Scientific publication T. M. Lange, M. Gültas, A. O. Schmitt & F. Heinrich (2025). optRF: Optimising random forest stability by determining the optimal number of trees. BMC bioinformatics, 26(1), 95.Follow this LINK to the original publication. Random Forest — A Powerful Tool for Anyone Working With Data What is Random Forest? Have you ever wished you could make better decisions using data — like predicting the risk of diseases, crop yields, or spotting patterns in customer behavior? That’s where machine learning comes in and one of the most accessible and powerful tools in this field is something called Random Forest. So why is random forest so popular? For one, it’s incredibly flexible. It works well with many types of data whether numbers, categories, or both. It’s also widely used in many fields — from predicting patient outcomes in healthcare to detecting fraud in finance, from improving shopping experiences online to optimising agricultural practices. Despite the name, random forest has nothing to do with trees in a forest — but it does use something called Decision Trees to make smart predictions. You can think of a decision tree as a flowchart that guides a series of yes/no questions based on the data you give it. A random forest creates a whole bunch of these trees (hence the “forest”), each slightly different, and then combines their results to make one final decision. It’s a bit like asking a group of experts for their opinion and then going with the majority vote. But until recently, one question was unanswered: How many decision trees do I actually need? If each decision tree can lead to different results, averaging many trees would lead to better and more reliable results. But how many are enough? Luckily, the optRF package answers this question! So let’s have a look at how to optimise Random Forest for predictions and variable selection! Making Predictions with Random Forests To optimise and to use random forest for making predictions, we can use the open-source statistics programme R. Once we open R, we have to install the two R packages “ranger” which allows to use random forests in R and “optRF” to optimise random forests. Both packages are open-source and available via the official R repository CRAN. In order to install and load these packages, the following lines of R code can be run: > install.packages(“ranger”) > install.packages(“optRF”) > library(ranger) > library(optRF) Now that the packages are installed and loaded into the library, we can use the functions that these packages contain. Furthermore, we can also use the data set included in the optRF package which is free to use under the GPL license (just as the optRF package itself). This data set called SNPdata contains in the first column the yield of 250 wheat plants as well as 5000 genomic markers (so called single nucleotide polymorphisms or SNPs) that can contain either the value 0 or 2. > SNPdata[1:5,1:5] Yield SNP_0001 SNP_0002 SNP_0003 SNP_0004 ID_001 670.7588 0 0 0 0 ID_002 542.5611 0 2 0 0 ID_003 591.6631 2 2 0 2 ID_004 476.3727 0 0 0 0 ID_005 635.9814 2 2 0 2 This data set is an example for genomic data and can be used for genomic prediction which is a very important tool for breeding high-yielding crops and, thus, to fight world hunger. The idea is to predict the yield of crops using genomic markers. And exactly for this purpose, random forest can be used! That means that a random forest model is used to describe the relationship between the yield and the genomic markers. Afterwards, we can predict the yield of wheat plants where we only have genomic markers. Therefore, let’s imagine that we have 200 wheat plants where we know the yield and the genomic markers. This is the so-called training data set. Let’s further assume that we have 50 wheat plants where we know the genomic markers but not their yield. This is the so-called test data set. Thus, we separate the data frame SNPdata so that the first 200 rows are saved as training and the last 50 rows without their yield are saved as test data: > Training = SNPdata[1:200,] > Test = SNPdata[201:250,-1] With these data sets, we can now have a look at how to make predictions using random forests! First, we got to calculate the optimal number of trees for random forest. Since we want to make predictions, we use the function opt_prediction from the optRF package. Into this function we have to insert the response from the training data set (in this case the yield), the predictors from the training data set (in this case the genomic markers), and the predictors from the test data set. Before we run this function, we can use the set.seed function to ensure reproducibility even though this is not necessary (we will see later why reproducibility is an issue here): > set.seed(123) > optRF_result = opt_prediction(y = Training[,1], + X = Training[,-1], + X_Test = Test) Recommended number of trees: 19000 All the results from the opt_prediction function are now saved in the object optRF_result, however, the most important information was already printed in the console: For this data set, we should use 19,000 trees. With this information, we can now use random forest to make predictions. Therefore, we use the ranger function to derive a random forest model that describes the relationship between the genomic markers and the yield in the training data set. Also here, we have to insert the response in the y argument and the predictors in the x argument. Furthermore, we can set the write.forest argument to be TRUE and we can insert the optimal number of trees in the num.trees argument: > RF_model = ranger(y = Training[,1], x = Training[,-1], + write.forest = TRUE, num.trees = 19000) And that’s it! The object RF_model contains the random forest model that describes the relationship between the genomic markers and the yield. With this model, we can now predict the yield for the 50 plants in the test data set where we have the genomic markers but we don’t know the yield: > predictions = predict(RF_model, data=Test)$predictions > predicted_Test = data.frame(ID = row.names(Test), predicted_yield = predictions) The data frame predicted_Test now contains the IDs of the wheat plants together with their predicted yield: > head(predicted_Test) ID predicted_yield ID_201 593.6063 ID_202 596.8615 ID_203 591.3695 ID_204 589.3909 ID_205 599.5155 ID_206 608.1031 Variable Selection with Random Forests A different approach to analysing such a data set would be to find out which variables are most important to predict the response. In this case, the question would be which genomic markers are most important to predict the yield. Also this can be done with random forests! If we tackle such a task, we don’t need a training and a test data set. We can simply use the entire data set SNPdata and see which of the variables are the most important ones. But before we do that, we should again determine the optimal number of trees using the optRF package. Since we are insterested in calculating the variable importance, we use the function opt_importance: > set.seed(123) > optRF_result = opt_importance(y=SNPdata[,1], + X=SNPdata[,-1]) Recommended number of trees: 40000 One can see that the optimal number of trees is now higher than it was for predictions. This is actually often the case. However, with this number of trees, we can now use the ranger function to calculate the importance of the variables. Therefore, we use the ranger function as before but we change the number of trees in the num.trees argument to 40,000 and we set the importance argument to “permutation” (other options are “impurity” and “impurity_corrected”).  > set.seed(123) > RF_model = ranger(y=SNPdata[,1], x=SNPdata[,-1], + write.forest = TRUE, num.trees = 40000, + importance="permutation") > D_VI = data.frame(variable = names(SNPdata)[-1], + importance = RF_model$variable.importance) > D_VI = D_VI[order(D_VI$importance, decreasing=TRUE),] The data frame D_VI now contains all the variables, thus, all the genomic markers, and next to it, their importance. Also, we have directly ordered this data frame so that the most important markers are on the top and the least important markers are at the bottom of this data frame. Which means that we can have a look at the most important variables using the head function: > head(D_VI) variable importance SNP_0020 45.75302 SNP_0004 38.65594 SNP_0019 36.81254 SNP_0050 34.56292 SNP_0033 30.47347 SNP_0043 28.54312 And that’s it! We have used random forest to make predictions and to estimate the most important variables in a data set. Furthermore, we have optimised random forest using the optRF package! Why Do We Need Optimisation? Now that we’ve seen how easy it is to use random forest and how quickly it can be optimised, it’s time to take a closer look at what’s happening behind the scenes. Specifically, we’ll explore how random forest works and why the results might change from one run to another. To do this, we’ll use random forest to calculate the importance of each genomic marker but instead of optimising the number of trees beforehand, we’ll stick with the default settings in the ranger function. By default, ranger uses 500 decision trees. Let’s try it out: > set.seed(123) > RF_model = ranger(y=SNPdata[,1], x=SNPdata[,-1], + write.forest = TRUE, importance="permutation") > D_VI = data.frame(variable = names(SNPdata)[-1], + importance = RF_model$variable.importance) > D_VI = D_VI[order(D_VI$importance, decreasing=TRUE),] > head(D_VI) variable importance SNP_0020 80.22909 SNP_0019 60.37387 SNP_0043 50.52367 SNP_0005 43.47999 SNP_0034 38.52494 SNP_0015 34.88654 As expected, everything runs smoothly — and quickly! In fact, this run was significantly faster than when we previously used 40,000 trees. But what happens if we run the exact same code again but this time with a different seed? > set.seed(321) > RF_model2 = ranger(y=SNPdata[,1], x=SNPdata[,-1], + write.forest = TRUE, importance="permutation") > D_VI2 = data.frame(variable = names(SNPdata)[-1], + importance = RF_model2$variable.importance) > D_VI2 = D_VI2[order(D_VI2$importance, decreasing=TRUE),] > head(D_VI2) variable importance SNP_0050 60.64051 SNP_0043 58.59175 SNP_0033 52.15701 SNP_0020 51.10561 SNP_0015 34.86162 SNP_0019 34.21317 Once again, everything appears to work fine but take a closer look at the results. In the first run, SNP_0020 had the highest importance score at 80.23, but in the second run, SNP_0050 takes the top spot and SNP_0020 drops to the fourth place with a much lower importance score of 51.11. That’s a significant shift! So what changed? The answer lies in something called non-determinism. Random forest, as the name suggests, involves a lot of randomness: it randomly selects data samples and subsets of variables at various points during training. This randomness helps prevent overfitting but it also means that results can vary slightly each time you run the algorithm — even with the exact same data set. That’s where the set.seed() function comes in. It acts like a bookmark in a shuffled deck of cards. By setting the same seed, you ensure that the random choices made by the algorithm follow the same sequence every time you run the code. But when you change the seed, you’re effectively changing the random path the algorithm follows. That’s why, in our example, the most important genomic markers came out differently in each run. This behavior — where the same process can yield different results due to internal randomness — is a classic example of non-determinism in machine learning. Taming the Randomness in Random Forests As we just saw, random forest models can produce slightly different results every time you run them even when using the same data due to the algorithm’s built-in randomness. So, how can we reduce this randomness and make our results more stable? One of the simplest and most effective ways is to increase the number of trees. Each tree in a random forest is trained on a random subset of the data and variables, so the more trees we add, the better the model can “average out” the noise caused by individual trees. Think of it like asking 10 people for their opinion versus asking 1,000 — you’re more likely to get a reliable answer from the larger group. With more trees, the model’s predictions and variable importance rankings tend to become more stable and reproducible even without setting a specific seed. In other words, adding more trees helps to tame the randomness. However, there’s a catch. More trees also mean more computation time. Training a random forest with 500 trees might take a few seconds but training one with 40,000 trees could take several minutes or more, depending on the size of your data set and your computer’s performance. However, the relationship between the stability and the computation time of random forest is non-linear. While going from 500 to 1,000 trees can significantly improve stability, going from 5,000 to 10,000 trees might only provide a tiny improvement in stability while doubling the computation time. At some point, you hit a plateau where adding more trees gives diminishing returns — you pay more in computation time but gain very little in stability. That’s why it’s essential to find the right balance: Enough trees to ensure stable results but not so many that your analysis becomes unnecessarily slow. And this is exactly what the optRF package does: it analyses the relationship between the stability and the number of trees in random forests and uses this relationship to determine the optimal number of trees that leads to stable results and beyond which adding more trees would unnecessarily increase the computation time. Above, we have already used the opt_importance function and saved the results as optRF_result. This object contains the information about the optimal number of trees but it also contains information about the relationship between the stability and the number of trees. Using the plot_stability function, we can visualise this relationship. Therefore, we have to insert the name of the optRF object, which measure we are interested in (here, we are interested in the “importance”), the interval we want to visualise on the X axis, and if the recommended number of trees should be added: > plot_stability(optRF_result, measure="importance", + from=0, to=50000, add_recommendation=FALSE) The output of the plot_stability function visualises the stability of random forest depending on the number of decision trees This plot clearly shows the non-linear relationship between stability and the number of trees. With 500 trees, random forest only leads to a stability of around 0.2 which explains why the results changed drastically when repeating random forest after setting a different seed. With the recommended 40,000 trees, however, the stability is near 1 (which indicates a perfect stability). Adding more than 40,000 trees would get the stability further to 1 but this increase would be only very small while the computation time would further increase. That is why 40,000 trees indicate the optimal number of trees for this data set. The Takeaway: Optimise Random Forest to Get the Most of It Random forest is a powerful ally for anyone working with data — whether you’re a researcher, analyst, student, or data scientist. It’s easy to use, remarkably flexible, and highly effective across a wide range of applications. But like any tool, using it well means understanding what’s happening under the hood. In this post, we’ve uncovered one of its hidden quirks: The randomness that makes it strong can also make it unstable if not carefully managed. Fortunately, with the optRF package, we can strike the perfect balance between stability and performance, ensuring we get reliable results without wasting computational resources. Whether you’re working in genomics, medicine, economics, agriculture, or any other data-rich field, mastering this balance will help you make smarter, more confident decisions based on your data. The post How to Set the Number of Trees in Random Forest appeared first on Towards Data Science.

AI startup Cohere has acquired Ottogrid, a Vancouver-based platform that develops enterprise tools for automating certain kinds of high-level market research. Sully Omar, one of the founders of Ottogrid, announced the deal Friday in a post on X. He didn’t disclose the terms. Ottogrid will sunset its product, according to Omar, but will give customers “ample notice” and “a reasonable transition period.” “We’re very excited to join the Cohere team and integrate Ottogrid into Cohere’s … platform,” Omar said in a statement. “Through our work with Cohere, we’re [going to] dramatically impact how people can automate their workflows, enrich their data, and scale their operations.” Cohere didn’t immediately respond to a request for comment. Cohere’s purchase of Ottogrid comes as the former experiences a bit of corporate turbulence. According to The Information, Cohere fell well short of revenue projections the company prepared in early 2023, missing its target for last year by 85%. The company told Reuters on Thursday that its annualized revenue recently reached $100 million, following a strategic shift with a focus on private AI deployments for customers in sectors like healthcare, government, and finance. 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 Ottogrid launched in 2023 as Cognosys, led by Omar and Homam Malkawi. It rebranded in October 2024 with a major platform redesign that introduced a number of new integrations, tools, and APIs. Today, Ottogrid offers a “native table interface” with AI-powered document analysis capabilities. Customers can use it to extract data from a website and save it directly to a spreadsheet, for example, or automatically enrich sales lead lists. Ottogrid managed to raise $2 million in venture capital from investors, including GV (Google Ventures), Untapped Capital, Replit CEO Amjad Masad, Vercel CEO Guillermo Rauch, Cohere co-founders Ivan Zhang and Aidan Gomez, and others prior to its exit, according to Crunchbase. As part of Cohere, Omar says that Ottogrid will focus primarily on North, Cohere’s recently launched ChatGPT-style application designed to assist knowledge workers with tasks such as summarizing documents.

UK-based artist Fern Hook accused Bungie of directly using her artwork as textures for Marathon's Alpha build.

AMD seems to plan to aggressively enter the AI market through its Instinct MI450 AI clusters, as it would be the firm's first rack-scale product. AMD's Instinct MI450 Rack-Scale Solutions To Offer Faster Performance Relative To NVIDIA's Vera Rubin AI Clusters When compared in terms of scalability, NVIDIA leads the frontier with its AI clusters, notably the GB200 Blackwell configurations, which have seen massive adoption by the market. Not just because of their performance, but there isn't a rack-scale solution in the industry that is comparable to NVIDIA's GB200/GB300 lineup, but that looks to change in the future, as according to SemiAnalysis, AMD plans to introduce its first rack, the MI450 IF128 cluster, by H2 2026, and it is said to overthrow Team Green's dominance in the segment, giving Team Red a chance in the AI race. It is claimed that AMD plans to drop 128-GPU and 64-GPU solutions as its first rack-scale products, and they are set to directly rival NVIDIA's "Vera Rubin" VR200 NVL144 architecture, well, at least on paper. One of the momentous technologies Team Red is expected to adopt with its rack-scale solution is the use of "Infinity Fabric" over Ethernet as the interconnect mechanism, which will allow the firm to deliver over 1.8TByte/s unidirectional bandwidth per GPU. Regarding scale-out communications, AMD is set to include 3x Pensando 800GbE network cards per GPU with the MI450 rack-scale solution, allowing them to provide a network bandwidth of 2.4 Tbit/s, 1.5x higher than NVIDIA's VR200 NVL144. There will also be a secondary configuration available for scale-out communication, which will involve the use of 2x custom Ethernet network cards per GPU, but with a PCIe interface. While on paper, AMD plans to take a lead with rack-scale solutions, there are some downsides, which we'll cover ahead. NVIDIA GB300 AI server The MI450 IF128 cluster is said to be one of the most complex designs, which will make high-volume production a big concern. Moreover, being AMD's first rack-scale venture, the firm would be met with competition that has already managed to dominate the segment, so Team Red needs to come up with a more solid option. For that, AMD plans to focus more on the "smaller" MI450 IF64 rack-scale solution, which will feature a simpler yet effective design. It is clear that AMD wants to scale up its position in the AI industry with the MI400 lineup, and it could very easily turn out to be a tipping point for the firm if it manages to deliver on performance and efficiency. Deal of the Day

Zach Lipovsky and Adam Stein were dying to make Final Destination Bloodlines, and we mean more seriously than you think. When the directing partners pitched their take on the film over Zoom, they created an illusion where, during the pitch, each was “killed” in a wild, Final Destination way. It was the spark that got them the job, and now their film is finally here. That type of innovation is to be expected from the duo, who first met as contestants on the Steven Spielberg-produced reality show On the Lot. They didn’t win, but they became friends and have since been working their way up in the industry. Since then, they’ve made films in the Dead Rising and Leprechaun franchises, an excellent subtle superhero movie called Freaks, and even the Disney film Kim Possible. But now, finally, almost 20 years after meeting, Lipovsky and Stein have their first big Hollywood release. And something tells us it won’t be their last. io9 spoke to the directors in Los Angeles last week about their long journey through Hollywood. We broke down what they were most excited about in the Final Destination franchise, what Spider-Man: No Way Home director Jon Watts had to do with it all, and, most importantly, some of the ways they took Bloodlines from a sequel that could have been easily dismissed—and turned it into what we think is the series’ best yet. Note: There are a few minor spoilers throughout the interview, especially regarding the opening and overall themes of the movie. If you’re worried, come back after you see the film. Stein and Lipovsky doing their best Sam Raimi impression – Warner Bros. Germain Lussier, io9: The first time I heard about you guys was On the Lot in 2007. Adam Stein: That’s where we met. io9: [Laughs] Yes, that’s true. You’ve made a lot of cool stuff since then. But if you guys now could talk to those guys then and say it’ll take almost 20 years to get our first big theatrical release, what would they have said? Stein: [Laughs] You know, it has been such a roller coaster navigating the industry. But it’s been amazing that we have gotten to do it together. It’s a really hard thing to try to become a director in Hollywood. And so to have your best friend there with you to bounce ideas off of and commiserate with along the way has really been—we’ve been so lucky to have that. But I would say for people who are out there, who are in the struggle themselves, it takes a lot of perseverance, it takes a lot of resilience. Zack Lipovsky: But also those ups and downs are the lessons you need to learn. Like, we couldn’t have made this movie if someone had handed it to us 20 years ago. We didn’t have the wisdom or the experience that you get from things falling apart, and failures, and trying things that don’t work, and making movies that aren’t good. And all those types of things that you have to do to kind of hone your craft and get there. io9: Tell me about the conversations you guys had when you found out this was available. I’ve heard all about the Zoom, which is so cool and I’m mad it doesn’t exist somewhere. [Both laugh] But tell me about even before that—like when you said, “My God, Final Destination 6, let’s pitch on it and figure it out.” © Warner Bros./New Line Lipovsky: Well, I think actually the first gut reaction was, “How are they gonna make another one, because the end of five is perfect.” Right? So, “Really? They’re gonna make another one? Like what could they possibly do?” And then we started to hear this idea that Jon Watts had come up with about making the movie about a family. Starting it with the origin of this family and death is coming for this bloodline. And how much that would enrich all the relationships and the stakes and all those types of things. That really made us start to lean in and go, “Okay, that’s pretty interesting.” And then from there, [we] started to work on the process of getting hired. And then, a many-year process of working with the writers and the team to just kind of bring that inherent new idea through the movie in a way to make it feel more fresh and surprising. io9: Right, the first time I heard about this, it was like “Jon Watts is doing it.” And obviously he came up with kind of the kernel idea. But how much was actually in place when you guys came on board, and how much were you able to make it your own? Stein: They had done one draft of the script based on Jon’s idea when we first got involved. And the core of it was there. The idea of the opening would take place with Iris, I think it was the 1950s originally? But in the past, and then flash forward to her granddaughter. That was there. The details of who each person was, what all the death scenes were, completely changed over the course of two years of us being involved. But the core thing of it being a family that was fractured, that then could come together in a way, even as death was chasing them. That got us really excited. That was present in the original. io9: I’ll come back to some more of that soon but just as guys who are film fans, filmmakers, film nerds, what was the most exciting thing about the freedom of, presumably having your biggest budget? Besides the food. Everyone says the food. Lipovsky: We never had time to eat food. [Laughs] I think for us, it’s working with the caliber of crew that we had. You’re suddenly working with some of the best people at their craft in the world. You have the most incredible stunt people, special effects people, makeup effects people, and incredible actors. All the toys you’ve ever wanted. You can be like, “I really need this very special lens for this one shot I want to do.” You have it. So that, as a filmmaker, is just so much fun. Now, of course, you never have enough time or money. Because as soon as you start going, “Oh we could do that too, and do this,” you max out all the ability. Stein: You build a 120-foot set, surrounded by a volume, and light it on fire, well, that costs a lot. [All laugh.] But I think as filmmakers, the thing that we really got a kick out of is Final Destination, what makes it special as a horror franchise, is that there’s no personified villain coming for the characters. There’s no guy with a knife and a mask. What comes for the characters is the filmmaking. It’s these close-up, insert shots connected to other shots that build this mouse trap in the perfect, suspenseful way. It’s really like the directors are coming for the characters. So that was so much fun to kind of relish that experience. © Warner Bros./New Line io9: Speaking of the effects, just knowing you guys, I know wanted to do practical as much as you can. So talk about when you’re coming up with the death scenes, when, you know, “Okay, this is as far as we can push the practical, and then we have to kind of finish it off with digital.” Lipovsky: Yeah, we call it “the Guillermo del Toro school of filmmaking” where you do as much as you can practically, knowing that visual effects can help with the magic trick. And then the visual effects are really removing the things that are supporting the practical element. So we have a practical body or a practical set or something like that, or even people with a whole bunch of green rods and blue screens and things around it that are making it so that that practical thing can happen right at the center of frame. And then you use the visual effects to extend the set or erase the rod or do those things so that it really becomes hard to tell at which point what’s real and what isn’t. So there’s always something real there. And then the visual effects just kind of make it magical and kind of impossible. We try to do that whenever possible. Even there’s people falling through the dance floor and they’re falling and grabbing onto things dropping 40 feet, they’re [actually] falling through a dance floor and grabbing onto things, dropping 40 feet. But then everything around them is extended and kind of beautifully done by the visual effects team. But at the heart of it, we always just wanted something real. Stein: To give you another example, there’s a character that gets torn apart in the opening sequence. And there was a full-size mannequin body of this person with puppeteers that rip it in pieces, but then you erase the puppeteers. But the real thing you’re looking at really happened. io9: What sequence changed the most during either pre-production or editing as you were thinking about the balance of the practical and digital? Anna Lore in the barbecue scene.  – Warner Bros./New Line Lipovsky: I think one sequence that changed dramatically was what became the backyard barbecue scene, which is a really great scene. It actually started as a kids’ play place, Chuck E. Cheese-type set piece with bumper cars and bowling and all that crazy stuff. But we kind of ultimately realized that the core of these movies, especially this movie, was family, and also very relatable things. And so by making it a barbecue, you can kind of do all the things that we all have in our house, which is even more terrifying. And so it kind of changed into that. io9: This film—almost more than the other ones, which I watched beforehand—really messes with you with the red herrings. And I love it so much. So how early do you have to think about that, to make sure you have all the coverage while editing to know that these are like 15 things that we’re going to use to mess with you? Stein: You plan it out way before editing, in storyboards and scripting. To plant things that come back much later or to misdirect the audience. Because the main thing we wanted to do was keep it surprising. So you have a character walking backwards into a road and everyone in the audience goes, “Oh my god, oh my god,” and then [the car] misses him. So the audience kind of knows that we’re playing with them. Death is toying with the characters, just like we’re playing with the audience. Because we always wanted you to feel on the edge of your seat, like you don’t know what’s about to happen. So we had to plan that very far in advance. We were inspired by some of the other movies—the guy in FD2 throws the spaghetti out the window, and then his house catches on fire, he almost gets his hand caught in the garbage disposal, and all this stuff happens, and he gets on the fire escape and then slips on the spaghetti. So we tried to do that kind of thing where, like, there’s something that happens with a vending machine that you kind of forget about, and then it comes back. Stefani and her mom are not having a scream. – Warner Bros./New Line io9: Oh it comes back, yes it does. Another thing that I loved about this is that it takes the formula and it tweaks it in a couple of really interesting ways. One that happens early, so it’s not a huge spoiler, is we don’t get the premonition right way. So how early was that in the process that you’re like, right from the beginning we’re making this a new Final Destination? Lipovsky: That was a designing principle which was not have the premonition working exactly the way that all the other movies had, so that fans of the franchise right away went, “Wait, this is different than I’m used to. What is going on here?” And almost worried that we’re jumping the shark or breaking the rules or doing something wrong. But then, slowly, as the movie continues on and you realize what’s going on, you realize, “Oh, this does fit within the canon. It obeys all these rules that we’ve come to love.” But it was, like Adam was saying, it was building on that idea that as early as possible, we wanted people to not know where this was going next. And it’s really difficult with a Final Destination movie because there’s basically a formula. It’s something that you know what’s going to happen. But that makes for a boring movie. And so you have to figure out ways to still be true to it, but also keep it fresh io9: And you guys crushed that. Another thing I loved about it is, I love lore. Star Wars, Marvel, I love all that. And here the whole thing is lore, basically, right? So how early on was it decided that all the other movies are kind of the baby of this one? Stein: I mean, I think that was kind of a fun aspect of having this character, you know, Grandma Iris, who’s been in this cabin for decades, who’s been researching all the ways that people have died in the other movies. And it was actually inspired by FD1 how Alex in that movie, Devon Sawa’s character, traps himself in the cabin and is opening tuna cans with oven mitts because he knows death is coming. Well, we always loved that scene. And we thought, “Okay, what if you did that for decades? How would your life be? What would you know? What would you be researching?” And so Grandma Iris’s book with all the lore from all the other movies, that’s how that came together. Lipovsky: We even put an Easter egg in there where in FD1, he says, “I see you, you fuck” to death. And then Iris says the same thing in her cabin. © Warner Bros./New Line io9: One thing we haven’t seen now in six movies, though, is, death always wins but something is giving them the premonition. Something is battling. Stein: Yes. We have a theory. io9: Okay. I would love to hear your theory but I was just wondering, is that something you think could be introduced into a future movie or would that ruin kind of the formula? Stein: There have been lots of fan theories about some force of good that’s providing the premonition. I mean, we don’t want to piss people off and that is very valid theory. Lipovsky: And we don’t have a canon answer, but we have theories. Stein: Our own personal answer is that it’s death is bringing the premonitions because death, you know, he gets bored. He wants to have fun. So he does this thing, gives them enough clues that they could escape it if they’re paying attention, knowing that he’ll get them in the end in an even more fun way later. Tony Todd with the cast and crew of Bloodlines. – Warner Bros./New Line io9: I like that and I think fans will like that. Before I finish, I have to ask about Tony Todd. The fact that he passed is obviously horrible, but having him with this role in your movie it gives it this aura and this heart that is totally unexpected. So tell me about like first shooting with him. It was like when you found out? Lipovsky: Yeah. Well, we knew from while we were developing, we knew that he was sick. So it was developed with that in mind with him. You know, at first we were worried that he wouldn’t want to be in it or he wouldn’t be able to be in it. But he was so excited to be in the movie. Stein: And he’s like, “Don’t write me out of this movie. I have to do this.” Lipovsky: And we were very grateful that we could write the movie with that in mind, knowing that at the very least, this would be probably his last Final Destination movie, maybe his last movie ever. But we didn’t know that for sure. But that this would be saying goodbye to Bludworth. And that was designed from the beginning and worked very closely with him. And then even beyond that, when we shot that scene that day he was full of so much joy. It was just this electric feeling of seeing him do what he does so well. And then when it came to doing that last moment where he says goodbye, even with all the preparation we had done, we just said, “Tony, just throw it all away and speak from the heart about what this has all been about and what’s on your mind.” And, you know, his own mortality was on his mind. And he just spoke about how precious life is and how to take advantage of it. And that’s what he had shown us all day on set. He had done that. And it really comes through on screen when he he’s just he says that. Stein: Yeah, he’s speaking directly to the audience there. And I think takes a lot of people by surprise how much emotion they feel in that moment. © Warner Bros./New Line io9: It’s wonderful. Now that your first studio movie is in the books, and I hope there are many more, but what do you say the biggest lesson you guys took out of it? Lipovsky: The biggest lesson… I think the biggest thing that we learned was just that you don’t ever want to lose that same enthusiasm you had when you started making small movies and trusting your gut on the things you think that the audience will love. And just kind of trying to ignore the studio-ness of it all and just make that movie for yourself who’s the fan and bring that to life. And hold on to that, that joy of cinema and making a movie with your friends in your backyard and just trying to do that with cooler toys and amazing actors. Io9: One last stupid final question. The little kid at the beginning. You hate him so much. Stein: By the way, I love the theater reaction in that moment. The first time we screened it for a test audience and we had someone jump up in their seat and pump their fits when [he dies]. So you’re on the right track. That was really fun. io9: But since that was in the premonition, did we ever see him die again in real life? Lipovsky: There is a there is a newspaper article on the wall and if you look for the headline, I think it says “Pickpocket hit by piano moving van.” Final Destination Bloodlines is now in theaters. Want more io9 news? Check out when to expect the latest Marvel, Star Wars, and Star Trek releases, what’s next for the DC Universe on film and TV, and everything you need to know about the future of Doctor Who.

When we look at the relationship between 73-year-old legendary football coach Bill Belichick and his girlfriend and business partner, 24-year-old Jordon Hudson, it’s hard to know exactly what we’re seeing. Two grown-ups in love forging a dynamic business partnership? Elder abuse, as some have wildly speculated? Or is it, as the vitriolic comments in Hudson’s social media posts would have it, good old-fashioned gold digging? Gold digging is a misogynistic and retro term, but this scandal is a retro one. It’s a bizarre, slightly off-putting mystery that fits remarkably well into our current age of newly regressive gender politics.Hudson and Belichick, the six-time Super Bowl-winning former NFL coach, met on a flight in 2021 and went public with their relationship last December. While their nearly 50-year age gap has raised eyebrows amid observers, the scandal has only grown as Hudson has taken on an increasingly central role in Belichick’s professional world. It’s a bizarre, slightly off-putting mystery that fits remarkably well into our current age of newly regressive gender politics.Belichick requested Hudson be cc’ed on all publicity and media emails about him at the University of North Carolina at Chapel Hill, where he is now a college football coach. In public records, she’s listed as the manager of several companies that appear to be tied to Belichick. Insider accounts say she essentially cast herself in a Dunkin Donuts commercial in which Belichick appeared, and that she blocked a docuseries about Belichick’s career. But speculation around the relationship reached a fever pitch when Hudson interrupted during Belichick’s interview on CBS Sunday Morning in April. “How did you guys meet?” host Tony Dokoupil asked Belichick, referring to Hudson. “We’re not talking about this,” Hudson cut in tersely from off-camera. “Jordon was a constant presence during our interview,” Dokoupil informs viewers in voiceover, in a moment that launched a thousand TikTok reaction videos. In the midst of this controversy, the couple has been manufacturing social media content about their relationship that isn’t alarming so much as it is lightly uncanny, especially given Belichick’s famously gruff public persona. On Hudson’s Instagram account, she has posted beachside pictures of herself balancing athletically on Belichick’s outstretched legs, and of Belichick, dressed as a fisherman, reeling in a mermaid-tailed Hudson from the surf. (“Ouucchhhh!!!” goes the caption on that one.)A few have defended the relationship between Belichick and Hudson. Sports media personality Colin Cowherd has said that Hudson’s choice to hop into Belichick’s interview was normal for PR directors, if “kind of a cringy thing.” Right-wing sports outlet The Outkick has rallied to Hudson’s side on the grounds that she triggers the libs, saying, “If Jordon wants to spend her weekends at Bill’s house on Nantucket, soaking up the sun and enjoying Cisco Brewery on Bill’s dime, I say GO FOR IT.” But the most common reaction to the spectacle of Hudson and Belichick’s relationship is the one outlined by sports media personality Katie Nolan on a podcast in February: This is weird and seems like it has sinister undertones.“We’re already going, ‘You’re how much younger than him?’ And then you show up in a commercial. And then you hear that she’s in charge in his career.” Nolan said. “It seems like you could be taking advantage of the guy. And he’s obviously taking advantage of the girl.” Notably, it doesn’t appear to be much of a head-scratcher why a powerful septuagenarian would choose to be with a recent college graduate and pageant contestant. But the nefarious reason implied about why Hudson wants to be with Belichick is the one people are misogynistically and freely throwing around: gold digger.The resurgence of the gold digger I don’t know anything about Hudson and Belichick’s private relationship, and most likely, neither do you. I don’t want to make any claims about who is using whom or what their private life is like. But as a feminist pop culture critic, I am interested to see the term “gold digger” swim back up from the collective unconsciousness again, ready and willing to go to work. There’s something almost old-fashioned about it. The last time “gold digger” was thrown around in pop culture so much was when teenager Courtney Stodden skyrocketed to fame in 2011 after they got married at age 16 to 51-year-old actor Doug Hutchison. Stodden would later describe their marriage as one characterized by grooming and sexual assault. Before that, the great pop culture gold digger was Anna Nicole Smith, the model turned paparazzi obsession who married oil tycoon J. Howard Marshall II in 1994. Since Smith’s untimely death in 2007, she’s found a place in the pantheon of wronged women of the 1990s, someone we came to believe over the last decade was publicly mistreated — in part by her tarring as a gold digger. “It’s a bit provincial to look down one’s nose at a woman wanting something from a man and give a pass to a man who arguably would have never given her a second glance had she not been a beautiful, blonde Playboy Playmate,” essayist and sex work activist Laura LeMoon argued in 2023, writing about Smith in Salon. “Rather than singling out and condemning Anna Nicole Smith, we should be pointing our fingers at the inequities and systemic failures that put people like me and Smith in positions where obtaining money and resources from men, directly or indirectly, is our best option for survival.”“Gold digger” is a euphemism that allows us as a culture to talk around those systemic failures. It’s a way to deride women when they take seriously the idea that their financial well-being should depend on their relationships with men. So it’s odd to see the term becoming so popular during a moment when popular culture has become rather infatuated with the idea that life is most pleasant, simple, and straightforward when women’s finances do depend on romantic relationships with men.TikTok is full of tradwives explaining how their lives became better once they got out of the 9-to-5 grind to make cereal from scratch for their children and let their husbands do the breadwinning. (That the most successful of the tradwife influencers are actually the primary breadwinners in their homes is an irony the influencers themselves seldom discuss.) Or there are the stay-at-home girlfriends cooing over how taking care of their boyfriends without even the financial safety net of a marriage contract has been their ticket to the soft life. The gold digger is the tradwife as seen through a funhouse mirror: a woman living off her ability to attract a man, only here done crassly.Meanwhile, some of President Donald Trump’s most vocal supporters are arguing that his tariffs will force women out of the workforce and make them once again financially dependent on men, either through marriage or through sex work, life paths these men treat as equivalent. Trump’s appeal to the dream of a lost American manufacturing economy is a nostalgic one. It’s an appealing fantasy to some of his more incel-adjacent fans: that under this economy, the sexes will revert to an older, allegedly more natural economic relationship, one in which women trade their sexuality and childrearing capabilities to men in exchange for financial security. The gold digger is the tradwife as seen through a funhouse mirror: a woman living off her ability to attract a man, only here done crassly, without the show of love to soften the crude edges of the transaction. The gold digger is simply a figure we can blame for how uncomfortable this dynamic makes us feel, without having to think through just what is so uncomfortable about it. So if we’re angry at Jordon Hudson, it’s worth asking the question: Are we angry with her, or with the fact that powerful people want to make gold digging one of a woman’s most viable career paths again?See More:

Ready or not–and despite a spotty safety record–the EV maker is racing to launch a pilot ride service in Austin to show off its self-driving chops. Elon Musk is rolling out a handful of Tesla robotaxis in Austin next month, where up to 20 self-driving electric Model Ys will be unleashed to ferry passengers around the Texas city’s streets. He’s betting the future of Tesla on their success, as the automaker’s electric vehicle revenue tanks thanks to faster-growing Chinese rivals and a political backlash against Musk’s right-wing politics and role as job-slasher-in-chief for the Trump Administration. But there’s a big hitch: Tesla hasn’t proven its self-driving taxis are safe enough to start delivering rides. Given its misleadingly named Autopilot and Full Self-Driving (FSD) software’s deadly track record, Musk’s failure to provide detailed safety and technical data about Tesla’s technology and his determination to rely on cheap cameras instead of more robust sensors to navigate complicated urban environments, the Austin rollout could be a debacle. For the latest in cleantech and sustainability news, sign up here for our Current Climate newsletter. “It's going to fail for sure,” Dan O’Dowd, a long-time critic of Musk’s autonomous driving claims who’s spent his own money on Super Bowl commercials to call out Autopilot and FSD safety flaws, told Forbes. His anti-Tesla initiative, The Dawn Project, tests every update of FSD, a more advanced version of which is powering Musk’s robotaxis in Austin, as soon as they’re available. That update is to roll out to all Tesla drivers who pay a $99 monthly subscription fee. A pre-production Tesla Cybercab at the Petersen Automotive Museum in Los Angeles.Copyright 2024 The Associated Press. All rights reserved. O’Dowd has been putting the current version of FSD through its paces. “We drove it around Santa Barbara for 80 minutes and there were seven failures,” said O’Dowd, whose company Green Hills Software supplies security tech to defense and aerospace industry customers. “If there had not been a driver sitting in the driver's seat, it would’ve hit something.” While the company hasn’t booked a dollar of robotaxi revenue, that hasn’t stopped the world’s wealthiest person from declaring victory. “I don’t see anyone being able to compete with Tesla at present,” Musk said on the company’s April 22 results call. His assessment may be premature.“I was looking for a signal this was ready. I didn't get that.”   The sole public demonstration of Tesla’s robotaxi chops was staged drives of its new “Cybercab” at Warner Brothers Studio in Los Angeles last October. The event included hauling invited Tesla fans around a fake cityscape–free of pedestrians but with lots of Tesla technicians keeping close tabs on vehicles. It struck safety researcher Noah Goodall, who published a technical analysis of Tesla’s safety data, independently from his role with the Virginia Department of Transportation, as more amusement park attraction than real-world test. “It was just operating vehicles on a closed track on a movie lot. It was not impressive at all,” he said. “Navigating a real urban environment with uncertainty, other parties moving around, situations where just braking is not enough, that’s difficult. I was looking for a signal this was ready. I didn't get that.” Autonomy Promises In the decade since Tesla began selling customers its Autopilot and FSD features–for which it currently charges $8,000–the software has been linked to several fatal accidents where human drivers trusted the tech to drive their car, only for it to crash. The National Highway Traffic Safety Administration has opened multiple probes of Tesla’s Autopilot feature since 2016, including one last year to determine if it needed additional safety features after linking Autopilot to those 13 deaths. Last October, NHTSA also began investigating problems with FSD linked to two fatalities. Despite the names, this software has always been designed to have a human behind the wheel. For the past decade, Musk has repeatedly claimed “full autonomy”–where a car can drive without human assistance–was only months or a year away, repeatedly missing his targets. Now, with Tesla’s EV sales down 13% in the first quarter, the company needs some buzz to reassure investors CEO Musk can turn things around. Robotaxis, as well as AI and humanoid robots, are exactly that, according to Musk. So it’s running extensive tests in downtown Austin. “There’s just always a convoy of Teslas going all over to Austin in circles,” Musk said on the call. But a recent Business Insider story, citing interviews with former Tesla test drivers, doesn’t inspire confidence. The program “feels very forced,” one former worker said. "It's this breakthrough moment for Tesla, but there is also this feeling of so many last-minute details being up in the air.”The downtown Austin skyline.Getty Images Tesla’s program will operate in a very limited area of Austin and rely heavily on remote operators to minimize accidents, according to an executive with another autonomous tech company, based on conversations with Texas officials, who asked not to be identified as the matter isn’t public. To back up the AI driving the vehicles, Tesla has also hired human staff to monitor and assist if they get into jams, taking full control if necessary. “As we iterate on the AI that powers them, we need the ability to access and control them remotely,” the company said in a posting for one such job. Alphabet Inc.’s Waymo, the leader in robotaxi tech, also uses remote operators to assist the vehicles by providing suggested solutions to tricky situations, but those people don’t actually drive them. Lag and latency in cellular networks make remote operations unsafe. Limited Data Along with limited tests, there’s a dearth of trustworthy data about how well Tesla’s self-driving software operates. The company does file occasional safety performance reports about the software, but it’s not peer-reviewed by outside technical experts and it frames the data as positively as possible, according to Goodall, a technical witness in a lawsuit against Tesla over the death of Walther Huang, killed in 2018 when his Model X slammed into a highway divider while running on Autopilot. “With Full Self Driving, when they first started publishing numbers on that, they neglected to share that they’d only rolled the software out to drivers who had a very high safety score of 90 or above,” he said. “So of course the data showed it was safer, as your safest drivers were the only ones that had it.” By contrast, Waymo frequently posts detailed reports on how its robotaxis are performing, claiming the data is peer-reviewed by experts. Tesla also hasn’t yet shared details with the public about where in Austin it will offer its robotaxi service or exactly how it will operate. The city’s police and fire departments told Forbes the company contacted Austin’s Autonomous Vehicle Task Force, which includes their staff, and the city provided Tesla with “maps of schools and school zones; information about traffic control for special events; and information about our fire and police vehicles and procedures.” But a request to see communications between Tesla and the city was denied. “The City of Austin is withholding responsive documents without a ruling from the Attorney General’s office, as permitted by law,” it said in an email. “All responsive information has been withheld due to 3rd party.” The city didn’t respond to a question to confirm Tesla is that third party. The company, Musk and Ashok Elluswamy, head of Tesla’s autonomous vehicle program, didn’t respond to emails about the Austin rollout. NHTSA this week requested details about Tesla’s Austin plans to understand how the vehicles perform in bad weather. It’s been investigating Tesla collisions involving Autopilot and FSD in poor visibility situations since last October. It’s not clear if the company has responded yet. Tesla has had a permit to test autonomous vehicles in California for a decade, which requires companies to share safety data. Numerous competitors, including Waymo, Amazon’s Zoox, which hopes to operate robotaxis this year, Nuro and even Apple, which abandoned its program, have all submitted data on test miles logged including “disengagements”–when a human driver has to take over–as well as accident reports. Tesla hasn’t. Not Just Driving It’s hard to talk about Musk’s robotaxi dreams without comparing his approach to Waymo’s. The Alphabet unit has spent 16 years and billions of dollars trying to master every aspect of what a robotaxi has to do. Long before it gave its first paid rides to customers in Phoenix in 2018, the company tested intensely on public roads, privately at the “Castle,” its test facility at a decommissioned Air Force base in Central California, and with endless miles in virtual simulation to train its AI.A Waymo robotaxi in San Francisco.dpa/picture alliance via Getty Images Recognizing that robotaxis aren’t just a technical challenge, it also recruited people from the airline industry and businesses specializing in customer service. For robotaxis to work, the cars have to be good at doing small things that can be tough to master but are critical, like picking up and dropping off passengers. “We've been working on that for a long time,” said Chris Ludwick, director of Waymo’s product management team. “The first challenge with PUDO (the company’s shorthand for pickup, drop-off) is that when you get there, the on-road scene is going to be somewhat different each time. You may encounter construction or a stopped delivery truck or something like that. This leads to a whole suite of challenges of what do you do when you can't do the exact thing that you said to the rider when they requested the ride.” That includes developing a sophisticated app to guide passengers to the safest, most convenient spots for them and other road users. “You can't just block traffic. That's unacceptable. If you do that the community gets upset,” Ludwick said. “There's just a lot of small details you have to get right.” As far as safety, Waymo has avoided major accidents, injuries and fatalities so far, but its AI-enabled driver isn’t flawless. The company just recalled software in its fleet to fix a flaw that could cause vehicles to hit chains, gates and other barriers, following a NHTSA investigation. Cheaper Robotaxis In all the years Musk has promised autonomous Teslas and a robotaxi service, he hasn’t talked about what it’s doing to master ride-service essentials. But he does talk a big game about Tesla’s cost advantage. “The issue with Waymo’s car is it costs way more money,” the billionaire said on Tesla’s results call. “Their car is very expensive, made in low volume. Teslas probably cost 25% or 20% of what a Waymo costs and are made in very high volume,” last month. A base Model Y with FSD software costs consumers about $55,000 before taxes. While Waymo doesn’t disclose the cost of its modified, electric Jaguar I-PACE robotaxis, the lidar, radar, computers and other sensors mean it’s likely double that of Tesla’s vehicles. Those costs should drop substantially over the next year or so as Waymo shifts to lower-cost sensors and cheaper vehicles, including Hyundai’s Ioniq 5 and a small electric van from China’s Zeekr. Boasts about cheaper Tesla robotaxis will be meaningless if they can’t safely pick up and drop off riders without causing traffic jams, yielding to pedestrians or avoiding collisions. That’s made harder by the fact that Tesla uses eight 5-megapixel cameras as the main sensors for its system–far lower resolution than the 48-megapixel system on Apple’s iPhone 16. They’re inexpensive, but struggle with sunlight glare and low light conditions. Musk denied that was the case on Tesla’s April 22 call, but tests by O’Dowd’s Dawn Project after that found FSD disengages when directly facing the sun.“He thinks having [lidar] does not add enough benefit to outweigh the cost. This is a pretty typical engineering argument in general but incorrect in this particular case.” “We went out and took the car and drove it directly into the setting sun and guess what: it gave up,” O’Dowd said. “It starts flashing and it starts panicking, red lights going, it starts making noises, says put your hands back on the wheel.” By contrast, Waymo uses multiple sensors, including the much more expensive lidar, to ensure its vehicles see all potential road hazards, in daylight or at night, in 3D.Elon Musk attends a Cabinet meeting with President Donald Trump on April 30, 2025.The Washington Post via Getty Images “Musk has repeatedly said lidar is expensive and not needed,” said Missy Cummings, an artificial intelligence expert who advised NHTSA on autonomous vehicles. “He thinks having it does not add enough benefit to outweigh the cost. This is a pretty typical engineering argument in general but incorrect in this particular case.” After the Austin rollout, Musk said last month the goal is to expand to other U.S. markets, China and Europe, “limited only by regulatory approvals.” And one day soon, he envisions every person who owns a Tesla flipping a switch and deploying their car while not in use to a Tesla robotaxi network, helping them make additional cash on the side (as long as they pay Tesla $99 per month).“It’s all lies.” The world’s wealthiest person has achieved remarkable things with Tesla’s EVs, SpaceX rockets and Starlink satellites. But for years he’s also repeatedly failed to deliver big ideas he touted as potential game-changers or massive moneymakers, including battery swapping stations, solar tile roofs, the Hyperloop and high-speed underground transportation networks created by his Boring Co. Whether self-driving vehicles join that list remains to be seen. After repeatedly promising and failing to deliver vehicles that safely drive themselves for the past decade, let alone pick up riders, his track record isn’t looking good. Critics have a harsher interpretation. “It's all lies, everything he says,” said O’Dowd. More from Forbes

May 16, 2025Ravie LakshmananMalware / Cyber Attack Cybersecurity researchers have shed light on a new malware campaign that makes use of a PowerShell-based shellcode loader to deploy a remote access trojan called Remcos RAT. "Threat actors delivered malicious LNK files embedded within ZIP archives, often disguised as Office documents," Qualys security researcher Akshay Thorve said in a technical report. "The attack chain leverages mshta.exe for proxy execution during the initial stage." The latest wave of attacks, as detailed by Qualys, employs tax-related lures to entice users into opening a malicious ZIP archive containing a Windows shortcut (LNK) file, which, in turn, makes use of mshta.exe, a legitimate Microsoft tool used to run HTML Applications (HTA). The binary is used to execute an obfuscated HTA file named "xlab22.hta" hosted on a remote server, which incorporates Visual Basic Script code to download a PowerShell script, a decoy PDF, and another HTA file similar to xlab22.hta called "311.hta." The HTA file is also configured to make Windows Registry modifications to ensure that "311.hta" is automatically launched upon system startup. Once the PowerShell script is executed, it decodes and reconstructs a shellcode loader that ultimately proceeds to launch the Remcos RAT payload entirely in memory. Remcos RAT is a well-known malware that offers threat actors full control over compromised systems, making it an ideal tool for cyber espionage and data theft. A 32-bit binary compiled using Visual Studio C++ 8, it features a modular structure and can gather system metadata, log keystrokes, capture screenshots, monitor clipboard data, and retrieve a list of all installed programs and running processes. In addition, it establishes a TLS connection to a command-and-control (C2) server at "readysteaurants[.]com," maintaining a persistent channel for data exfiltration and control. This is not the first time fileless versions of Remcos RAT have been spotted in the wild. In November 2024, Fortinet FortiGuard Labs detailed a phishing campaign that filelessly deployed the malware by making use of order-themed lures. What makes the attack method attractive to threat actors is that it allows them to operate undetected by many traditional security solutions as the malicious code runs directly in the computer's memory, leaving very few traces on the disk. "The rise of PowerShell-based attacks like the new Remcos RAT variant demonstrates how threat actors are evolving to evade traditional security measures," J Stephen Kowski, Field CTO at SlashNext, said. "This fileless malware operates directly in memory, using LNK files and MSHTA.exe to execute obfuscated PowerShell scripts that can bypass conventional defenses. Advanced email security that can detect and block malicious LNK attachments before they reach users is crucial, as is real-time scanning of PowerShell commands for suspicious behaviors." The disclosure comes as Palo Alto Networks Unit 42 and Threatray detailed a new .NET loader that's used to detonate a wide range of commodity information stealers and RATS like Agent Tesla, NovaStealer, Remcos RAT, VIPKeylogger, XLoader, and XWorm. The loader features three stages that work in tandem to deploy the final-stage payload: A .NET executable that embeds the second and third stages in encrypted form, a .NET DLL that decrypts and loads the next stage, and a .NET DLL that manages the deployment of the main malware. "While earlier versions embedded the second stage as a hardcoded string, more recent versions use a bitmap resource," Threatray said. "The first stage extracts and decrypts this data, then executes it in memory to launch the second stage." Unit 42 described the use of bitmap resources to conceal malicious payloads a a steganography technique that can bypass traditional security mechanisms and evade detection. The findings also coincide with the emergence of several phishing and social engineering campaigns that are engineered for credential theft and malware delivery - Use of trojanized versions of the KeePass password management software – codenamed KeeLoader – to drop a Cobalt Strike beacon and steal sensitive KeePass database data, including administrative credentials. The malicious installers are hosted on KeePass typosquat domains that are served via Bing ads. Use of ClickFix lures and URLs embedded within PDF documents and a series of intermediary dropper URLs to deploy Lumma Stealer. Use of booby-trapped Microsoft Office documents that are used to deploy the Formbook information stealer protected using a malware distribution service referred to as Horus Protector. Use of blob URIs to locally loads a credential phishing page via phishing emails, with the blob URIs served using allow-listed pages (e.g., onedrive.live[.]com) that are abused to redirect victims to a malicious site that contains a link to a threat actor-controlled HTML page. Use of RAR archives masquerading as setup files to distribute NetSupport RAT in attacks targeting Ukraine and Poland. Use of phishing emails to distribute HTML attachments that contain malicious code to capture victims' Outlook, Hotmail, and Gmail credentials and exfiltrate them to a Telegram bot named "Blessed logs" that has been active since February 2025 The developments have also been complemented by the rise in artificial intelligence (AI)-powered campaigns that leverage polymorphic tricks that mutate in real-time to sidestep detection efforts. These include modifying email subject lines, sender names, and body content to slip past signature-based detection. "AI gave threat actors the power to automate malware development, scale attacks across industries, and personalize phishing messages with surgical precision," Cofense said. "These evolving threats are increasingly able to bypass traditional email filters, highlighting the failure of perimeter-only defenses and the need for post-delivery detection. It also enabled them to outmaneuver traditional defenses through polymorphic phishing campaigns that shift content on the fly. The result: deceptive messages that are increasingly difficult to detect and even harder to stop." Found this article interesting? Follow us on Twitter  and LinkedIn to read more exclusive content we post. SHARE    

OpenAI has introduced Codex, a cloud-native software engineering agent integrated into ChatGPT, signaling a new era in AI-assisted software development. Unlike traditional coding assistants, Codex is not just a tool for autocompletion—it acts as a cloud-based agent capable of autonomously performing a wide range of programming tasks, from writing and debugging code to running tests and generating pull requests. A Shift Toward Parallel, Agent-Driven Development At the core of Codex is codex-1, a fine-tuned version of OpenAI’s reasoning model, optimized specifically for software engineering workflows. Codex can handle multiple tasks simultaneously, operating inside isolated cloud sandboxes that are preloaded with the user’s codebase. Each request is handled in its own environment, allowing users to delegate different coding operations in parallel without disrupting their local development environment. This architecture introduces a fundamentally new approach to software engineering—developers now interact with an agent that behaves more like a collaborative teammate than a static code tool. You can ask Codex to “fix a bug,” “add logging,” or “refactor this module,” and it will return a verifiable response, including diffs, terminal logs, and test results. If the output looks good, you can copy the patch directly into your repository—or ask for revisions. Embedded Within ChatGPT, Accessible to Teams Codex lives in the ChatGPT interface, currently available to Pro, Team, and Enterprise users, with broader access expected soon. The interface includes a dedicated sidebar where developers can describe what they want in natural language. Codex then interprets the intent and handles the coding behind the scenes, surfacing results for review and feedback. This integration offers a significant boost to developer productivity. As OpenAI notes, Codex is designed to take on many of the repetitive or boilerplate-heavy aspects of coding—allowing developers to focus on architecture, design, and higher-order problem solving. In one case, an OpenAI staffer even “checked in two bug fixes written entirely by Codex,” all while working on unrelated tasks. Codex Understands Your Codebase What makes Codex more than just a smart code generator is its context-awareness. Each instance runs with full access to your project’s file structure, coding conventions, and style. This allows it to write code that aligns with your team’s standards—whether you’re using Flask or FastAPI, React or Vue, or a custom internal framework. Codex’s ability to adapt to a codebase makes it particularly useful for large-scale enterprise teams and open-source maintainers. It supports workflows like branch-based pull request generation, test suite execution, and static analysis—all initiated by simple English prompts. Over time, it learns the nuances of the repository it works in, leading to better suggestions and more accurate code synthesis. Broader Implications: Lowering the Barrier to Software Creation OpenAI frames Codex as a research preview, but its long-term vision is clear: AI will increasingly take over much of the routine work involved in building software. The aim isn’t to replace developers but to democratize software creation, allowing more people—especially non-traditional developers—to build working applications using natural language alone. In this light, Codex is not just a coding tool, but a stepping stone toward a world where software development is collaborative between humans and machines. It brings software creation closer to the realm of design and ideation, and further away from syntax and implementation details. What’s Next? Codex is rolling out gradually, with usage limits in place during the preview phase. OpenAI is gathering feedback to refine the agent’s capabilities, improve safety, and optimize its performance across different environments and languages. Whether you’re a solo developer, part of a DevOps team, or leading an enterprise platform, Codex represents a significant shift in how code is written, tested, and shipped. As AI agents continue to mature, the future of software engineering will be less about writing every line yourself—and more about knowing what to build, and asking the right questions. Check out the Details here. 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. Asif RazzaqWebsite |  + postsBioAsif Razzaq is the CEO of Marktechpost Media Inc.. As a visionary entrepreneur and engineer, Asif is committed to harnessing the potential of Artificial Intelligence for social good. His most recent endeavor is the launch of an Artificial Intelligence Media Platform, Marktechpost, which stands out for its in-depth coverage of machine learning and deep learning news that is both technically sound and easily understandable by a wide audience. The platform boasts of over 2 million monthly views, illustrating its popularity among audiences.Asif Razzaqhttps://www.marktechpost.com/author/6flvq/A Step-by-Step Guide to Build an Automated Knowledge Graph Pipeline Using LangGraph and NetworkXAsif Razzaqhttps://www.marktechpost.com/author/6flvq/Google DeepMind Introduces AlphaEvolve: A Gemini-Powered Coding AI Agent for Algorithm Discovery and Scientific OptimizationAsif Razzaqhttps://www.marktechpost.com/author/6flvq/Rime Introduces Arcana and Rimecaster (Open Source): Practical Voice AI Tools Built on Real-World SpeechAsif Razzaqhttps://www.marktechpost.com/author/6flvq/A Step-by-Step Guide to Build a Fast Semantic Search and RAG QA Engine on Web-Scraped Data Using Together AI Embeddings, FAISS Retrieval, and LangChain

Just in time to make the most of the warmer weather, REI is slashing prices on a bunch of our favorite outdoor gadgets and goods as a part of its annual anniversary sale. Now through May 26th, the outdoor retailer is taking up to 30 percent off a large number of items, from Garmin watches and water filters to stoves, satellite communicators, and one of the best camping / backpacking chairs around.RelatedBelow, we’ve rounded up some of the best deals to help you gear up for your summer adventures. In some instances, other retailers like Amazon are matching REI’s pricing, giving you additional ways to save. Also, keep in mind that REI members can save an additional 20 percent on one REI Outlet item or get 20 percent off on one qualifying full-price item when they use coupon ANNIV2025 at checkout.Garmin Forerunner 265 (42mm)$350$45022% off$350Garmin’s midrange Forerunner watch adds a new OLED display, a week’s worth of battery life, and dual-frequency GPS for better accuracy. Read our review.Many of our favorite Garmin smartwatches are also on sale right now, including the Garmin Forerunner 265, which is our top pick for runners. Normally $450, you can buy the smartwatch in either its 42mm or 46mm sizing at REI and Amazon for $349.99 ($100 off), which matches its best price to date. The 265 features a vibrant OLED display, up to 15 days of battery life, and multi-band GPS for more accurate tracking. It also provides insight into a number of running metrics, including stride length and cadence, while supporting Garmin’s PacePro, Training Readiness, and Race Predictor features.Garmin’s Fenix 8, our top pick for endurance athletes, is available in its 43mm configuration from REI, Amazon, and Garmin starting at $799.99 ($200 off). The platform-agnostic Fenix 8 is an excellent training watch with exceptional battery life that easily surpasses rivals like the Apple Watch Ultra 2. Garmin’s rugged, top-of-the-line wearable also includes a sharp OLED display and a built-in speaker / mic, so you can make calls or carry out voice commands using the onboard voice assistant (you can also connect it to your phone’s voice assistant for more complex queries). That being said, the Fenix 8 doesn’t offer built-in LTE, so your phone will need to stay within Bluetooth range for those features to work. Read our review.Garmin Fenix 8$800$100020% off$800The Garmin Fenix 8 adds smart features like the ability to take calls on the wrist, and an on-board voice assistant.Garmin’s 41mm Venu 3 is currently available for an all-time low of $349.99, down from its regular price of $449.99, at Amazon and REI. The Venu 3 is our favorite Garmin smartwatch, one that offers a mix of smart features and advanced health tracking at an affordable price. You can make and receive calls when paired with your phone, use your phone’s voice assistant via Bluetooth, and pay with contactless payments. It also includes a broad range of health features, including heart rate monitoring, period tracking, abnormal heart rate alerts, fall detection, and more. Just note: there’s no cellular connectivity, so you’ll need your phone nearby for full functionality. Read our review.Garmin inReach Mini 2$300$40025% off$300Garmin’s palm-size inReach Mini 2 satellite communicator provides off-the-grid contact with subscription plans starting at $14.95 a month.If your summer plans involve hiking or backpacking, the Garmin inReach Mini 2 can provide some peace of mind. And right now, it’s available from REI, Amazon, and Garmin for $299.99 ($100 off), which is just $15 shy of its all-time low. The rugged, palm-sized satellite communicator — which requires a monthly subscription — allows you to send texts, share your location, check the weather, and even trigger SOS alerts. It can easily outlast your average smartphone with satellite connectivity, too, so you’re free to enjoy the outdoors without having to worry about running out of power.The Helinox Chair Zero — a personal favorite of several Verge staffers — is on sale at REI and Amazon for $104.98 ($35 off), which is its best price to date. Weighing just a pound, the collapsible backpacking chair is exceptionally lightweight and easy to pack. It’s also built to support up to 265 pounds and is both comfortable and stable, despite its portable design.LifeStraw Peak Series Straw$18$2528% off$18LifeStraw’s basic filter removes unwanted viruses, bacteria, and microplastics as you drink, giving you a safe way to hydrate no matter where you are.LifeStraw’s portable water filters make for helpful travel companions, regardless of whether you’re heading into the backcountry or touring cities with questionable tap water. Thankfully, Verge favorites like the LifeStraw Peak Series Straw are on sale at Amazon and REI starting at $17.58, down from $24.95. The standalone straw is perfect for filtering out bacteria, parasites, and microplastics from streams and rivers. The Peak Series model features an improved flow rate and a more durable design than earlier models as well, and thanks to its versatile build, you can drink directly from a water source, attach it to a water bottle, or connect it to a gravity hose system.If you’re looking for a more convenient, all-in-one solution, the 22-ounce LifeStraw Go Series Water Bottle is also on sale for an all-time low of around $34 ($11 off) at REI and Amazon. The BPA-free bottle features a built-in two-stage filtration system, including a carbon filter that’s designed to reduce odors and improve taste, making it a solid pick for everything from travel to your daily commute.The MSR PocketRocket 2 Stove is on sale for $44.89, down from $55.95, at REI. The single-burner canister stove can boil water in just 3.5 minutes. It’s also lightweight, foldable, and compatible with most isobutane-propane fuel canisters, making it a great pick for backpackers looking to shed weight.Black Diamond Spot 400$37$5026% off$37Black Diamond’s Spot 400 headlamp offers great battery life and simple, easy-to-use controls. Best off all, it’s plenty bright, making it a great option for both indoor and outdoor use.$37 at REIIf you’re looking for a reliable and comfortable headlamp, Black Diamond’s Spot 400 is great for everything from camping to hiking to late-night reading sessions. It’s currently available from REI for $37.39 (about $13 off), which remains one of its better prices to date. The 400-lumen headlamp offers adjustable brightness, a red night vision mode, and an IPX8 waterproof rating, so it can handle rainy nights. It’s also dual-fuel compatible, meaning you can use standard AAA batteries or use it in tandem with Black Diamond’s rechargeable BD 1500 Li-ion battery for more flexibility.The pocket-sized Black Diamond Moji Lantern is down to $18.69 (about $6 off) at REI in multiple colors. The rugged Moji lantern offers 200 lumens of warm, dimmable light and adjustable lightness, along with an IPX4 rating for water resistance. It features a double-hook design, too, so you can use it as a table lantern or hang it up. The JBL Flip 6 is selling for $99.99 at REI and Amazon, down from $129.95 and about $20 shy of its all-time low. The portable Bluetooth speaker delivers powerful, clear sound and an IP67 waterproof and dustproof rating. The Flip 6 also offers up to 12 hours of playtime and features JBL’s PartyBoost feature, which allows you to pair multiple JBL devices together for a more powerful audio experience.Garmin’s Varia RTL515 Radar Taillight is on sale at REI and Amazon for $149.99 ($50 off), nearly matching its lowest price to date. The headlight combines a bright LED light with a rearview radar that, when paired with a compatible device, warns you of cars approaching from behind from up to 153 yards away.Coleman Cascade Classic Camp Stove$97$13025% off$97Coleman’s basic Cascade Classic Camp Stove features enough cooking power for most meals, two burners, and a matchless igniter.$97 at Coleman$97 at REIIf you want a straightforward stove that gets the job done, Coleman’s Cascade Classic Camp Stove is a good option that starts at $96.99 (about $33 off) when purchased directly from Coleman or at REI. The compact stove is pretty basic, but it offers everything you need for simple, reliable camp cooking. It’s got enough cooking power to handle most meals and two burners, so you can cook multiple things at once. It also features a matchless push-button igniter, along with built-in wind guards for better flame control and adjustable heat knobs for more precise cooking. It’s easy to carry and set up as well, with a durable steel cooktop and a removable chrome-plated grate for easy cleanup.You can buy MSR’s TrailShot Pocket-Sized Water Filter for $48.69, down from $64.95, at REI. The squeeze-and-sip filter removes bacteria, protozoa, and particulates, so you can drink directly from natural water sources or refill bottles. The filter is highly portable at just 5.2 ounces, too, and can easily fit in your pocket.The BioLite SolarHome 620 Plus kit is on sale for $139.93, down from $199.95, at REI. The compact, solar-powered system is designed to bring lights, charging, and audio to off-grid locations. Along with a 6W solar panel, it features an integrated FM radio / MP3 player, three hanging 120-lumen LED lights, and a pair of USB ports for charging various devices.See More: