are still dependent on classical simulations due to limited access to usable quantum hardware.</p><p>Another point is the dataset size. These experiments are based on a few data points. This should not be universal proof that quantum is worse than classical. Instead, they show how quantum models can behave under small-scale conditions.</p><p>Currently, quantum AI is more valuable as a research field than as a practical problem-solving tool, for now. While the quantum kernel has shown the ability to represent complex relationships, the computational costs and the instability show that major scalability challenges exist. However, this shows why researchers are interested in quantum AI. It has the potential to outperform classical AI models and can work with high-dimensional datasets that classical kernels cannot capture or work with. For small datasets, classical systems are already so optimized that quantum AI cannot keep up. However, when there are larger datasets, researchers believe that quantum AI can easily take the lead.</p><p>I will start by showing code segments from the Iris model, then explaining the code and what it does.</p><pre>start_c = time.time()<br>c_svc = SVC(kernel=&#39;rbf&#39;)<br>c_svc.fit(X_train, y_train)<br>c_time = time.time() - start_c<br>c_acc = accuracy_score(y_test, c_svc.predict(X_test))</pre><p>This code trains the classical model and measures training time and accuracy. The RBF kernel is used because it can identify nonlinear relationships efficiently.</p><p>Unlike the classical model, the quantum model cannot directly process normal data. The data has to be transformed into quantum states using a feature map.</p><pre>feature_dim = 2<br>f_map = ZZFeatureMap(<br>    feature_dimension=feature_dim,<br>    reps=2,<br>    entanglement=&#39;linear&#39;<br>)</pre><p>The ZZFeatureMap converts the classical values into quantum values while factoring entanglement. This creates a quantum version fo the data, and increasing reps will increase the complexity of the quantum state.</p><pre>q_kernel = FidelityQuantumKernel(feature_map=f_map)<br><br>q_svc = SVC(kernel=q_kernel.evaluate)<br>q_svc.fit(X_train, y_train)</pre><p>After making the classical model and encoding the data, I used a Fidelity Quantum Kernel to find similarities between quantum states generated by the feature map. These similarities are used by the model to determine decision boundaries.</p><pre>X = iris.data[:100, :2]<br>y = iris.target[:100]<br><br>X = MinMaxScaler(<br>    feature_range=(0, 2 * np.pi)<br>).fit_transform(X)</pre><p>Furthermore, the data for the Iris set needed to be scaled to meet the quantum circuit requirements. I only chose 2 features and 2 classes to make it easier to understand. The data was scaled to between 0 and 2π. This is the standard for the ZZFeatureMap and helps ensure the quantum circuit can represent individual input data properly.</p><pre>print(f&quot;{&#39;Classical&#39;:&lt;15} | {c_acc*100:&gt;8.1f}% | {c_time:&gt;9.4f}&quot;)<br>print(f&quot;{&#39;Quantum&#39;:&lt;15} | {q_acc*100:&gt;8.1f}% | {q_time:&gt;9.4f}&quot;)</pre><p>This code outputs the final results with accuracy and runtime.</p><p>Here is the output and the graph:</p><pre>Computing Classical SVM...<br>Computing Quantum Kernel (Simulated)...<br>C:\Users\nishk\AppData\Local\Temp\ipykernel_22612\3922429699.py:44: DeprecationWarning: The class ``qiskit.circuit.library.data_preparation._zz_feature_map.ZZFeatureMap`` is deprecated as of Qiskit 2.1. It will be removed in Qiskit 3.0. Use the zz_feature_map function as a replacement. Note that this will no longer return a BlueprintCircuit, but just a plain QuantumCircuit.<br>  f_map = ZZFeatureMap(feature_dimension=feature_dim, reps=2, entanglement=&#39;linear&#39;)<br><br>========================================<br>MODEL           | ACCURACY   | TIME (s)  <br>----------------------------------------<br>Classical       |    100.0% |    0.0285<br>Quantum         |     70.0% |   10.3575<br>========================================</pre><figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/1*aufr8DdZ3p8ngGbrDQWQZw.png" /></figure><p>On the Iris dataset, the classical model acheived 100% accuracy in 0.0285 seconds, while the quantum kernel achieved 70% accuracy and needed 10.36 seconds to train. Even though the dataset was easy, the quantum model classifies based on fragmented boundaries, which overcomplicates a simple dataset like the Iris dataset. Furthermore, the additional cost comes from repeatedly simulating quantum circuits when computing kernal values. This dataset demonstrates an important lesson about quantum machine learning: more expensive does not mean better.</p><p>Now I will use screenshots and the code for the synthetic quantum machine learning dataset. The NumPy set is a less predictable and synthetic dataset.</p><pre>from qiskit_machine_learning.datasets import ad_hoc_data<br>train_data, train_labels, test_data, test_labels = ad_hoc_data(<br>    training_size=20,<br>    test_size=10,<br>    n=feature_dim,<br>    gap=0.3,<br>    plot_data=False<br>)</pre><p>This makes a complex classification dataset with patterns that are difficult to separate. The purpose is to observe how classical and quantum models behave when the decision boundary becomes more complicated.</p><pre>train_labels_1d = np.argmax(train_labels, axis=1)<br>test_labels_1d = np.argmax(test_labels, axis=1)</pre><p>These dataset labels are initially stored in a 1D format. Since the model expects the labels as a single value, they needed to be converted into a 1D presentation before training the models.</p><p>Here are the results:</p><pre>Running Classical SVM... Done.<br>Running Quantum Kernel (this will be slower)... Done.<br><br>========================================<br>MODEL           | ACCURACY   | TIME (s)  <br>----------------------------------------<br>Classical       |     70.0% |    0.0060<br>Quantum         |    100.0% |    1.5493<br>========================================</pre><figure><img alt="" src="https://cdn-images-1.medium.com/max/968/1*rxd3x2dWjQDnaef6NOuJ4w.png" /></figure><p>In this dataset, the quantum model outperformed the classical model in accuracy. While the classical model achieved roughly 70% accuracy, the quantum kernel reached 100%. However, this came at the cost of longer training times, showing that increased complexity is not efficient.</p><p>First, I compared accuracy across both models. The classical model was consistently 100% accurate on the iris dataset in about 0.03 seconds, but struggled on the NumPy set, with around 70% in about 0.006 seconds.</p><p>The accuracy is inverted on the quantum model, with about 70% on the Iris set and 100% on the NumPy set. However, the time difference is shocking. For the Iris set, it took 10.35 seconds to train the Iris set and 1.54 seconds for the NumPy set.</p><p>Both show that the quantum model has longer training times and greater difficulty achieving stability.</p><p>Right now, Quantum ML is in experiment mode rather than practical mode. It is nowhere near as efficient as classical models. However, building these models taught me a lot and showed where the issues are: data encoding and optimization. The most important takeaway is that quantum AI is not a failure but is in the growing and experimental stage. If future quantum hardware becomes more scalable and more fault-tolerant, these issues could be less severe or disappear. Classical machine learning is currently more practical today, but who knows what tomorrow will bring?</p><img src="https://medium.com/_/stat?event=post.clientViewed&referrerSource=full_rss&postId=e1af5ae4bb9f" width="1" height="1" alt="">]]></content:encoded>
        </item>
        <item>
            <title><![CDATA[Quantum AI Review]]></title>
            <link>https://medium.com/@nishkalspandya/quantum-ai-review-7f0ba53f2171?source=rss-3a0d887ba305------2</link>
            <guid isPermaLink="false">https://medium.com/p/7f0ba53f2171</guid>
            <category><![CDATA[quantum-computing]]></category>
            <category><![CDATA[ai]]></category>
            <dc:creator><![CDATA[Nishkal Pandya]]></dc:creator>
            <pubDate>Mon, 30 Mar 2026 22:57:15 GMT</pubDate>
            <atom:updated>2026-03-30T22:57:15.686Z</atom:updated>
            <content:encoded><![CDATA[<h3><strong>A Comparative Review of Quantum and Classical Artificial Intelligence: Evaluating Computational Accuracy, Predictive Precision, and Runtime Dynamics</strong></h3><h3><strong>Abstract</strong></h3><p>The number of transistors on computer chips <strong>doubles</strong> every year. This is known as Moore’s Law and has fueled decades of innovation. However, transistors are currently close to the size of a singular atom (Havlíček et al., 2019). This means we have a limit that needs a change in the chips. This paper explores quantum advantage and how the exponential increase in qubits can redefine how models learn and form connections. A comparison was made between Variational Quantum Circuits (VQC) and Quantum Long Short-Term Memory (QLSTM) networks against deep learning models. The results showed that a quantum-trained AI achieved approximately 97% accuracy in a financial context, whereas a classically trained AI led in runtime efficiency and stability (Su, Li, &amp; Qiu, 2025). These results suggest that while Quantum AI currently is task-specific, it points to a future of hybrid quantum-classical integration.</p><h3><strong>Introduction</strong></h3><p>Modern AI relies on a singular bit, which is limited to the values of a 0 and a 1. This was the foundation that came from the first computer, but the only way we can increase storage is by including more bits, taking up more energy.</p><p>However, Quantum AI uses the qubit, leveraging superposition and entanglement to hold more states: a 0, a 1, or both (IBM Technology, 2022; AI Uncovered, 2023). A qubit is based on probability, which allows different data to be held at the same time without adding more qubits (Marr, 2024; Bosch, 2022).</p><p>Implementation of quantum-enhanced architectures will represent a significant shift in these four fields:</p><ol><li>Environmental Sustainability: Training classical AI models uses large data centers that consume more electricity and water than small cities (International Energy Agency [IEA], 2024; MIT News, 2025). Quantum AI is more efficient and can handle more complex data for a fraction of the cost.</li><li>Market Stability: Financial markets are more connected than ever. Quantum LSTMs (Long Short-Term Memory) can reduce error by about 13.1%, leading to accurate risk assessments and increased global security.</li><li>Drug Discovery and Molecular Models: Simulating molecular interactions in great detail is something classically trained AIs sometimes struggle to do. Quantum AI can simulate it and figure out the tiny placements that are based on probability quickly, reducing the development time of new medicines from years to months (Chen et al., 2025; Emmanoulidis et al., 2024).</li><li>Technological Advancement: Currently, transistors are approaching the size of an atom, and it is not possible to make the hardware smaller than that. We need to use a different approach, which is what Quantum AI can achieve (Pandya, 2024; Taghandiki, 2024).</li></ol><p>Even though classical AIs are faster and more stable, quantum-trained AIs are more accurate and efficient at solving problems.</p><h3><strong>Literature Review</strong></h3><p>Early research into Quantum Machine Learning (QML) was mostly theoretical. In 2018/2019, papers like Nature’s <em>Supervised Learning with Quantum-Enhanced Feature Spaces </em>and Physical Review Letters’ <em>Quantum Machine Learning in Feature Hilbert Spaces</em> showed that quantum computers could map data into something called a Hilbert Space (Biamonte et al., 2017; Schuld &amp; Killoran, 2019). Quantum feature mapping embeds classical data into high-dimensional Hilbert spaces, allowing complex correlations to become linearly separable (Schuld &amp; Killoran, 2019). However, the theory only worked to fake data that was too simple to have any meaningful applications.</p><p>However, as time went on to 2024/2025, the field became more tangible, with applications such as the stock market. The focus shifted from physics to AI (Su, Li, &amp; Qiu, 2025; Bosch, 2022).</p><p>The reason Quantum AI works is because of Effective Dimension. The term is coined by IBM’s paper <em>The Power of Quantum Neural Networks</em> in 2021. Effective Dimension is the minimum number of independent variables needed to describe a complex system/model (Abbas et al., 2021). Instead of counting the total number of variables, having fewer independent variables that can find correlations and can consider noise allows for more efficient learning. Classical AI has variables that don’t contribute, while a high effective dimension makes sure all variables are used to identify complex patterns. IBM discovered that quantum models have a higher effective dimension, which is why the model is so accurate (Abbas et al., 2021; Zhang, 2025)..</p><p>Everyone is excited about the accuracy increase, but there is a gap in the studies. Studies focus on the accuracy percentage, but they don’t take the data encoding phase into account. Classical data cannot just simply be plugged into a quantum computer (Caleb Writes Code, 2025; Ranga et al., 2024). The bits have to turn into qubits, but the process is slow and takes up a lot of energy. This is a major limitation currently for real-world use.</p><p>Furthermore, there are debates about where quantum models have the advantage. Classical AI wins on small datasets since quantum hardware is too “noisy”, which means it has too many errors (Bowles et al., 2024; Marr, 2024). The paper <em>Better than Classical? The Subtle Art of Benchmarking Quantum Machine Learning Models </em>found that classical models beat quantum models in 60% of small-scale tests. However, for massive, multi-dimensional problems like predicting protein folds or markets, quantum models are years faster than classical ones.</p><h3><strong>Methodology</strong></h3><p>This paper is a comparative meta-analysis to see if Quantum Advantage has potential across different tests. This shows how different researchers use quantum processors to see if they can solve the same problem. For the meta-analysis, studies were selected based on 3 criteria:</p><ol><li>Use of quantum hardware (IBM processors) over pure simulators,</li><li>Use of direct Classical benchmark (High-end GPUs)</li><li>Focus on high-dimensional datasets.</li></ol><p>Studies based on synthetic, low-dimensional datasets were excluded to ensure the analysis reflects real-world applications.</p><p>The 2 tests include stock price prediction and a general test from over 160 datasets (Bowles et al., 2024). 4 primary variables are common throughout all studies. The independent variable is which type of learning model is used: Classical Deep Learning (LSTM) or Quantum Deep Learning (QLSTM/VQC), while the dependent variables are the accuracy percentage, the RMSE (root mean square error), and the runtime (Su, Li, &amp; Qiu, 2025; Zhang, 2025).</p><p>Stock market prices are not always linear and are “noisy,” so random fluctuations in the data can confuse classical AI. Past stock price data (Open, High, Low, Close, Volume) needs the models to remember long-term trends while internalizing short-term changes. This study used IBM’s Eagle processor, which has 127 qubits (IBM Technology, 2022; Wang et al., 2025). This was chosen because the Eagle has advanced error reduction abilities. The classical AI was run on NVIDIA H100 GPUs, which were the most powerful classical AI hardware available at that time.</p><p>The 2nd test was to make sure the findings have more applications than finance. This test generalized over 160 different datasets, from medical images to logic puzzles. Here, the dependent variables are slightly different. The variables measured are effective dimension, runtime stability, and the predictive gap. The processors used were IBM’s Osprey processor, which has 433 qubits, which allows researchers to test larger datasets than in previous years. PennyLane and Qiskit software are the connection between quantum and classical (ResearchGate, 2019; Li &amp; Zhang, 2025). Testing all 160 datasets can help identify which types of data become too complex for a GPU and are perfect for a quantum system.</p><figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/1*sFPwn7sCaE2wsFmhgXP5Yg.png" /></figure><p>The 97.36% accuracy from the quantum model can be attributed to entanglement. (Su, Li, &amp; Qiu, 2025; Chen et al., 2025). In classical models, variables are independent and have no connection to one another. However, qubits can be correlated and find hidden relationships between data points (AI Uncovered, 2023). The NVIDIA H100 provides superior runtime stability, yet it encounters a representation ceiling when processing non-linear, multi-dimensional financial time series. Conversely, the quantum models excelled because of their higher effective dimension, which allows them to map the “noise” of the stock market into a Hilbert space where correlations become visible. An example of this is how the price of a commodity in a market can be related to a stock price. Using this ability, the model can predict with greater accuracy, resulting in a 13.1% reduction in RMSE (Chen et al., 2025; Wang et al., 2025).</p><figure><img alt="" src="https://cdn-images-1.medium.com/max/899/1*hHddneXi01AdJdRK9lsApA.png" /></figure><p>Even with these great numbers, Quantum AI suffers from decoherence and noise. The paper, <em>Better than classical? The subtle art of benchmarking quantum machine learning models </em>found a 60% fail rate when quantum models were tested with small and simple datasets (Bowles et al., 2024). This 60% fail rate is directly linked to the limits of current quantum processors. While the IBM Eagle (127 qubits) and Osprey (433 qubits) have a high number of qubits, they are susceptible to thermal decoherence. For simple datasets, the signal is lost to hardware-based noise. In these small-scale tests, the classical error correction of the NVIDIA H100 ensures a level of reliability that current quantum circuits cannot match.</p><p>This happens because today’s quantum hardware is very sensitive. Heat and vibration can cause the qubits to “flip,” changing their state and leading to errors that classical models don’t have. Even though Quantum AI is powerful, it is too unstable for general data and needs specific, complex problems to be worth the noise risk.</p><figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/1*NtIyw2CYUffhm0P9loAbfg.png" /></figure><p>Looking at the runtime data, there seems to be an issue: Fewer steps are required, but hardware latency is higher. They require fewer steps because they can explore multiple solutions at once. However, classical data needs to be converted into quantum states through data encoding (IBM Technology, 2022; Ranga et al., 2024). This takes up a massive amount of time. Furthermore, resetting the IBM chips in between runs (qubit initialization) creates a delay that doesn’t exist on an NVIDIA chip. This means the speed of a quantum computer is trapped between hardware being slow.</p><figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/1*rbJads2aTfna5ZMRjPVE2Q.png" /></figure><h3><strong>Concluding Remarks</strong></h3><p>This data shows that Quantum AI is not an instant boost for speed (Pandya, 2024; AI Uncovered, 2023). This is because of hardware issues and qubit initialization taking more time than a regular GPU. However, the 97.36% accuracy and 13.1% reduction in RMSE show that Quantum AI has superior precision for high-dimensional datasets. Using Entanglement and the Effective Dimension of Quantum AI, quantum models can find correlations in complex data like stock markets that classical AI cannot process with the same level of speed and accuracy.</p><p>In 2017, Quantum ML was purely theoretical in the form of math proofs in the paper “Quantum Machine Learning” (Biamonte et al., 2017; Bosch, 2022). In 2025, we have companies like IBM making processors and testing them.</p><p>In the next 5 years, the best AI system will not be purely classical or purely quantum. Instead, it will be a combination of classical and quantum (ResearchGate, 2019; Li &amp; Zhang, 2025). In this architecture, classical GPUs serve as the processors for data encoding and validation, while the QPUs (quantum processing units) are reserved for complex calculations. This integration can remove the 60% error rates that the noisy quantum hardware has while using the effective dimension of quantum circuits for specialized tasks (Pandya, 2024; Su, Li, &amp; Qiu, 2025).</p><h3>References</h3><p>Abbas, A., Sutter, D., Zoufal, C., Lucchi, A., Figalli, A., &amp; Woerner, S. (2021). The power of quantum neural networks. <em>Nature Computational Science</em>, 1(6), 403–409.<a href="https://doi.org/10.1038/s43588-021-00084-1"> https://doi.org/10.1038/s43588-021-00084-1</a></p><p>AI Uncovered. (2023, December 8). <em>Here’s what will happen when we combine Quantum Computing with AI!</em> [Video]. YouTube.<a href="https://www.youtube.com/watch?v=P7_SfxRrXTE"> https://www.youtube.com/watch?v=P7_SfxRrXTE</a></p><p>Biamonte, J., Wittek, P., Pancotti, N., Rebentrost, P., Wiebe, N., &amp; Lloyd, S. (2017). Quantum machine learning. <em>Nature</em>, 549(7671), 195–202.<a href="https://doi.org/10.1038/nature23474"> https://doi.org/10.1038/nature23474</a></p><p>Bosch, S. (2022, February 26). <em>Quantum Artificial Intelligence | My PhD at MIT</em> [Video]. YouTube.<a href="https://www.youtube.com/watch?v=aFDbO0CZFto"> https://www.youtube.com/watch?v=aFDbO0CZFto</a></p><p>Bowles, J., Ahmed, S., &amp; Schuld, M. (2024). Better than classical? The subtle art of benchmarking quantum machine learning models. <em>arXiv:2403.07059</em>.<a href="https://arxiv.org/abs/2403.07059"> https://arxiv.org/abs/2403.07059</a></p><p>Caleb Writes Code. (2025, October 19). <em>Quantum Computing and AI</em> [Video]. YouTube.<a href="https://www.youtube.com/watch?v=sQSQBYHR0ms"> https://www.youtube.com/watch?v=sQSQBYHR0ms</a></p><p>Chen, J., et al. (2025). Quantum-enhanced temporal convolutional networks for volatility forecasting. <em>arXiv:2505.23860</em>.<a href="https://arxiv.org/abs/2505.23860"> https://arxiv.org/abs/2505.23860</a></p><p>El-Refaey, M. (2024). Magic matrices: Machine learning &amp; recommender systems. <em>LinkedIn Pulse</em>.<a href="https://www.linkedin.com/pulse/magic-matrices-machine-learning-recommender-systems-mohamed-el-refaey/"> https://www.linkedin.com/pulse/magic-matrices-machine-learning-recommender-systems-mohamed-el-refaey/</a></p><p>IBM Technology. (2022, October 19). <em>Quantum Machine Learning Explained</em> [Video]. YouTube. <a href="https://www.youtube.com/watch?v=NqHKr9CGWJ0">https://www.youtube.com/watch?v=NqHKr9CGWJ0</a></p><p>International Energy Agency. (2024). <em>Energy demand from AI: Analysis and forecast</em>. IEA Reports. <a href="https://www.iea.org/reports/energy-and-ai/energy-demand-from-ai">https://www.iea.org/reports/energy-and-ai/energy-demand-from-ai</a></p><p>Li, X., &amp; Zhang, Q. (2025). Hybrid quantum-classical optimization for large-scale logistics. <a href="https://arxiv.org/abs/2505.14295"><em>arXiv:2505.14295</em>. https://arxiv.org/abs/2505.14295</a></p><p>Marr, B. (2024, October 8). <em>What is Quantum AI? An Easy Explanation For Anyone</em> [Video]. YouTube. <a href="https://www.youtube.com/watch?v=8I7IKrLP6Hk">https://www.youtube.com/watch?v=8I7IKrLP6Hk</a></p><p>MIT News. (2025, January 17). <em>Explained: Generative AI’s environmental impact</em>. <a href="https://news.mit.edu/2025/explained-generative-ai-environmental-impact-0117">https://news.mit.edu/2025/explained-generative-ai-environmental-impact-0117</a></p><p>Pandya, N. S. (2024). In a world where GPUs cannot keep up, can quantum computing step up? <em>Medium</em>. <a href="https://medium.com/@nishkalspandya/in-a-world-where-gpus-cannot-keep-up-can-quantum-computing-step-up-98b5e55078bd">https://medium.com/@nishkalspandya/in-a-world-where-gpus-cannot-keep-up-can-quantum-computing-step-up-98b5e55078bd</a></p><img src="https://medium.com/_/stat?event=post.clientViewed&referrerSource=full_rss&postId=7f0ba53f2171" width="1" height="1" alt="">]]></content:encoded>
        </item>
        <item>
            <title><![CDATA[A Comparative Meta-Analysis of Quantum and Classical Artificial Intelligence: Evaluating…]]></title>
            <link>https://medium.com/@nishkalspandya/a-comparative-meta-analysis-of-quantum-and-classical-artificial-intelligence-evaluating-14808c4a4c4b?source=rss-3a0d887ba305------2</link>
            <guid isPermaLink="false">https://medium.com/p/14808c4a4c4b</guid>
            <category><![CDATA[quantum-computing]]></category>
            <category><![CDATA[ai]]></category>
            <category><![CDATA[quantum-ai]]></category>
            <dc:creator><![CDATA[Nishkal Pandya]]></dc:creator>
            <pubDate>Sun, 15 Feb 2026 16:08:42 GMT</pubDate>
            <atom:updated>2026-02-15T16:08:42.464Z</atom:updated>
            <content:encoded><![CDATA[<h3>A Comparative Meta-Analysis of Quantum and Classical Artificial Intelligence: Evaluating Computational Accuracy, Predictive Precision, and Runtime Dynamics</h3><p>I wrote this as part of a research capstone for my school’s SNHS</p><p>Nishkal Pandya</p><p>Monroe Township High School</p><p>Science National Honor Society — Research Capstone #1</p><p>February 2nd, 2026</p><h3><strong>Abstract</strong></h3><p>The number of transistors on computer chips <strong>double</strong> every year. This is known as Moore’s Law and has fueled decades of innovation. However, transistors are currently close to the size of a singular atom (Havlíček et al., 2019). This means we have a limit that needs a change in the chips. This paper explores Quantum Advantage, and using the exponential increase of qubits can redefine how models learn and make connections. A comparison was made between Variational Quantum Circuits (VQC) and Quantum Long Short-Term Memory (QLSTM) networks against deep learning models. The results showed quantum training an AI can achieve an accuracy rate of about 97%, but classically-trained AI leads in runtime efficiency and stability (Su, Li, &amp; Qiu, 2025). These results show while Quantum AI can be a solution to the limits of transistors, the future relies on hybrid systems.</p><h3><strong>Introduction</strong></h3><p>Modern AI relies on a singular bit, which is limited to the value of a 0 and a 1. This was the foundation that came from the first computer, but the only way we can increase storage is by including more bits, taking up more energy.</p><p>However, Quantum AI uses the qubit, leveraging superposition and entanglement to hold more states: a 0, a 1, or both (IBM Technology, 2022; AI Uncovered, 2023). A qubit is based on probability, which allows different data to be held at the same time without adding more qubits (Marr, 2024; Bosch, 2022).</p><p>Quantum-trained AI is not just a puzzle to mess around with. the benefits are shown in these four fields:</p><ol><li>Environmental Sustainability: Training classical AI models uses large data centers that consume more electricity and water than small cities (International Energy Agency [IEA], 2024; MIT News, 2025). Quantum AI is more efficient and can handle more complex data for a fraction of the cost.</li><li>Market Stability: Financial markets are more connected than ever. Quantum LSTMs (Long Short-Term Memory) can reduce error by about 13.1%, leading to accurate risk assessments and increased global security.</li><li>Drug Discovery and Molecular Models: Simulating molecular interactions in great detail is something classically trained AIs sometimes struggle to do. Quantum AI can simulate it and figure out the tiny placements that are based on probability quickly, reducing the development time of new medicines from years to months (Chen et al., 2025; Emmanoulidis et al., 2024).</li><li>Technological Advancement: Currently, transistors are approaching the size of an atom, and it is not possible to make the hardware smaller than that. We need to use a different approach, which is what Quantum AI can achieve (Pandya, 2024; Taghandiki, 2024).</li></ol><p>Even though classical AIs are faster and more stable, quantum-trained AIs are more accurate and efficient at solving problems.</p><h3><strong>Literature Review</strong></h3><p>Early research into Quantum Machine Learning (QML) was mostly theoretical. In 2018/2019, papers like Nature’s <em>Supervised Learning with Quantum-Enhanced Feature Spaces </em>and Physical Review Letters’ <em>Quantum Machine Learning in Feature Hilbert Spaces</em> showed that quantum computers could map data into a something called a Hilbert Space (Biamonte et al., 2017; Schuld &amp; Killoran, 2019). A Hilbert Space is a world where complex patterns are easier to see. However, the theory only worked to fake data that was too simple to have any meaningful applications.</p><p>However, as time went on to 2024/2025, the field became more tangible, with applications such as the stock market. The focus shifted from physics to AI (Su, Li, &amp; Qiu, 2025; Bosch, 2022).</p><p>The reason Quantum AI works is because of Effective Dimension. The term is coined by IBM’s paper <em>The Power of Quantum Neural Networks</em> in 2021. Effective Dimension is the minimum number of independent variables needed to describe a complex system/model (Abbas et al., 2021). Instead of counting the total number of variables, having fewer independent variables that can find correlations and can consider noise allows for more efficient learning. Classical AI has variables that don’t contribute, while a high effective dimension makes sure all variables are used to identify complex patterns. IBM discovered that quantum models have a higher effective dimension, which is why the model is so accurate (Abbas et al., 2021; Zhang, 2025)..</p><p>Everyone is excited about the accuracy increase, but there is a gap in the studies. Studies focus on the accuracy percentage, but they don’t take the data encoding phase into account. Classical data cannot just simply be plugged into a quantum computer (Caleb Writes Code, 2025; Ranga et al., 2024). The bits have to turn into qubits, but the process is slow and takes up a lot of energy. This is a major limitation currently for real-world use.</p><p>Furthermore, there are debates about where quantum models have the advantage. Classical AI wins on small datasets since quantum hardware is too “noisy”, which means it has too many errors (Bowles et al., 2024; Marr, 2024). The paper <em>Better than Classical? The Subtle Art of Benchmarking Quantum Machine Learning Models </em>found that classical models beat quantum models in 60% of small-scale tests. However, for massive, multi-dimensional problems like predicting protein folds or markets, quantum models are years faster than classical ones.</p><h3><strong>Methodology</strong></h3><p>This paper is a comparative meta-analysis to see if Quantum Advantage has potential across different tests. This shows how different researchers use quantum processors to see if they can solve the same problem.</p><p>The 2 tests include stock price prediction and a general test from over 160 datasets (Bowles et al., 2024). 4 primary variables are common throughout all studies. The independent variable is which type of learning model is used: Classical Deep Learning (LSTM) or Quantum Deep Learning (QLSTM/VQC), while the dependent variables are the accuracy percentage, the RMSE (Root mean square error), and the runtime (Su, Li, &amp; Qiu, 2025; Zhang, 2025).</p><p>Stock market prices are not always linear and “noisy”, so random fluctuations in the data can confuse classical AI. Past stock price data (Open, High, Low, Close, Volume) needs the models to remember long-term trends while internalizing short-term changes. This study used IBM’s Eagle processor, which has 127 qubits (IBM Technology, 2022; Wang et al., 2025). This was chosen because the Eagle has advanced error reduction abilities. The classical AI was run on NVIDIA H100 GPUs, which were the most powerful classical AI hardware available at that time.</p><p>The 2nd test was to make sure the findings have more applications than finance. This test generalized over 160 different datasets, from medical images to logic puzzles. Here, the dependent variables are slightly different. The variables measured are effective dimension, runtime stability, and the predictive gap. The processors used were IBM’s Osprey processor, which has 433 qubits, which allows researchers to test larger datasets than the previous years. PennyLane and Qiskit software, which is the connection between quantum and classical (ResearchGate, 2019; Li &amp; Zhang, 2025). Testing all 160 datasets can help identify which types of data become too complex for a GPU and perfect for a quantum system.</p><h3><strong>Analysis and Data</strong></h3><figure><img alt="" src="https://cdn-images-1.medium.com/max/686/1*OxXeArby-K9LcQaFevrA9Q.png" /></figure><p>The reason the Quantum model achieved 97.36% accuracy is due to entanglement (Su, Li, &amp; Qiu, 2025; Chen et al., 2025). In classical models, variables are independent and have no connection to one another. However, qubits can be correlated and find hidden relationships between data points (AI Uncovered, 2023). An example of this is how the price of a commodity in a market can be related to a stock price. Using this ability, the model can predict with greater accuracy, resulting in a 13.1% reduction in RMSE (Chen et al., 2025; Wang et al., 2025).</p><figure><img alt="" src="https://cdn-images-1.medium.com/max/594/1*PiQJ6zYRDuyTdRFdVhDkxw.png" /></figure><p>Even with these great numbers, Quantum AI suffers from decoherence and noise. The paper, <em>Better than classical? The subtle art of benchmarking quantum machine learning models </em>found a 60% fail rate when quantum models were tested with small and simple datasets (Bowles et al., 2024).</p><p>This happens because today’s quantum hardware is very sensitive. Heat and vibration can cause the qubits to “flip,” changing their state and leading to errors that classical models don’t have. Even though Quantum AI is powerful, it is too unstable for general data and needs specific, complex problems to be worth the noise risk.</p><figure><img alt="" src="https://cdn-images-1.medium.com/max/707/1*6ReWjMCnYge_SGzLzvw5IA.png" /></figure><p>Looking at the runtime data, there seems to be an issue: Fewer steps are required, but hardware latency is higher. They require fewer steps because they can explore multiple solutions at once. However, classical data needs to be converted into quantum states through data encoding (IBM Technology, 2022; Ranga et al., 2024). This takes up a massive amount of time. Furthermore, resetting the IBM chips in between runs (Qubit Initialization) creates a delay that doesn’t exist on an NVIDIA chip. This means the speed of a quantum computer is trapped between hardware being slow.</p><h3><strong>Concluding Remarks</strong></h3><p>This data shows that Quantum AI is not an instant boost for speed (Pandya, 2024; AI Uncovered, 2023). This is because of hardware issues and qubit initialization taking more time than a regular GPU. However, the 97.36% accuracy and 13.1% reduction in RMSE show that Quantum AI is smarter for certain tasks. Using Entanglement and the Effective Dimension of Quantum AI, quantum models can find correlations in complex data like stock markets that classical AI cannot process with the same level of speed and accuracy.</p><p>In 2017, Quantum ML was purely theoretical in the form of math proofs in the paper “Quantum Machine Learning” (Biamonte et al., 2017; Bosch, 2022). In 2025, we have companies like IBM making processors and testing them.</p><p>In the next 5 years, the best AI system will not be purely classical or purely quantum. Instead, it will be a combination of Classical and Quantum (ResearchGate, 2019; Li &amp; Zhang, 2025). This model has a GPU handle the data conversion and checks the Quantum output. The Quantum Processor does the most complex math. This can remove the 60% fail rate that the noisy quantum hardware has (Pandya, 2024; Su, Li, &amp; Qiu, 2025).</p><h3><strong>References</strong></h3><p><a href="https://doi.org/10.1038/s43588-021-00084-1">https://doi.org/10.1038/s43588-021-00084-1</a></p><p><a href="https://www.youtube.com/watch?v=P7_SfxRrXTE">https://www.youtube.com/watch?v=P7_SfxRrXTE</a></p><p><a href="https://doi.org/10.1038/nature23474">https://doi.org/10.1038/nature23474</a></p><p><a href="https://www.youtube.com/watch?v=aFDbO0CZFto">https://www.youtube.com/watch?v=aFDbO0CZFto</a></p><p><a href="https://arxiv.org/abs/2403.07059">https://arxiv.org/abs/2403.07059</a></p><p><a href="https://www.youtube.com/watch?v=sQSQBYHR0ms">https://www.youtube.com/watch?v=sQSQBYHR0ms</a></p><p><a href="https://arxiv.org/abs/2505.23860">https://arxiv.org/abs/2505.23860</a></p><p><a href="https://www.linkedin.com/pulse/magic-matrices-machine-learning-recommender-systems-mohamed-el-refaey/">https://www.linkedin.com/pulse/magic-matrices-machine-learning-recommender-systems-mohamed-el-refaey/</a></p><p><a href="https://www.youtube.com/watch?v=NqHKr9CGWJ0">https://www.youtube.com/watch?v=NqHKr9CGWJ0</a></p><p><a href="https://www.iea.org/reports/energy-and-ai/energy-demand-from-ai">https://www.iea.org/reports/energy-and-ai/energy-demand-from-ai</a></p><p><a href="https://arxiv.org/abs/2505.14295">https://arxiv.org/abs/2505.14295</a></p><p><a href="https://www.youtube.com/watch?v=8I7IKrLP6Hk">https://www.youtube.com/watch?v=8I7IKrLP6Hk</a></p><p><a href="https://news.mit.edu/2025/explained-generative-ai-environmental-impact-0117">https://news.mit.edu/2025/explained-generative-ai-environmental-impact-0117</a></p><p><a href="https://medium.com/@nishkalspandya/in-a-world-where-gpus-cannot-keep-up-can-quantum-computing-step-up-98b5e55078bd">https://medium.com/@nishkalspandya/in-a-world-where-gpus-cannot-keep-up-can-quantum-computing-step-up-98b5e55078bd</a></p><p><a href="https://physicsworld.com/a/physicists-entangle-qubits-in-a-semiconductor-at-room-temperature/">https://physicsworld.com/a/physicists-entangle-qubits-in-a-semiconductor-at-room-temperature/</a></p><p><a href="https://www.researchgate.net/figure/Hybrid-quantum-classical-algorithm-for-data-driven-quantum-circuit-learning-DDQCL_fig1_330576706">https://www.researchgate.net/figure/Hybrid-quantum-classical-algorithm-for-data-driven-quantum-circuit-learning-DDQCL_fig1_330576706</a></p><p><a href="https://arxiv.org/abs/2504.07396">https://arxiv.org/abs/2504.07396</a></p><p><a href="https://doi.org/10.1103/PhysRevLett.122.040504">https://doi.org/10.1103/PhysRevLett.122.040504</a></p><p><a href="https://arxiv.org/abs/2411.06863">https://arxiv.org/abs/2411.06863</a></p><p><a href="https://arxiv.org/abs/2510.01797">https://arxiv.org/abs/2510.01797</a></p><p><a href="https://www.mdpi.com/2673-2688/6/8/175">https://www.mdpi.com/2673-2688/6/8/175</a></p><img src="https://medium.com/_/stat?event=post.clientViewed&referrerSource=full_rss&postId=14808c4a4c4b" width="1" height="1" alt="">]]></content:encoded>
        </item>
        <item>
            <title><![CDATA[In a world where GPUs cannot keep up, can Quantum Computing Step Up?]]></title>
            <link>https://medium.com/@nishkalspandya/in-a-world-where-gpus-cannot-keep-up-can-quantum-computing-step-up-98b5e55078bd?source=rss-3a0d887ba305------2</link>
            <guid isPermaLink="false">https://medium.com/p/98b5e55078bd</guid>
            <category><![CDATA[quantum-computing]]></category>
            <category><![CDATA[quantum-ai]]></category>
            <category><![CDATA[artificial-intelligence]]></category>
            <dc:creator><![CDATA[Nishkal Pandya]]></dc:creator>
            <pubDate>Sun, 09 Nov 2025 22:51:33 GMT</pubDate>
            <atom:updated>2025-11-09T22:51:33.138Z</atom:updated>
            <content:encoded><![CDATA[<h3>Llama 3.1 and Scaling Issues</h3><p>Picture this: It’s launch day for Meta’s new AI, Llama 3.1. You have spent months collecting data and optimizing code to outperform OpenAI with a sophisticated model. You load up your computer, start running the training… and you see the timeline: 4,486 years. Never fear, you are part of a multibillion-dollar company, so you can get 15,999 more GPUs to lower the timeline to 3 months via parallel computing. However, the energy required to train the AI is comparable to powering a small city. This is not an exaggeration that would make sense in a sci-fi novel. This is real life, and we are hitting the limits of our GPUs’ physical capabilities. There will be a point in the near future where we have so much data that we run out of GPUs to train AI models to write the next line of code or generate images more accurately.</p><p>Why do we need so much time to train these AI models? Training an AI model means we need to convert datasets into matrices. Then, once we have another one of these matrices, we multiply them together, as shown below.</p><figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/1*SEg51NYJHvS3OjLGYgDm1w.png" /><figcaption>(Source: LinkedIn)</figcaption></figure><p>As you can see, they can get really long. For example, we might need to multiply a 1000 × 1000 matrix by another matrix of a similar size. This means about a billion calculations (FLOPs) for a single layer. Now imagine what happens when we have thousands of layers. This is when the fans sound like jet engines.</p><p>Here are a few statistics of power costs and their effect on the environment. In 2024, global data center consumption was about 415 Terawatt-Hours, which is 1.5% of global electricity use. Nearly 40% of the electricity is for cooling. For each kilowatt-hour needed for cooling, we need about 2 liters of water. Doing the math, that is about 332 billion liters of water for cooling.</p><p>We can make AI smarter. But based on the above information, throwing more GPUs at it won’t work forever. We need to rethink how we process information. Enter quantum computing.</p><h3>Quantum Computing</h3><p>How is quantum computing different from a regular computer? Classical computers use bits, while quantum computers use something called a qubit. A qubit can be a 0 or a 1 at the same time, like deciding a coin flip while it is still in the air. This is called superposition. If qubits become linked, they can influence each other instantly, like twins. This means quantum computers can explore many solutions simultaneously.</p><figure><img alt="" src="https://cdn-images-1.medium.com/max/221/1*VnzUaV8hN9FA9Hdbxv119w.png" /><figcaption>(Source: IEA)</figcaption></figure><p>Remember those giant matrix multiplications? Instead of using fixed matrices, quantum computing can encode it into quantum states using something called quantum feature maps (see image below).</p><figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/1*cx6oeEixHymuCjh-iCrJAA.jpeg" /><figcaption>(Source: PhysicsWorld)</figcaption></figure><p>The data is embedded into a Hilbert space, which has more dimensions than any GPU can handle. The results are patterns that look random in the classical world are crystal clear. Imagine stacking 2D images vs. projecting the images into a 3D hologram: A much more efficient way of examining the data. Before we explore what happens if we succeed, we need to figure out what is in the way today. This all sounds amazing and all, but why do we not have this already? In 1 word: noise.</p><h3>Today’s Reality</h3><p>Quantum computing is a powerful tool, but today’s versions are far from perfect. The biggest challenge so far is noise, which are tiny disturbances to qubits because of heat, vibration, or electromagnetic interference. These issues may seem small, but they can change a qubits state and cause the system to be inaccurate. If this happens, the qubit can lose its quantum state, or it “decoheres.” This is a limiting factor on the amount of time a quantum computer can run before information goes away. Error rates are another major issue. In classical computers, a transistor flipping by mistake rarely happens. However, quantum computing is based on probability, which includes a chance of error. This probability is a feature instead of a bug, and the longer the algorithm runs, the more likely it is to fail. Another issue is stability. For qubits to be stable, they need to be at near absolute zero. The temperatures make scaling hard. For example, building 1,000 reliable qubits means a lot more than simply adding more, it requires keeping every single qubit perfectly synchronized and protected.</p><p>Since we do not have perfect quantum computers, researchers combined imperfect quantum computers with perfect classical computers to make hybrid models. The quantum components handle the hard math, while the classical computer adjusts parameters and fixes errors.</p><figure><img alt="" src="https://cdn-images-1.medium.com/max/690/1*P-i9_lblMuImTMA7ipcmRw.png" /><figcaption>(Source: ResearchGate)</figcaption></figure><p>Variational Quantum Algorithms (VQAs) such as Variational Quantum Eigensolvers (VQEs) and Quantum Approximate Optimization Algorithms (QAOAs) can tolerate noise, dealing with 1 of the many issues of quantum computing. VQEs are mostly used to find the lowest energy state of molecules for chemistry and material science, and QAOA solves large-scale optimization problems. Another vocab word: quantum annealing, which is a technique that uses quantum mechanics to find the best solution for complex problems (basically QAOA).</p><p>Here is a real-world example. D-Wave’s Quantum Annealing System has up to 40% better performance in optimization than classical computers. The system uses 1000s of connected qubits to explore many possible combinations, and then uses a classical system to find the best results. It is not going to be the best method forever, but it shows that quantum computing can have an application in real-world tasks.</p><p>A solution that IonQ uses is trapped-ion processors to increase coherence time, so more calculations can be made for longer periods of time. Another solution comes from IBM and Rigetti, which uses superconducting qubits (qubits from superconducting circuits to do math) to make reliable gate-based systems. Gate-based systems are systems that use quantum gates/operations in the Hilbert space to change qubits and do math.</p><p>I found a researcher named Samuel Bosch, who did part of a PhD at MIT in quantum machine learning. He used MIT supercomputers to simulate quantum computing and showed that it can outperform classical computers for small datasets. In one project, he used quantum kernel methods, which are types of algorithms that map data into the Hilbert space by encoding classical data into quantum states. The result was that the model could separate data points that classical machines couldn’t. This experiment showed the potential of quantum computing, which can also be shown in this quote by Jensen Huang, the CEO of NVIDIA.</p><figure><img alt="" src="https://cdn-images-1.medium.com/max/346/1*IhQ5tn3Z1Fxrc1C5dKRWsA.png" /></figure><p>This quote implies that even with NVIDIA producing the most powerful GPUs in the world, it will soon be replaced. The only logical possibility of that is quantum computers. He also talked about how the quantum rush is similar to the GPU rush, where companies had to compete for limited funding of something that was obviously going to be the next big thing. However, what separated NVIDIA was that it was able to leverage the need for GPUs in the gaming industry. Quantum computing has to do the same thing, which is leveraging a popular industry. Now, as promised, let’s see what happens if quantum computing actually delivers in the future.</p><h3>What Happens if we Succeed?</h3><p>When we reach fully functional quantum AI with stable hardware and faultless qubits, we will have more possibilities than ever before in human history. Instead of using millions of GPU hours, terawatts of power, and large cooling farms, using up the limited fresh water on Earth, we can have quantum AI do the same amount of training using a small amount of energy. What can happen in months can happen in hours. Since quantum systems can represent exponentially large states, we can use less data to increase complexity. That means that the future of AI models can go from growing “bigger” with more data to smarter with relationships with less data.</p><p>With quantum AI, we can access entirely new fields of science that normal AI cannot hope to achieve today. We can simulate molecular interactions down to the electron, design new materials for batteries or superconductors, or even model biological interactions down to the molecule. We can train AI on data from the past, but we can also use simulated realities to discover physics laws, new drugs, and brand-new energy sources that don’t even exist yet.</p><p>This is another technological arms race. The winner will redefine the possibilities in every field that requires predictions and math, such as research, defense, finance, and technology. More AI models can be trained with minimal costs, increasing innovation in more fields than the ones listed above.</p><p>This power can create a societal imbalance. If only a few nations/corporations control access to high-level quantum technology, they control the next phases of innovations and intelligence. Models can be cheap to train, but only for people who have quantum access. For everyone else, it will be very expensive and impossible to replicate the sophisticated levels of the quantum class. This is why governments, startups, and research labs are all competing to change who controls the innovation.</p><p>AI isn’t limited by intelligence. It’s limited by hardware. Right now, we can say “add more GPUs,” but that uses too much energy, water, and time. We need a different thought process.</p><p>Quantum AI is the different thought process. It mirrors the uncertainty of nature through probability. We can rebuild intelligence and innovation from the ground up in a new realm of math. Intelligence could be more efficient, more sustainable, and more human.</p><p>Sources:</p><p><a href="https://arxiv.org/abs/2505.23860?utm_source=chatgpt.com">https://arxiv.org/abs/2505.23860?</a></p><p><a href="https://www.iea.org/reports/energy-and-ai/energy-demand-from-ai?utm_source=chatgpt.com">https://www.iea.org/reports/energy-and-ai/energy-demand-from-ai</a></p><p><a href="https://arxiv.org/abs/2411.06863">https://arxiv.org/abs/2411.06863</a></p><p><a href="https://arxiv.org/abs/2505.14295?utm_source=chatgpt.com">https://arxiv.org/abs/2505.14295?</a></p><p><a href="https://arxiv.org/abs/2504.07396?utm_source=chatgpt.com">https://arxiv.org/abs/2504.07396?</a></p><p><a href="https://arxiv.org/abs/2510.01797?utm_source=chatgpt.com">https://arxiv.org/abs/2510.01797?</a></p><p><a href="https://www.mdpi.com/2673-2688/6/8/175">https://www.mdpi.com/2673-2688/6/8/175?</a></p><p><a href="https://www.youtube.com/watch?v=sQSQBYHR0ms">https://www.youtube.com/watch?v=sQSQBYHR0ms</a></p><p><a href="https://www.youtube.com/watch?v=P7_SfxRrXTE&amp;t=554s">https://www.youtube.com/watch?v=P7_SfxRrXTE&amp;t=554s</a></p><p><a href="https://www.youtube.com/watch?v=8I7IKrLP6Hk">https://www.youtube.com/watch?v=8I7IKrLP6Hk</a></p><p><a href="https://www.youtube.com/watch?v=NqHKr9CGWJ0">https://www.youtube.com/watch?v=NqHKr9CGWJ0</a></p><p><a href="https://www.youtube.com/watch?v=aFDbO0CZFto&amp;t=2s">https://www.youtube.com/watch?v=aFDbO0CZFto&amp;t=2s</a></p><p><a href="https://www.linkedin.com/pulse/magic-matrices-machine-learning-recommender-systems-mohamed-el-refaey/">https://www.linkedin.com/pulse/magic-matrices-machine-learning-recommender-systems-mohamed-el-refaey/</a></p><p><a href="https://news.mit.edu/2025/explained-generative-ai-environmental-impact-0117">https://news.mit.edu/2025/explained-generative-ai-environmental-impact-0117</a></p><p><a href="https://physicsworld.com/a/physicists-entangle-qubits-in-a-semiconductor-at-room-temperature/">https://physicsworld.com/a/physicists-entangle-qubits-in-a-semiconductor-at-room-temperature/</a></p><p><a href="https://www.researchgate.net/figure/Hybrid-quantum-classical-algorithm-for-data-driven-quantum-circuit-learning-DDQCL_fig1_330576706">https://www.researchgate.net/figure/Hybrid-quantum-classical-algorithm-for-data-driven-quantum-circuit-learning-DDQCL_fig1_330576706</a></p><img src="https://medium.com/_/stat?event=post.clientViewed&referrerSource=full_rss&postId=98b5e55078bd" width="1" height="1" alt="">]]></content:encoded>
        </item>
    </channel>
</rss>