Large Language Models (LLMs) can produce varied, creative, and sometimes surprising outputs even when given the same prompt. This randomness is not a bug but a core feature of how the model samples its next token from a probability distribution. In this article, we break down the key sampling strategies and demonstrate how parameters such as temperature, top-k, and top-p influence the balance between consistency and creativity. In this tutorial, we take a hands-on approach to understand: How logits become probabilities How temperature, top-k, and top-p sampling work How different sampling strategies shape the model’s next-token distribution By the end, you will understand the mechanics behind LLM inference and be able to adjust the creativity or determinism of the output. Let’s get started. How LLMs Choose Their Words: A Practical Walk-Through of Logits, Softmax and SamplingPhoto by Colton Duke. Some rights reserved. Overview This article is divided into four parts; they are: How Logits Become Probabilities Temperature Top-k Sampling Top-p Sampling How Logits Become Probabilities When you ask an LLM a question, it outputs a vector of logits. Logits are raw scores the model assigns to each possible next token in its vocabulary. If the model has a vocabulary of $V$ tokens, it will output a vector of $V$ logits for each next word position. A logit is a real number. It is converted into a probability by the softmax function: $$p_i = \frac{e^{x_i}}{\sum_{j=1}^{V} e^{x_j}}$$ where $x_i$ is the logit for token $i$ and $p_i$ is the corresponding probability. Softmax transforms these raw scores into a probability distribution. All $p_i$ are positive, and their sum is 1. Suppose we give the model this prompt: Today’s weather is so ___ The model considers every token in its vocabulary as a possible next word. For simplicity, let’s say there are only 6 tokens in the vocabulary: wonderful cloudy nice hot gloomy delicious wonderful cloudy nice hot gloomy delicious The model produces one logit for each token. Here’s an example set of logits the model might output and the corresponding probabilities based on the softmax function: Token Logit Probability wonderful 1.2 0.0457 cloudy 2.0 0.1017 nice 3.5 0.4556 hot 3.0 0.2764 gloomy 1.8 0.0832 delicious 1.0 0.0374 You can confirm this by using the softmax function from PyTorch: import torch import torch.nn.functional as F vocab = [“wonderful”, “cloudy”, “nice”, “hot”, “gloomy”, “delicious”] logits = torch.tensor([1.2, 2.0, 3.5, 3.0, 1.8, 1.0]) probs = F.softmax(logits, dim=-1) print(probs) # Output: # tensor([0.0457, 0.1017, 0.4556, 0.2764, 0.0832, 0.0374]) import torch import torch.nn.functional as F   vocab = [“wonderful”, “cloudy”, “nice”, “hot”, “gloomy”, “delicious”] logits = torch.tensor([1.2, 2.0, 3.5, 3.0, 1.8, 1.0]) probs = F.softmax(logits, dim=–1) print(probs) # Output: # tensor([0.0457, 0.1017, 0.4556, 0.2764, 0.0832, 0.0374]) Based on this result, the token with the highest probability is “nice”. LLMs don’t always select the token with the highest probability; instead, they sample from the probability distribution to produce a different output each time. In this case, there’s a 46% probability of seeing “nice”. If you want the model to give a more creative answer, how can you change the probability distribution such that “cloudy”, “hot”, and other answers would also appear more often? Temperature Temperature ($T$) is a model inference parameter. It is not a model parameter; it is a parameter of the algorithm that generates the output. It scales logits before applying softmax: $$p_i = \frac{e^{x_i / T}}{\sum_{j=1}^{V} e^{x_j / T}}$$ You can expect the probability distribution to be more deterministic if $T<1$, since the difference between each value of $x_i$ will be exaggerated. On the other hand, it will be more random if $T>1$, as the difference between each value of $x_i$ will be reduced. Now, let’s visualize this effect of temperature on the probability distribution: import matplotlib.pyplot as plt import torch import torch.nn.functional as F vocab = [“wonderful”, “cloudy”, “nice”, “hot”, “gloomy”, “delicious”] logits = torch.tensor([1.2, 2.0, 3.5, 3.0, 1.8, 1.0]) # (vocab_size,) scores = logits.unsqueeze(0) # (1, vocab_size) temperatures = [0.1, 0.5, 1.0, 3.0, 10.0] fig, ax = plt.subplots(figsize=(10, 6)) for temp in temperatures: # Apply temperature scaling scores_processed = scores / temp # Convert to probabilities probs = F.softmax(scores_processed, dim=-1)[0] # Sample from the distribution sampled_idx = torch.multinomial(probs, num_samples=1).item() print(f”Temperature = {temp}, sampled: {vocab[sampled_idx]}”) # Plot the probability distribution ax.plot(vocab, probs.numpy(), marker=”o”, label=f”T={temp}”) ax.set_title(“Effect of Temperature”) ax.set_ylabel(“Probability”) ax.legend() plt.show() 1 2 3 4 5 6 7 8 9 10 11 12 13 14 15 16 17 18 19 20 21 22 23 24 25 import matplotlib.pyplot as plt import torch import torch.nn.functional as F   vocab = [“wonderful”, “cloudy”, “nice”, “hot”, “gloomy”, “delicious”] logits = torch.tensor([1.2, 2.0, 3.5, 3.0, 1.8, 1.0]) # (vocab_size,) scores = logits.unsqueeze(0) # (1, vocab_size) temperatures = [0.1, 0.5, 1.0, 3.0, 10.0]   fig, ax = plt.subplots(figsize=(10, 6)) for temp in temperatures:     # Apply temperature scaling     scores_processed = scores / temp     # Convert to probabilities     probs = F.softmax(scores_processed, dim=–1)[0]     # Sample from the distribution     sampled_idx = torch.multinomial(probs, num_samples=1).item()     print(f“Temperature = {temp}, sampled: {vocab[sampled_idx]}”)     # Plot the probability distribution     ax.plot(vocab, probs.numpy(), marker=‘o’, label=f“T={temp}”)   ax.set_title(“Effect of Temperature”) ax.set_ylabel(“Probability”) ax.legend() plt.show() This code generates a probability distribution over each token in the vocabulary. Then it samples a token based on the probability. Running this code may produce the following output: Temperature = 0.1, sampled: nice Temperature = 0.5, sampled: nice Temperature = 1.0, sampled: nice Temperature = 3.0, sampled: wonderful Temperature = 10.0, sampled: delicious Temperature = 0.1, sampled: nice Temperature = 0.5, sampled: nice Temperature = 1.0, sampled: nice Temperature = 3.0, sampled: wonderful Temperature = 10.0, sampled: delicious and the following plot showing the probability distribution for each temperature: The effect of temperature to the resulting probability distribution The model may produce the nonsensical output “Today’s weather is so delicious” if you set the temperature to 10! Top-k Sampling The model’s output is a vector of logits for each position in the output sequence. The inference algorithm converts the logits to actual words, or in LLM terms, tokens. The simplest method for selecting the next token is greedy sampling,

New Delhi: India’s annual telecom exports have jumped by 72 per cent to over Rs 18,406 crore in the last five years, Union Communications Minister Jyotiraditya M. Scindia informed Parliament on Wednesday. Responding to a question in the Lok Sabha, the minister stated that India’s telecom exports have risen from Rs 10,000 crore in 2020–21 to Rs 18,406 crore in 2024–25, while imports have remained capped at around Rs 51,000 crore. He said that under the leadership of Prime Minister Narendra Modi, India is not only moving rapidly towards self-reliance in the telecom sector but is also preparing itself for global leadership. In response to a supplementary question, Scindia highlighted India’s achievements in 5G deployment. He informed that out of 778 districts in the country, 767 districts have already been connected to the 5G network. He further stated that India currently has 36 crore 5G subscribers, a number expected to rise to 42 crore by 2026 and reach 100 crore by 2030. About SATCOM, Scindia said that the worldwide experience shows that areas which cannot be connected through conventional BTS or backhaul, or through broadband connectivity using optical fibre cable, can only be served through satellite communications. In this context, India has taken a decisive step to ensure that SATCOM services are made available to customers across the length and breadth of the country. Add Zee News as a Preferred Source He noted that the objective of the government is to offer a full bouquet of telecom services to every customer, enabling individuals to make informed choices based on their needs and preferred price points. The Union Minister highlighted that the SATCOM policy framework is firmly in place, with spectrum slated for administrative assignment. Three licences have already been issued — to Starlink, OneWeb, and Reliance. He further said that two key aspects must be addressed before operators can commence commercial services. The first pertains to spectrum assignment, including the determination of administrative spectrum charges, which falls under the purview of the Telecom Regulatory Authority of India (TRAI). The second aspect relates to security clearances from enforcement agencies. To facilitate this process, operators have been provided with a sample spectrum to conduct demonstrations, and all three licensees are currently undertaking the required compliance activities. Once the operators demonstrate adherence to prescribed security norms — including the requirement to host international gateways within India — the necessary approvals will be granted, enabling the rollout of SATCOM services to customers, the minister added.

In this article, you will learn how to build, train, and compare an LSTM and a transformer for next-day univariate time series forecasting on real public transit data. Topics we will cover include: Structuring and windowing a time series for supervised learning. Implementing compact LSTM and transformer architectures in PyTorch. Evaluating and comparing models with MAE and RMSE on held-out data. All right, full steam ahead. Transformer vs LSTM for Time Series: Which Works Better?Image by Editor Introduction From daily weather measurements or traffic sensor readings to stock prices, time series data are present nearly everywhere. When these time series datasets become more challenging, models with a higher level of sophistication — such as ensemble methods or even deep learning architectures — can be a more convenient option than classical time series analysis and forecasting techniques. The objective of this article is to showcase how two deep learning architectures are trained and used to handle time series data — long short term memory (LSTM) and the transformer. The main focus is not merely leveraging the models, but understanding their differences when handling time series and whether one architecture clearly outperforms the other. Basic knowledge of Python and machine learning essentials is recommended. Problem Setup and Preparation For this illustrative comparison, we will consider a forecasting task on a univariate time series: given the temporally ordered previous N time steps, predict the (N+1)th value. In particular, we will use a publicly available version of the Chicago rides dataset, which contains daily recordings for bus and rail passengers in the Chicago public transit network dating back to 2001. This initial piece of code imports the libraries and modules needed and loads the dataset. We will import pandas, NumPy, Matplotlib, and PyTorch — all for the heavy lifting — along with the scikit-learn metrics that we will rely on for evaluation. import pandas as pd import numpy as np import matplotlib.pyplot as plt import torch import torch.nn as nn from sklearn.metrics import mean_squared_error, mean_absolute_error url = “https://data.cityofchicago.org/api/views/6iiy-9s97/rows.csv?accessType=DOWNLOAD” df = pd.read_csv(url, parse_dates=[“service_date”]) print(df.head()) import pandas as pd import numpy as np import matplotlib.pyplot as plt   import torch import torch.nn as nn from sklearn.metrics import mean_squared_error, mean_absolute_error   url = “https://data.cityofchicago.org/api/views/6iiy-9s97/rows.csv?accessType=DOWNLOAD” df = pd.read_csv(url, parse_dates=[“service_date”]) print(df.head()) Since the dataset contains post-COVID real data about passenger numbers — which may severely mislead the predictive power of our models due to being very differently distributed than pre-COVID data — we will filter out records from January 1, 2020 onwards. df_filtered = df[df[‘service_date’] <= ‘2019-12-31’] print(“Filtered DataFrame head:”) display(df_filtered.head()) print(“\nShape of the filtered DataFrame:”, df_filtered.shape) df = df_filtered df_filtered = df[df[‘service_date’] <= ‘2019-12-31’]   print(“Filtered DataFrame head:”) display(df_filtered.head())   print(“\nShape of the filtered DataFrame:”, df_filtered.shape) df = df_filtered A simple plot will do the job to show what the filtered data looks like: df.sort_values(“service_date”, inplace=True) ts = df.set_index(“service_date”)[“total_rides”].fillna(0) plt.plot(ts) plt.title(“CTA Daily Total Rides”) plt.show() df.sort_values(“service_date”, inplace=True) ts = df.set_index(“service_date”)[“total_rides”].fillna(0)   plt.plot(ts) plt.title(“CTA Daily Total Rides”) plt.show() Chicago rides time series dataset plotted Next, we split the time series data into training and test sets. Importantly, in time series forecasting tasks — unlike classification and regression — this partition cannot be done at random, but in a purely sequential fashion. In other words, all training instances come chronologically first, followed by test instances. This code takes the first 80% of the time series as a training set, and the remaining 20% for testing. n = len(ts) train = ts[:int(0.8*n)] test = ts[int(0.8*n):] train_vals = train.values.astype(float) test_vals = test.values.astype(float) n = len(ts) train = ts[:int(0.8*n)] test = ts[int(0.8*n):]   train_vals = train.values.astype(float) test_vals = test.values.astype(float) Furthermore, raw time series must be converted into labeled sequences (x, y) spanning a fixed time window to properly train neural network-based models upon them. For example, if we use a time window of N=30 days, the first instance will span the first 30 days of the time series, and the associated label to predict will be the 31st day, and so on. This gives the dataset an appropriate labeled format for supervised learning tasks without losing its important temporal meaning: def create_sequences(data, seq_len=30): X, y = [], [] for i in range(len(data)-seq_len): X.append(data[i:i+seq_len]) y.append(data[i+seq_len]) return np.array(X), np.array(y) SEQ_LEN = 30 X_train, y_train = create_sequences(train_vals, SEQ_LEN) X_test, y_test = create_sequences(test_vals, SEQ_LEN) # Convert our formatted data into PyTorch tensors X_train = torch.tensor(X_train).float().unsqueeze(-1) y_train = torch.tensor(y_train).float().unsqueeze(-1) X_test = torch.tensor(X_test).float().unsqueeze(-1) y_test = torch.tensor(y_test).float().unsqueeze(-1) def create_sequences(data, seq_len=30):     X, y = [], []     for i in range(len(data)–seq_len):         X.append(data[i:i+seq_len])         y.append(data[i+seq_len])     return np.array(X), np.array(y)   SEQ_LEN = 30 X_train, y_train = create_sequences(train_vals, SEQ_LEN) X_test, y_test = create_sequences(test_vals, SEQ_LEN)   # Convert our formatted data into PyTorch tensors X_train = torch.tensor(X_train).float().unsqueeze(–1) y_train = torch.tensor(y_train).float().unsqueeze(–1) X_test = torch.tensor(X_test).float().unsqueeze(–1) y_test = torch.tensor(y_test).float().unsqueeze(–1) We are now ready to train, evaluate, and compare our LSTM and transformer models! Model Training We will use the PyTorch library for the modeling stage, as it provides the necessary classes to define both recurrent LSTM layers and encoder-only transformer layers suitable for predictive tasks. First up, we have an LSTM-based RNN architecture like this: class LSTMModel(nn.Module): def __init__(self, hidden=32): super().__init__() self.lstm = nn.LSTM(1, hidden, batch_first=True) self.fc = nn.Linear(hidden, 1) def forward(self, x): out, _ = self.lstm(x) return self.fc(out[:, -1]) lstm_model = LSTMModel() class LSTMModel(nn.Module):     def __init__(self, hidden=32):         super().__init__()         self.lstm = nn.LSTM(1, hidden, batch_first=True)         self.fc = nn.Linear(hidden, 1)       def forward(self, x):         out, _ = self.lstm(x)         return self.fc(out[:, –1])   lstm_model = LSTMModel() As for the encoder-only transformer for next-day time series forecasting, we have: class SimpleTransformer(nn.Module): def __init__(self, d_model=32, nhead=4): super().__init__() self.embed = nn.Linear(1, d_model) enc_layer = nn.TransformerEncoderLayer(d_model=d_model, nhead=nhead, batch_first=True) self.transformer = nn.TransformerEncoder(enc_layer, num_layers=1) self.fc = nn.Linear(d_model, 1) def forward(self, x): x = self.embed(x) x = self.transformer(x) return self.fc(x[:, -1]) transformer_model = SimpleTransformer() class SimpleTransformer(nn.Module):     def __init__(self, d_model=32, nhead=4):         super().__init__()         self.embed = nn.Linear(1, d_model)         enc_layer = nn.TransformerEncoderLayer(d_model=d_model, nhead=nhead, batch_first=True)         self.transformer = nn.TransformerEncoder(enc_layer, num_layers=1)         self.fc = nn.Linear(d_model, 1)       def forward(self, x):         x = self.embed(x)         x = self.transformer(x)         return self.fc(x[:, –1])   transformer_model = SimpleTransformer() Note that

OnePlus 15R Price In India: OnePlus has launched the OnePlus 15R along with the OnePlus Pad Go 2 in the Indian market. The Chinese smartphone also launched the ‘Ace’ variant of the OnePlus 15R. The phone comes in Charcoal Black, Mint Breeze, and Electric Violet colour options. The dual SIM handset runs on Android 16-based OxygenOS 16 and the company promises four OS upgrades and six years of security updates for the smartphone.   OnePlus 15R Specifications The OnePlus 15R features a 6.83-inch Full-HD+ AMOLED display with a resolution of 1,272×2,800 pixels, offering a smooth 165Hz refresh rate, 450ppi pixel density, and a tall 19.8:9 aspect ratio. Add Zee News as a Preferred Source The smartphone is powered by the Qualcomm’s octa-core 3nm Snapdragon 8 Gen 5 chipset with a peak clock speed of 3.8GHz, paired with 12GB LPDDR5x Ultra RAM, up to 512GB of UFS 4.1 storage, and an Adreno 8-series GPU. The device packs a large 7,400mAh silicon carbon battery with 80W wired fast charging support, and OnePlus claims it will retain 80 percent of its capacity even after four years of use. It also comes with IP66, IP68, IP69, and IP69K ratings for dust and water resistance. On the photography front, the OnePlus 15R houses a 50-megapixel primary rear camera with OIS, alongside an 8-megapixel ultra-wide lens with a 112-degree field of view, supporting up to 4K video recording at 120fps, cinematic mode, multi-view video, and video zoom. For quality selfies and videos, there is a 32-megapixel selfie camera capable of shooting 4K videos at 30fps. On the connectivity front, the smartphone supports 5G, 4G LTE, Wi-Fi 7, Bluetooth 6.0, NFC, USB Type-C, GPS, GLONASS, BDS, Galileo, QZSS, and NavIC. OnePlus 15R Bank Offers Buyers can avail attractive bank offers on the OnePlus 15R, with HDFC Bank credit card users getting an instant discount of up to Rs 3,000 along with easy EMI options. Axis Bank credit card customers can also enjoy an instant Rs 3,000 discount and flexible EMI plans. In addition, consumers can opt for up to six months of no-cost EMI on major credit cards. As part of other benefits, OnePlus is offering a 180-day phone replacement plan along with a Lifetime Display Warranty for all OnePlus 15R users.  OnePlus 15R Price In India And Availability The smartphone is priced at Rs 47,999 for the base variant with 12GB RAM and 256GB storage. The 512GB storage variant costs Rs 52,999. Pre-orders begin today, while sales start next week. In India, the handset will go on sale on December 22 at 12 PM (noon) via Amazon, the OnePlus India online store, and select offline retail stores. 

New Delhi: India has emerged as the world’s largest and most active market for large language model (LLM) adoption, according to an analysis by Bank of America (BofA) on Wednesday. The country now leads globally in the number of users for popular AI apps such as ChatGPT, Gemini and Perplexity, both in terms of monthly active users (MAUs) and daily active users (DAUs). BofA said India’s rapid rise as a key AI market is driven by a combination of scale, affordability and demographics. India has the second-largest online population in the world, with more than 700–750 million mobile internet users. Affordable data plans have made AI access easier, with users able to consume 20–30 GB of data a month at around $2. In addition, more than 60 per cent of Indian internet users are under the age of 35, and a large part of this population is English-speaking and quick to adopt new technologies. Add Zee News as a Preferred Source AI adoption in India is also getting a push from telecom companies. BofA noted that telcos such as Jio and Bharti Airtel are offering complimentary subscriptions to paid versions of AI apps like Gemini and Perplexity. This, according to the report, is creating a win-win situation for users, AI companies and telecom operators. For consumers, access to advanced AI tools at a low cost is helping create a level playing field. Indian users are using these tools to improve learning outcomes and boost productivity. The availability of AI models in multiple Indian languages is also helping bridge the digital divide and reduce language barriers, leading to what BofA described as the democratisation of AI. The report also highlighted that India could emerge as a testing ground for the next phase of AI, known as agentic AI. These are AI applications that can reason, plan and execute tasks independently. Given India’s massive and diverse user base, BofA believes the country is well suited to stress-test such applications in real-world conditions before they are rolled out globally. It also suggested that global AI companies could partner with Indian firms for fulfilment services, similar to how AI agents in the US have tied up with travel platforms like Booking and Expedia.

Google Translate In Real Time App: In a country like India, every journey is a conversation in itself, one that shifts languages with every mile. From a chai-side chat in a small town to a boardroom discussion in a bustling metro, words often change faster than people can keep up. That is where Google Translate quietly steps in, turning confusion into clarity. Whether you are travelling across states, attending meetings, or connecting with someone from a different culture, language barriers no longer have to slow you down. Now, Google is taking this experience a step further by offering real-time translation through the Translate app, powered by Gemini and seamlessly accessible across all headphones, making communication feel natural, instant, smooth, and effortless. More interesting part is that you do not need costly smart earbuds or special translation devices. Any regular wired or wireless headphones will do the job. If you are wondering how to use Google Translate for real-time translations on any headphones, this article will guide you step by step. Add Zee News as a Preferred Source How To Use Google Translate On Any Headphones In Real Time Step 1: Open the Google Translate app on your smartphone. Step 2: At the top, select your spoken language on the left and the language you want to understand on the right. Step 3: Tap the Conversation option located at the bottom left of the main screen. Step 4: When the prompt appears, tap the Start button to begin real-time translation. Step 5: Keep your phone close to the person who is speaking and ask them to speak clearly at a normal pace. Step 6: The app will listen through your phone’s microphone, and the translated audio will automatically play in your connected headphones. Google Translate Vs Apple Live Translation: What Makes Google’s Real-Time App Different Google’s real-time translation keeps things simple and easy to use. You do not need to buy special devices or costly earbuds. An Android phone, the Google Translate app, and any headphones, wired or wireless, are enough. Once you turn it on, you just listen while the other person speaks, and the translated voice plays in your headphones. The feature supports more than 70 languages in beta and even keeps the speaker’s voice style and pauses, so it feels natural. (Also Read: Google Pixel 10 Pro Gets Huge Discount On THIS Platform Under Rs 1,00,000; Check Camera, Display, Battery, Price And Other Specs) On the other hand, Apple takes a different path with Live Translation. It is deeply tied to the Apple ecosystem and works only with certain AirPods and an iPhone running the latest iOS with Apple Intelligence. The big advantage is two-way conversation, where both people can talk and hear translations. Apple also offers text transcripts and replay options, making it helpful but limited to Apple devices. Conclusion: Since this is a beta feature, you may notice a few bugs or minor translation errors, which are common with AI tools. Tech giant Google is actively improving the feature to make it more accurate and reliable. iPhone users are expected to get access sometime in 2026, along with a wider rollout to more countries. It is important to note that Google Translate does not work in offline mode, so you must ensure a stable network connection while using the app. 

New Delhi: In a bid to support India’s research ecosystem, Google.org, the philanthropic arm of Google, on Tuesday announced a funding of $8 million for four AI Centres of Excellence for health, agriculture, education, and sustainable cities. The centres were established by the Government, aligning with the vision to “Make AI in India and Make AI work for India”. The centres include TANUH at IISc Bangalore, which will focus on developing scalable AI solutions for effective treatment of non-communicable diseases; Airawat Research Foundation at IIT Kanpur, which will focus on pioneering research on AI to transform urban governance. The AI Centre of Excellence for Education at IIT Madras will focus on developing solutions to enhance learning and teaching outcomes, while ANNAM.AI at IIT Ropar will focus on developing data-driven solutions for agriculture and farmer welfare. Add Zee News as a Preferred Source In addition, Google announced a $2 million founding contribution to establish the new Indic Language Technologies Research Hub at IIT Bombay. The hub, set up in memory of Professor Pushpak Bhattacharyya, a pioneer in Indic language technologies and a Visiting Researcher at Google DeepMind, will aim to ensure that global AI advancements serve India’s linguistic diversity. “India is approaching artificial intelligence as a strategic national capability, not as a short-term technology trend. The four AI Centres of Excellence have been conceived as a coordinated national research mission, advancing foundational research, responsible AI, and applied solutions that serve public purpose, and contributing to our larger aspiration of Viksit Bharat 2047,” said Dharmendra Pradhan, Minister of Education. “Building a globally competitive AI ecosystem requires not only public investment, but also strong institutional leadership and long-term partnerships with industry. This effort is supported by Google and Google.org through their $8 million contribution to the AI Centres of Excellence and a $2 million founding contribution to the Indic Language Technologies Research Hub at IIT Bombay,” he added. At Google’s “Lab to Impact” dialogue, supported by the India AI Impact Summit 2026, the company also committed $400,000 to support the development of India’s Health Foundation model using Google’s MedGemma — the specialised AI model designed for healthcare. As a first step, Ajna Lens will work with experts from the All India Institute of Medical Sciences (AIIMS) to build models that will support India-specific use cases in Dermatology and OPD Triaging. The resulting models will contribute to India’s Digital Public Infrastructure, and their outcomes will be made accessible to the ecosystem. Google is also working with India’s National Health Authority (NHA) to deploy its advanced AI to convert millions of fragmented, unstructured medical records, such as doctors’ clinical and progress notes, into the international, machine-readable FHIR standard. “From foundational research to ecosystem deployment to scaled impact, our full-stack approach is equipping the country to lead a global AI-powered future, with innovations from India’s labs benefiting billions across the world,” said Dr Manish Gupta, Senior Research Director, Google DeepMind.

India’s First Dual-Core Microprocessor DHRUV64: India has taken a crucial step in its semiconductor journey by launching DHRUV64, the country’s first 1.0 GHz, 64-bit dual-core microprocessor, according to a government release. This processor has been fully designed in India by the Centre for Development of Advanced Computing (C-DAC) under the Microprocessor Development Programme. DHRUV64 is part of India’s plan to reduce its dependence on processors made in other countries. The government says the chip can be used for both defence and commercial purposes and helps build a strong, homegrown technology system. DHRUV64: Dual-Core Microprocessor Features Add Zee News as a Preferred Source The processor uses modern technology to work faster, handle multiple tasks at the same time, and remain reliable. It can easily connect with different hardware systems, making it useful for 5G networks, vehicles, electronic devices, factories, and Internet of Things (IoT) applications. DHRUV64: Future Semiconductor Projects The release highlights that India contributes nearly 20 per cent of global microprocessor production, showing why homegrown chip development is important for long-term security. Even before DHRUV64, India had started building its processor ecosystem with chips such as SHAKTI, AJIT, VIKRAM, and THEJAS64. These initiatives helped create a strong base for domestic chip design. DHRUV64 gives startups, researchers, and industries a local platform to develop and test products without depending on foreign processors. It also allows prototype development at a lower cost. With India’s large pool of chip design talent, the processor is expected to help train more skilled professionals for future semiconductor projects. Its success is also likely to speed up the development of the upcoming Dhanush and Dhanush+ processors. DHRUV64: India’s First 1.0 GHz, 64-bit dual-core Microprocessor DHRUV64 is India’s first homegrown 1.0 GHz, 64-bit dual-core microprocessor, strengthening the indigenous processor pipeline. National Programmes like Digital India RISC-V support the design, testing and… pic.twitter.com/i5mfZOY3BZ — PIB India (@PIB_India) December 15, 2025 DHRUV64 Is Part Of Digital India Programme The rollout of DHRUV64 is part of the Digital India RISC-V (DIR-V) Programme, which aims to make India a global hub for electronics system design and manufacturing. The release says DHRUV64 is the third chip fabricated under DIR-V, following THEJAS32 and THEJAS64. Work on DHANUSH64 and DHANUSH64+ System-on-Chip variants is now in progress. The Ministry of Electronics and Information Technology (MeitY) guides national programmes such as the Microprocessor Development Programme, DIR-V, Chips to Startup (C2S), the India Semiconductor Mission, the Design Linked Incentive Scheme, and INUP-i2i. These initiatives strengthen the design ecosystem and support research and training across academia, startups, and industry. C-DAC continues to lead processor design efforts and is now developing the next processors in the RISC-V roadmap. “The progress from THEJAS32 to DHRUV64, and the ongoing development of Dhanush and Dhanush+, illustrates a confident national pathway toward indigenous processor innovation and technological self-reliance,” the release said.(With IANS Inputs)

Motorola Edge 70 Price In India: Motorola has expanded its premium mid-range portfolio in India with the launch of the ultra-slim Motorola Edge 70 today in the Indian market. The newly launched smartphone succeeds the popular Edge 60 and brings a focus on design, cameras and durability, alongside modern software and AI features. The Motorola Edge 70 comes with a thin form factor, that is, 5.99mm thickness and weighs around 159 grams. It is available in three colour options which includes the Lily Pad, Gadget Grey and Bronze Green. Adding further, the Motorola Edge 70 sticks to the classic Edge-series design, with a squircle-shaped rear camera module and a premium vegan leather back. Measuring just 5.99mm in thickness, it is one of the slimmest smartphones in its segment. Despite its ultra-slim body, the phone is claimed to meet MIL-STD-810H military-grade durability standards. The new Motorola Edge 70 runs on Android 16, with Hello UI. It will get three major Android upgrades and four years of security updates. Add Zee News as a Preferred Source Motorola Edge 70 Specifications The smartphone features a 6.7-inch 1.5K AMOLED display with a 120Hz refresh rate, up to 4,500 nits of peak brightness, Gorilla Glass 7i protection, Dolby Vision, and HDR10+ support. It is powered by Qualcomm’s Snapdragon 7 Gen 4 chipset, paired with 8GB of LPDDR5x RAM and 256GB of UFS 3.1 internal storage. The smartphone packs a 5,000mAh silicon carbon battery that is claimed to deliver up to 31 hours of continuous video playback, along with 68W wired and 15W wireless fast charging support. Just 5.99mm THIN, the motorola edge 70 redefines ultra-slim engineering finished in Pantone™ curated colours Sale starts 23rd December at ₹28,999* on Flipkart, https://t.co/azcEfy2uaW, and leading retail stores.#MotorolaEdge70 #Motorola #ImpossiblyThinYetUncompromised pic.twitter.com/3QI2S76fNR — Motorola India (@motorolaindia) December 15, 2025 On the photography front, it comes with a triple rear camera setup led by a 50-megapixel primary sensor with optical image stabilisation (OIS), a 50-megapixel ultra-wide camera, and a three-in-one light sensor, while the front houses a 50-megapixel selfie camera. (Also Read: Apple Rolls Out iOS 26.2 Stable Update For iPhone Users: Check New Features, Eligible Models; Follow THESE Steps To Install) The device supports video recording at up to 4K resolution at 60fps. It also ships with Moto AI features such as Next Move, Catch Me Up 2.0, Pay Attention 2.0, Remember This + Recall, and Co-pilot, along with AI Video Enhancement, AI Action Shot, and AI Photo Enhancement tools. Motorola Edge 70 Price In India, Availability The smartphone is priced in India at Rs 29,999 for the single variant with 8GB RAM and 256GB storage. Buyers can also avail a Rs 1,000 bank discount on select bank cards. The smartphone will be available for purchase through Flipkart, the Motorola India website, and authorised offline retail stores across the country.  Motorola Edge 70 India Launch: Alternative Options Apart from the Motorola Edge 70, consumers can also consider several alternative smartphones in a similar price range. These include the Vivo V60e, priced at Rs 27,990 on Amazon; the Realme 15 Pro 5G, available for Rs 25,999 on Amazon; the Motorola Edge 60 Pro, listed at around Rs 28,000 on Amazon; and the Nothing Phone 3a Pro, which is currently priced at Rs 27,900 on Amazon. 

Google And Apple Emergency Security Updates: Apple And Google have rolled out urgent security updates to protect against a hacking campaign targeting an unknown number of their users. Both tech giants are discovering that hackers were actively exploiting previously unknown vulnerabilities in their devices. These flaws, known as zero-day bugs, were being used in real-world attacks, prompting both tech giants to act swiftly to protect users. Google Releases Security Bugs In Chrome Browser Due to the serious nature of the threat, Google quickly released fixes for several security flaws in its Chrome browser. The company confirmed that at least one of these bugs was already being used by hackers before a patch could be issued. Google later updated its disclosure to clarify that the vulnerability was discovered by Apple’s security engineering team along with Google’s Threat Analysis Group. Add Zee News as a Preferred Source This team mainly tracks government-linked hackers and commercial spyware groups, suggesting that the attack may have been carried out by state-backed actors rather than ordinary cybercriminals. Apple Security Updates In Flagship Products Apple also rolled out critical security updates for its major products, including iPhones, iPads, Macs, Vision Pro, Apple TV, Apple Watches, and the Safari browser. In its security advisory for iPhones and iPads, the company said it fixed two vulnerabilities. The Cupertino-based tech giant Apple added that it was aware these flaws may have been used in an “extremely sophisticated attack” targeting specific individuals who were using devices running versions older than iOS 26. (Also Read: Global AI Rankings: India Becomes The World’s Third Most Competitive Country In AI Power; China Grabs Second Spot, United States Ranks At…) Such attacks are often associated with government surveillance programs or spyware companies like NSO Group and Paragon Solutions, whose tools are commonly used to target journalists, activists, lawyers, and political dissidents worldwide. Notably, users across the globe have been advised to update their devices immediately, as the exploits were reportedly targeting select individuals. How To Avoid Zero-Day Bugs Keep Everything Updated: Enable automatic updates on OS, apps, browsers, plugins, and firmware to close zero-days quickly once patches drop. Use Smart Protection: Run a reputable security suite/EDR with behavior-based detection, not just classic antivirus signatures. Reduce Attack Surface: Disable macros, remove unused plugins/software, avoid risky sites, and don’t use admin accounts for daily work Practice Safe Clicks: Treat emails, links, and attachments with suspicion, verify senders/URLs, and avoid downloading from untrusted sources.