Estimating the time required to complete a task is riddled with pitfalls.Let’s start with two key concepts you must understand — after managing and designing projects for over 20 years, I’ve seen these patterns play out again and again.Hofstadter’s lawHofstadter’s law is a self-referential adage about time that goes like this:“It always takes longer than you expect, even when you take into account Hofstadter’s law.”This law highlights how difficult it is to accurately estimate the time required to complete complex tasks. Its recursive nature reflects the widespread experience that, no matter how complex a task appears, calculating the time needed is hard, even with our best efforts. (See the header image.)Why is that?Optimism biasThe main reason stems from the tendency known as optimism bias — the inclination to be overly optimistic and thus overestimate favorable outcomes. In fact, we tend to overestimate our own abilities as well.It’s worth noting that this bias is actually quite necessary for humans: without it, we wouldn’t take risks, start businesses, or grow. However, when it comes to estimation, it can seriously backfire. This phenomenon naturally leads to…Planning fallacyOverly optimistic forecasts about the outcomes of plans are almost everywhere. Daniel Kahneman and Amos Tversky introduced the term planning fallacy to describe how we consistently underestimate the necessary resources (money, time) when planning complex projects. This phenomenon pushes us toward optimistic planning. Unsurprisingly, we’re much more realistic when reviewing someone else’s plan.In short:We overestimate favorable outcomes and our abilities, and we underestimate the required resources and constraints.What can we do?As I mentioned earlier, I’ll share how I personally handle this. The phenomenon is not new, and there are countless methods available.2x + 15 ruleFor smaller tasks (~1–20 hours), I use this technique. I look at the task, estimate the number of hours, double it, and add 15 minutes. The extra 15 minutes is to get started and immerse yourself in the task. Most of the time, it works quite well and is incredibly simple.Hofstadter multiplierMy teammates track and record the actual time spent on tasks, which is useful for several reasons. First, it helps everyone gain experience with how long typical tasks take, which they can then compare to their own estimates, improving their estimation skills. Second, you can determine your own — or a particular person’s — so-called Hofstadter multiplier. We’re all different, so this is a more advanced, personalized version of the 2x + 15 rule.Complex projectsThe earlier solutions work well for simpler tasks, but complex projects require additional considerations. But first, it’s important to clarify one question:How much energy can you, or do you want to, invest in the estimate/proposal?From there, more questions arise:Is it a rough indicative proposal or a detailed, itemized one?What inputs, expertise, and professionals are needed for the estimate?Is all necessary information available?How much capacity do you have (if any) for the estimation and/or execution?Do you realistically have a chance of winning the work? :)Your answers to these questions will influence the path you choose.Going back to the planning fallacy for a moment: it can be mitigated by involving a third party or by relying on real data from similar past projects.That’s why it’s crucial to keep precise records of the time spent on every task so that the data is available later.You will log the time for every task. — Obi-Wan KenobiMoving on…In an ideal world, the best approach is for the team that will actually execute the project to estimate it together. But consider how much it costs if a 4–5 person team individually reviews and interprets the materials, discusses them together, collects missing information, breaks down the project into items, and estimates each one using some method. It’s incredibly time-consuming and expensive.Therefore, as a substitute, it’s good to have an experienced person who’s seen many projects and/or to rely on actual hours from similar past projects. In this case, you only need to account for the differences and unknowns (the risks) at the time of the proposal. Estimation becomes much simpler, faster, and cheaper.For indicative proposals, I almost always use this latter method, with one important addition: always assume a ±20% variation.Whoa, be careful!It may seem like you’ve done something similar before, but hidden behind the scenes are major differences. This can lead to nasty surprises if you rely on data from nearly similar projects. In such cases, it’s wise to consult multiple external teams to benchmark the work. This helps refine the estimate and creates a healthy competitive environment.And when a detailed, itemized proposal is necessary… There’s no shortcut: you’ll need to apply a methodology. I won’t even attempt to cover this topic here, as there are entire books written about the relevant methodologies.FinallyI’ve been in the product design profession for over 20 years, and I often play the game where, after hearing the requirements, I throw out a number almost off the top of my head. It works surprisingly often, but of course, I wouldn’t even call it indicative. Still, it’s fun.Estimating time for complex projects was originally published in UX Collective on Medium, where people are continuing the conversation by highlighting and responding to this story.

The new ES90 electric car is the flagship of Volvo's latest digital safety tech.TT News Agency/AFP via Getty Images For decades, the Volvo brand has been synonymous with safety. But keeping passengers secure is no longer just about a strong cabin or cleverly designed crumple zones. Increasingly, safety is about semi-autonomous driving technology that can mitigate collisions or even avoid them entirely. Volvo intends to be ahead of the game in this era too. Its secret weapon? Something called “Gaussian Splatting”. I asked Volvo’s Head of Software Engineering Alwin Bakkenes and subsidiary Zenseact’s VP Product Erik Coelingh exactly what this is and why it’s so important. Volvo: Early Application Of Safety Data “We have a long history of innovations based on data,” says Bakkenes. “The accident research team from the 70s started with measuring tapes. Now in the digital world we’re collecting millions of real-life events. That data has helped us over the years to develop a three-point safety belt and the whiplash protection system. Now, we can see from the data we collect from fleets that a very large portion of serious accidents happen in the dark on country roads where vulnerable road users are involved. That’s why, with the ES90 that we just launched, we are also introducing a function called lighter AES where we have enabled the car to steer away from pedestrians walking on the side of the road or cyclists, which in the dark you can’t see even if you have your high beam on. This technology picks that up earlier than a human driver.” The Volvo EX90 SUV will also benefit from this technology. Volvo Cars uses AI and virtual worlds with the aim to create safer carsVolvo “If you want to lead in collision avoidance and self-driving, you need to have the best possible data from the real world,” adds Bakkenes. “But everyone is looking also at augmenting that with simulated data. The next step is fast automation, so we’re using state-of-the-art end-to-end models to achieve speed in iterations. But sometimes these models hallucinate. To avoid that, we use our 98 years of safety experience and these millions of data points as guardrails to make sure that the car behaves well because we believe that when you start to automate it needs to be trusted. For us every kilometer driven with Pilot Assist or Pilot Assist Plus needs to be safer than when you've driven it yourself. In the world of AI data is king. We use Gaussian Splatting to enhance our data set.” What Is Volvo’s Gaussian Splatting? “Cars are driven all around the world in different weather and traffic conditions by different people,” says Coelingh. “The variation is huge. We collect millions of data points, but it’s still a limited amount compared to reality. Gaussian Splatting is a new technology that some of our PhD students have been developing the last few years into a system where you can take a single data point from the real world where you have all the sensor, camera, radar and LIDAR sequences and then blow it up into thousands or tens of thousands of different scenarios. In that way, you can get a much better representation of the real world because we can test our software against this huge variation. If you do it in software, you can test much faster, so then you can iterate your software much more quickly and improve our product.” “Gaussian Splatting is used in different areas of AI,” continues Coelingh. “It comes from the neural radiance fields (NeRFs).” The original version worked with static images. “The first academic paper was about a drum kit where somebody took still pictures from different angles and then the neural net was trained on those pictures to create a 3D model. It looked perfect from any angle even though there was only a limited set of pictures available. Later that technology was expanded from 3D to 4D space-time, so you could also do it on the video set. We now do this not just with video data, but also with LiDAR and radar data.” A real-world event can be recreated from every angle. “We can start to manipulate other road users in this scenario. We can manipulate real world scenarios and do different simulations around this to make sure that our system is robust to variations.”Gaussian Splatting allows multiple scenario variations to be created from one real event.Volvo Volvo uses this system particularly to explore how small adjustments could prevent accidents. “Most of the work that we do is not about the crash itself,” says Coelingh. “It’s much more about what's happening 4-5 seconds before the crash or potential crash. The data we probe is from crashes, but it's also from events where our systems already did an intervention and in many cases those interventions come in time to prevent an accident and in some cases they come late and we only mitigated it. But all these scenarios are relevant because they happen in the real world, and they are types of edge case. These are rare, but through this technology of Gaussian Splatting, we can go from a few edge cases to suddenly many different edge cases and thereby test our system against those in a way that we previously could not.” Volvo’s Global Safety Focus This is increasingly important for addressing the huge variation in global driving habits and conditions a safety system will be expected to encounter. “Neural Nets are good at learning these types of patterns,” says Coelingh. “Humans can see that because of the behavior of a car the driver is talking into their phone, either slowing down or wiggling in the lane. If you have an end-to-end neural network using representations from camera images, LiDAR and radar, it will anticipate those kinds of things. We are probing data from cars all around the world where Volvo Cars are being driven.” The system acts preemptively, so it can perform a safety maneuver for example when a pedestrian appears suddenly in the path of the vehicle. “You have no time to react,” says Coelingh. Volvo’s safety system will be ready, however. “Even before that, the car already detects free space. It can do an auto steer and it’s a very small correction. It doesn't steer you out of lane. It doesn't jerk you around. It slows down a little bit and it does the correction. It's undramatic, but the impact is massive. Oncoming collisions are incredibly severe. Small adjustments can have big benefits.”Volvo's safety tech can detect pedestrians the human driver may not have seen.Volvo Volvo has developed one software platform to cover both safety and autonomy. “The software stack that we develop is being used in different ways,” says Coelingh. “We want the driver to drive manually undisturbed unless there’s a critical situation. Then we try to assist in the best possible way to avoid collision, either by warning, steering, auto braking or a combination of those. Then we also do cruising or L2 automation.” Volvo demonstrated how it has been using Gaussian Splatting at NVIDIA’s GTC in April. “We went deeply into the safe automation concept,” says Bakkenes. “Neural nets are good at picking up things that you can’t do in a rule-based system. We're developing one stack based on good fleet data which has end-to-end algorithms to achieve massive performance, and it has guard rails to make sure we manage hallucinations. It's not like we have a collision avoidance stack and then we have self-driving stack.” “There was a conscious decision that if we improve performance, then we want the benefits of that to be both for collision avoidance in manual driving and for self-driving,” says Coelingh. “We build everything from the same stack, but the stack itself is scalable. It’s one big neural network that we can train. But then there are parts that we can deploy separately to go from our core premium ADAS system all the way to a system that can do unsupervised automation. Volvo’s purpose is to get to zero collisions, saving lives. We use AI and all our energy to get there.”

Reuters reports that Elon Musk’s annoying chatbot, Grok, is now being used by the U.S. government. While the extent and nature of that usage is unclear, sources interviewed by the news outlet have expressed alarm at the implications of the chatbot’s access to government data. Grok was launched by xAI, an AI company founded by Musk in 2023,  and has since become integrated into Musk’s social media platform, X. The chatbot is known to summarize information in the most cringe-inducing manner possible, and was originally fashioned as an “anti-woke” antidote to ChatGPT and other more politically correct applications (though it’s turned out to be too woke for conservatives anyway). Musk’s Department of Government Efficiency team is now using a customized version of Grok, with the apparent goal of sorting and analyzing tranches of data. The team may also be using the chatbot to prepare reports, sources told the outlet. Aside from the very obvious data privacy concerns raised by Grok’s integration with government data, it appears that, once again, Musk is at the center of a conflict-of-interest violation. In fact, Reuters characterizes the promotion of Grok as a potentially criminal transgression of federal regulations. The outlet writes: If Musk was directly involved in decisions to use Grok, it could violate a criminal conflict-of-interest statute which bars officials — including special government employees — from participating in matters that could benefit them financially, said Richard Painter, ethics counsel to former Republican President George W. Bush and a University of Minnesota professor. “This gives the appearance that DOGE is pressuring agencies to use software to enrich Musk and xAI, and not to the benefit of the American people,” said Painter. The statute is rarely prosecuted but can result in fines or jail time. Yes, but how many times have we heard that one before? Elon has conflicts of interest up the wazoo. He is a walking conflict of interest, at this point. To my knowledge, he’s never seen the interior of a courtroom and, unless he gets caught with a dead body or something, it seems doubtful he ever will. Ever since Musk helped Trump get re-elected with hundreds of millions from his own piggybank, he’s has been treating the U.S. government like his personal plaything to destroy. Everywhere you look, the billionaire appears to be benefiting from his work with the government, whether it’s the White House bullying tariffed countries to adopt services from the billionaire’s satellite internet company, Starlink, or a new report that shows the billionaire’s companies may have saved nearly $2.37 billion from federal fines and penalties that were active under Biden but have since been “neutralized” in the Trump era. As far as DOGE’s mandate goes, the organization has been an unmitigated failure. It has barely saved a fraction of the money that Musk initially claimed that it would and, in the long term, the cuts are likely to cost Americans money, since many of them have been to important agencies that dispense key services to Americans.

Continuous Integration and Continuous Delivery/Deployment (CI/CD) refers to practices that automate how code is developed and released to different environments. CI/CD pipelines are fundamental in modern software development, ensuring code is consistently tested, built, and deployed quickly and efficiently. While CI/CD automation accelerates software delivery, it can also introduce security risks. Without proper security measures, CI/CD workflows can be vulnerable to supply chain attacks, insecure dependencies, and insider threats. To mitigate these risks, organizations must integrate measures for continuous monitoring and enforcing security best practices at every pipeline stage. Securing CI/CD workflows preserves the software delivery process's confidentiality, integrity, and availability. Security challenges and risks in CI/CD workflows While CI/CD workflows offer benefits in terms of automation and speed, they also bring unique security challenges that must be addressed to maintain the integrity of the development process. Some common challenges and risks include: Lack of visibility and inadequate security monitoring: CI/CD workflows involve multiple tools and stages, which make it challenging to maintain security visibility into potential threats. Vulnerabilities, especially in third-party libraries or containerized applications, can introduce security risks that go undetected if not correctly managed. Without centralized monitoring, real-time threat detection and response become difficult. Manual, reactive incident response increases the risk of exploitation. Compliance requirements: Meeting regulatory standards such as GDPR or HIPAA while maintaining fast deployment cycles can be challenging. Organizations must balance enforcing security policies, data protection, and compliance requirements without slowing down their CI/CD workflows.Code and dependency vulnerabilities: Unpatched or outdated dependencies in the workflow can introduce significant security risks. Third-party libraries or outdated packages can become attack vectors if not regularly updated and monitored for vulnerabilities. These risks are increased by the fast pace of CI/CD, where vulnerabilities may go untreated.Container vulnerabilities and image security: While containers are mainly used in CI/CD workflows, they are not safe from security risks. Vulnerabilities in container images, such as outdated software versions, misconfigurations, or insecure base images, present a risk in CI/CD workflows and can be exploited by attackers. Without proper scanning and validation, these weaknesses can propagate through the pipeline.Misconfiguration of CI/CD tools: Improper configuration of CI/CD tools can leave the workflow open to unauthorized access or unintentionally expose sensitive code. Misconfigurations in access control settings can increase the likelihood of privilege escalation or code exposure. Additionally, hardcoded credentials or mismanaged environment variables introduce a risk of being extracted by attackers, which could lead to data breaches.Supply chain attacks: Compromised third-party dependencies can introduce malicious packages or vulnerabilities into the workflow. These vulnerabilities can spread throughout the entire pipeline and infect production environments, primarily when third-party tools or libraries are not sufficiently validated.Insider threats: Insider threats in CI/CD workflows involve authorized users such as developers, DevOps engineers, system administrators, or third-party contractors, who may intentionally or unintentionally compromise the pipeline. Weak authentication mechanisms, inadequate access controls, and a lack of monitoring can increase the risk of unauthorized changes, credential theft, or the introduction of malicious code into the workflow.Enhancing CI/CD workflow security with Wazuh Wazuh is an open source security platform that offers unified XDR and SIEM capabilities for on-premises, containerized, virtualized, and cloud-based environments. Wazuh provides flexibility in threat detection, compliance, incident handling, and third-party integration. Organizations can implement Wazuh to address the challenges and mitigate the risks associated with CI/CD workflow security. Below are some ways Wazuh helps improve security in CI/CD workflows. Log collection and system monitoring Wazuh provides log collection and analysis capabilities to ensure the components of your CI/CD environment are continuously monitored for security threats. It collects and analyzes logs from various CI/CD pipeline components, including servers, containerization and orchestration tools such as Docker and Kubernetes, and version control systems like GitHub. This allows security teams to monitor for unusual activities, unauthorized access, or security breaches across the CI/CD environment. Additionally, the Wazuh File Integrity Monitoring (FIM) capability can detect unauthorized changes in code or configuration files. By monitoring files in real time or on a schedule, Wazuh generates alerts for security teams about file activities like creation, deletion, or modification. Figure 1: Wazuh dashboard showing File Integrity Monitoring (FIM) alerts. Custom rules and streamlined security monitoring Wazuh allows users to create custom rules and alerts that align with a pipeline's security requirements. Organizations can create custom rules matching their specific security needs, such as monitoring code changes, server configurations, or container images. This flexibility allows organizations to enforce granular security controls tailored to their CI/CD workflow. For instance, the Center for Internet Security (CIS) Docker Benchmark provides guidelines for securing Docker environments. Organizations can automate the compliance checks against CIS Docker Benchmark v1.7.0 using the Wazuh Security Configuration Assessment (SCA) capability. Figure 2: Wazuh dashboard showing Wazuh Security configuration assessment (SCA) results. Integration with third-party security tools Wazuh can integrate with various security tools and platforms, including container vulnerability scanners and CI/CD orchestration systems. This is particularly important in CI/CD workflows, where multiple tools may be used to manage the development lifecycle. Wazuh can pull in data from various sources, which helps to provide a centralized view of security across the pipeline. For instance, Wazuh integrates with container vulnerability scanning tools Trivy and Grype, which are commonly used to scan container images for vulnerabilities, insecure base images, or outdated software versions. By scanning container images before they are deployed into production, organizations can ensure that only secure, up-to-date images are used in the deployment processes. You can configure the Wazuh Command module to run a Trivy scan on an endpoint hosting container images and display any detected vulnerabilities in the Wazuh dashboard. This helps to ensure that insecure images are identified and prevented from being pushed into production. Figure 3: Wazuh dashboard displaying vulnerabilities discovered on container images from a Trivy scan. Automated incident response The speed of CI/CD workflows means that threats must be detected and mitigated quickly to minimize the risk of breaches or downtime. Wazuh provides incident response capabilities that help organizations respond to security incidents as soon as they occur. The Wazuh Active Response module can automatically take action when a security threat is detected. For example, suppose a malicious IP address is detected trying to access a system that runs CI/CD processes. In that case, Wazuh can automatically block the IP address and trigger predefined remediation actions. This automation ensures fast response, reduces manual intervention, and prevents potential threats from escalating. Conclusion Securing CI/CD workflows is important for maintaining a reliable and safe software development process. By using Wazuh, organizations can detect vulnerabilities early, monitor for anomalies, enforce compliance, and automate security responses while maintaining the speed and efficiency of CI/CD workflows. Integrating Wazuh into your CI/CD workflow ensures that security keeps pace with development speed. Found this article interesting? This article is a contributed piece from one of our valued partners. Follow us on Twitter  and LinkedIn to read more exclusive content we post.

UK-based engineering equipment supplier Kingsbury and metal additive manufacturing company Additure have been appointed by the UK Atomic Energy Authority (UKAEA) to supply additive manufacturing technology and expertise as part of the UK’s ongoing efforts to advance fusion energy research. The partnership will support the development of components designed to endure the extreme conditions within fusion reactors, with a focus on innovative materials and design approaches. A key area of focus involves the use of tungsten—layered with materials such as copper—to achieve the necessary durability. To support this work, Kingsbury and Additure will deliver and install a Nikon SLM Solutions SLM 280 2.0 Laser Powder Bed Fusion (LPBF) system at UKAEA’s facilities. “We are excited to support the team at the UKAEA as they scale, not just with the SLM 280’s LPBF capability, but with all the key elements of the AM ecosystem to make this a robust manufacturing solution for UKAEA and the UK’s fusion programme,” said Will Priest, Business Development Manager at Additure. The SLM 280 Production Series system. Image via Nikon SLM Solutions. About UKAEA The UK Atomic Energy Authority (UKAEA) is the United Kingdom’s national fusion energy research organisation. It operates as an executive non-departmental public body, sponsored by the Department for Energy Security and Net Zero. A key part of its mission involves fostering industrial fusion capability by working with manufacturers and supply chains to introduce and scale the technologies required for commercial fusion energy deployment. “The UKAEA aims to develop the commercialisation of additive manufacturing and support UK industry in the transition into the fusion energy sector. We conduct the complex areas of research and development to the point where it becomes commercially viable, the advice and support of our supply chain is hugely valuable in expediting this process,” said Roy Marshall, Head of Operations for Fabrication, Installation, and Maintenance at UKAEA. JET interior with super imposed plasma. Image via UK Atomic Energy Authority. Additure’s Role and Technology Contribution At the center of this initiative is the SLM 280 2.0, an LPBF system designed for high-performance applications, including the development of refractory metals. The system offers build speeds up to 80% faster than single-laser alternatives and includes integrated safety features such as a powder sieve module and system cooling enhancements. Beyond equipment delivery, Additure is also providing comprehensive technical training to UKAEA’s research, materials, and design teams. This includes detailed guidance on machine setup, build optimization, and specialized functions—such as a heated reduced build volume. “The applications training from Additure will provide our engineers with new ways to design some of the complex structures required by fusion and allow them to do this using some of the most challenging materials to work with. For additive manufacture to contribute to fusion energy, more designers need to think, ‘What process is most suitable for the desired thermal or structural performance?’ And ‘how do I create a design that is best optimised for additive manufacture?’”, said Mr. Marshall. Advancing Laser Beam Shaping 3D Printing  Given its notable advantages for industrial metal 3D printing, beam shaping capabilities are being developed and commercialized by several players in the research and LPBF 3D printing spheres. In 2024, German research organization Fraunhofer Institute for Laser Technology ILT showcased its new 3D printing beam shaping technology. Working with the Chair of Technology of Optical Systems (TOS) at RWTH Aachen University, the new platform, the Fraunhofer team is developing a test system for investigating complex laser beam profiles.  This platform can create customized beam profiles for laser powder bed fusion (LPBF) 3D printing, enhancing part quality, process stability and productivity, while minimizing material waste.  In 2022, Equispheres and Aconity3D used laser beam-shaping 3D printing to achieve build rates nearly nine times higher than industry norms. Equispheres’ NExP-1 aluminum powder was used with Aconity3D’s AconityMIDI+ LPBF 3D printer to unlock speeds exceeding 430 cm3/hr for a single laser.  The system was modified to employ a PG YLR 3000/1000-AM laser with beam-shaping capabilities. By using a shaped beam over a zoomed Gaussian profile, the team reduced overheating and mitigated spatter formation during high-speed 3D printing.  Take the 3DPIReader Survey — shape the future of AM reporting in under 5 minutes. Who won the 2024 3D Printing Industry Awards? Subscribe to the3D Printing Industry newsletter to keep up with the latest 3D printing news. You can also follow us on LinkedIn, and subscribe to the 3D Printing Industry Youtube channel to access more exclusive content. Featured image shows JET interior with super imposed plasma. Image via UK Atomic Energy Authority.

BySamuel MorenoPublished5 minutes agoWe may earn a commission from links on this page.Screenshot: Capcom / Samuel Moreno / KotakuJump ToMonster Hunter Wilds’ first title update introduced a lot of new content, but a standout addition is Mizutsune. This serpentine behemoth is a one-of-a-kind threat packed with unique mechanics and an aggressive attitude. The Tempered version takes it to another level and is considered by many to be stronger than anything encountered in the campaign or after. No matter which variant you’re having trouble with, we can help you beat this frothy threat.Suggested ReadingThe 3 Best And 3 Worst Korok Challenges In Tears Of The Kingdom Share SubtitlesOffEnglishview videoSuggested ReadingThe 3 Best And 3 Worst Korok Challenges In Tears Of The Kingdom Share SubtitlesOffEnglishYou’ll first take on a Mizutsune in the “Spirit in the Moonlight” side mission. The only requirements are being at Hunter Rank 21 and having already completed the “Fishing: Life, In Microcosm” side mission for Kanya. Once you’ve hunted the monster down one time, it will start spawning in both the Scarlet Forest and Ruins of Wyveria.Tempered Mizutsune can spawn in these same areas, albeit only once you’ve leveled up more and at least reached chapter six of the main quest line. I can say from personal experience that the Tempered version spawned more often during a Fallow season or an Inclemency. Feel free to use the Rest function a couple of times to make it show up.Screenshot: Capcom / Samuel Moreno / KotakuMizutsune’s most distinctive aspect are its bubbles, which are dispersed during many of its attacks. Getting hit by most of these will inflict you with the unique Bubbleblight ailment. This ailment is split into a minor and a major stage. Minor Bubbleblight isn’t too bad, as it’s a buff that enhances your evasion. Getting hit with another bubble will change it to the more frustrating Major Bubbleblight, which causes you to slip while running and be sent farther away by large attacks or explosions.A Nulberry is unfortunately not able to cure this status ailment, although they’re still worth bringing since Mizutsune can additionally inflict both Waterblight and Fireblight. The quickest way to cure Bubbleblight is to use a Cleanser. Alternatively, if you have fought this monster before and are just farming, equipping Mizutsune-forged armor is a big help. Its Bubbly Dance skill will prevent Major Bubbleblight so that you can focus more on dealing damage. If you don’t have the armor or run out of Cleanser, your next best bet is waiting 30 seconds for it to disappear.Some bubbles apply different effects, which makes things more complicated. Thankfully, they’re color-coded; however, you’ll still need to be quick on your feet when they’re coming right at you. Here are the different-colored bubbles and what they do:Clear: Deals damage and inflicts BubbleblightGreen: Provides healing and inflicts BubbleblightRed: Provides a temporary attack boost and inflicts BubbleblightFiery Blue: Deals damage and inflicts FireblightTrying to dodge these bubbles isn’t always the best use of your time. Both Slinger ammo and attacks from your weapon can pop them, although slower melee weapons can require precise timing. Using ranged weapons like Heavy or Light Bowguns is a lot more convenient.Screenshot: Capcom / Samuel Moreno / KotakuMizutsune is a highly mobile monster that can hit hard and use its water-based attacks to trigger a variety of status ailments. Start your hunts with the following weaknesses in mind because you’ll want to take advantage of them:Elemental Weaknesses: Thunder, DragonWeapon Type Weaknesses: Cut, BluntBreakable Parts: Head x2, Claws x2, Tail (can also be severed), Dorsal FinWeak Point: MouthSusceptible Status Ailments: Blastblight, Exhaust, Paralysis, Poison, Sleep, StunScreenshot: Capcom / Samuel Moreno / KotakuEven though you can tackle this thing as early as Hunter Rank 21, I suggest holding off for a bit. Mizutsune dishes out massive damage that can be mitigated with better armor and weapons. Wait until you’ve finished chapter five and can farm Gore Magala hunts. Weapons crafted from Gore Magala parts dish out Dragon element damage, while its armor offers great Water resistance. These are huge advantages to have when fighting this creature.Any weapon type can work, but slower ones will feel extra cumbersome against Mizutsune. Between the erratic movements and seemingly endless bubbles, it’s convenient to have quicker weapons like Dual Blades or Sword and Shield. Their multi-hitting nature will help with applying elemental damage as well. I advise pinning it down using traps and different ailments if you’re still having trouble landing hits.You should also watch out for its Waterblight-inflicting jet stream attacks. Mizutsune has a handful of moves that involve vertical or horizontal sweeping water beams. Thankfully, they’re easily telegraphed and have small hitboxes. Make sure to exploit Mizutsune’s long recovery periods after using these attacks.Mizutsune’s long-reaching tail attacks are another notable characteristic. The most deadly of these are the tail slams, which come out fast and can take out all of your health if your defense is low. There is a backflip variant to hit anyone behind and another where it twists its body in the air to slam those in front. I’ve seen the latter countered with an Offset Attack, but dodging it is the less risky solution.All of the above is amplified when the monster enters its unique enraged state. Breaking its head will enable it to transition into a powered-up mode akin to Soulseer Mizutsune from prior entries, complete with blue fire flaring from its left eye. Mizutsune will start shooting fire-covered bubbles in addition to using attacks more rapidly and aggressively when this state is triggered. While you can try to avoid this by not breaking the head, the trade-off is that you’ll be inflicting less damage. The one positive to this enraged state is that it will tire quicker and eventually become exhausted. That’s your cue to start dealing as much damage as possible.Don’t feel bad if these hunts leave you frustrated. There is a lot to keep track of with little time for reaction. Tempered Mizutsune is even more challenging and might just be the toughest fight in Monster Hunter Wilds yet. Still, everything we’ve mentioned will apply all the same. Memorizing which animations initiate which attacks will go a long way. Otherwise, bring your best gear, outfit them with appropriate decorations, and carry the Armorcharm and Powercharm items for good measure.Screenshot: Capcom / Samuel Moreno / KotakuNothing makes a tough fight feel worth it more than some good loot. Flipping through the monster’s Detailed Info tabs will break down all the various drop rates for the end of a hunt, destroyed wounds, and body part carvings. It’s a lot to take in, but worth perusing to narrow down what parts you need. I’ve provided a simple list of the attainable Mizutsune materials below that’s sorted by the overall drop frequency, with the most common parts at the top.Mizutsune Fin+ (100% chance for breaking the Head or Dorsal Fin)Mizutsune Claw+ (100% chance for breaking either Claw)Mizutsune Purplefur+ (100% chance for breaking the Tail)Mizutsune TailMizutsune Scale+Bubblefoam+Mizutsune Certificate SMizutsune Water Orb The community’s pleas for harder hunts certainly seem to have been heard. With the addition of Arch-Tempered monsters on the horizon, I can imagine these are only going to get tougher..

New owners of VPNSecure have been taking major flak recently after unexpectedly cancelling all lifetime subscriptions for users of the VPN service. The owners claimed that they didn’t know about the lifetime subscriptions when they first bought VPNSecure and are now unable to continue honoring them. Per Ars Technica, the first complaints from users losing their lifetime access began in March of this year. In April, users with lifetime subscriptions received an email stating, “To continue providing a secure and high-quality experience for all users, Lifetime Deal accounts have now been deactivated as of April 28th, 2025.” In response to a sudden wave—24 pages in total—of negative 1-star customer reviews on TrustPilot, a representative for the provider clarified, “In 2023, we acquired only the infrastructure and brand in a distressed asset sale after that company shut down. No contracts, payment data, or customer obligations were transferred.” VPNSecure has apologized to customers who were caught off guard by the sudden cancellations but reinforced its new position by saying it “never will” sell lifetime subscriptions. The whole ordeal seems to have been bungled from the start, but the company drew further ire with the way in which it so abruptly cut subscriptions. While it may have mitigated some of these issues by giving more advanced notice of the cancellations, it was never going to be an easy pill to swallow for previous lifetime plan holders. Adding insult to injury, there is no evidence that VPNSecure is offering refunds to those who purchased lifetime subscriptions. Instead, it’s offering affected users “an exclusive deal starting at just $1.87/month to continue using the service.” Insert eyeroll here. The whole ordeal is a great reminder that lifetime subscriptions are rarely as trustworthy as advertisements lead you to believe. Should the company ever go out of business or be sold, these previous ‘guarantees’ can be nothing but hot air. If you happen to be someone affected by these cancellations or are just looking for a VPN service that you can trust, refer to our list of the best VPNs.

Author Hello devs. Apologies in advance if this is not the correct forum for this. I've recently been spending a lot of my free time working on an old project of mine. It originated in 2013, and after several years-long hiatuses and complete rewrites, it's now in a state that I think is worth sharing.The goal is to simulate spheroids with real-time, procedurally generated terrain on a planetary scale and beyond. My inspiration for this project are old voxel games like Comanche. I wanted to do something like that, but instead of the finite play area being enforced through artificial boundaries, have it be a natural consequence of the world's geometry.Here's a short demo I recorded: https://www.youtube.com/watch?v=8S3MPTzymX8The planet in this demo has a radius of 2^23 meters which makes it about 2.28x the size of Earth by volume. The surface can be traversed freely without loading areas as it's all sampled in real time. Admittedly, the detail in this video is not that great. I've only recently begun working on the terrain sampler, so this is in a very early stage of development.To speak briefly on the geometry; the world is an icosahedron formed by 20 recursively subdivided tetrahedra. The tetrahedra subdivide into an oct-tet truss: a lattice of octahedra and tetrahedra. This structure is also called a spaceframe, and it's the cause for the shape of the terrain.If this sort of thing intrigues you, please let me know. I have been working on generating realistic planet-scale terrain for the last 2.5 years. You can read about it on my blog (start at the bottom).I decided against the icosahedron subdivision for my planet engine because it makes lots of things more difficult. I use the cube-sphere subdivision instead, because it leads to a quadtree subdivision of each cube face. This makes it much easier to do things like determine what terrain tiles are near to a given location, or to implement erosion. Everything exists on a rectilinear grid, which makes hydraulic erosion much easier.It seems like you should increase the distance at which you subdivide the terrain, to reduce the LOD popping. As it is, terrain features materialize to near to the camera. You ideally should implement a system so that the terrain is subdivided until a certain screen-space resolution is achieved (e.g. size of a triangle in pixels). Advertisement Author Your project looks very nice, but we have chosen very different paths in our implementations of planetary terrain. There is no grid or mesh spaceframe. The surface map is a direct result of sampling. The space is recursively sampled with a binary function and vertices are emitted based on the boundary resolved between filled and empty space. The detail popping is a consequence of the size difference between shapes in finer and coarser LODs. In theory this could be mitigated with higher LODs, but that is not a practical option given how CPU intensive real time, full planet binary space partitioning and boundary resolution is. Maybe one day I'll write a GPU implementation, though. What I'm saying is that your approach, while interesting, cannot use various optimizations which can be applied to rectilinear coordinates, which may explain why it is slow and can't increase the detail as much. My approach, running entirely on a 12-year old 4-core CPU can in real time generate triangles down to 5mm resolution at an angular LOD of 4 pixels per triangle, covering an earth-sized planet (and much bigger is not a problem), using around 4GB RAM. Furthermore, I'm not just generating terrain by fractal noise, I'm also applying various erosion processes which use the majority of the compute. This is only possible by a careful choice of spatial representation of the terrain, asynchronous multithreaded generation, and by highly optimized code (nearly all uses SIMD). Since the terrain is on a 2D rectangular grid, all operations aside from final mesh generation are essentially image processing, and make good use of CPU SIMD capabilities, which provides up to 8x speedup with AVX. Unless your data is laid out in a similar compact way you will have a hard time making it fast on modern CPUs. I don't think the “space frame” is a good fit for modern CPU architectures.

The Kids Online Safety Act (KOSA) has been reintroduced into Congress. If passed into law, this bill could impose some of the most significant legislative changes that the internet has seen in the U.S. since the Children’s Online Privacy Protection Act (COPPA) of 1998. As it currently stands, KOSA would be able to hold social media platforms legally accountable if it’s proven that these companies aren’t doing enough to protect minors from harm. The bill includes a long list of possible harms, such as eating disorders, sexual exploitation, substance abuse, and suicide. Though it overwhelmingly passed through the Senate last year, the bill was stifled in the House. KOSA has faced much backlash since its introduction in 2022. Human rights groups like the ACLU raised concerns that the bill could be weaponized as a tool for censorship and surveillance. While amendments to KOSA have mitigated some of these concerns, groups like the Electronic Frontier Foundation and Fight for the Future have remained against the bill. “The bill’s authors have claimed over and over that this bill doesn’t impact speech. But the Duty of Care is about speech: it’s about blocking speech that the government believes is bad for kids,” Fight for the Future wrote in a statement. “And the people who will be determining what speech is harmful? They are the same ones using every tool to silence marginalized communities and attack those they perceive as enemies.” However, KOSA has garnered support from companies like Microsoft, Snap, and X; X CEO Linda Yaccarino even worked with Senators Marsha Blackburn (R-TN) and Richard Blumenthal (D-CT) on the most recent draft of the bill. Google and Meta have remained opposed to the bill, but Apple announced today that it will support the legislation. “Apple is pleased to offer our support for the Kids Online Safety Act (KOSA). Everyone has a part to play in keeping kids safe online, and we believe [this] legislation will have a meaningful impact on children’s online safety,” Timothy Powderly, Apple’s Senior Director of Government Affairs, said in a statement. Techcrunch event Join us at TechCrunch Sessions: AI Secure your spot for our leading AI industry event with speakers from OpenAI, Anthropic, and Cohere. For a limited time, tickets are just $292 for an entire day of expert talks, workshops, and potent networking. 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 REGISTER NOW

Summary of This Study Hardware choices – specifically hardware type and its quantity – along with training time, have a significant positive impact on energy, water, and carbon footprints during AI model training, whereas architecture-related factors do not. The interaction between hardware quantity and training time slows the growth of energy, water, and carbon consumption slightly by 0.00002%. Overall energy efficiency during AI model training has improved slightly over the years, around 0.13% per year. Longer training time can gradually “drain” the overall energy efficiency by 0.03% per hour. Outline Introduction Research Question 1: Architectural and Hardware Choices vs Resource Consumption Research Question 2: Energy Efficiency over Time Methods Estimation methods Analysis methods Results RQ1: Architecture Factors Don’t Hold Much Predictive Power as Hardware Ones Final Model Selection Coefficients Interpretation RQ2 Discussion 1. Introduction Ever since the 1940s, when the first digital computers were invented, scientists have always dreamed of creating machines as smart as humans, what now became Artificial Intelligence (AI). Fast forward to November 2022, when ChatGPT — an AI model capable of listening and answering instantly — was released, it felt like a dream come true. Afterward, hundreds of new AI models have rushed into the race (take a look at the timeline here). Today, every single day, one billion messages are sent through ChatGPT (OpenAI Newsroom, 2024), highlighting the rapid AI adoption by users. Yet, few people stop to ask: What are the environmental costs behind this new convenience? Before users can ask AI questions, these models must first be trained. Training is the process where models, or algorithms, are fed datasets and try to find the best fit. Imagine a simple regression y = ax + b: training means feeding the algorithm x and y values and allowing it to find the best parameters a and b. Of course, AI models typically would not be as simple as a linear regression. They would contain tons of parameters, thus requiring massive amounts of computation and datasets. Moreover, they would need to run a substantial amount of specialized hardware that can handle that sheer amount of computation and complexity. All of that combined made AI consume much more energy than traditional software. In addition, AI training requires a stable and uninterrupted energy supply, which primarily comes from non-renewable energy sources like natural gas or coal-based, because solar and wind energy can fluctuate based on weather conditions (Calvert, 2024). Moreover, due to the high intensity of energy use, data centers — buildings that store AI models — heat up rapidly, emitting significant carbon footprints and requiring large amounts of water for cooling. Therefore, AI models have broad environmental impacts that include not only energy usage but also water consumption and carbon emissions. Unfortunately, there is not much official and disclosed data regarding energy, water, and carbon footprints of AI models. The public remains largely unaware of these environmental impacts and thus has not created strong pressure or motivations for tech companies to take more systematic changes. Furthermore, while some improvements have been made — especially in hardware energy efficiency — there remains little systematic or coordinated effort to effectively reduce the overall environmental impacts of AI. Therefore, I am hoping to increase public awareness of these hidden environmental costs and to explore whether recent improvements in energy efficiency are substantial. More particularly, I’m seeking to address two research questions in this study: RQ1: Is there a significant relationship between AI models’ architectural and hardware choices and their resource consumption during training? RQ2: Has AI training become energy-efficient over time? 2. Methods:  The paper used a dataset called Notable AI Models from Epoch AI (Epoch AI, 2025), a research institute that investigates the trends of AI development. The models included were either historically relevant or represent cutting-edge advances in AI. Each model was recorded with key training information such as the number of parameters, dataset size, total compute, hardware type, and hardware quantity, all collected from various sources, including literature reviews, publications, and research papers. The dataset also reported the confidence level for these attributes. To produce a reliable analysis, I evaluated only models with a confidence rating of “Confident” or “Likely”. As noted earlier, there was limited data regarding direct resource consumption. Fortunately, the dataset authors have estimated Total Power Draw (in watts, or W) based on several factors, including hardware type, hardware quantity, and some other data center efficiency rates and overhead. It is important to note that power and energy are different: power (W) refers to the amount of electricity used per unit of time, while energy (in kilowatt-hours, or kWh) measures the total cumulative electricity consumed over time. Since this study investigated resource consumption and energy efficiency during the training phase of AI models, I constructed and estimated four environmental metrics: total energy used (kWh), total water used (liters, or L), total carbon emissions (kilograms of CO2e, or kgCO2e), and energy efficiency (FLOPS/W, to be explained later). a. Estimation methods First, this study estimated energy consumption by selecting models with available total power draw (W) and training times (hours). Energy was computed as follows: \[\text{Energy (kWh)} = \frac{\text{Total Power Draw (W)}}{1000} \times \text{Training Time (h)}\] Next, water consumption and carbon emissions were estimated by rearranging the formulas of two standard rates used in data centers: Water Usage Effectiveness (WUE, in L/kWh) and Carbon Intensity (CI, in kgCO2e/kWh): \[\text{WUE (L/kWh)} = \frac{\text{Water (L)}}{\text{Energy (kWh)}}\ \Longrightarrow\ \text{Water (L)} = \text{WUE (L/kWh)} \times \text{Energy (kWh)}\] This study used the average WUE of 0.36 L/kWh in 2023, reported by Lawrence Berkeley National Laboratory (2024). \[\mathrm{CI\ \left( \frac{\mathrm{kgCO_2e}}{\mathrm{kWh}} \right)} = \frac{\mathrm{Carbon\ (kgCO_2e)}}{\mathrm{Energy\ (kWh)}}\ \Longrightarrow\ \mathrm{Carbon\ (kgCO_2e)} = \mathrm{CI\ \left( \frac{\mathrm{kgCO_2e}}{\mathrm{kWh}} \right)} \times \mathrm{Energy\ (kWh)}\] This study used an average carbon intensity of 0.548 kg CO₂e/kWh, reported by recent environmental research (Guidi et al, 2024). Finally, this study estimated energy efficiency using the FLOPS/W metric. A floating-point operation (FLOP) is a basic arithmetic operation (e.g., addition or multiplication) with decimal numbers. FLOP per second (FLOPS) measures how many such operations a system can perform each second, and is commonly used to evaluate computing performance. FLOPS per Watt (FLOPS/W) measures how much computing performance is achieved per unit of power consumed: \[\text{Energy Efficiency (FLOPS/W)} = \frac{\text{Total Compute (FLOP)}}{\text{Training Time (h)} \times 3600 \times \text{Total Power Draw(W)}}\] It is important to note that FLOPS/W is typically used to measure hardware-level energy efficiency. However, it’s possible that the actual efficiency during AI training may be different from the thereotical efficiency reported for the hardware used. I would like to investigate whether any of the training-related factors, beyond hardware alone, may contribute significantly to overall energy efficiency. b. Analysis methods: RQ1: Architectural and Hardware Choices vs Resource Consumption  Among energy, water, and carbon consumption, I focused on modeling energy consumption, as both water and carbon are derived directly from energy using fixed conversion rates and all three response variables shared identical distributions. As a result, I believe we could safely assume that the best-fitting model of energy consumption can be applied to water and carbon. While the statistical models were the same, I would still report the results of all three to quantify how many kilowatt-hours of energy, liters of water, and kilograms of carbon are wasted for every unit increase in each significant factor. That way, I am hoping to communicate the environmental impacts of AI in a more holistic, concrete, and tangible terms. Figure 2a. Histogram of Energy Consumption (kWh) Figure 2b. Histogram of log of Energy Consumption (kWh) Based on Figure 1, the histogram of energy showed extreme right skew and the presence of some outliers. Therefore, I performed a log transformation on energy data, aiming to stabilize variance and move the distribution closer to normality (Fig. 2). A Shapiro-Wilk test confirmed the log-transformed energy data is approximately normal (p-value = 0.5). Based on this, two types of distributions were considered: the Gaussian (normal) and the Gamma distribution. While the Gaussian distribution is approriate for symmetric and normal data, the Gamma distribution is more suited for positive, skewed data — commonly used in engineering modeling where small values occur more frequently than larger values. For each distribution, the paper compared two approaches for incorporating the log transformation: directly log transforming the response variable versus using a log link function within a generalized linear model (GLM). I identified the best combination of distribution and log approach by evaluating their Akaike Information Criterion (AIC), diagnostic plots, along with prediction accuracy. The candidate predictors included Parameters, Training Compute, Dataset Size, Training Time, Hardware Quantity, and Hardware Type. Architecture-related variables comprised Parameters, Training Compute, and Dataset Size, while hardware-related variables consisted of Hardware Quantity and Hardware Type. Training Time didn’t fall neatly into either category but was included due to its central role in training AI models. After fitting all candidate predictors into the selected GLM specification, I tested for multicollinearity to determine whether any variables should be excluded. Following this, I explored interaction terms, as each resource consumption may not have responded linearly to each independent variable. The following interactions were considered based on domain knowledge and various sources: Model Size and Hardware Type: Different hardware types have different memory designs. The larger and more complex the model is, the more memory it requires (Bali, 2025). Energy consumption can be different depending on how the hardware handles memory demands. Dataset Size and Hardware Type: Similarly, with different memory designs, hardware may access and read data at different data size (Krashinsky et al, 2020). As dataset size increases, energy consumption can vary depending on how the hardware handles large volumes of data. Training Time with Hardware Quantity: Running multiple hardware units at the same time adds extra overhead, like keeping everything in sync (HuggingFace, 2025). As training goes on, these coordination costs can grow and put more strain on the system, leading to faster energy drain. Training Time with Hardware Type: As training time increases, energy use may vary across hardware types since some hardware types may manage heat better or maintain performance more consistently over time, while others may slow down or consume more energy. RQ2: Energy Efficiency over Time Figure 2c. Histogram of Energy Efficiency (FLOPS/W) Figure 2d. Histogram of Energy Efficiency (FLOPS/W) The distribution of energy efficiency was highly skewed. Even after a log transformation, the distribution remained non-normal and overdispersed. To reduce distortion, I removed one extreme outlier with exceptionally high efficiency, as it was not a frontier model and likely less impactful. A Gamma GLM was then fitted using Publication Date as the primary predictor. If models using the same hardware exhibited wide variation in efficiency, it would suggest that other factors beyond the hardware may contribute to these differences. Therefore, architecture and hardware predictors from the first research question would be used to assess which variables significantly influence energy efficiency over time. 3. Results RQ1: Architectural and Hardware Choices vs Resource Consumption I ultimately used a Gamma GLM with a log link to model resource consumption. This combination was chosen because it had a lower AIC value (1780.85) than the Gaussian log-link model (2005.83) and produced predictions that matched the raw data more closely than models using a log-transformed response variable. Those log-transformed models generated predictions that substantially underestimated the actual data on the original scale (see this article on why log-transforming didn’t work in my case). Architecture Factors Don’t Hold Much Predictive Power as Hardware Ones After fitting all candidate explanatory variables to a Gamma log-link GLM, we found that two architecture-related variables — Parameters and Dataset Size — do not exhibit a significant relationship with resource consumption (p > 0.5). A multicollinearity test also showed that Dataset Size and Training Compute were highly correlated with other predictors (GVIF > 6). Based on this, I hypothesized that all three architecture variables—Parameters, Dataset Size, and Training Compute) may not hold much predictive power. I then removed all three variables from the model and an ANOVA test confirmed that simplified models (Models 4 and 5) are not significantly worse than the full model (Model 1), with p > 0.05: Model 1: Energy_kWh ~ Parameters + Training_compute_FLOP + Training_dataset_size + Training_time_hour + Hardware_quantity + Training_hardware + 0 Model 2: Energy_kWh ~ Parameters + Training_compute_FLOP + Training_time_hour + Hardware_quantity + Training_hardware Model 3: Energy_kWh ~ Parameters + Training_dataset_size + Training_time_hour + Hardware_quantity + Training_hardware Model 4: Energy_kWh ~ Parameters + Training_time_hour + Hardware_quantity + Training_hardware + 0 Model 5: Energy_kWh ~ Training_time_hour + Hardware_quantity + Training_hardware + 0 Resid. Df Resid. Dev Df Deviance Pr(>Chi) 1 46 108.28 2 47 111.95 -1 -3.6700 0.07809 . 3 47 115.69 0 -3.7471 4 48 116.09 -1 -0.3952 0.56314 5 49 116.61 -1 -0.5228 0.50604 Moving on with Model 5, I found that Training Time and Hardware Quantity showed significant positive relationships with Energy Consumption (GLM: training time, t = 9.70, p-value < 0.001; hardware quantity, t = 6.89, p-value < 0.001). All hardware types were also statistically significant (p-value < 0.001), indicating strong variation in energy use across different types. Detailed results are presented below: glm(formula = Energy_kWh ~ Training_time_hour + Hardware_quantity + Training_hardware + 0, family = Gamma(link = "log"), data = df) Coefficients: Estimate Std. Error t value Pr(>|t|) Training_time_hour 1.351e-03 1.393e-04 9.697 5.54e-13 *** Hardware_quantity 3.749e-04 5.444e-05 6.886 9.95e-09 *** Training_hardwareGoogle TPU v2 7.213e+00 7.614e-01 9.474 1.17e-12 *** Training_hardwareGoogle TPU v3 1.060e+01 3.183e-01 33.310 < 2e-16 *** Training_hardwareGoogle TPU v4 1.064e+01 4.229e-01 25.155 < 2e-16 *** Training_hardwareHuawei Ascend 910 1.021e+01 1.126e+00 9.068 4.67e-12 *** Training_hardwareNVIDIA A100 1.083e+01 3.224e-01 33.585 < 2e-16 *** Training_hardwareNVIDIA A100 SXM4 40 GB 1.084e+01 5.810e-01 18.655 < 2e-16 *** Training_hardwareNVIDIA A100 SXM4 80 GB 1.149e+01 5.754e-01 19.963 < 2e-16 *** Training_hardwareNVIDIA GeForce GTX 285 3.065e+00 1.077e+00 2.846 0.00644 ** Training_hardwareNVIDIA GeForce GTX TITAN X 6.377e+00 7.614e-01 8.375 5.13e-11 *** Training_hardwareNVIDIA GTX Titan Black 6.371e+00 1.079e+00 5.905 3.28e-07 *** Training_hardwareNVIDIA H100 SXM5 80GB 1.149e+01 6.825e-01 16.830 < 2e-16 *** Training_hardwareNVIDIA P100 5.910e+00 7.066e-01 8.365 5.32e-11 *** Training_hardwareNVIDIA Quadro P600 5.278e+00 1.081e+00 4.881 1.16e-05 *** Training_hardwareNVIDIA Quadro RTX 4000 5.918e+00 1.085e+00 5.455 1.60e-06 *** Training_hardwareNVIDIA Quadro RTX 5000 4.932e+00 1.081e+00 4.563 3.40e-05 *** Training_hardwareNVIDIA Tesla K80 9.091e+00 7.760e-01 11.716 8.11e-16 *** Training_hardwareNVIDIA Tesla V100 DGXS 32 GB 1.059e+01 6.546e-01 16.173 < 2e-16 *** Training_hardwareNVIDIA Tesla V100S PCIe 32 GB 1.089e+01 1.078e+00 10.099 1.45e-13 *** Training_hardwareNVIDIA V100 9.683e+00 4.106e-01 23.584 < 2e-16 *** --- Signif. codes: 0 ‘***’ 0.001 ‘**’ 0.01 ‘*’ 0.05 ‘.’ 0.1 ‘ ’ 1 (Dispersion parameter for Gamma family taken to be 1.159293) Null deviance: 2.7045e+08 on 70 degrees of freedom Residual deviance: 1.1661e+02 on 49 degrees of freedom AIC: 1781.2 Number of Fisher Scoring iterations: 25 Final Model Selection To better capture possible non-additive effects, various interaction terms were explored and their respective AIC scores (Table 1). The table below summarizes the tested models and their respective AIC scores: ModelPredictorsAIC5Training Time + Hardware Quantity + Hardware Type350.786Training Time + Hardware Quantity + Hardware Type * Parameters357.977Training Time + Hardware Quantity + Hardware Type * Dataset Size335.898Training Time * Hardware Quantity + Hardware Type345.399Training Time * Hardware Type + Hardware Quantity333.03Table 1. Summary of different GLM models and their respective AIC scores. Although AIC scores did not vary drastically, meaning their model fits are similar, Model 8 was preferred as it was the only one with significant effects in both main terms and interaction. Interactions involved Hardware Type were not significant despite some exhibiting better AIC, likely due to limited sample size across 18 hardware types. In Model 8, both Training Time and Hardware Quantity showed a significant positive relationship with energy consumption (GLM: t = 11.09, p < 0.001), and between hardware quantity and energy consumption (GLM: training time, t = 11.09, p < 0.001; hardware quantity, t = 7.32, p < 0.001; Fig. 3a). Their interaction term was significantly negative (GLM: t = –4.32, p < 0.001), suggesting that energy consumption grows more slowly when training time increases alongside with a higher number of hardware units. All hardware types remained significant (p < 0.001). Detailed results are as below: glm(formula = Energy_kWh ~ Training_time_hour * Hardware_quantity + Training_hardware + 0, family = Gamma(link = "log"), data = df) Coefficients: Estimate Std. Error t value Pr(>|t|) Training_time_hour 1.818e-03 1.640e-04 11.088 7.74e-15 *** Hardware_quantity 7.373e-04 1.008e-04 7.315 2.42e-09 *** Training_hardwareGoogle TPU v2 7.136e+00 7.379e-01 9.670 7.51e-13 *** Training_hardwareGoogle TPU v3 1.004e+01 3.156e-01 31.808 < 2e-16 *** Training_hardwareGoogle TPU v4 1.014e+01 4.220e-01 24.035 < 2e-16 *** Training_hardwareHuawei Ascend 910 9.231e+00 1.108e+00 8.331 6.98e-11 *** Training_hardwareNVIDIA A100 1.028e+01 3.301e-01 31.144 < 2e-16 *** Training_hardwareNVIDIA A100 SXM4 40 GB 1.057e+01 5.635e-01 18.761 < 2e-16 *** Training_hardwareNVIDIA A100 SXM4 80 GB 1.093e+01 5.751e-01 19.005 < 2e-16 *** Training_hardwareNVIDIA GeForce GTX 285 3.042e+00 1.043e+00 2.916 0.00538 ** Training_hardwareNVIDIA GeForce GTX TITAN X 6.322e+00 7.379e-01 8.568 3.09e-11 *** Training_hardwareNVIDIA GTX Titan Black 6.135e+00 1.047e+00 5.862 4.07e-07 *** Training_hardwareNVIDIA H100 SXM5 80GB 1.115e+01 6.614e-01 16.865 < 2e-16 *** Training_hardwareNVIDIA P100 5.715e+00 6.864e-01 8.326 7.12e-11 *** Training_hardwareNVIDIA Quadro P600 4.940e+00 1.050e+00 4.705 2.18e-05 *** Training_hardwareNVIDIA Quadro RTX 4000 5.469e+00 1.055e+00 5.184 4.30e-06 *** Training_hardwareNVIDIA Quadro RTX 5000 4.617e+00 1.049e+00 4.401 5.98e-05 *** Training_hardwareNVIDIA Tesla K80 8.631e+00 7.587e-01 11.376 3.16e-15 *** Training_hardwareNVIDIA Tesla V100 DGXS 32 GB 9.994e+00 6.920e-01 14.443 < 2e-16 *** Training_hardwareNVIDIA Tesla V100S PCIe 32 GB 1.058e+01 1.047e+00 10.105 1.80e-13 *** Training_hardwareNVIDIA V100 9.208e+00 3.998e-01 23.030 < 2e-16 *** Training_time_hour:Hardware_quantity -2.651e-07 6.130e-08 -4.324 7.70e-05 *** --- Signif. codes: 0 ‘***’ 0.001 ‘**’ 0.01 ‘*’ 0.05 ‘.’ 0.1 ‘ ’ 1 (Dispersion parameter for Gamma family taken to be 1.088522) Null deviance: 2.7045e+08 on 70 degrees of freedom Residual deviance: 1.0593e+02 on 48 degrees of freedom AIC: 1775 Number of Fisher Scoring iterations: 25 Figure 3a. Relationship between hardware quantity and log of energy consumption across training time groups. Training time was originally a continuous variable. For the sake of visualization, training time was divided into three equal-sized levels and labeled as high, mid, and low. Coefficients Interpretation To further interpret the coefficients, we can exponentiate each coefficient and subtract one to estimate the percent change in the response variable for each additional unit in the predictor (Popovic, 2022). For energy consumption, each additional hour of training would increase energy use by 0.18%, each additional hardware unit would add 0.07%, and their interaction reduced their combined main effects by 0.00002%. Similarly, since water and carbon were directly proportional with energy, the percent change in training time, hardware quantity, and their interaction remained the same (Fig. 3b, Fig. 3c). However, since hardware types were categorical variables and functioned as baseline intercepts, their values differed across energy, water, and carbon models to reflect differences in overall scale. Figure 3b. Relationship between hardware quantity and log of water consumption across training time groups. Figure 3c. Relationship between hardware quantity and log of carbon emissions across training time groups. RQ2: Energy Efficiency over Time I also used a log-linked Gamma model to examine the relationship between Energy Efficiency and Publication Date, as the Shapiro-Wilk test indicated that the log-transformed data was not normally distributed (p < 0.001). There was a positive relationship between Publication Date and Energy Efficiency, with an estimated improvement of 0.13% per year (GLM: t = 8.005, p < 0.001, Fig. 3d). Figure 3d. Relationship between publication year and log of energy efficiency (FLOPS/W). Each point represents a model, and the blue line shows a fitted trend using a linear model. To further investigate, we examined the trends by individual hardware type and observed noticeable variation in efficiency among AI models using the same hardware (Fig. 3e). Among all architecture and hardware choices, Training Time was the only statistically significant factor influencing energy efficiency (GLM: t = 8.581, p < 0.001), with longer training time decreases energy efficiency by 0.03% per hour. Figure 3e. Trends in log of energy efficiency (FLOPS/W) by hardware type over time. Each panel represents a specific hardware model, showing individual data points and fitted linear trends. Only hardware types used in at least three models are included. 4. Discussion This study found that hardware choices — including Hardware Type and Hardware Quantity — along with Training Time, have a significant relationship with each resource consumption during AI Model Training, while architecture variables do not. I suspect that Training Time may have implicitly captured some of the underlying effects of those architecture-related factors. In addition, the interaction between Training Time and Hardware also contributes to the resource usage. However, this analysis is constrained by the small dataset (70 valid models) across 18 hardware types, which likely limits the statistical power of hardware-involved interaction terms. Further research could explore these interactions with larger and more diverse datasets. To illustrate how resource-intensive AI training can be, we use Model 8 to predict the baseline energy consumption for a single hour of training on one NVIDIA A100 chip. Here are the predictions for each type of resource under this simple setup: Energy: The predicted energy use is 29,213 kWh, nearly three times the annual energy consumption of an average U.S. household (10,500 kWh/year) (U.S. Energy Information Administration, 2023), with each extra hour adding 5258 kWh more and each extra chip adding 2044 kWh. Water: Similarly, the same training session would consume 10,521 liters of water, almost ten times the average U.S. household’s daily water use (300 gallons or 1135 liters/day) (United States Environmental Protection Agency, 2024), with each extra hour adding 1,894 liters and each chip adding 736 liters. Carbon: the predicted carbon emission is 16,009 kg, about four times the annual emissions of a U.S. household (4000kg/year) (University of Michigan, 2024), with each extra hour adding 2881 kg and each extra chip adding 1120 kg. This study also found that AI models have become more energy-efficient over time, but only slightly, with an estimated improvement of 0.13% per year. This suggests that while newer hardware is more efficient, its adoption has not been widespread. While the environmental impact of AI may be mitigated over time as hardware hardware has become more efficient, this focus on hardware alone may overlook other contributors to overall energy consumption. In this dataset, both Training Compute and Total Power Draw are often estimated values and may include some system-level overhead beyond hardware alone. Therefore, the efficiency estimates in this study may reflect not just hardware performance, but potentially other training-related overhead. This study observed substantial variation in energy efficiency even among models using the same hardware. One key finding is that longer training time can “drain” energy efficiency, reducing it by approximately 0.03%. Further studies should explore how training practices, beyond hardware selection, impact the environmental costs of AI development. References Calvert, B.. 2024. AI already uses as much energy as a small country. It’s only the beginning. Vox. https://www.vox.com/climate/2024/3/28/24111721/climate-ai-tech-energy-demand-rising OpenAI Newsroom. 2024. Fresh numbers shared by @sama earlier today: 300M weekly active ChatGPT users. 1B user messages sent on ChatGPT every day 1.3M devs have built on OpenAI in the US. Tweet via X. 2024. https://x.com/OpenAINewsroom/status/1864373399218475440 Epoch AI. 2025. Data on Notable AI Models. Epoch AI. https://epoch.ai/data/notable-ai-models Shehabi, A., S.J. Smith, A. Hubbard, A. Newkirk, N. Lei, M.A.B. Siddik, B. Holecek, J. Koomey, E. Masanet, and D. Sartor. 2024. 2024 United States Data Center Energy Usage Report. Lawrence Berkeley National Laboratory, Berkeley, California. LBNL-2001637. Guidi, G., F. Dominici, J. Gilmour, K. Butler, E. Bell, S. Delaney, and F.J. Bargagli-Stoffi. 2024. Environmental Burden of United States Data Centers in the Artificial Intelligence Era. arXiv abs/2411.09786. Bali, S.. 2025. GPU Memory Essentials for AI Performance. NVIDIA Developer. https://developer.nvidia.com/blog/gpu-memory-essentials-for-ai-performance/ Krashinsky, R., O. Giroux, S. Jones, N. Stam, and S. Ramaswamy. 2020. NVIDIA Ampere Architecture In-Depth. NVIDIA Developer. https://developer.nvidia.com/blog/nvidia-ampere-architecture-in-depth/ HuggingFace. 2025. Performance Tips for Training on Multiple GPUs. HuggingFace Documentation. https://huggingface.co/docs/transformers/en/perf_train_gpu_many Popovic, G.. 2022. Interpreting GLMs. Environmental Computing. Environment Computing. https://environmentalcomputing.net/statistics/glms/interpret-glm-coeffs/ U.S. Energy Information Administration. 2023. Use of Energy Explained: Electricity Use in Homes. https://www.eia.gov/energyexplained/use-of-energy/electricity-use-in-homes.php United States Environmental Protection Agency. 2024. How We Use Water. https://www.epa.gov/watersense/how-we-use-water Center for Sustainable Systems, University of Michigan. 2024. Carbon Footprint Factsheet. Pub. No. CSS09–05. The post Rethinking the Environmental Costs of Training AI — Why We Should Look Beyond Hardware appeared first on Towards Data Science.