Including, the fire-nado. An excerpt from issue #31 of befores & afters magazine. Replicating natural phenomena was ILM’s central task on Twisters. The visual effects studio would need to analyze the reference, then consider the science behind storms and tornadoes, and finally develop new tools for simulation and rendering to bring what were key ‘characters’ to life. The ILM Art Department also had an early role in crafting keyframe art to help visualize some of the more epic scenes in Twisters. Senior concept artist Brett Northcutt then continued the design process for specific tornadoes. “At first,” he says, “I wasn’t sure how much help I would be as there is a ton of amazing real storm chaser footage online. At that point, ILM visual effects supervisor Florian Witzel told me that director Lee Isaac Chung wanted each tornado to be like a monster, each with its own characteristics.  That really resonated with me as I could play with size, lighting and visibility for each tornado.” Concept art. “We also looked at a lot of real footage and I quickly learned how much variation they can have aesthetically,” adds Northcutt. “For scenes where our heroes are scientifically chasing storms we could use clarity and front lighting to show them as less threatening. However, for scenes where our heroes are in peril, we could backlight them or hide them in weather so the actual threat level wasn’t entirely clear.  The greatest movie monsters, like in Jaws or Alien, aren’t seen very clearly and are therefore much more scary.” ILM began the 3D tornado building process by breaking down the different elements that make up a storm and a tornado, while also giving distinct personalities to each of the ten tornadoes that appear in the film. This look and feel to the weather also came directly from the director, as production visual effects supervisor Ben Snow points out. “At the beginning we knew there were these six major sequences and that tornadoes went through different phases. I said to Isaac, ‘Look, I’d really like to get one or two key images from all of this amazing material that we’ve collected. Let’s just choose a couple of images for each tornado to get the character. So we went through and did that, and we ended up with a pretty good focused set of material.” Plate. Roto. Layout. Final. “In addition,” says Snow, “there were key beats in the film where Isaac would have looked at YouTube footage or footage from the storm-chasers, and he’d say, ‘This is the sort of feeling I want for this sequence, this is the quality of light I want. Not just for visual effects, but also for special effects and for the DP Dan Mindel and production designer Patrick Sullivan. Here are the key feelings I want, based on footage that we found.’” That gave Snow and the team at ILM two sets of creative inputs to start crafting the look of the storms and tornadoes, alongside the storm-chaser footage and the imagery produced by Giles Hancock. “What the VFX artists would have to do is start by trying to get it to match that reference and that look,” states Snow. “We ultimately had a development cycle for each tornado, with iterations. Even though we were using science and weather-based systems to do the simulations, there was still an artistic component to all this. We still got to sculpt, shape and add to them. And actually what that ended up giving us, too, was a library of clouds shapes and sky pieces and other parts of weather that could be combined together. I think it became key to both achieving realism and achieving the creative side to those tornadoes as well.” The outcome of this analysis of the tornadoes was a break-up of the general structure that needed to be depicted on screen. Tornadoes had a funnel, a thin layer of vapor surrounding the funnel, a debris field of dirt picked up from the ground, rigid debris objects and a shelf cloud connecting to the sky. “Our simulations are not just a tornado alone, there’s a lot more than that,” details ILM associate visual effects supervisors Charles Lai. “We have the wall cloud that connects from the top of the funnel to the sky, but that also has to be connected to the shelf cloud. So, we also built that shelf cloud, and that is moving a little bit slower, and that has to be connected to the HDRI panel that’s in the background. So, there’s a whole lot of little pieces that go into it. And then, not only are we simulating the tornado and those separate pieces, we’ve got a lot of ground effects that showcase that spiral inflow, and we pretty much had to do it for all our sequences.” “There’s so much more to it than just the tornadoes,” notes Witzel. “We’re essentially inside this dense soup of weather. There are varying levels of moisture and fine particulates in the air that diffuse the background, so you can’t just drop in an element and expect it to work. You constantly have to balance how the object or background is being affected and diffused. That was always an important thing to get right in each shot.” After the initial layout phase, the animation team was equipped with a flexible rig and toolset to define the base motion of each tornado. “The anim department blocked out the overall performance without breaking physical accuracy,” Witzel explains. “With the twin tornadoes, for instance, we could choreograph the timing and when they split apart. How they spiral around each other, how wide are they, how they relate to each other or how fast they move. All of these foundational choices help guide the simulation and effects teams. Once that framework is there, we hand it off to the fluid solver to do its simulation, and then you refine it from there.” “With the tornadoes, they were not always perfectly vertical,” adds Lai. “When you see these massive wedges on screen, you could believe, sure, it is just a massive vertical column. But, when we worked on some of the smaller tornadoes that were just trying to touch down, they often curved around and touched down. I think that’s where it was really valuable to be able to have the animation rigged that way and to show animatics in that form and just get quick buy-offs on, screen-space wise. It let us art direct them.” ILM used the latest iteration of its proprietary fluid solver, ILM Pyro, to simulate the tornadoes. “ILM Pyro started out several years ago as a wrapper around Houdini’s Pyro, just to make things easier for artists,” explains Witzel. “Over time, it evolved. Today its core is completely replaced with proprietary microsolvers and custom tools. The idea was always to create a system that gave artists both an easy entry point and deep artistic control.” With ILM Pyro, artists could manage a wide range of forces – as well as detailed emission and dissipation, pressure, buoyancy, gravity, and temperature bands – to sculpt the look and behavior of each tornado. The resulting velocity fields were then used to advect additional environmental effects like debris, smoke, and fire. “We’re essentially guiding the volumetrics,” says Witzel. “But when you do that, it’s easy to lose energy and detail. Other solvers can get soft and blurry very quickly. With ILM Pyro, we put a lot of emphasis on maintaining that energy. We use techniques like reflection solvers to preserve vorticity and confinement. It’s all about keeping the tornado’s energy and details alive.” The scale of the weather systems, especially supercells and the larger storm structures, often exceeded what a single simulation could handle efficiently. To manage this, ILM adopted a wedging workflow, splitting simulations into smaller blocks that were reassembled at render time. To retain fine detail, a secondary narrowband particle simulation was layered in, adding localized noise and complexity. “We’d run a core sim first,” Witzel explains, “then layer in another sim with added wavelet turbulence. Or we’d run a narrowband sim focused on the tornado’s edge, where the core was coarse, but the outer layers had much higher resolution.” Lighting the synthetic clouds proved compute-intensive, especially allocating for volumetric rendering time (rendering was generally handled in Mantra and Katana, with Katana XPU rendering done towards the end of production). To speed up the renders for clouds, ILM developed a workflow that pre-baked the light scatter on the clouds while being sampled directly at render time in the shader. This gave renders a six-time speed increase but preserved the desired look. One particular tornado seen in the film was unlike anything previously witnessed in Twister. This is the ‘fire-nado’ at the oil refinery. Snow worked directly with the previs team from The Third Floor, led by James Willingham III, on the refinery sequence. “We did a lot of work on what it was going to look like,” says Snow, “and we’d printed out little diagrams of what the imagined landscape might be like to take with us to Oklahoma. But we actually hadn’t come to a final agreement about what the scene should be.” The intervening Hollywood strikes in 2023 provided an opportunity to revisit the fire-nado sequence. “What we were trying to bring was a sense of mystery to the whole thing,” relates Snow. “We thought, let’s have them drive into a fog bank so they lose visibility, all of their GPS has gone out because of the storm, they don’t have knowledge that they’re next to this big refinery, which would make it more scary and creepy.” Anim. Creature dev. FX. Final. Owing to the time of the year that they could finally shoot the refinery scenes, the plates tended to be in bright daylight, necessitating a complex layering in of atmosphere and fog in visual effects. “One of our compositing supervisors, Ben O’Brien, came up with a really beautiful fire-lit fog look that we were able to bring into play in that end sequence.” Bluescreen photography for the sequence of the actors in their vehicles saw Scott Fisher fire off a series of pyro events for interactive lighting. “This really helped tie in the bluescreen plates to what we had filmed already,” notes Snow. “Scott also built these really cool rotisserie rigs and ways to flip the cars over with the actors inside that could then be dragged around, which happens in that storm.” Meanwhile, in the ILM Art Department, Brett Nortcutt worked with a 3D model of the refinery developed by model supervisor Bruce Holcomb to help art direct the scenes. “Even though the model was a work in progress, it was already an amazingly detailed model that I could start with,” notes Nortcutt. “For the low-angle image, I started with plate photography of the caravan on the road but ended up replacing everything but the road. The real challenge was showing a massive tornado interacting with an oil refinery as well as heavy fog. I imagined the wind forces of the tornado twisting and pulling apart the tanks and pipes of the oil refinery while also sucking fire and fog into its funnel. To add more chaos, I added sparks, explosions, and an indication of wind forces pulling everything into it across the terrain from the foreground.” “While I rendered Bruce’s 3D model of the refinery for these concepts,” continues Northcutt, “I knew that the only way I was going to be able to quickly create tornado variations for each concept was to digitally paint them. I developed a technique in Photoshop using three different brushes that allowed me to quickly vary the edges and thickness of each tornado. I then had great success adding a long exposure waterfall texture on top at a very low opacity that looked very similar to the texture of real tornadoes. It’s always fun to find something that you wouldn’t expect helps more than the real thing.” For the final fire-nado, the concept was that the nearby oil refinery accelerates the tornado, making it grow larger and more intense. “That meant,” says Witzel, “we had to layer in all kinds of elements. Crude oil fires, butane tanks exploding. We imagined a mix of chemicals being sucked into the atmosphere, creating this real Frankenstein’s monster of a storm. We had to build the entire refinery asset in detail, then destroy it, and on top of that, the whole scene was blanketed in fog. It was incredibly complex.” The post ‘Twisters’: what went into a storm appeared first on befores & afters.

Back in March, OpenAI announced a major upgrade to image generation in ChatGPT. Instead of relying on a separate model like DALL·E, ChatGPT began using GPT-4o's native image generation capabilities to deliver more accurate and visually appealing images based on text prompts. The GPT-4o image generation model accurately renders text and follows prompts precisely by leveraging both its knowledge base and the ongoing chat context. It also enables users to edit uploaded images or generate entirely new images using an uploaded photo as visual inspiration. This vastly improved image generation capability went viral, with over 130 million users creating more than 700 million images in just one week. 4o Image Generation is live! Bring your creativity to life and share your creations by using #MakeItWithCopilot. Here’s what I can do:​ — Microsoft Copilot (@Copilot) ⚡Render accurate, readable text​ ⚡Edit what you created​ ⚡Follow complex directions​ ⚡Transform an existing image’s style​ ⚡Make… pic.twitter.com/3ZhXB19J5gMay 15, 2025 Nearly 50 days after ChatGPT’s viral success, Microsoft is bringing the same image generation technology to Copilot users. With this update, Copilot users can now generate images with improved accuracy, better text rendering, the ability to edit generated images using text prompts, and more. While it's commendable that Microsoft is bringing this enhanced capability to Copilot, the company needs to accelerate its pace to remain competitive against formidable rivals like ChatGPT (OpenAI) and Gemini (Google). During its recent 50th Anniversary celebration event, Microsoft shared several Copilot updates—but many of them appear to be catching up to features already available on ChatGPT and Gemini for months. Microsoft AI CEO Mustafa Suleyman has promised to make Copilot deeply personal for users. It remains to be seen how the company will deliver on that vision in the future.

Solar storms batter Earth every year, creating the occasional aurora and sometimes even paralyzing power grids. But these common phenomena pale in comparison to a monstrous event that inundated the planet with particles from the Sun around 14,300 years ago, identified as the strongest solar storm ever recorded in a recent study.The study, published in Earth and Planetary Science Letters, has shed light on the extreme solar particle event that Earth experienced in 12,350 B.C.E. Such extreme events have only occurred a handful of times throughout history, and this is the first time one has been observed before the Holocene, around the end of the last Ice Age. The Impacts of Solar StormsOccasionally, the Sun launches a burst of accelerated particles into space, prompting what is called a solar particle event. These are often too weak to have considerable impacts here on Earth, although stronger events can pierce the planet’s magnetic field and disrupt technological systems.   Extreme solar particle events (ESPEs), however, are in a class of their own. Only eight of these events have been identified during the Holocene (the current epoch, ongoing for almost 12,000 years) — notable storms took place around 994 A.D., 663 B.C.E., 5259 B.C.E., and 7176 B.C.E. A separate 2024 study states that ESPEs are “up to three orders of magnitude stronger than” any solar particle event that has been observed directly by satellites in the modern age. The same study predicts that if an ESPE were to hit Earth during a period when its magnetic field is weakened, it could cause DNA damage in humans and impair aquatic ecosystems.Measuring the Strongest Solar StormThe ESPE that occurred in 12,350 B.C.E. takes the word “extreme” to a new level, having been much stronger than solar storms that followed in the Holocene.“Compared to the largest event of the modern satellite era — the 2005 particle storm — the ancient 12,350 B.C.E. event was over 500 times more intense, according to our estimates,” said Kseniia Golubenko, a postdoctoral researcher at the University of Oulu in Finland, in a statement.The researchers gleaned the details of this storm — including its strength, timing, and terrestrial effects — through a model they developed, called SOCOL:14C-Ex. The data used to assess the 12,350 B.C.E. event comes from a radiocarbon deposit in a tree from Southwestern Europe; tree rings act like records for solar particle storms, preserving radiocarbon spikes in the atmosphere (called Miyake events) that were once caused by ESPEs.  The model was first tested on an ESPE that occurred around 775 A.D., the previous titleholder of the strongest solar storm. It then determined the intensity of the 12,350 B.C.E. event, confirming that it was 18 percent stronger than the 775 A.D. event. Preparing for Solar StormsThe newly developed model has allowed researchers to better understand radiocarbon data in ancient glacial conditions, different from the relatively stable and warm conditions of the Holocene. In the present day, we likely won’t have to worry about an ESPE as intense as the 12,350 B.C.E. event in our lifetime; less intense, yet still formidable solar storms are still a cause for concern, however. For example, the Gannon solar storm that hit Earth in May 2024 caused a series of short-lived communication and navigation problems. Although the current solar cycle reached its peak near the end of 2024 (meaning an overall decline in solar activity is soon to follow), scientists are still determined to improve solar storm predictions. Details surrounding the 12,350 B.C.E. event may even assist preparation efforts in the future.“This event establishes a new worst-case scenario,” said Golubenko. “Understanding its scale is critical for evaluating the risks posed by future solar storms to modern infrastructure like satellites, power grids, and communication systems.”Article SourcesOur writers at Discovermagazine.com use peer-reviewed studies and high-quality sources for our articles, and our editors review for scientific accuracy and editorial standards. Review the sources used below for this article:Earth and Planetary Science Letters. New SOCOL:14C-Ex model reveals that the Late-Glacial radiocarbon spike in 12350 BC was caused by the record-strong extreme solar stormJack Knudson is an assistant editor at Discover with a strong interest in environmental science and history. Before joining Discover in 2023, he studied journalism at the Scripps College of Communication at Ohio University and previously interned at Recycling Today magazine.

Text-to-audio generation has emerged as a transformative approach for synthesizing sound directly from textual prompts, offering practical use in music production, gaming, and virtual experiences. Under the hood, these models typically employ Gaussian flow-based techniques such as diffusion or rectified flows. These methods model the incremental steps that transition from random noise to structured audio. While highly effective in producing high-quality soundscapes, the slow inference speeds have posed a barrier to real-time interactivity. It is particularly limiting when creative users expect an instrument-like responsiveness from these tools. Latency is the primary issue with these systems. Current text-to-audio models can take several seconds or even minutes to generate a few seconds of audio. The core bottleneck lies in their step-based inference architecture, requiring between 50 and 100 iterations per output. Previous acceleration strategies focus on distillation methods where smaller models are trained under the supervision of larger teacher models to replicate multi-step inference in fewer steps. However, these distillation methods are computationally expensive. They demand large-scale storage for intermediate training outputs or require simultaneous operation of several models in memory, which hinders their adoption, especially on mobile or edge devices. Also, such methods often sacrifice output diversity and introduce over-saturation artifacts. While a few adversarial post-training methods have been attempted to bypass the cost of distillation, their success has been limited. Most existing implementations rely on partial distillation for initialization or do not scale well to complex audio synthesis. Also, audio applications have seen fewer fully adversarial solutions. Tools like Presto integrate adversarial objectives but still depend on teacher models and CFG-based training for prompt adherence, which restricts their generative diversity. Researchers from UC San Diego, Stability AI, and Arm introduced Adversarial Relativistic-Contrastive (ARC) post-training. This approach sidesteps the need for teacher models, distillation, or classifier-free guidance. Instead, ARC enhances an existing pre-trained rectified flow generator by integrating two novel training objectives: a relativistic adversarial loss and a contrastive discriminator loss. These help the generator produce high-fidelity audio in fewer steps while maintaining strong alignment with text prompts. When paired with the Stable Audio Open (SAO) framework, the result was a system capable of generating 12 seconds of 44.1 kHz stereo audio in only 75 milliseconds on an H100 GPU and around 7 seconds on mobile devices. With ARC methodology, they introducedStable Audio Open Small, a compact and efficient version of SAO tailored for resource-constrained environments. This model contains 497 million parameters and uses an architecture built on a latent diffusion transformer. It consists of three main components: a waveform-compressing autoencoder, a T5-based text embedding system for semantic conditioning, and a DiT (Diffusion Transformer) that operates within the latent space of the autoencoder. Stable Audio Open Small can generate stereo audio up to 11 seconds long at 44.1 kHz. It is designed to be deployed using the ‘stable-audio-tools’ library and supports ping-pong sampling, enabling efficient few-step generation. The model demonstrated exceptional inference efficiency, achieving generation speeds of under 7 seconds on a Vivo X200 Pro phone after applying dynamic Int8 quantization, which also cut RAM usage from 6.5GB to 3.6 GB. This makes it especially viable for on-device creative applications like mobile audio tools and embedded systems. The ARC training approach involves replacing the traditional L2 loss with an adversarial formulation where generated and real samples, paired with identical prompts, are evaluated by a discriminator trained to distinguish between them. A contrastive objective teaches the discriminator to rank accurate audio-text pairs higher than mismatched ones to improve prompt relevance. These paired objectives eliminate the need for CFG while achieving better prompt adherence. Also, ARC adopts ping-pong sampling to refine the audio output through alternating denoising and re-noising cycles, reducing inference steps without compromising quality. ARC’s performance was evaluated extensively. In objective tests, it achieved an FDopenl3 score of 84.43, a KLpasst score of 2.24, and a CLAP score of 0.27, indicating balanced quality and semantic precision. Diversity was notably strong, with a CLAP Conditional Diversity Score (CCDS) of 0.41. Real-Time Factor reached 156.42, reflecting outstanding generation speed, while GPU memory usage remained at a practical 4.06 GB. Subjectively, ARC scored 4.4 for diversity, 4.2 for quality, and 4.2 for prompt adherence in human evaluations involving 14 participants. Unlike distillation-based models like Presto, which scored higher on quality but dropped to 2.7 on diversity, ARC presented a more balanced and practical solution. Several Key Takeaways from the Research by Stability AI on Adversarial Relativistic-Contrastive (ARC) post-training and  Stable Audio Open Small include:  ARC post-training avoids distillation and CFG, relying on adversarial and contrastive losses. ARC generates 12s of 44.1 kHz stereo audio in 75ms on H100 and 7s on mobile CPUs. It achieves 0.41 CLAP Conditional Diversity Score, the highest among tested models. Subjective scores: 4.4 (diversity), 4.2 (quality), and 4.2 (prompt adherence). Ping-pong sampling enables few-step inference while refining output quality. Stable Audio Open Small offers 497M parameters, supports 8-step generation, and is compatible with mobile deployments. On Vivo X200 Pro, inference latency dropped from 15.3s to 6.6s with half the memory. ARC and SAO Small provide real-time solutions for music, games, and creative tools. In conclusion, the combination of ARC post-training and Stable Audio Open Small eliminates the reliance on resource-intensive distillation and classifier-free guidance, enabling researchers to deliver a streamlined adversarial framework that accelerates inference without compromising output quality or prompt adherence. ARC enables fast, diverse, and semantically rich audio synthesis in high-performance and mobile environments. With Stable Audio Open Small optimized for lightweight deployment, this research lays the groundwork for integrating responsive, generative audio tools into everyday creative workflows, from professional sound design to real-time applications on edge devices. Check out the Paper, GitHub Page and Model on Hugging Face. All credit for this research goes to the researchers of this project. Also, feel free to follow us on Twitter and don’t forget to join our 90k+ ML SubReddit. Mohammad AsjadAsjad is an intern consultant at Marktechpost. He is persuing B.Tech in mechanical engineering at the Indian Institute of Technology, Kharagpur. Asjad is a Machine learning and deep learning enthusiast who is always researching the applications of machine learning in healthcare.Mohammad Asjadhttps://www.marktechpost.com/author/mohammad_asjad/Meta AI Introduces CATransformers: A Carbon-Aware Machine Learning Framework to Co-Optimize AI Models and Hardware for Sustainable Edge DeploymentMohammad Asjadhttps://www.marktechpost.com/author/mohammad_asjad/Enterprise AI Without GPU Burn: Salesforce’s xGen-small Optimizes for Context, Cost, and PrivacyMohammad Asjadhttps://www.marktechpost.com/author/mohammad_asjad/ServiceNow AI Released Apriel-Nemotron-15b-Thinker: A Compact Yet Powerful Reasoning Model Optimized for Enterprise-Scale Deployment and EfficiencyMohammad Asjadhttps://www.marktechpost.com/author/mohammad_asjad/Researchers from Fudan University Introduce Lorsa: A Sparse Attention Mechanism That Recovers Atomic Attention Units Hidden in Transformer Superposition

AlphaEvolve imagined as a genetic algorithm coupled to a large language model. Picture created by the author using various tools including Dall-E3 via ChatGPT. Large Language Models have undeniably revolutionized how many of us approach coding, but they’re often more like a super-powered intern than a seasoned architect. Errors, bugs and hallucinations happen all the time, and it might even happen that the code runs well but… it’s not doing exactly what we wanted. Now, imagine an AI that doesn’t just write code based on what it’s seen, but actively evolves it. To a first surprise, this means you increase the chances of getting the right code written; however, it goes far beyond: Google showed that it can also use such AI methodology to discover new algorithms that are faster, more efficient, and sometimes, entirely new. I’m talking about AlphaEvolve, the recent bombshell from Google DeepMind. Let me say it again: it isn’t just another code generator, but rather a system that generates and evolves code, allowing it to discover new algorithms. Powered by Google’s formidable Gemini models (that I intend to cover soon, because I’m amazed at their power!), AlphaEvolve could revolutionize how we approach coding, mathematics, algorithm design, and why not data analysis itself. How Does AlphaEvolve ‘Evolve’ Code? Think of it like natural selection, but for software. That is, think about Genetic Algorithms, which have existed in data science, numerical methods and computational mathematics for decades. Briefly, instead of starting from scratch every time, AlphaEvolve takes an initial piece of code – possibly a “skeleton” provided by a human, with specific areas marked for improvement – and then runs on it an iterative process of refinement. Let me summarize here the procedure detailed in Deepmind’s white paper: Intelligent prompting: AlphaEvolve is “smart” enough to craft its own prompts for the underlying Gemini Llm. These prompts instruct Gemini to act like a world-class expert in a specific domain, armed with context from previous attempts, including the points that seemed to have worked correctly and those that are clear failures. This is where those massive context windows of models like Gemini (even you can run up to a million tokens at Google’s AI studio) come into play. Creative mutation: The LLM then generates a diverse pool of “candidate” solutions – variations and mutations of the original code, exploring different approaches to solve the given problem. This parallels very closely the inner working of regular genetic algorithms. Survival of the fittest: Again like in genetic algorithms, but candidate solutions are automatically compiled, run, and rigorously evaluated against predefined metrics. Breeding of the top programs: The best-performing solutions are selected and become the “parents” for a next generation, just like in genetic algorithms. The successful traits of the parent programs are fed back into the prompting mechanism. Repeat (to evolve): This cycle – generate, test, select, learn – repeats, and with each iteration, AlphaEvolve explores the vast search space of possible programs thus gradually homing in on solutions that are better and better, while purging those that fail. The longer you let it run (what the researchers call “test-time compute”), the more sophisticated and optimized the solutions can become. Building on Previous Attempts AlphaEvolve is the successor to earlier Google projects like AlphaCode (which tackled competitive Programming) and, more directly, of FunSearch. FunSearch was a fascinating proof of concept that showed how LLMs could discover new mathematical insights by evolving small Python functions. AlphaEvolve took that concept and “injected it with steroids”. I mean this for various reasons… First, because thanks to Gemini’s huge token window, AlphaEvolve can grapple with entire codebases, hundreds of lines long, not just tiny functions as in the early tests like FunSearch. Second, because like other LLMs, Gemini has seen thousands and thousands of code in tens of programming languages; hence it has covered a wider variety of tasks (as typically different languages are used more in some domains than others) and it became a kind of polyglot programmer. Note that with smarter LLMs as engines, AlphaEvolve can itself evolve to become faster and more efficient in its search for solutions and optimal programs. AlphaEvolve’s Mind-Blowing Results on Real-World Problems Here are the most interesting applications presented in the white paper: Optimizing efficiency at Google’s data centers: AlphaEvolve discovered a new scheduling heuristic that squeezed out a 0.7% saving in Google’s computing resources. This may look small, but Google’s scale this means a substantial ecological and monetary cut! Designing better AI chips: AlphaEvolve could simplify some of the complex circuits within Google’s TPUs, specifically for the matrix multiplication operations that are the lifeblood of modern AI. This improves calculation speeds and again contributes to lower ecological and economical costs. Faster AI training: AlphaEvolve even turned its optimization gaze inward, by accelerating a matrix multiplication library used in training the very Gemini models that power it! This means a slight but sizable reduction in AI training times and again lower ecological and economical costs! Numerical methods: In a kind of validation test, AlphaEvolve was set loose on over 50 notoriously tricky open problems in mathematics. In around 75% of them, it independently rediscovered the best-known human solutions! Towards Self-Improving AI? One of the most profound implications of tools like AlphaEvolve is the “virtuous cycle” by which AI could improve AI models themselves. Moreover, more efficient models and hardware make AlphaEvolve itself more powerful, enabling it to discover even deeper optimizations. That’s a feedback loop that could dramatically accelerate AI progress, and lead who knows where. This is somehow using AI to make AI better, faster, and smarter – a genuine step on the path towards more powerful and perhaps general artificial intelligence. Leaving aside this reflection, which quickly gets close to the realm of science function, the point is that for a vast class of problems in science, engineering, and computation, AlphaEvolve could represent a paradigm shift. As a computational chemist and biologist, I myself use tools based in LLMs and reasoning AI systems to assist my work, write and debug programs, test them, analyze data more rapidly, and more. With what Deepmind has presented now, it becomes even clearer that we approach a future where AI doesn’t just execute human instructions but becomes a creative partner in discovery and innovation. Already for some months we have been moving from AI that completes our code to AI that creates it almost entirely, and tools like AlphaFold will push us to times where AI just sits to crack problems with (or for!) us, writing and evolving code to get to optimal and possibly entirely unexpected solutions. No doubt that the next few years are going to be wild. References and Related Reads Deepmind’s blog post and white paper on AlphaEvolve A Google Colab notebook with the mathematical discoveries of AlphaEvolve outlined in Section 3 of the paper! Powerful Data Analysis and Plotting via Natural Language Requests by Giving LLMs Access to Functions New DeepMind Work Unveils Supreme Prompt Seeds for Language Models www.lucianoabriata.com I write about everything that lies in my broad sphere of interests: nature, science, technology, programming, etc. Subscribe to get my new stories by email. To consult about small jobs check my services page here. You can contact me here. You can tip me here. The post Google’s AlphaEvolve Is Evolving New Algorithms — And It Could Be a Game Changer appeared first on Towards Data Science.

Join our daily and weekly newsletters for the latest updates and exclusive content on industry-leading AI coverage. Learn More To date, vibe coding platforms have largely relied on existing large language models (LLMs) to help write code. However, writing code is only one of many different tasks developers need to perform to build a full enterprise-grade production platform. Other tasks in the complete software engineering workflow require using different tools to help review, commit and maintain code over time. It’s a challenge Windsurf (formerly Codeium) is taking on with a series of new frontier AI models it calls SWE-1 (software engineer 1) as part of the company’s Wave 9 update. The news comes as Windsurf is reportedly in the midst of being acquired by AI leader OpenAI for as much as $3 billion. That deal has not yet formally closed, and Windsurf is not currently publicly commenting on the deal. SWE-1 is a family of frontier-class AI models specifically designed to accelerate the entire software engineering process. Unlike general-purpose AI models that have been adapted for coding tasks, the SWE-1 family was built to address the full spectrum of software engineering activities. The new models aim to support developers through multiple surfaces, incomplete work states and long-running tasks that characterize real-world software development. Available immediately to Windsurf users, SWE-1 marks the company’s entry into frontier model development with performance competitive to established foundation models, but with a focus on software engineering workflows. “Our main goal here is to accelerate all software engineering by 99%,” Anshul Ramachandran, head of product and strategy at Windsurf, told VentureBeat.  Enterprise developers need more than just coding-capable models The core innovation behind SWE-1 is Windsurf’s recognition that coding represents only a fraction of what software engineers actually do. This approach addresses a critical limitation in current AI coding LLMs. Many different models can be used today to write application code, including OpenAI’s GPT-4.1, Anthropic Claude 3.7 and Google’s Gemini 2.5 Pro I/O edition.  Windsurf has a modular interface that can enable use of multiple different models. Ramachandran explained that Windsurf users have given the company feedback that existing coding models tend to do well with user guidance, but over time tend to miss things. This limitation stems from a fundamental difference in task structure. While code generation is often a single-shot task, real software engineering involves navigating multiple tools, working with incomplete code and maintaining context across long-running projects. The SWE-1 family: Purpose-built for different engineering tasks Rather than creating a one-size-fits-all solution, Windsurf has developed three specialized models: SWE-1: Full-size model designed for advanced reasoning and tool use, available to all paid users. SWE-1-lite: A smaller but powerful model replacing Windsurf’s existing Cascade Base, available to all users (both free and paid). SWE-1-mini: A lightweight model powering passive code predictions in Windsurf Tab, unlimited for all users. The SWE models were built through an extensive in-house training process focused specifically on software engineering tasks. Ramachandran said that the company used a new data model with sequential steps for training. Performance benchmarks: How SWE-1 compares  While SWE-1 isn’t positioned to replace foundation models from major labs, Windsurf claims it achieves frontier-class performance specifically for software engineering tasks. The company reports that it substantially outperforms mid-sized foundation models and open-weight models. However, Windsurf is careful not to oversell these initial results.  “Even our benchmark shows it’s not objectively better than all the other models,” Ramachandran acknowledged. Instead, the goal is to position SWE-1 as the first step toward purpose-built models that will eventually surpass general-purpose ones for specific engineering tasks — and potentially at a lower cost. What makes Windsurf’s approach technically distinctive is its implementation of the flow awareness concept. The basic idea is that a flow of steps need to happen as part of enterprise development. Rather than just writing code for one specific step, flow awareness is about being aware of the broader context. Flow awareness is centered on creating a shared timeline of actions between humans and AI in software development. The core idea is to progressively transfer tasks from human to AI by understanding where AI can most effectively assist. This approach creates a continuous improvement loop for the models.  “As we continue to improve the models, more of the steps in that shared timeline will be flipped from human to AI,” said Ramachandran. “The AI will be able to do more things that the human had to do before because the AI wasn’t right.” What this means for technical decision-makers For enterprises building or maintaining software, SWE-1 represents an important evolution in AI-assisted development. Rather than treating AI coding assistants as simply autocomplete tools, this approach promises to accelerate the entire development lifecycle. The potential impact extends beyond just writing code more quickly. The recognition that application development is more involved will help mature the vibe coding paradigm to be more applicable for stable enterprise software development. While it’s still early days for SWE-1, this move is important. If and when OpenAI completes the acquisition of Windsurf, the new models could become even more important as they intersect with the larger model research and development resources that will become available. Technical leaders should consider how much of their development workflow could benefit from AI assistance beyond mere code generation. Teams spending significant time on code reviews, debugging and managing technical debt might see more substantial benefits from tools like SWE-1 than those primarily focused on generating new code. Daily insights on business use cases with VB Daily If you want to impress your boss, VB Daily has you covered. We give you the inside scoop on what companies are doing with generative AI, from regulatory shifts to practical deployments, so you can share insights for maximum ROI. Read our Privacy Policy Thanks for subscribing. Check out more VB newsletters here. An error occured.

Not so fast The top fell off Australia’s first orbital-class rocket, delaying its launch The Australian startup behind the Eris rocket says the rest of the vehicle was undamaged. Stephen Clark – May 15, 2025 8:03 pm | 35 Gilmour's Eris rocket stands on its launch pad earlier this week at Bowen Orbital Spaceport in the Australian state of Queensland. Credit: Gilmour Space Gilmour's Eris rocket stands on its launch pad earlier this week at Bowen Orbital Spaceport in the Australian state of Queensland. Credit: Gilmour Space Story text Size Small Standard Large Width * Standard Wide Links Standard Orange * Subscribers only   Learn more The payload fairing at the top of Gilmour Space's first Eris rocket was supposed to deploy a few minutes after lifting off from northeastern Australia. Instead, the nose cone fell off the rocket hours before it was supposed to leave the launch pad Thursday. Gilmour, the Australian startup that developed the Eris rocket, announced the setback in a post to the company's social media accounts Thursday. "During final launch preparations last night, an electrical fault triggered the system that opens the rocket’s nose cone (the payload fairing)," Gilmour posted on LinkedIn. "This happened before any fuel was loaded into the vehicle. Most importantly, no one was injured, and early checks show no damage to the rocket or the launch pad." Gilmour was gearing up for a launch attempt from a privately-owned spaceport in the Australian state of Queensland early Friday, local time (Thursday in the United States). The company's Eris rocket, which was poised for its first test flight, stands about 82 feet (25 meters) tall with its payload fairing intact. It's designed to haul a payload of about 670 pounds (305 kilograms) to low-Earth orbit. While Gilmour didn't release any photos of the accident, a company spokesperson confirmed to Ars that the payload fairing "deployed" after the unexpected electrical issue triggered the separation system. Payload fairings are like clamshells that enclose the satellites mounted to the top of their launch vehicle, protecting them from weather on the launch pad and from airflow as the rocket accelerates to supersonic speeds. Once in space, the rocket releases the payload shroud, usually in two halves. There were no satellites aboard the rocket as Gilmour prepared for its first test flight. This was unusual Payload fairing problems have caused a number of rocket failures, usually because they don't jettison during launch, or only partially deploy, leaving too much extra weight on the launch vehicle for it to reach orbit. Gilmour said it is postponing the Eris launch campaign "to fully understand what happened and make any necessary updates." The company was founded by two brothers—Adam and James Gilmour—in 2012, and has raised approximately $90 million from venture capital firms and government funds to get the first Eris rocket to the launch pad. The astronauts on NASA's Gemini 9A mission snapped this photo of a target vehicle they were supposed to dock with in orbit. But the rocket's nose shroud only partially opened, inadvertently illustrating the method in which payload fairings are designed to jettison from their rockets in flight. Credit: NASA The Eris rocket was aiming to become the first all-Australian launcher to reach orbit. Australia hosted a handful of satellite launches by US and British rockets more than 50 years ago. Gilmour is headquartered in Gold Coast, Australia, about 600 miles south of the Eris launch pad near the coastal town of Bowen. In a statement, Gilmour said it has a replacement payload fairing in its factory in Gold Coast. The company will send it to the launch site and install it on the Eris rocket after a "full investigation" into the cause of the premature fairing deployment. "While we’re disappointed by the delay, our team is already working on a solution and we expect to be back at the pad soon," Gilmour said. Officials did not say how long it might take to investigate the problem, correct it, and fit a new nose cone on the Eris rocket. This setback follows more than a year of delays Gilmour blamed primarily on holdups in receiving regulatory approval for the launch from the Australian government. Like many rocket companies have done before, Gilmour set modest expectations for the first test flight of Eris. While the rocket has everything needed to fly to low-Earth orbit, officials said they were looking for just 10 to 20 seconds of stable flight on the first launch, enough to gather data about the performance of the rocket and its unconventional hybrid propulsion system. Stephen Clark Space Reporter Stephen Clark Space Reporter Stephen Clark is a space reporter at Ars Technica, covering private space companies and the world’s space agencies. Stephen writes about the nexus of technology, science, policy, and business on and off the planet. 35 Comments

Join our daily and weekly newsletters for the latest updates and exclusive content on industry-leading AI coverage. Learn More To date, vibe coding platforms have largely relied on existing large language models (LLMs) to help write code. However, writing code is only one of many different tasks developers need to perform to build a full enterprise-grade production platform. Other tasks in the complete software engineering workflow require using different tools to help review, commit and maintain code over time. It’s a challenge Windsurf (formerly Codeium) is taking on with a series of new frontier AI models it calls SWE-1 (software engineer 1) as part of the company’s Wave 9 update. The news comes as Windsurf is reportedly in the midst of being acquired by AI leader OpenAI for as much as $3 billion. That deal has not yet formally closed, and Windsurf is not currently publicly commenting on the deal. SWE-1 is a family of frontier-class AI models specifically designed to accelerate the entire software engineering process. Unlike general-purpose AI models that have been adapted for coding tasks, the SWE-1 family was built to address the full spectrum of software engineering activities. The new models aim to support developers through multiple surfaces, incomplete work states and long-running tasks that characterize real-world software development. Available immediately to Windsurf users, SWE-1 marks the company’s entry into frontier model development with performance competitive to established foundation models, but with a focus on software engineering workflows. “Our main goal here is to accelerate all software engineering by 99%,” Anshul Ramachandran, head of product and strategy at Windsurf, told VentureBeat.  Enterprise developers need more than just coding-capable models The core innovation behind SWE-1 is Windsurf’s recognition that coding represents only a fraction of what software engineers actually do. This approach addresses a critical limitation in current AI coding LLMs. Many different models can be used today to write application code, including OpenAI’s GPT-4.1, Anthropic Claude 3.7 and Google’s Gemini 2.5 Pro I/O edition.  Windsurf has a modular interface that can enable use of multiple different models. Ramachandran explained that Windsurf users have given the company feedback that existing coding models tend to do well with user guidance, but over time tend to miss things. This limitation stems from a fundamental difference in task structure. While code generation is often a single-shot task, real software engineering involves navigating multiple tools, working with incomplete code and maintaining context across long-running projects. The SWE-1 family: Purpose-built for different engineering tasks Rather than creating a one-size-fits-all solution, Windsurf has developed three specialized models: SWE-1: Full-size model designed for advanced reasoning and tool use, available to all paid users. SWE-1-lite: A smaller but powerful model replacing Windsurf’s existing Cascade Base, available to all users (both free and paid). SWE-1-mini: A lightweight model powering passive code predictions in Windsurf Tab, unlimited for all users. The SWE models were built through an extensive in-house training process focused specifically on software engineering tasks. Ramachandran said that the company used a new data model with sequential steps for training. Performance benchmarks: How SWE-1 compares  While SWE-1 isn’t positioned to replace foundation models from major labs, Windsurf claims it achieves frontier-class performance specifically for software engineering tasks. The company reports that it substantially outperforms mid-sized foundation models and open-weight models. However, Windsurf is careful not to oversell these initial results.  “Even our benchmark shows it’s not objectively better than all the other models,” Ramachandran acknowledged. Instead, the goal is to position SWE-1 as the first step toward purpose-built models that will eventually surpass general-purpose ones for specific engineering tasks — and potentially at a lower cost. What makes Windsurf’s approach technically distinctive is its implementation of the flow awareness concept. The basic idea is that a flow of steps need to happen as part of enterprise development. Rather than just writing code for one specific step, flow awareness is about being aware of the broader context. Flow awareness is centered on creating a shared timeline of actions between humans and AI in software development. The core idea is to progressively transfer tasks from human to AI by understanding where AI can most effectively assist. This approach creates a continuous improvement loop for the models.  “As we continue to improve the models, more of the steps in that shared timeline will be flipped from human to AI,” said Ramachandran. “The AI will be able to do more things that the human had to do before because the AI wasn’t right.” What this means for technical decision-makers For enterprises building or maintaining software, SWE-1 represents an important evolution in AI-assisted development. Rather than treating AI coding assistants as simply autocomplete tools, this approach promises to accelerate the entire development lifecycle. The potential impact extends beyond just writing code more quickly. The recognition that application development is more involved will help mature the vibe coding paradigm to be more applicable for stable enterprise software development. While it’s still early days for SWE-1, this move is important. If and when OpenAI completes the acquisition of Windsurf, the new models could become even more important as they intersect with the larger model research and development resources that will become available. Technical leaders should consider how much of their development workflow could benefit from AI assistance beyond mere code generation. Teams spending significant time on code reviews, debugging and managing technical debt might see more substantial benefits from tools like SWE-1 than those primarily focused on generating new code. Daily insights on business use cases with VB Daily If you want to impress your boss, VB Daily has you covered. We give you the inside scoop on what companies are doing with generative AI, from regulatory shifts to practical deployments, so you can share insights for maximum ROI. Read our Privacy Policy Thanks for subscribing. Check out more VB newsletters here. An error occured.