Transform in this co-op adventure!​

Digital Artist and cel-shading specialist known online as Hotty 3D continues their series of stylized beauties by presenting a fantastic NPR take on an old-school Volkswagen Transporter Type 2 from the 1950s.Featuring appealing outlines and a neat gloss effect on the windows, the model was set up using Hotty's go-to digital art software Blender, which they previously used to create a cozy Japanese bar diorama with adorable animal customers and a cool 3D model of a cat driver. Within Blender, the artist noted, they relied heavily on nodes and the Grease Pencil tool to produce the piece.Also, check out some of Hotty's earlier works:We highly encourage you to visit Hotty's Twitter and Instagram pages to see more of their fantastic projects.Also, don't forget to join our 80 Level Talent platform and our new Discord server, follow us on Instagram, Twitter, LinkedIn, Telegram, TikTok, and Threads, where we share breakdowns, the latest news, awesome artworks, and more.

Your initiation is about to begin. Commit to the family by pre-ordering Mafia: The Old Country: http://xbx.lv/4d4Z4V7

Make your laptop into a desktop with this Thunderbolt 5 dock

schwit1 shares a report from Behind The Black: Almost immediately after India's government issued this week new tightened regulations for allowing private satellite constellations to sell their services in India, it also apparently completed negotiations with SpaceX to allow it to sell Starlink in India based on these rules. Business Today reports: "According to sources, the DoT [Department of Transportation] granted the LoI [Letter of Intent] after Starlink accepted 29 strict security conditions, including requirements for real-time terminal tracking, mandatory local data processing, legal interception capabilities, and localisation of at least 20% of its ground segment infrastructure within the first few years of operation. Starlink's nod came amid heightened national security sensitivities, coinciding with India's pre-dawn Operation Sindoor strikes on terror camps across the border in response to the Pahalgam massacre. However, DoT officials clarified that the decision to approve Starlink was independent of these military developments." At the moment SpaceX's chief competitors, OneWeb and Amazon's Kuiper constellation, have not yet obtained the same permissions. This allows SpaceX to grab a large portion of the market share in India before either of these other companies. Read more of this story at Slashdot.

When you purchase through links on our site, we may earn an affiliate commission. Here’s how it works. Using generative AI at work makes co-workers question your competence, study claims David Uzondu Neowin · May 9, 2025 04:00 EDT Image by Yan Krukau via Pexels Recently Google came out with a report sharing findings from a study where customer service agents used an optional conversational AI assistant. Google found this tool could significantly boost productivity, reporting that agents using the AI saw an average 14% increase in efficiency. This gain, according to Google's calculations, could save a full-time worker approximately 122 hours per year, surpassing Google's initial estimate. The study also noted the AI had a particularly large impact on lower-performing agents, helping them handle more difficult tasks and boosting their output by 35%, compared to a more modest 7% gain for higher performers. However, these potential efficiency gains may come with a hidden social cost, according to a new study from Duke University, published recently in the Proceedings of the National Academy of Sciences. This research claims that despite AI's productivity benefits, using tools like ChatGPT, Claude, or Gemini might lead to your coworkers and managers viewing you as less competent. The research, titled "Evidence of a social evaluation penalty for using AI," involved four experiments with over 4,400 participants. Researchers Jessica A. Reif, Richard P. Larrick, and Jack B. Soll from Duke's Fuqua School of Business found a consistent pattern: employees who use AI tools tend to face negative judgments from colleagues and managers regarding their competence and motivation. In the first experiment, participants imagined using either an AI tool or a standard dashboard creation tool for a work task. Those in the AI group anticipated being seen as lazier, less competent, less diligent, and more replaceable. They also indicated they would be less willing to tell their managers or colleagues about their AI use. A supplemental study replicated these core findings, showing that participants in the AI condition expected significantly lower competence ratings from colleagues compared to those using a dashboard tool. In the table below, higher scores indicate a more positive outcome for "Disclosure," "Competence," and Diligence. On the other hand, higher scores for "Lazy" and "Replaceable" represent a more negative outcome. The second experiment seemed to confirm these anxieties. When participants evaluated descriptions of employees, those who received help from AI were consistently rated as lazier, less competent, less diligent, less independent, and less self-assured compared to those who got similar help from non-AI sources or no help at all. Another supplemental study investigated whether these social penalties changed if AI use was described as common versus uncommon in the workplace. Interestingly, the perceived norm of AI use did not significantly alter these negative social evaluations, suggesting the penalty is quite robust. The researchers also found that this bias can influence real-world business decisions. In a hiring simulation, managers who did not personally use AI frequently were less likely to hire candidates who reported regular AI tool use. Conversely, managers who were frequent AI users themselves showed a preference for AI-using candidates. This aligns with findings from another part of the study, indicating that the perception of laziness in an AI-using candidate is stronger among evaluators who themselves use AI less frequently. The final experiment identified perceived laziness as a primary driver for this negative evaluation. However, this penalty could be lessened if the AI tool was clearly beneficial and appropriate for the specific task at hand. When AI use made obvious sense for the job, the negative perceptions were significantly reduced. For example, the study detailed that for manual tasks, using AI had a negative direct effect on perceived task fit, even beyond the laziness factor. In contrast, for digital tasks where AI could be seen as more useful, AI use had a positive direct effect on perceived task fit, which helped to partially counteract the negative impact of perceived laziness. Tags Report a problem with article Follow @NeowinFeed

AMD's Zen 5 Ryzen 5 9600X paired with a free 500GB Patriot P400 Lite M.2 SSD for just $208 - an absolute bargain.

TL;DR: Get lifetime access to ad-free, high-quality documentary streaming with MagellanTV for $149.97 (reg. $999). If your current streaming lineup leans too heavily on reality TV or endless reboots, it might be time to try something that feels good on your brain. MagellanTV is a documentary-focused streaming platform with over 4,000 titles across history, science, nature, space, crime, and more. And right now now, you can get lifetime access for $149.97 (reg. $999). MagellanTV has new content added every week and no ads ever. You can stream on up to five devices at once, including smart TVs, mobile apps, and streaming boxes like Roku, Fire TV, or Apple TV. It also supports casting, so you can start a show on your phone and send it to the big screen without missing a beat. This platform doesn’t just skim the surface. It has curated playlists and deep dives into specific people, events, and discoveries, covering topics like ancient civilizations, military history, environmental science, and true crime. If you’re looking for a break from binge-watching the same old sitcoms, it’s a solid way to learn something new. Get a lifetime of documentaries with a lifetime subscription to MagellanTV on sale for $149.97. MagellanTV Documentary Streaming Service: Lifetime SubscriptionSee Deal StackSocial prices subject to change.

The new memory improvements are now fully rolled out to Plus and Pro users in the EEA, UK, Switzerland, Norway, Iceland, and Liechtenstein.OpenAI: Starting today, memory in ChatGPT can now reference all of your past chats to provide more personalized responses, drawing on your preferences and interests to make it even more helpful for writing, getting advice, learning, and beyond.

Reading Time: 6 minutes You invest heavily in driving traffic, but if visitors aren’t converting – signing up, making purchases, or engaging deeply – you might be missing out on significant growth. The culprit is the generic, one-size-fits-all experience that fails to resonate. You did everything right, maybe even worked with a CRO expert, but did not achieve personalizing your website for every visitor.  That is where Website/App Personalization and Experimentation Software come in useful. This category of tools includes well-known names like Dynamic Yield, Optimizely, VWO, Adobe Target, and Salesforce Marketing Cloud Personalization. Our own product, MoEngage Personalize, also competes in the same category and is gaining traction quickly with our customers.  These tools empower marketers to craft tailored experiences that dramatically improve engagement and conversions. Early platforms like Optimizely and Adobe Target focused on A/B testing and basic segmentation during the initial days, but that was over a decade ago. Over time, a lot of tools started to leverage AI, real-time data, predictive modelling and omnichannel integration to deliver dynamic experiences. So let’s dive into what these tools do, why they’re vital for CRO, and see some real-world examples. What are Personalization & Experimentation Platforms? To describe them simply, these powerful CRO tools are designed to transform your digital properties from static pages into dynamic, responsive environments. They generally have capabilities to: Understand User Behavior in Real Time: Go beyond basic analytics. Track clicks, scrolls, hovers, time-on-page, navigation paths, and more to see precisely how users interact with your site or app while they are on it. Implement Sophisticated Audience Segmentation: Group visitors based on a multitude of criteria – demographics, geotargeting, device type, traffic source, past purchase history, real-time behavior (like pages viewed or cart contents), and imported CRM data. Run Data-Driven Experiments (A/B Testing & More): Scientifically test variations of nearly any element on the page. This can include A/B testing headlines, calls-to-action (CTAs), images, layouts, forms, or even entire user flows using multivariate testing to pinpoint which variations really drive results. A lot of modern tools use AI and predictions to automatically distribute traffic based on results. Deliver Dynamic, Personalized Content: Automatically adjust website or app content in real-time. Show targeted banners, serve relevant product recommendations, customize offers, personalize text, and even alter navigation based on the specific user segment or individual behavior. Think of these platforms as your CRO command center, enabling you to hypothesize, test, learn, and adapt your digital experience continuously. How Personalization & A/B Testing Tools Drive CRO Success Now let’s talk about how all this fits into your CRO strategy. Conversion Rate Optimization (CRO) is the systematic process of increasing the percentage of visitors who take a desired action. Personalization and experimentation platforms are fundamental to achieving and executing these strategies. Here’s what can be achieved: Reduce Bounce Rates: By tailoring content and offers, you meet users where they are in their journey, instantly making your site feel more relevant and valuable. This is key for reducing bounce rates. Higher Conversions: Personalized pathways guide users quickly towards the final conversion goals, removing confusing steps or providing critical information proactively. A smoother journey equals higher conversion likelihood. Improve Loyalty: Relevant and easy-to-navigate journeys lead to greater user satisfaction, building trust and encouraging repeat visits and loyalty. Eliminate Guesswork: Marketers can stop relying on intuition for content and designs. A/B testing tools provide concrete, statistical evidence about what design, copy, or flow performs best.  In summary, leveraging these tools for CRO translates directly into tangible business results: more leads, higher sales, increased engagement, and better ROI on your acquisition spend. Real-World Examples: Personalization & Experimentation Use Cases That was all a little bit theoretical, so let’s also look at some practical examples. The following examples are in increasing order of complexity and businesses can try them out depending on the maturity and readiness of the organization: Example 1. Dynamic Content & Offers (Simple Banner Replacements): Business Scenario: An e-commerce site selling apparel wants to increase revenue. Application: Instead of showing everyone the same homepage banner, the platform identifies user segments. New visitors will see a “Welcome – 10% off your first order” banner. Returning customers who previously browsed women’s dresses will see a banner showcasing the latest dress arrivals or a special promotion on that category. Visitors from colder climates could see winter wear prioritized. And so on… CRO Impact: Increased banner click-through rates, higher engagement with relevant categories, and ultimately, more purchases. Example 2. A/B Testing for Improving Conversions (Simple Experimentation): Scenario: An e-learning platform wants to boost free trial sign-ups from their homepage. Hypothesis: Changing the main CTA button text from “Start Free Trial” to “Start Learning for Free” might resonate better. Action: Split homepage traffic where 50% see the original button (Control), 50% see the new variation (Variant). CRO Impact: The tool tracks clicks on each button and subsequent trial sign-ups. If “Start Learning for Free” yields a statistically significant (>18%) uplift in sign-ups, it’s declared the winner. Learning: The benefit-oriented language performed better. The change is rolled out to 100% of traffic. (Platforms like VWO and Optimizely frequently feature case studies centered around CTA and headline testing.) Example 3: Behavioral Targeting to Activate Subscriptions (Simple Behavior Based Offers) Scenario: An online publication wants to convert anonymous readers into subscribers. Application: The platform identifies users who have visited more than 5 articles on the website in the past 1 month but haven’t subscribed. Display a targeted pop-up or banner offering a special introductory subscription rate, emphasizing the value they are already getting. CRO Impact: Higher subscription rates, recovery of potentially lost sales, and tailored nurturing based on user engagement levels. Example 4: Geofencing based App & Web personalization (Complex Online-Offline Journey using Location Data) Scenario: A Quick Service Restaurant Chain wants to drive traffic into their outlets Action: A user who has previously opted-in for location sharing via your app enters a 1-mile geofence around your downtown retail store. They get a push notification to see the closest outlets and personalized offers in the app. Or for an anonymous user, detect they are on a mobile device and broadly locate them in the city using IP-based geo-targeting (less accurate). On the website, show a generic banner for “nearby stores.” Result: Users immediately see options relevant to their specific area, increasing the likelihood of browsing menus and placing an order in a nearby outlet compared to seeing generic and irrelevant menus or offers. Example 5: Dynamic Product Recommendations Based on Behavior & Attributes to Boost Conversions or AOV (Product Recommendations and 1:1 Personalization) Scenario: An online electronics retailer wants to increase the average order value (AOV) and help users discover relevant accessories. Action: A recommendation engine tracks real-time browsing (products viewed, added to cart) and uses historical data (past purchases, loyalty status) for every single user to predict individual outcomes. Personalization Ideas: On a laptop product page: Show a “Customers Also Bought” carousel featuring compatible items like laptop bags, mice, and docking stations. On the cart page: If a user adds a camera, display a recommendation block for memory cards or tripods (“Don’t Forget These!”). Homepage for returning customers who previously bought gaming gear: Prioritize showing new gaming peripherals or special offers on games. Result: Users are intelligently guided towards relevant or complementary products, increasing the chance of adding more items to their cart and boosting AOV. (Tools like Dynamic Yield and MoEngage Personalize have in built sophisticated e-commerce recommendation capabilities) Getting Started Embarking on a personalization and experimentation journey as a large organization can be overwhelming. There are a lot of use cases that are possible and constantly building and improving journeys requires a mindset shift.  The best approach is to start small. Identify a key conversion goal and a specific user segment. Formulate a hypothesis about what might improve the experience for that segment and design a simple A/B test. Or, implement a basic personalization rule, like showing a targeted banner to returning visitors. Some complex use cases like geofencing and product recommendations can come later once the muscle for building personalized journeys has been established within the organization. And at a foundational level, the key is to begin gathering data, testing your ideas, and learning what resonates with your specific audience. Over time, these incremental improvements, powered by the right tools, can lead to significant gains in conversion rates and overall business growth.  Talk to a personalization expert. Disclaimer: This blog post discusses the general category of Personalization and Experimentation SaaS tools. Mention of specific company names is for illustrative purposes only and does not constitute an endorsement. The post Your Guide to Top CRO Tools & Strategies (Using Website/App Personalization & Experimentation) appeared first on MoEngage.

We found the version of Office that doesn’t charge you up the wazoo Credit: InterPlein Deal pricing and availability subject to change after time of publication. TL;DR: Grab Microsoft Office 2024 Home and Business for PC or Mac for just £120.54 Opens in a new window Credit: InterPlein Microsoft Office 2024 Home & Business for Mac or PC Lifetime License £120.54 £188.37 Save £67.83 You wouldn’t keep paying for Netflix if you could own your favorite shows, right? So why are you still subscribing to Office apps you use every day? Microsoft 365’s price keeps going up, but there’s finally a way to break free — and it'll cost you way less in the long run.Microsoft Office 2024 is the answer you've been looking for. Instead of monthly payments, simply pay £120.54 once and be set for life (reg. £188.37). It's that simple. And, yes, this lifetime download works for PC or Mac.What's included?This license comes with: WordExcelPowerPointOutlookOneNote The newest version of Microsoft Office is a little different from Microsoft 365. But just because you're switching to a lifetime license doesn't mean you'll miss out on some of the most recent updates. Word and Excel both still have AI integrations for text suggestions and smart data analysis, and PowerPoint still has improved tools for recorded presentations. Once you've redeemed your purchase, you can install your apps on one computer. After that, they're yours to use however you want. No more subscription fees or sudden price hikes to worry about. Why rent when you can own?  Mashable Deals Want more hand-picked deals from our shopping experts? Sign up for the Mashable Deals newsletter. By clicking Sign Me Up, you confirm you are 16+ and agree to our Terms of Use and Privacy Policy. Thanks for signing up! Get a Microsoft Office lifetime license on sale for £120.54 with no coupon needed.StackSocial prices subject to change. Topics Microsoft

Économie ChatGPT Fidji Simo, une Française de 39 ans dans la direction d’OpenAI, l’entreprise d’intelligence artificielle La Sétoise, ancienne de Meta et d’Instacart, n’a pas officiellement le poste de numéro deux. Elle épaulera néanmoins le PDG, Sam Altman, auquel « elle reportera directement ». Article réservé aux abonnés Fidji Simo, lors du Sommet pour l’action sur l’intelligence artificielle, au Grand Palais, à Paris, le 11 février 2025. JOEL SAGET/AFP Une Française de 39 ans dans la direction d’OpenAI, la start-up d’intelligence artificielle (IA) la plus en vue du monde. Le parcours de Fidji Simo a de quoi frapper les esprits. La Sétoise a été nommée, jeudi 8 mai, par le PDG Sam Altman, « CEO of applications » – soit « directrice générale des applications » – de l’entreprise créatrice du célèbre assistant conversationnel ChatGPT. Elle prendra ses fonctions dans quelques semaines, le temps de lui trouver un remplaçant à son poste actuel de PDG de la start-up d’e-commerce Instacart. Lire aussi | Article réservé à nos abonnés Pour OpenAI, une levée de fonds de 40 milliards de dollars, mais des défis Si elle n’a pas officiellement le poste de numéro deux, Fidji Simo épaulera Sam Altman, auquel « elle reportera directement ». Selon ce dernier, elle supervisera « les fonctions “traditionnelles” de l’entreprise », c’est-à-dire « les équipes opérationnelles et commerciales responsables de rendre accessibles notre recherche en IA au monde entier, d’une façon bénéfique ». Elle s’occupera de l’assistant ChatGPT aux 500 millions d’utilisateurs hebdomadaires, de sa diffusion sur des plateformes tierces comme les iPhone, de la vente des modèles d’IA aux entreprises, en direct et par le biais des hébergeurs en ligne dans le cloud, etc. Elle dirigera le chef des opérations, Brad Lightcap, la directrice financière, Sarah Friar, et le responsable des produits, Kevin Weil, selon l’agence Bloomberg. Il vous reste 61.15% de cet article à lire. La suite est réservée aux abonnés.

Wandercraft, start-up française mettant au point des exosquelettes pour les personnes en situation de handicap, s'appuie désormais sur...

Introduction Whether you’re preparing for interviews or building Machine Learning systems at your job, model compression has become a must-have skill. In the era of LLMs, where models are getting larger and larger, the challenges around compressing these models to make them more efficient, smaller, and usable on lightweight machines have never been more relevant. In this article, I will go through four fundamental compression techniques that every ML practitioner should understand and master. I explore pruning, quantization, low-rank factorization, and Knowledge Distillation, each offering unique advantages. I will also add some minimal PyTorch code samples for each of these methods. I hope you enjoy the article! Feel free to connect on LinkedIn! Follow me on GitHub and visit my website for more content. Model pruning Pruning is probably the most intuitive compression technique. The idea is very simple: remove some of the weights of the network, either randomly or remove the “less important” ones. Of course, when we talk about “removing” weights in the context of neural networks, it means setting the weights to zero. Model pruning (Image by the author and ChatGPT | Inspiration: [3]) Structured vs unstructured pruning Let’s start with a simple heuristic: removing weights smaller than a threshold. \[ w’_{ij} = \begin{cases} w_{ij} & \text{if } |w_{ij}| \ge \theta_0 \\0 & \text{if } |w_{ij}| < \theta_0\end{cases} \] Of course, this is not ideal because we would need to find a way to find the right threshold for our problem! A more practical approach is to remove a specified proportion of weights with the smallest magnitudes (norm) within one layer. There are 2 common ways of implementing pruning in one layer: Structured pruning: remove entire components of the network (e.g. a random row from the weight tensor, or a random channel in a convulational layer) Unstructured pruning: remove individual weights regardless of their positions and of the structure of the tensor We can also use global pruning with either of the two above methods. This will remove the chosen proportion of weights across multiple layers, and potentially have different removal rates depending on the number of parameters in each layer. PyTorch makes this pretty straightforward (by the way, you can find all code snippets in my GitHub repo). import torch.nn.utils.prune as prune # 1. Random unstructured pruning (20% of weights at random) prune.random_unstructured(model.layer, name="weight", amount=0.2) # 2. L1‑norm unstructured pruning (20% of smallest weights) prune.l1_unstructured(model.layer, name="weight", amount=0.2) # 3. Global unstructured pruning (40% of all weights by L1 norm across layers) prune.global_unstructured( [(model.layer1, "weight"), (model.layer2, "weight")], pruning_method=prune.L1Unstructured, amount=0.4 ) # 4. Structured pruning (remove 30% of rows with lowest L2 norm) prune.ln_structured(model.layer, name="weight", amount=0.3, n=2, dim=0) Note: if you have taken statistics classes, you probably learned regularization-induced methods that also implicitly prune some weights during training, by using L0 or L1 norm regularization. Pruning differs from that because it is applied as a post-Model Compression technique Why does pruning work? The Lottery Ticket Hypothesis Image generated by ChatGPT I would like to conclude that section with a quick mention of the Lottery Ticket Hypothesis, which is both an application of pruning and an interesting explanation of how removing weights can often improve a model. I recommend reading the associated paper ([7]) for more details. Authors use the following procedure: Train the full model to convergence Prune the smallest-magnitude weights (say 10%) Reset the remaining weights to their original initialization values Retrain this pruned network Repeat the process multiple times After doing this 30 times, you end up with only 0.930 ~ 4% of the original parameters. And surprisingly, this network can do as well as the original one. This suggests that there is important parameter redundancy. In other words, there exists a sub-network (“a lottery ticket”) that actually does most of the work! Pruning is one way to unveil this sub-network. I recommend this very good video that covers the topic! Quantization While pruning focuses on removing parameters entirely, Quantization takes a different approach: reducing the precision of each parameter. Remember that every number in a computer is stored as a sequence of bits. A float32 value uses 32 bits (see example picture below), whereas an 8-bit integer (int8) uses just 8 bits. An example of how float32 numbers are represented with 32 bits (Image by the author and ChatGPT | Inspiration: [2]) Most deep learning models are trained using 32-bit floating-point numbers (FP32). Quantization converts these high-precision values to lower-precision formats like 16-bit floating-point (FP16), 8-bit integers (INT8), or even 4-bit representations. The savings here are obvious: INT8 requires 75% less memory than FP32. But how do we actually perform this conversion without destroying our model’s performance? The math behind quantization To convert from floating-point to integer representation, we need to map the continuous range of values to a discrete set of integers. For INT8 quantization, we’re mapping to 256 possible values (from -128 to 127). Suppose our weights are normalized between -1.0 and 1.0 (common in deep learning): \[ \text{scale} = \frac{\text{float_max} – \text{float_min}}{\text{int8_max} – \text{int8_min}} = \frac{1.0 – (-1.0)}{127 – (-128)} = \frac{2.0}{255} \] Then, the quantized value is given by \[\text{quantized_value} = \text{round}(\frac{\text{original_value}}{\text{scale}} \] + \text{zero_point}) Here, zero_point=0 because we want 0 to be mapped to 0. We can then round this value to the nearest integer to get integers between -127 and 128. And, you guessed it: to get integers back to float, we can use the inverse operation: \[\text{float_value} = \text{integer_value} \times \text{scale} – \text{zero_point} \] Note: in practice, the scaling factor is determined based on the range values we quantize. How to apply quantization? Quantization can be applied at different stages and with different strategies. Here are a few techniques worth knowing about: (below, the word “activation” refers to the output values of each layer) Post-training quantization (PTQ): Static Quantization: quantize both weights and activations offline (after training and before inference) Dynamic Quantization: quantize weights offline, but activations on-the-fly during inference. This is different from offline quantization because the scaling factor is determined based on the values seen so far during inference. Quantize-aware training (QAT): simulate quantization during training by rounding values, but calculations are still done with floating-point numbers. This makes the model learn weights that are more robust to quantization, which will be applied after training. Under the hood, the idea is to add “fake” operations: x -> dequantize(quantize(x)): this new value is close to x, but it still helps the model tolerate the 8-bit rounding and clipping noise. import torch.quantization as tq # 1. Post‑training static quantization (weights + activations offline) model.eval() model.qconfig = tq.get_default_qconfig('fbgemm') # assign a static quantization config tq.prepare(model, inplace=True) # we need to use a calibration dataset to determine the ranges of values with torch.no_grad(): for data, _ in calibration_data: model(data) tq.convert(model, inplace=True) # convert to a fully int8 model # 2. Post‑training dynamic quantization (weights offline, activations on‑the‑fly) dynamic_model = tq.quantize_dynamic( model, {torch.nn.Linear, torch.nn.LSTM}, # layers to quantize dtype=torch.qint8 ) # 3. Quantization‑Aware Training (QAT) model.train() model.qconfig = tq.get_default_qat_qconfig('fbgemm') # set up QAT config tq.prepare_qat(model, inplace=True) # insert fake‑quant modules # [here, train or fine‑tune the model as usual] qat_model = tq.convert(model.eval(), inplace=False) # convert to real int8 after QAT Quantization is very flexible! You can apply different precision levels to different parts of the model. For instance, you might quantize most linear layers to 8-bit for maximum speed and memory savings, while leaving critical components (e.g. attention heads, or batch-norm layers) at 16-bit or full-precision. Low-Rank Factorization Now let’s talk about low-rank factorization — a method that has been popularized with the rise of LLMs. The key observation: many weight matrices in neural networks have effective ranks much lower than their dimensions suggest. In plain English, that means there is a lot of redundancy in the parameters. Note: if you have ever used PCA for dimensionality reduction, you have already encountered a form of low-rank approximation. PCA decomposes large matrices into products of smaller, lower-rank factors that retain as much information as possible. The linear algebra behind low-rank factorization Take a weight matrix W. Every real matrix can be represented using a Singular Value Decomposition (SVD): \[ W = U\Sigma V^T \] where Σ is a diagonal matrix with singular values in non-increasing order. The number of positive coefficients actually corresponds to the rank of the matrix W. SVD visualized for a matrix of rank r (Image by the author and ChatGPT | Inspiration: [5]) To approximate W with a matrix of rank k < r, we can select the k greatest elements of sigma, and the corresponding first k columns and first k rows of U and V respectively: \[ \begin{aligned} W_k &= U_k\,\Sigma_k\,V_k^T \\[6pt] &= \underbrace{U_k\,\Sigma_k^{1/2}}_{A\in\mathbb{R}^{m\times k}} \underbrace{\Sigma_k^{1/2}\,V_k^T}_{B\in\mathbb{R}^{k\times n}}. \end{aligned} \] See how the new matrix can be decomposed as the product of A and B, with the total number of parameters now being m * k + k * n = k*(m+n) instead of m*n! This is a huge improvement, especially when k is much smaller than m and n. In practice, it’s equivalent to replacing a linear layer x → Wx with 2 consecutive ones: x → A(Bx). In PyTorch We can either apply low-rank factorization before training (parameterizing each linear layer as two smaller matrices – not really a compression method, but a design choice) or after training (applying a truncated SVD on weight matrices). The second approach is by far the most common one and is implemented below. import torch # 1. Extract weight and choose rank W = model.layer.weight.data # (m, n) k = 64 # desired rank # 2. Approximate low-rank SVD U, S, V = torch.svd_lowrank(W, q=k) # U: (m, k), S: (k, k), V: (n, k) # 3. Form factors A and B A = U * S.sqrt() # [m, k] B = V.t() * S.sqrt().unsqueeze(1) # [k, n] # 4. Replace with two linear layers and insert the matrices A and B orig = model.layer model.layer = torch.nn.Sequential( torch.nn.Linear(orig.in_features, k, bias=False), torch.nn.Linear(k, orig.out_features, bias=False), ) model.layer[0].weight.data.copy_(B) model.layer[1].weight.data.copy_(A) LoRA: an application of low-rank approximation LoRA fine-tuning: W is fixed, A and B are trained (source: [1]) I think it’s crucial to mention LoRA: you have probably heard of LoRA (Low-Rank Adaptation) if you have been following LLM fine-tuning developments. Though not strictly a compression technique, LoRA has become extremely popular for efficiently adapting large language models and making fine-tuning very efficient. The idea is simple: during fine-tuning, rather than modifying the original model weights W, LoRA freezes them and learn trainable low-rank updates: $$W’ = W + \Delta W = W + AB$$ where A and B are low-rank matrices. This allows for task-specific adaptation with just a fraction of the parameters. Even better: QLoRA takes this further by combining quantization with low-rank adaptation! Again, this is a very flexible technique and can be applied at various stages. Usually, LoRA is applied only on specific layers (for example, Attention layers’ weights). Knowledge Distillation Knowledge distillation process (Image by the author and ChatGPT | Inspiration: [4]) Knowledge distillation takes a fundamentally different approach from what we have seen so far. Instead of modifying an existing model’s parameters, it transfers the “knowledge” from a large, complex model (the “teacher”) to a smaller, more efficient model (the “student”). The goal is to train the student model to mimic the behavior and replicate the performance of the teacher, often an easier task than solving the original problem from scratch. The distillation loss Let’s explain some concepts in the case of a classification problem: The teacher model is usually a large, complex model that achieves high performance on the task at hand The student model is a second, smaller model with a different architecture, but tailored to the same task Soft targets: these are the teacher’s model predictions (probabilities, and not labels!). They will be used by the student model to mimic the teacher’s behaviors. Note that we use raw predictions and not labels because they also contain information about the confidence of the predictions Temperature: in addition to the teacher’s prediction, we also use a coefficient T (called temperature) in the softmax function to extract more information from the soft targets. Increasing T softens the distribution and helps the student model give more importance to wrong predictions. In practice, it is pretty straightforward to train the student model. We combine the usual loss (standard cross-entropy loss based on hard labels) with the “distillation” loss (based on the teacher’s soft targets): $$ L_{\text{total}} = \alpha L_{\text{hard}} + (1 – \alpha) L_{\text{distill}} $$ The distillation loss is nothing but the KL divergence between the teacher and student distribution (you can see it as a measure of the distance between the 2 distributions). $$ L_{\text{distill}} = D{KL}(q_{\text{teacher}} | | q_{\text{student}}) = \sum_i q_{\text{teacher}, i} \log \left( \frac{q_{\text{teacher}, i}}{q_{\text{student}, i}} \right) $$ As for the other methods, it is possible and encouraged to adapt this framework depending on the use case: for example, one can also compare logits and activations from intermediate layers in the network between the student and teacher model, instead of only comparing the final outputs. Knowledge distillation in practice Similar to the previous techniques, there are two options: Offline distillation: the pre-trained teacher model is fixed, and a separate student model is trained to mimic it. Both models are completely separate, and the teacher’s weights remain frozen during the distillation process. Online distillation: both models are trained simultaneously, with knowledge transfer happening during the joint training process. And below, an easy way to apply offline distillation (the last code block of this article ): import torch.nn.functional as F def distillation_loss_fn(student_logits, teacher_logits, labels, temperature=2.0, alpha=0.5): # Standard Cross-Entropy loss with hard labels student_loss = F.cross_entropy(student_logits, labels) # Distillation loss with soft targets (KL Divergence) soft_teacher_probs = F.softmax(teacher_logits / temperature, dim=-1) soft_student_log_probs = F.log_softmax(student_logits / temperature, dim=-1) # kl_div expects log probabilities as input for the first argument! distill_loss = F.kl_div( soft_student_log_probs, soft_teacher_probs.detach(), # don't calculate gradients for teacher reduction='batchmean' ) * (temperature ** 2) # optional, a scaling factor # Combine losses according to formula total_loss = alpha * student_loss + (1 - alpha) * distill_loss return total_loss teacher_model.eval() student_model.train() with torch.no_grad(): teacher_logits = teacher_model(inputs) student_logits = student_model(inputs) loss = distillation_loss_fn(student_logits, teacher_logits, labels, temperature=T, alpha=alpha) loss.backward() optimizer.step() Conclusion Thanks for reading this article! In the era of LLMs, with billions or even trillions of parameters, model compression has become a fundamental concept, essential in almost every scenario to make models more efficient and easily deployable. But as we have seen, model compression isn’t just about reducing the model size – it’s about making thoughtful design decisions. Whether choosing between online and offline methods, compressing the entire network, or targeting specific layers or channels, each choice significantly impacts performance and usability. Most models now combine several of these techniques (check out this model, for instance). Beyond introducing you to the main methods, I hope this article also inspires you to experiment and develop your own creative solutions! Don’t forget to check out the GitHub repository, where you’ll find all the code snippets and a side-by-side comparison of the four compression methods discussed in this article. Feel free to connect on LinkedIn Follow me on GitHub for more content Visit my website: maximewolf.com Check out my previous articles: LLaDA: The Diffusion Model That Could Redefine Language Generation Training Large Language Models: From TRPO to GRPO References [1] Hu, E., et al.  (2021). Low-rank Adaptation of Large Language Models. arXiv preprint arXiv:2106.09685. [2] Lightning AI. Accelerating Large Language Models with Mixed Precision Techniques. Lightning AI Blog. [3] TensorFlow Blog. Pruning API in TensorFlow Model Optimization Toolkit. TensorFlow Blog, May 2019. [4] Toward AI. A Gentle Introduction to Knowledge Distillation. Towards AI, Aug 2022. [5] Ju, A. ML Algorithm: Singular Value Decomposition (SVD). LinkedIn Pulse. [6] Algorithmic Simplicity. THIS is why large language models can understand the world. YouTube, Apr 2023. [7] Frankle, J., & Carbin, M. (2019). The Lottery Ticket Hypothesis: Finding Sparse, Trainable Neural Networks. arXiv preprint arXiv:1803.03635. The post Model Compression: Make Your Machine Learning Models Lighter and Faster appeared first on Towards Data Science.

Meta is developing what’s described as “super-sensing” vision software that can recognize people by name, according to The Information. The company is reportedly working on two pairs of smart glasses for next year, internally called Aperol and Bellini, and has recently reworked its processes for assessing privacy and safety risks to release products more quickly. In addition to facial recognition, Meta’s AI software — which is activated with a “Hey Meta, start live AI” command — could eventually do things like remind you to grab your keys if it saw that you didn’t, or remind you to pick up groceries as you walk home. The super-sensing feature keeps cameras and sensors always on so the AI can keep track of everything you do. Meta reportedly is already testing live AI on already released models. However, it cuts the battery down to just 30 minutes. Meta is aiming for its upcoming glasses, and possibly a set of earphones that have cameras, to run the software for hours on a single charge.

TOH is using Microsoft’s DAX Copilot to capture physician-patient conversations and generate draft clinical notes in real time.Read More

Square Enix's Symbiogenesis Web3 game was slated to be discontinued in July 2025, but Sony said the game will now expand to Sony's Soneium blockchain.Read More

Google is rolling out a feature in its Gemini API that the company claims will make its latest AI models cheaper for third-party developers. Google calls the feature “implicit caching” and says it can deliver 75% savings on “repetitive context” passed to models via the Gemini API. It supports Google’s Gemini 2.5 Pro and 2.5 Flash models. That’s likely to be welcome news to developers as the cost of using frontier models continues to grow. Caching, a widely adopted practice in the AI industry, reuses frequently accessed or pre-computed data from models to cut down on computing requirements and cost. For example, caches can store answers to questions users often ask of a model, eliminating the need for the model to re-create answers to the same request. Google previously offered model prompt caching, but only explicit prompt caching, meaning devs had to define their highest-frequency prompts. While cost savings were supposed to be guaranteed, explicit prompt caching typically involved a lot of manual work. Some developers weren’t pleased with how Google’s explicit caching implementation worked for Gemini 2.5 Pro, which they said could cause surprisingly large API bills. Complaints reached a fever pitch in the past week, prompting the Gemini team to apologize and pledge to make changes. In contrast to explicit caching, implicit caching is automatic. Enabled by default for Gemini 2.5 models, it passes on cost savings if a Gemini API request to a model hits a cache. Techcrunch event Exhibit at TechCrunch Sessions: AI Secure your spot at TC Sessions: AI and show 1,200+ decision-makers what you’ve built — without the big spend. Available through May 9 or while tables last. Exhibit at TechCrunch Sessions: AI Secure your spot at TC Sessions: AI and show 1,200+ decision-makers what you’ve built — without the big spend. Available through May 9 or while tables last. Berkeley, CA | June 5 BOOK NOW “[W]hen you send a request to one of the Gemini 2.5 models, if the request shares a common prefix as one of previous requests, then it’s eligible for a cache hit,” explained Google in a blog post. “We will dynamically pass cost savings back to you.” The minimum prompt token count for implicit caching is 1,024 for 2.5 Flash and 2,048 for 2.5 Pro, according to Google’s developer documentation, which is not a terribly big amount, meaning it shouldn’t take much to trigger these automatic savings. Tokens are the raw bits of data models work with, with a thousand tokens equivalent to about 750 words. Given that Google’s last claims of cost savings from caching ran afoul, there are some buyer-beware areas in this new feature. For one, Google recommends that developers keep repetitive context at the beginning of requests to increase the chances of implicit cache hits. Context that might change from request to request should be appended at the end, the company says. For another, Google didn’t offer any third-party verification that the new implicit caching system would deliver the promised automatic savings. So we’ll have to see what early adopters say.

Hi there! I’m Adam Scott (@adamscott on GitHub), the Godot Engine Web team lead, live from the GodotCon Boston 2025!On May 6, during my State of Godot and the Web talk, I revealed a world premiere to the GodotCon attendees. It was something some thought impossible, and others were resigned to not seeing during their lifetime.Please note that the VOD isn’t available yet. The link to the recording will be added to the article once available. Meanwhile, you can access the slides of the presentation.The .NET team lead, Raul Santos (@raulsntos on GitHub), managed to create a prototype running C# on the Web.Click here to see the prototype!Wasn’t it impossible?We tried dotnet.js. It didn’t work.We tried NativeAOT-LLVM. It didn’t work.We (initially) tried statically linking Mono. It didn’t work.Well, Raul gave it another shot recently and managed to make it work. Here’s how he explains his accomplishment.As mentioned in my last update, the option of statically linking Mono seemed the most promising so we went ahead with that.I have opened a draft PR with the latest changes, but keep in mind it’s still a work in progress:[.NET] Add web export support #106125As a reminder, this approach still seems very brittle to me and has some limitations. The C# project must match the WASM features supported by the Godot template (this includes things like the threading model, exception handling, SIMD support, etc.). The TargetFramework version of the C# project must also match the one that was used to build the template.Additionally, since we currently don’t load any globalization data, we only support invariant mode. This is not a big problem because most users will likely rely on Godot’s localization features, so it’s not a blocker.On the last update the issue we run into was retrieving function pointers. We were able to workaround that by declaring stub C# methods on the project we use to retrieve the Mono runtime. When building the Godot templates, this ensures these methods are included in the generated table which seems to be enough to let us retrieve the function pointer at runtime.In a regular C# application built for the web platform you’d use dotnet.js which acts as a loader for the WASM file and other required assets. In Godot we have our own way of doing this, and as a result we’re currently missing a key part of the process.The Mono runtime exports some JavaScript functions but at the time of building the templates these are stubs. The dotnet.js loader takes care of replacing these stubs with their real implementation in dotnet.runtime.js. We’re currently missing this step so they remain stubs.This means some .NET APIs that rely on these exported JavaScript functions will not work, resulting in unexpected behavior. This includes things that must be implemented using browser APIs such as cryptography.Nonetheless, I think this is significant progress and has allowed us to build a demo running on the browser.What’s next for C#/.NET on the WebRaul has published his pull request as a draft so far. It means that there’s a lot of work and testing to do.We’re working as fast as we can to enable you to export C#/.NET to the Web, but we cannot commit to a specific timeline yet. If everything works great, it should be available in the next Godot release.What about file size? My .pck size has exploded!Currently, we do have to include some .dll files in your main .pck file in order to be able to run a C# project. For the simple prototype, the .pck file size is at an astounding 72 MiB. Fortunately, when Brotli-compressed, the file can be reduced to 23.8 MiB.It’s not perfect, as we often suggest targeting the smallest size when deploying to the Web. But the following announcement should help at least somewhat concerning this issue.Introducing built-in (pre)compressionI also revealed my upcoming built-in (pre)compression PR to the public.It will do the following:add an option to compress template files when compiling the engineadd export options to compress exported filesif the target server doesn’t support serving pre-compiled files, the preloading script will take over as a fallback, ensuring to download compressed files instead of non-compressed ones in order to save bandwidth.The future is brightC#/.NET support on the Web and file compression are two important features that will come to the platform. These features will push the limits of what can be done by Godot users on the Web. I do hope that you are as excited as we are!

To our community:I’m Marc Whitten, and I lead Unity Create which includes the Unity engine and editor teams.I want to start with this: I am sorry.We should have spoken with more of you and we should have incorporated more of your feedback before announcing our new Runtime Fee policy. Our goal with this policy is to ensure we can continue to support you today and tomorrow, and keep deeply investing in our game engine.You are what makes Unity great, and we know we need to listen, and work hard to earn your trust. We have heard your concerns, and we are making changes in the policy we announced to address them.Our Unity Personal plan will remain free and there will be no Runtime Fee for games built on Unity Personal. We will be increasing the cap from $100,000 to $200,000 and we will remove the requirement to use the Made with Unity splash screen.No game with less than $1 million in trailing 12-month revenue will be subject to the fee.For those creators on Unity Pro and Unity Enterprise, we are also making changes based on your feedback.The Runtime Fee policy will only apply beginning with the next LTS version of Unity shipping in 2024 and beyond. Your games that are currently shipped and the projects you are currently working on will not be included – unless you choose to upgrade them to this new version of Unity.We will make sure that you can stay on the terms applicable for the version of Unity editor you are using – as long as you keep using that version.For games that are subject to the runtime fee, we are giving you a choice of either a 2.5% revenue share or the calculated amount based on the number of new people engaging with your game each month. Both of these numbers are self-reported from data you already have available. You will always be billed the lesser amount.We want to continue to build the best engine for creators. We truly love this industry and you are the reason why.I’d like to invite you to join me for a live fireside chat hosted by Jason Weimann today at 4:00 pm ET/1:00 pm PT, where I will do my best to answer your questions. In the meantime, here are some more details.*Thank you for caring as deeply as you do, and thank you for giving us hard feedback.Marc Whitten*We are working to localize translations of this web page.

Usually in an RPG, when you think “absolute assbeating swordwielder” in a video game, you picture the likes of Cecil from Final Fantasy IV, Auron from FFX, or even someone like Hulkenberg from Metaphor ReFantazio. What you’re usually not thinking of is a tiny befreckled orphan girl who wins the Mia Goth soundalike contest.Suggested ReadingThis Innovative Hades-Like Feels Built For Speedrunners Share SubtitlesOffEnglishview videoSuggested ReadingThis Innovative Hades-Like Feels Built For Speedrunners Share SubtitlesOffEnglishBut don’t get it twisted, once the rapier comes out, everything that ain’t part of the Expedition is about to get their life expectancy cut real short. Maelle is Clair Obscur’s indisputable five tool player, and we’re gonna show you how to start running shit the second she’s in your party.Screenshot: Sandfall Interactive / Justin Clark / KotakuMaelle’s weapon of choice is a fencing rapier; and you best believe she knows how to use it. Actually, so should you, since nearly all of her additional moves are, basically, fencing strategies disguised as stat changes. With only a couple of late game exceptions, and a couple of stray Fire abilities we’ll get into, the vast majority of her obtainable skills are about doing either heavy Physical damage, or applying some diabolical shit to enemies in order to do heavy Physical damage. Even if she’s not the one to do the damage, she’s got the skills to set someone else up for success. “Oh ho ho ho,” you might say, “but those are all 1-3 hits and Fire skills. She still has her work cut out dealing with shields.” Au fucking contraire, mon ami, she’s got a skill called Breaking Rules that’s not even terribly far into her skill tree that literally nullifies shields, no matter how many the enemy’s got slapped on. You know that one Nevron with the floating orbs, and shooting one puts six shields on everybody else on the field? Light work. The game should literally come up with a prompt that asks “Are you sure you want to make the biggest mistake of your stupid life” if you ever try to remove Maelle from your party? Screenshot: Sandfall Interactive / Claire Jackson / KotakuBut of course, there’s the real reason you’re here, which is to figure out how to effectively use the Stance system. And the key word is “effectively,” because it’s extremely easy to get caught out there in Offense Mode, wondering why Maelle just got merked by a leaf hitting her face in a stiff breeze. To be fair, the game explains her mechanic way too quick, so let’s go over bascis again.. Every one of Maelle’s abilities puts her into one of four Stances each with their own perks/drawbacks:Stanceless: Where you start in every battle. No effects. Defensive Stance: Take less damage; gain 1 extra AP on every successful dodge or parry.Offensive Stance: Deal 50% more damage, but also take 50% more damage.Virtuose Stance: Deal 200% damage.Now, while Maelle’s skill set can absolutely slay a field of enemies, you still have to think of her in fencing terms. Overextension can get you stabbed, while being a turtle on constant defense doesn’t get things done. Meanwhile, the game actually incentivizes you to keep moving by giving you an additional AP every time you change stances. So, figuring out what to put on her to give you the least vulnerability while still being as effective as she needs to be is the trick of it, but there’s a lot of flexibility therein. The dream, of course, is to get in Virtuose Mode, where Maelle is become death, destroyer of worlds. But even then, you only get two turns to do the most damage with that, and your team may have other needs, like with one of her skills that grants Shell, or if, say, they’re all dead, and you need to use Phoenix Flame to rez them all. For much of the game, the only elemental damage Maelle will have access to is all Fire-based. You get Void damage way down the line. You’ll figure out how that plays out on your own by then. And even then, Fire is going to be enough, especially since the easiest way to get Maelle into Virtuose Mode most of the time is Burning an enemy and hitting them with Swift Stride. Yes, that means go get that off your Skill Board ASAP; it’s gonna be a mainstay through much of the first half of the game. So, naturally, giving Maelle or her companions Fire abilities to back her up and take advantage of the constant Burn is gonna be an imperative. But, even if you decide otherwise, Maelle’s got plenty of her own. She’ll start with Spark which will get the job done, just know that you’ll have to blow a turn to cast it and it’s already gonna cost you another one to cast Swift Stride. Screw that, let somebody else get that ball rolling, least till you pick up the far more devastating Rain of Fire skill for her later. Both are effective, though worth noting that while Spark will put Maelle in Defensive Mode, Rain of Fire ends with her in Offensive. Be prepared for whatever’s coming next. Once she’s Virtuosed out, though, you’ll need to capitalize as best you can, and in the early going, the Critical Burn Picto/Lumina (Critical Hit chance goes up 25% on Burning enemies) is a pretty easy place to start. It’d be nice if one of your companions can slap a Mark or Defenseless on your prey, but the irony there is that if the enemy is Burning already, that first hit will remove the status. Be mindful of the play order on the left of the battle screen so you can line that up just right and have Maelle go right after an enemy’s been afflicted. Percee and Fleuret Fury are usually a death sentence for enemies after that, and since she’s likely gonna be the last thing a lot of lesser enemies see, Maelle can get a lot of use out of the Dead Energy Picto (3 AP after killing an enemy). But in the event those two skills don’t put a Nevron out of its misery, Fleuret Fury will at least re-up on Virtuose. The main issue here is that Fleuret Fury’s expensive. Keeping Maelle topped off on AP is crucial but, especially for a character like Lune who’s less useful in a fight where the elements aren’t as effective, it’s worthwhile having someone whose primary job it is to keep Maelle juiced up if her Pictos haven’t gotten to kick in.Once she’s tapped out, though, your new job is making sure Maelle can build up to a full bar of AP again. And early on, that’s not as easy as it sounds. A Picto like Energising Pain (1 AP after taking damage) can help, but this is the exact situation Defensive Mode is good for, making sure she can take a hit and likely survive longer than her friends.Nail these principles and our little fencer will be an absolute menace against anything that threatens the Expedition.Clair Obscur: Expedition 33 is available now on PS5, Xbox Series X/S, and Windows PCs.

Senior Animator – Temp (SFD / Cinematics)Blizzard EntertainmentVancouver British Columbia V6A 4C4 Canada1 hour agoApplyJob Title:Senior Animator – Temp (SFD / Cinematics)Requisition ID:R025208Job Description:Blizzard Entertainment is looking for a dedicated 3D animator to assist our Cinematics team in crafting the next generation of Epic animated films.What You Bring to the TablePriorities can often change in a fast-paced environment like ours, so this role includes, but is not limited to, the following responsibilities:The cinematic animator will be skilled at animating characters at a level that lives up to the ever-increasing Blizzard Entertainment standard of quality. The ideal candidate has a passion for animation and film making, with experience animating a diverse array of characters and creatures in different styles, from cartoony to realistic and anywhere in-between. A solid grasp of body mechanics and character acting with a mastery of the 12 principles of animation are critical. You must enjoy working on a highly collaborative team of peers as passionate about making phenomenal movies and games as we are. Blizzard Entertainment offers a fun, creative, and technically challenging environment with excellent compensation.This role is anticipated to be a hybrid work position, with some work on-site and some work-from-home. The potential home studio for this role is Irvine, CA. However, we understand each person’s circumstances may be unique and will work with you to explore other possible options, including remote work arrangements.We have multiple temporary roles open with 5-month and 6-month terms. Start dates will be the beginning of June 2025.Minimum RequirementsExperience6+ years' experience in game development, television, or film industry using Maya or equivalent 3D programExperience working with Directors, Animation Supervisors and Leads developing, defining, and maintaining high-end cinematic animationsKnowledge & SkillsA demo reel demonstrating high quality hand-keyed animations through solid body mechanics and engaging performancesA solid grasp of body mechanics and character acting with a mastery of the 12 principles of animation are essentialKey AttributesPassion for film, animation, and video gamesSelf-motivation, good communication skills, and a great team-player attitudeAssist in supporting and evolving the Cinematics team with constructive feedback both technically and artisticallyExtra PointsArtistic skills: storyboarding, gesture drawing, 2D animation, compositionTechnical skills: rigging, scripting, and understanding of pipelinePhysical skills: acting, martial arts, and choreographyFilm making skills: cinematography, pre-visualization, editing, and story writingUnderstanding of FX, particles, cloth, and hair as they contribute to a character’s performance and other facets of both 2D and 3D animation is essentialYour PlatformBest known for iconic video game universes including Warcraft®, Overwatch®, Diablo®, and StarCraft®, Blizzard Entertainment, Inc. (www.blizzard.com), a division of Activision Blizzard, which was acquired by Microsoft (NASDAQ: MSFT), is a premier developer and publisher of entertainment experiences. Blizzard Entertainment has created some of the industry’s most critically acclaimed and genre-defining games over the last 30 years, with a track record that includes multiple Game of the Year awards. Blizzard Entertainment engages tens of millions of players around the world with titles available on PC via Battle.net®, Xbox, PlayStation, Nintendo Switch, iOS, and Android.Our WorldActivision Blizzard, Inc., is one of the world's largest and most successful interactive entertainment companies and is at the intersection of media, technology and entertainment. We are home to some of the most beloved entertainment franchises including Call of Duty ®, World of Warcraft® , Overwatch ®, Diablo ®, Candy Crush ™ and Bubble Witch ™. Our combined entertainment network delights hundreds of millions of monthly active users in 196 countries, making us the largest gaming network on the planet!Our ability to build immersive and innovate worlds is only enhanced by diverse teams working in an inclusive environment. We aspire to have a culture where everyone can thrive in order to connect and engage the world through epic entertainment. We provide a suite of benefits that promote physical, emotional and financial well-being for ‘Every World’ - we’ve got our employees covered!The videogame industry and therefore our business is fast-paced and will continue to evolve. As such, the duties and responsibilities of this role may be changed as directed by the Company at any time to promote and support our business and relationships with industry partners.We love hearing from anyone who is enthusiastic about changing the games industry. Not sure you meet all qualifications? Let us decide! Research shows that women and members of other under-represented groups tend to not apply to jobs when they think they may not meet every qualification, when, in fact, they often do! We are committed to creating a diverse and inclusive environment and strongly encourage you to apply.We are committed to working with and providing reasonable assistance to individuals with physical and mental disabilities. If you are a disabled individual requiring an accommodation to apply for an open position, please email your request to accommodationrequests@activisionblizzard.com. General employment questions cannot be accepted or processed here. Thank you for your interest.We are an equal opportunity employer and value diversity at our company. We do not discriminate on the basis of race, religion, color, national origin, gender, sexual orientation, gender identity, age, marital status, veteran status, or disability status, among other characteristics.RewardsWe provide a suite of benefits that promote physical, emotional and financial well-being for 'Every World' - we've got our employees covered! Subject to eligibility requirements, the Company offers comprehensive benefits including:Extended health plan, including medical, dental, vision, life and AD&D insurance, for employees and their dependents, as well as disability insurance;RRSP with Company match, tuition reimbursement, charitable donation matching;Paid holidays and vacation, paid sick time, compassion and bereavement leaves, parental leave;Mental health & wellbeing programs, fitness programs, free and discounted games, and a variety of other benefit programs like supplemental life & critical illness, legal service, and others;If the Company requires that you move geographic locations for the job, then you may also be eligible for relocation assistance.Eligibility to participate in these benefits may vary for part time and temporary full-time employees and interns with the Company.The standard base pay range for this role is $79,800.00 - $168,800.00 CAD. These values reflect the expected annualized base pay range of new hires in British Columbia, Canada. Actual amounts will vary depending on experience, performance and location. In addition to a competitive base pay, employees in this role may be eligible for incentive compensation. Incentive compensation is not guaranteed. Create Your Profile — Game companies can contact you with their relevant job openings. Apply

In a Japanese press release reported by 4Gamer, SEGA (the game's publisher in Japan) and Avalanche Software have revealed some technical details of the Nintendo Switch 2 port of Hogwarts Legacy. Enhanced Visuals Experience deeper immersion with significantly upgraded graphics, including NVIDIA DLSS support, improved lighting and shadow animation quality, and anti-aliasing. Experience up to 1440p resolution in TV mode and 1080p and HDR in handheld mode. Smooth Movement Move seamlessly through the open world, including Hogsmeade and its shops, without loading screens. Smooth Flight Thanks to the improved performance of the Nintendo Switch 2, not only is the frame rate higher than the Nintendo Switch version, but the loading times are also shorter. You can fly around the Hogwarts grounds and beyond comfortably. Supports mouse operation with Joy-Con 2 Supports mouse operation with Joy-Con 2 controller. Enjoy your preferred play style. The Nintendo Switch 2 hardware supports NVIDIA DLSS Super Resolution, but developers can choose whether to actually use it or not. It honestly seems like a no-brainer, though, especially for games that have already implemented the feature on the PC version, like Hogwarts Legacy or Cyberpunk 2077. The Switch 2 also technically supports ray tracing, but there's no mention of it being implemented in this game, even though the PlayStation 5 and Xbox Series X versions (and, of course, the PC version) support it. Indeed, it's highly unlikely that open world games will implement ray tracing in any capacity on the hybrid console, as it would simply be too taxing on the hardware. It could, however, be supported for simpler, smaller games. Anyway, the Nintendo Switch 2 version of Hogwarts Legacy will be available at the console's launch (June 5). The game size is 24 GB. For our original review of the game, go here. For Wccftech's info roundup article on the Nintendo Switch 2, click here.

Vex Member Oct 25, 2017 25,520 ...And one of them comes out next year. So that leaves one left. Are you ok??? .... GTA SA 2004 GTA 4 2008 GTAV 2013 GTAVI 2026 ...that's a 13 year difference give or take for the latest ones. Most of us are in our mid thirties. And some of y'all pushing 40 and up. Source: ERA age, how old are you? I was just curious on the overall age range of this forum and all. I'm 22, how about you guys? www.resetera.com Get rekt. Are you ok with the thought that you may not live to even see GTA7? We are getting up there. And even if you do, you may not live long after? Let's face it, some of us aren't even going to make it to the maximum age of ~72. Source: Life expectancy - Wikipedia en.wikipedia.org So don't try to argue with me. Just answer the question. Assuming the next GTA takes longer than 13 years (since each mainline GTA takes progressively longer to develop), that would mean 20+ years until the next game (GTA7) at the very least? This would put most of us at age 50+. You're done. Let me do the math for you who don't understand; 35+13 = 48. And that's being nice about it. Again, they take longer each time. Let's say... 20 years this time for gta7. 35+20= 55 for gta7. Think on that. And that's a conservative number. For example: I'm 37. So GTA7 basically represents death for YOU. GTA 7 is the grim reaper game because it represents the end of your life. It's the last GTA you will ever see. You're not going to see GTA8 20+ years later. It is YOUR DEMON DOOM GAME OF DEATH. TLDR; GTA7 = GRIM REAPER Wow, -Vex  Strafer The Flagpole is Wider Member Oct 25, 2017 30,747 Sweden Im going to freeze myself demolition man   KZXcellent One Winged Slayer Member Oct 25, 2017 4,168 I'm just trying to make it to VI personally.   Ant_17 Member Oct 28, 2017 2,870 Greece Don't fear the reaper.   BAW Member Oct 27, 2017 2,116 I wouldn't be so pessimistic, now that they have the tech in place they will be able to develop them much more quickly.   Patitoloco Member Oct 27, 2017 27,959 I'm extremely okay with this. Also, I don't plan to die at 43?  OhhEldenRing Member Aug 14, 2024 2,712 Bit of a shit framing. You know you're gonna die too someday. Signed Older fella  Killyoh Member Oct 28, 2017 1,775 Paris, France I'm 33, if it's exponential the next one should be out when I'm around 60, then the one after hmmm if I do the math... Yeah I'll be dead.   Bansai Teyvat Traveler Member Oct 28, 2017 14,048 Keep talking sheep and see what happens!   LiK Member Oct 25, 2017 34,396 GTA is just another video game.   Kyusha Member Jun 14, 2022 533 Ant_17 said: Don't fear the reaper. Click to expand... Click to shrink... hell yea   MessyLessy Member Oct 30, 2017 408 Silly of you to think the earth will be anything but a nuclear wasteland by then.   Buttonbasher Member Dec 4, 2017 5,716 Yeah, I don't think about anything this way. I don't care.   pksu Member Oct 27, 2017 1,627 Finland There are other, more interesting series already dead I'd swap places with GTA immediately   OP OP Vex Member Oct 25, 2017 25,520 BAW said: I wouldn't be so pessimistic, now that they have the tech in place they will be able to develop them much more quickly. Click to expand... Click to shrink... We said that about GTAV!!!!   Patitoloco Member Oct 27, 2017 27,959 Ant_17 said: Don't fear the reaper. Click to expand... Click to shrink... but i'm not seasons   JusDoIt ▲ Legend ▲ Member Oct 25, 2017 43,395 South Central gonna get blood transfusions from my son just to prove you wrong   Mary Celeste Member Oct 25, 2017 14,215 I'm more concerned with how few Kojima/Kamiya/Mikami games we have left   ara Member Oct 26, 2017 13,690 Sure   Chaserjoey Keeper of the White Materia Member Oct 27, 2017 9,841 That's true for anything. View: https://www.youtube.com/watch?v=N9GpgSQZ4J4&ab_channel=ThingsICantFindOtherwise Killyoh said: I'm 33, if it's exponential the next one should be out when I'm around 60, then the one after hmmm if I do the math... Yeah I'll be dead. Click to expand... Click to shrink... 33!? You should have been gommaged.   J75 Member Sep 29, 2018 7,823 Too distant and uncertain to care. I feel worst for the GTA fans that were around in 2013 but won't/didn't make it for when VI releases.   Killyoh Member Oct 28, 2017 1,775 Paris, France Chaserjoey said: 33!? You should have been gommaged. Click to expand... Click to shrink... Putain   Wallace Member Oct 25, 2017 27,865 Midwest I'd say earth has roughly 6 years before total annihilation so I'm ok with it.   Jiraiya Member Oct 27, 2017 10,846 This is the last one. And it'll be future proof up scalable. Online will eventually include multiple full scale cities from all their previous titles. It's modder scene will rival and/or surpass the most capable titles seen today. No story dlc though.  zeknurn Member Oct 25, 2017 1,483 You will be too busy fighting in the great fresh water wars to be caring about grand theft auto 7 and its sequels.   Blackpuppy Member Oct 28, 2017 5,019 Ya know, I've never really played a GTA game before apart from the 1st couple of missions in GTA: Chinatown on the DS… I think I'll survive   Shadow One Winged Slayer Member Oct 28, 2017 4,779 I went without playing GTAV until it was free on Epic (ok I played the first mission my brothers gameshare on Xbox in 2019, but that was more like a demo). I think I'll be okay!  Mekanos ▲ Legend ▲ Member Oct 17, 2018 48,405 zeknurn said: You will be too busy fighting in the great fresh water wars to be caring about grand theft auto 7 and its sequels. Click to expand... Click to shrink... Yet people will still be waiting for Half-Life 3   ClivePwned Member Oct 27, 2017 4,180 Australia I'm more of a CoD fan and so the lack of GTA games released for the rest of my life will more than be offsett by the next 20 Call of Duty games.   Javier23 Member Oct 28, 2017 3,184 I really don't mind, not to downplay the vast amounts of work and artistry behind every GTA release and how monumental of an achievement they are, but games that I find far more interesting to engage with keep being released at a neck breaking speed. I'd be far more heartbroken over needing to wait that long before a hidden gem like Blue Prince comes out of nowhere again, or replace that with any one of many other indie titles like Disco Elysium, Outer Wilds, etc etc that have unexpectedly become one of my favorite games ever in the last however many years. BAW said: I wouldn't be so pessimistic, now that they have the tech in place they will be able to develop them much more quickly. Click to expand... Click to shrink... Realistically speaking, what they're setting up with GTA 6 is a base platform for an ever expanding new GTA Online. You're not gonna get GTA 7 any faster than you got 6, and I wouldn't even put money on single-player DLC (but who knows).   RaySpencer Member Oct 27, 2017 5,710 Whatever. GTA is fun, I've been playing them since the first one, and I'm looking forward to playing 6. But if I never play one (even including 6) ever again, I don't really care. Way too many other games out there. I'll survive.  TissueBox Member Oct 25, 2017 12,394 Urinated States of America You think GTAVII is bad, imagine Elder Scrolls VII. (Reference: - 13 years between GTAV and GTAVI. - 14 years since Skyrim, and counting - On that logic, Elder Scrolls VII will be 2045+, and Elder Scrolls VIII will be 2060's)  Gavalanche Prophet of Regret Member Oct 21, 2021 25,595 There is no escape from deaths embrace, though you lead it on a merry chase.   Fevaweva Member Oct 30, 2017 7,511 I think, if we survive, society would've gone beyond Grand Theft Auto type games and we'd be in a world filled with peace and tranquility.   Dogui Member Oct 28, 2017 10,833 Brazil Really don't care about GTA, but I don't think I do this with series I care, either. The worst thing will be dying between the announcement of a game you're hyped and its release. I imagine sometimes, Capcom announces Breath of Fire VII to release in 2 years, then I get something and doctor says I have only 1 year to live. There's a lot of time until GTA VI, though. Who knows we won't all die in some nuclear winter thing in 2 months?  Dervius Member Oct 28, 2017 6,301 UK Too many unsupported assumptions in here so I reject the premise. I know it's largely meant in jest about just how long this game has taken o come out, but it also feels a bit mean spirited in a way I'm just not giving with. World's in a tough place atm, so, joke or not, having a 'you're gunna die soon ya old fucks!' thread just doesn't hit quite the same.  noffles Uncle Works at Nintendo Member Nov 10, 2017 1,219 if i started thinking about my life expectancy in terms of videogames it would be a grim day indeed however, fuck you, i feel old now  Aarglefarg Member Oct 27, 2017 4,199 I'm approaching 40 and there are 4 more GTAs on a 13 year schedule (including next year's) between my age and the life expectancy.   Gelf Member Oct 27, 2017 6,118 Look I've had some fun playing the previous games but I could very easily live without any more of them. I didn't play GTAV until many years after it's first release anyway. There's so much more to play out there. Maybe I'll get to VI in 2030 or something.  Doctor_Thomas Member Oct 27, 2017 11,424 Meh. There's other, better, games to play. GTA for me is something that is always something I enjoy to a degree but I don't get hyped or excited or whatever.  dreamfall Member Oct 25, 2017 7,297 I just gotta make it to this one! Survival is key, counting down the days!   Friendly Wizard Member Oct 31, 2017 143 It's just a game, relax.   tiebreaker "This guy are sick" Member Oct 25, 2017 14,467 What if GTA 7 never releases? Will I live forever?   The Quentulated Mox Corrupted by Vengeance Member Jun 10, 2022 6,531 you have no idea how many more cars i can steal before death takes me   JCal Member Oct 27, 2017 9,544 Los Alfheim Perfectly okay with it. What I'm not okay with is leaving this world without playing Mega Man X9 I won't... I WON'T, DAMN YOU  m_shortpants Member Oct 25, 2017 15,226 I posted this in the trailer thread. If I'm lucky, I got two more in me after 6. If they take 20 years that is. Again, assuming I still have my mental faculties and any desire to play games in my 70s, if society hasn't collapsed.  Blade30 Member Oct 26, 2017 5,943 I played a couple of them, they are alright. I'm ok with no GTA.   P-MAC Member Nov 15, 2017 5,334 I don't necessarily care about GTA specifically, but the more general version of this problem is definitely depressing. For example, there is zero possibility I will ever clear my backlog. There are games I'd love to try NOW, that I will die without ever getting the chance to. Even if I dedicated myself to work through my backlog in 100% of my free time, I don't think it would ever be possible to play every game I want to. But then again, this goes for every other experience in life too so it is what it is I guess  7thFloor Member Oct 27, 2017 7,171 U.S. TissueBox said: You think GTAVII is bad, imagine Elder Scrolls VII. (Reference: - 13 years between GTAV and GTAVI. - 14 years since Skyrim, and counting - On that logic, Elder Scrolls VII will be 2045+, and Elder Scrolls VIII will be 2060's) Click to expand... Click to shrink... No way are they going to repeat that kind of gap again   unicornKnight ▲ Legend ▲ Member Oct 27, 2017 14,641 Athens, Greece I don't think I've ever played GTA. I once downloaded V from GP but deleted before I even booted it up.  

News Death Stranding 2: On the Beach Gameplay Showcases Combat, Stealth, and New BTs Check out some of the new bosses that Sam will have to fight, including a giant mech launching cars with its octopus-like tendrils. Posted By Ravi Sinha | On 09th, May. 2025 Kojima Productions’ Death Stranding 2: On the Beach received tons of new details courtesy of various hands-on previews (which spent a hefty 30 hours with the game). However, there’s also new gameplay, highlighting several features and improvements. Check out a 15-minute showcase below, courtesy of Famitsu. The visuals are immediately striking, especially the mountain ranges, which Sam can roughly navigate (and observe phenomena like landslides). Other hazards like a duststorm and sudden water level changes must also be dealt with carefully. Then there’s the combat, with headshots triggering brief slow-motion for lining up your next shot. If you’re more keen on sneaking, non-lethal bolas guns are available. Alternatively, chuck an electric rod at an enemy, dealing damage in an area to stun them (though it will make noise and draw attention). The new BT, Watcher, is also briefly showcased, visually spotting Sam and pursuing him through the tar. There’s also a giant desiccated skull with tendrils that serves as a boss (with a weak point that’s tricky to hit) and an octopus-like mech that chucks cars at you. Death Stranding 2: On the Beach is available on June 26th for PS5. Head here to learn more about other new features or check out its limited edition DualSense. Tagged With: Elden Ring: Nightreign Publisher:Bandai Namco Developer:FromSoftware Platforms:PS5, Xbox Series X, PS4, Xbox One, PCView More DOOM: The Dark Ages Publisher:Bethesda Softworks Developer:id Software Platforms:Xbox Series X, PS5, PCView More Death Stranding 2: On the Beach Publisher:Sony Developer:Kojima Productions Platforms:PS5View More Amazing Articles You Might Want To Check Out! Frostpunk 2 Update Adds New Sandbox Map, Heating Overhaul, Serenity Mode, and More A classic map from Frostpunk 1 returns, Tales from the Frostland adds narrative quests for Utopia Builder and ... Death Stranding 2: On the Beach Gameplay Showcases Combat, Stealth, and New BTs Check out some of the new bosses that Sam will have to fight, including a giant mech launching cars with its o... 15 Brutal Shooters to Play Before DOOM: The Dark Ages Drops What are some of the best Doom-like titles to play before DOOM: The Dark Ages arrives on May 15th? Check out o... The Elder Scrolls 5: Skyrim – Should Bethesda Remaster It Like Oblivion? After the success of Oblivion Remastered, should Bethesda give its most successful Elder Scrolls title the sam... Blades of Fire – Everything You Need to Know MercurySteam's newest title arrives on May 22nd and pays homage to a cult classic. Check out everything to kno... Mafia: The Old Country’s Combat Has a “Life-or-Death” Vibe, Encourages More Tactical Play Protagonist Enzo Favara is "proving himself with guns," which means a "scarcity of resources" and the need for... View More

The Trump AI chip policy reversal announced this week signals a shift in how advanced computing technologies will flow in global markets, as the administration prepares to dismantle Biden’s complex three-tier regulatory framework.Set to take effect on May 15, 2025, the Biden administration’s Framework for Artificial Intelligence Diffusion would have created a stratified global technology landscape with significant implications for international trade, innovation, and geopolitical relationships.According to a Commerce Department on Wednesday, the Trump administration views the existing approach as fundamentally flawed. “The Biden AI rule is overly complex, overly bureaucratic, and would stymie American innovation,” a Commerce Department spokeswoman told Reuters.“We will be replacing it with a much simpler rule that freees American innovation and ensures American AI dominance,” they said. The policy shift comes months after the Biden administration finalised an export control framework during its final week in office.That framework represented the culmination of a four-year effort to restrict China’s access to cutting-edge chips while maintaining US leadership in artificial intelligence technology. The decision to rescind the rule reflects the Trump administration’s different approach to balancing national security concerns with commercial interests.<Understanding the three-tier systemThe soon-to-be-eliminated rule had established a hierarchical structure for global technology access. In the first tier, 17 countries plus Taiwan would have enjoyed unlimited access to advanced AI chips.A second tier of approximately 120 countries would have operated under strict numerical caps limiting their imports. The third and final tier – including China, Russia, Iran, and North Korea – would have been completely blocked from accessing these technologies.The structured approach aimed to prevent advanced technologies from reaching countries of concern through intermediaries while still allowing access for allies and neutral nations. However, critics argued the complexity of the system would create significant compliance burdens and push international partners toward alternative suppliers.<The new approach taking shapeInstead of the tiered system, sources cited by Reuters indicate the Trump administration is considering implementing a global licensing regime supported by inter-governmental agreements. The approach would potentially offer more flexibility and maintain controls over sensitive technology.The timing of the announcement appears strategically significant. Bloomberg reported the changes are developing as President Trump prepares for a trip to the Middle East, where countries including Saudi Arabia and the United Arab Emirates have expressed frustration over existing restrictions on their acquisition of AI chips.The Commerce Department’s decision could be announced as soon as Thursday, according to a source familiar with the matter.<Market reaction and industry impactNews of the policy reversal has already sent ripples through financial markets. Shares of Nvidia, the dominant manufacturer of chips used for training AI models, ended 3% higher on May 7 following the announcement, though they dipped 0.7% in after-hours trading, according to Reuters.The company has consistently opposed the growing number of US restrictions. Nvidia CEO Jensen Huang argues that American companies should be able to sell into China, which, he predicts, will become a $50 billion market for AI chips in the next couple of years.However, it’s important to note that the Trump AI chip policy shift does not signal a complete abandonment of export controls. The administration has already demonstrated its willingness to take strong action against China, specifically, by banning Nvidia from selling its H20 chip there – a move that cost the company $5.5 billion in writedowns, according to Bloomberg.<Global winners and losersThe policy reversal creates a complex map of potential winners and losers in the global technology landscape. Countries like India and Malaysia, which hadn’t faced chip restrictions before the Biden rule was unveiled in January, will see temporary relief. In Malaysia’s case, this could particularly benefit Oracle Corporation, which has plans for a massive data centre expansion that would have exceeded limits established by rules governing AI hardware distribution.Middle Eastern nations also stand to gain. The UAE and Saudi Arabia, which have faced chip export controls since 2023, may now be able to negotiate more favourable terms.Trump has expressed interest in easing restrictions for the UAE specifically and could announce the beginning of work on a government-to-government AI chip agreement during his upcoming visit to the region from May 13 to 16.The UAE’s aggressive pursuit of such an agreement, backed by its pledge to invest up to $1.4 trillion in US technology and infrastructure over the next decade, exemplifies how high-stakes these negotiations have become for countries seeking to establish themselves as AI powerhouses.<According to Axios, the Trump administration is currently developing a new control scheme, which could emerge as either a new rule or an executive order. The transition period creates significant uncertainty for companies like Nvidia regarding the regulatory environment they’ll face in the coming months.While the new framework takes shape, the administration has indicated it will continue enforcing existing chip export controls. One potential element of the new approach might involve imposing controls specifically on countries that have diverted chips to China, including Malaysia and Thailand, according to a source familiar with the matter.Industry stakeholders remain divided on the issue. While chip manufacturers have lobbied aggressively against strict export controls, some AI companies, including Anthropic, have advocated for maintaining protections that safeguard US intellectual property and technological advantages.<Balancing competing prioritiesThe Biden administration’s export controls were designed to limit access to chips needed for cutting-edge AI development, with a particular focus on preventing Chinese firms from finding indirect routes to technology that existing export controls prevented them from importing directly.Creating a balanced approach that addresses national security concerns while promoting US commercial interests presents significant challenges. Establishing agreements with a wide range of countries eager to purchase advanced AI chips would require navigating complex diplomatic relationships and potentially creating dozens of separate policy frameworks.The Commerce Department has not provided a specific timeline for when any new rules are to be finalised or implemented, only indicating that debate continues on the optimal approach forward.The shift in Trump AI chip policy reflects the administration’s broader emphasis on American competitiveness and innovation while still maintaining control over technologies with national security implications.As officials work to craft a replacement framework, the global AI chip market remains in flux, with profound implications for technological development, international relations, and corporate strategies in the evolving artificial intelligence landscape. also: US-China AI chip race: Cambricon’s first profit landsWant to learn more about AI and big data from industry leaders? Check out AI & Big Data Expo taking place in Amsterdam, California, and London. The comprehensive event is co-located with other leading events including Intelligent Automation Conference, BlockX, Digital Transformation Week, and Cyber Security & Cloud Expo.Explore other upcoming enterprise technology events and webinars powered by TechForge here.

Google released a new report on Thursday highlighting how it has been combating online scams across its different products using artificial intelligence (AI) technology. The report specifically focuses on how AI technology is being implemented in Search, Google Chrome, and Android. The Mountain View-based tech giant said that generative AI is enabling it to detect and stop even previously unheard-of scams. Notably, the company is using the on-device Gemini Nano AI model in the Chrome browser to deal with website-based scams.New Ways Google Is Fighting Scams Using AIIn a blog post, the tech giant announced that it has published a new Fighting Scams in Search report. The report details how Google has improved its AI-powered scam detection systems and classifiers, and claims that the new system can spot 20x the number of potential scamming websites compared to before.Website-based scams typically use fake or compromised websites to trick users into giving away money, personal information, or login credentials. While the way these scams are conducted can differ, their goal is generally to deceive users through imitation, urgency, or offers that seem too good to be true. Some examples of such scams include phishing sites, online shopping scams, investment and crypto scams, tech support scams, clone websites, and subscription traps.In Search, Google says it has improved its classifiers that detect and classify scams using AI. With these improvements, the tech giant can now de-rank these scams before any users can see them. Highlighting an example, the post mentioned that Search detected an increase in scammers impersonating airline customer service providers. The company claimed that these scams have been reduced by more than 80 percent in Search.Google Chrome has also improved the Enhanced Protection mode of Safe Browsing. The mode now supports Gemini Nano, the on-device large language model (LLM) on desktop, to add an extra layer of protection for users. The model can detect potential phishing sites and inform Google to take action on them. The company said Gemini Nano can even detect scams that have not been previously seen by distilling the complex nature of websites.“Our goal is to expand this protection to Android devices and even more types of scams in the future,” Google said.However, this does not mean the company has not added any new threat protection system on Android. Google Chrome on Android is getting a new feature to protect users from malicious website-based notifications. If the browser detects any notification from a website that can be a scam, it will instead show a warning labelled “Potential scam detected,” with an option to either unsubscribe or view the notification.If users trust the website and believe the AI-powered feature made a mistake, they can view the content that was blocked. They can also choose to allow future notifications from the website.