Liquid Sword makes "deeply regrettable" job cuts for "the studio's long-term sustainability""This is incredibly difficult, but we remain committed to our vision" News by Vikki Blake Contributor Published on Feb. 17, 2025 NetEase-backed studio Liquid Swords has confirmed a round of cuts, calling the redundancies "deeply regrettable" but necessary for "the studio's long-term sustainability."Founded by Avalanche Studios' co-founder, Christofer Sundberg, Liquid Swords revealed earlier today plans to "reduce its workforce," attributing the cuts to "part of an effort to scale back operations and adapt to current business conditions."The Swedish studio said it "fully recognizes the difficulty of the situation for its valued team" and said it would be providing all colleagues with "as much support as possible," but stopped short of detailing what that support will entail.It also did not clarify how many employees had been impacted."I promised our employees and shareholders success and to build a studio redefining game development with a small, expert-driven team and a sustainable work model," Sundberg said. "While we achieved much of this, shifting market conditions prevented us from succeeding within our timeframe."As a result, we must part ways with talented individuals who have been instrumental in our journey. This is incredibly difficult, but we remain committed to our vision and will continue to work on our IP and first game with a dedicated team."Liquid Sword says it remains committed to Sundberg's vision of "simplifying the game development process."In the first six weeks of 2025 alone, over 900 developers have lost their jobs, with cuts and closures at Freejam, Splash Damage, Piranha Games, Jar of Sparks, as well as 185 jobs cut by Ubisoft. We also reported that ProbablyMonsters initiated another round of layoffs, as well further job cuts at Iron Galaxy and Sumo Group.

Its been nearly four years since Spotify announced a HiFi tier for its music streaming service that would support lossless audio. That wait could end this year, Bloomberg reports, as the company works to finalize details, including streaming rights.Spotify could charge as much as $5 or $6 extra per month for the new tier said to be named Music Pro which is in line with estimates CEO Daniel Ek shared last year. Spotifys Premium tier currently starts at $11.99 per month before family or student discounts, the result of two price hikes in as many years.Competitors like Apple Music include lossless streaming in their services base price, but Spotify reportedly plans to justify the extra charge by including more features in Music Pro, including song remixing and concert ticket sales that would grant subscribers exclusive deals and early access.Earlier this month, Spotify signed a new multi-year licensing agreement with Warner Music Group to secure future streaming rights and help shape the future of audio-visual streaming, with HiFi presumably included in that vision. It inked a similar deal with Universal Music Group in late January, with UMG specifically teasing HiFi streaming and offering further ideas of what a superfan music service could offer. Notable suggestions included priority access to deluxe edition content and early music releases, as well as exclusive listening party invites and artist Q&A sessions.Its taken a little longer than expected, but it looks like the worlds biggest music streaming service may finally be ready to turn up the volume on HiFi in 2025.

A team from Elon Musks SpaceX is visiting the Air Traffic Control Command Center in Virginia Monday to help overhaul the system in the wake of last months deadly air disaster in Washington, DC, US Secretary of Transportation Sean Duffy announced. The news comes after CNN reported that the Federal Aviation Administration fired hundreds of probationary employees who maintain critical air traffic control infrastructure.The exact number of workers losing their jobs is unknown, but the union representing them said it was in the hundreds. The Trump administration is in the process of trying to eliminate thousands of federal employees as it works with Congressional Republicans on a massive tax cutting bill that is said to favor mostly corporations and the wealthy.Elon Musk, the richest man in the world, is playing a key role in the mass terminations from his perch at the Department of Government Efficiency. And as critics have noted, Musks status as a major government contractor mostly through his company SpaceX represents a massive conflict of interest that both he and President Donald Trump have repeatedly attempted to downplay.In a post on X, Duffy said the team from SpaceX went to Virginia to get a firsthand look at the current system, learn what air traffic controllers like and dislike about their current tools, and envision how we can make a new, better, modern and safer system. Previously, Duffy said that Musks DOGE team would plug in to the FAA to help upgrade our aviation system.Duffy also dismissed criticism about opening the door to a Musk-led team to another sensitive area of the federal government. Because I know the media (and Hillary Clinton) will claim Elons team is getting special access, let me make clear that the @FAANews regularly gives tours of the command center to both media and companies, Duffy said. (Clinton has previously criticized the DOGE teams lack of experience.)I know the media (and Hillary Clinton) will claim Elons team is getting special accessThe FAA is under heightened scrutiny three weeks after a midair collision over the Potomac River resulted in the deaths of 67 people. The tragedy highlighted shortages of air traffic controllers as well as congestion at major hubs like Ronald Reagan National Airport. The FAA has fielded hundreds of complaints from air traffic workers describing dangerous conditions from staff shortages to dilapidated buildings. And the agency itself lacked a permanent head at the time of the crash mostly because Musk had a hand in ousting the last administrator after the FAA fined SpaceX for failing to submit safety data.Duffys post doesnt mention Musks role in the ouster, nor the hundreds of workers who were just laid off. CNN says the probationary employees were likely targeted because theyve been employed for less than a year and lack the right to appeal.This draconian action will increase the workload and place new responsibilities on a workforce that is already stretched thin, David Spero, National President of the Professional Aviation Safety Specialists, AFL-CIO, said in a statement. This decision did not consider the staffing needs of the FAA, which is already challenged by understaffing. See More:

Author(s): Reelfy Originally published on Towards AI. How AI-powered agents are reshaping content creation, engagement, and automation giving small businesses the power to compete like never before.A group of AI Agents collaborating to take actions. As imagined by Dall-E 3The Small Business Dilemma: The High Cost of Staying RelevantPicture this: A small business owner is juggling product development, customer service, sales, and operations and on top of that, theyre expected to maintain an engaging social media presence.Meanwhile, large brands have entire marketing teams crafting daily content, analyzing trends, and ensuring their audience stays engaged. The result? Small brands get drowned out in an algorithm-driven world that rewards consistency, creativity, and engagement.The harsh reality, the stats paint a clear picture: The average internet user spends 143 minutes per day on social media, yet organic reach is declining. Businesses must now take a multi-channel approach to stay relevant, but small brands lack the resources to produce constant, engaging content. (Sprout Social, 2024) 44% of consumers prefer to learn about a new product or service via short-form video content, but without AI-powered automation, small brands struggle to produce enough engaging video content to stay relevant. (Sprout Social, 2024 & HubSpot, 2025) 87% of marketers report increased sales from video marketing, yet many small businesses lack the time, skills, or budget to create professional videos. (Sprout Social, 2024 & HubSpot, 2025)Despite the obvious benefits of an active social presence, the cost of hiring a social media manager or agency is often too high for small businesses. Theyre left with two options:1.Do it themselves, burning time and energy that could be spent growing their business.2.Stay inactive, missing out on new customers and brand awareness.But what if we could automate social media strategy, content creation, and posting without sacrificing quality?AI Agents: Making Social Media Automation a RealityFor years, AI lacked the reasoning capabilities to replace human strategists in creative fields like marketing. But thanks to recent advancements, AI agents can now analyze, plan, execute, and adapt just like a human social media manager.Why Now? The Three Key AI Breakthroughs Cost Reduction: Open-source AI models like DeepSeek, Mistral, and LLaMA provide high-quality reasoning at a fraction of the cost of proprietary AI. Advanced Multi-Step Reasoning: AI agents can now think step-by-step, just like a human strategist, making them capable of adapting content to trends and audience behavior. Creative AI Pipelines: AI-powered tools like Reelfy enable end-to-end video creation, turning text prompts into engaging social media videos something only human designers could do before.The result? AI is no longer just an assistant it can be the entire social media team, handling everything from strategy to content creation, scheduling, and even performance analysis. By observing how posted videos perform, AI agents can refine their approach, adapting future content based on real-time insights.Understanding AI Agents: The Thought-Action-Observation CycleAt the core of our AI-driven social media agent is a concept called the Thought-Action-Observation (TAO) Cycle. This is how modern AI agents can reason, interact with the world, and learn from their actions just like a human strategist.This gif was introduced as part of the free Agent Course given by HuggingFace Step 1: Thought Internal Reasoning & Strategy (ReAct Approach)Before an AI agent can act, it needs to think. Using a ReAct (Reasoning + Acting) approach, our AI agent will: Analyze the brands messaging, past posts, and audience behavior using LlamaIndex. Research industry trends and competitor strategies using LangChain. Develop a content plan based on engagement patterns and social media trends.In this phase, the AI is essentially a strategist, planning what to post, when to post, and how to engage users. Example of AI Thought Process:This brand focuses on eco-friendly fashion. Based on recent social media trends, sustainable clothing recycling challenges are a hot topic. I should create content that highlights how this brand supports sustainable fashion. Step 2: Action Engaging with the WorldOnce the AI has a plan, it executes actions. Using SmolAgents, our agent will: Generate AI-powered video content with Reelfys story-to-video pipeline. Schedule and post content on Instagram, ensuring consistency. Monitor engagement, tracking likes, comments, and shares.At this stage, the AI isnt just a thinker its also a content creator and social media manager. Example of AI Action:I will generate a 15-second Instagram Reel showcasing sustainable fashion tips, using a story-to-video pipeline. Then, I will post it at 6 PM when engagement is highest. Step 3: Observation Learning & AdaptingAfter posting, the AI doesnt just move on it analyzes performance and adjusts future content strategies.Using real-time analytics, the agent: Tracks which content performs best. Analyzes engagement metrics to determine optimal posting times. Adjusts the content style and messaging for better performance.Over time, this means the AI improves its social media strategy, just like a human marketing expert would. Example of AI Observation & Adaptation:This video received 3x more engagement than last weeks. Users commented that they liked the storytelling format. I should create more narrative-driven content for the next post.Conclusion: The Foundation for AI-Driven Social MediaIn this first part, we laid the groundwork for how AI can fully automate social media marketing for small businesses. What we covered: The problem small brands face in maintaining a consistent online presence. Why AI agents are now capable of replacing human strategists at a fraction of the cost. The Thought-Action-Observation (TAO) cycle, which allows AI to reason, act, and learn.Whats Coming in Part 2?In the next part, well take these concepts and build a working AI social media agent step-by-step. Youll see: How the agent extracts brand identity and audience insights. How AI-powered storytelling generates engaging videos. How SmolAgents automate posting and engagement tracking. How AI refines future content based on performance metrics.With open-source AI, small businesses no longer need a full marketing team they just need the right AI agent. Stay tuned for Part 2, where we bring this AI-powered social media manager to life!Written by Garry NewballReferences[1] LangChain Conceptual Guide. Retrieved From: https://python.langchain.com/docs/concepts/[2] ReAct: Synergizing Reasoning and Acting in Language Models. Retrieved From: https://arxiv.org/abs/2210.03629[3] Llama Index Documentation. Retrieved From: https://docs.llamaindex.ai/en/stable/[4] DeepSeek Reasoning Model. Retrieved From: https://huggingface.co/deepseek-ai/DeepSeek-R1[5] Conceptual Guide to SmolAgents. Retrieved From: https://huggingface.co/docs/smolagents/conceptual_guides/intro_agents[6] Understanding the Importance of Social Media Marketing. Retrieved from: https://sproutsocial.com/insights/importance-of-social-media-marketing-in-business[7] 52 Visual Content Marketing Statistics You Should Know. Retrieved from: https://blog.hubspot.com/marketing/visual-content-marketing-strategy[8] 50+ Must-Know Social Media Marketing Statistics for 2024. Retrieved from: https://sproutsocial.com/insights/social-media-statisticsJoin thousands of data leaders on the AI newsletter. Join over 80,000 subscribers and keep up to date with the latest developments in AI. From research to projects and ideas. If you are building an AI startup, an AI-related product, or a service, we invite you to consider becoming asponsor. Published via Towards AI

Author(s): Yasin Yousif Originally published on Towards AI. A starter tutorial of its basic structure, capabilities, and main workflowImages are rendered from xml source in menagerie repo under BSD-3-Clause for Trossen, and Apache license for Franka and ApptronikMujoCo is a physics simulator for robotics research developed by Google DeepMind and written in C++ with a Python API. The advantage of using MujoCo lies in its various implemented models along with full dynamic and physics properties, such as friction, inertia, elasticity, etc. This realism allows researchers to rigorously test reinforcement learning algorithms in simulations before deployment, mitigating risks associated with real-world applications. Simulating exact replicas of robot manipulators is particularly valuable, as it enables training in a safe virtual environment and then seamless transition to production. Notable examples of these models include popular brands like ALOHA, FRANKA, and KUKA readily available within MujoCo.Table of Content:OverviewMJCF FormatThe TaskContinuous Proximal Policy OptimizationTraining ResultsConclusionOverviewBeyond the core MujoCo library (installable via pip install mujoco), two invaluable packages enhance its capabilities: dm_control (https://github.com/google-deepmind/dm_control) and mujoco_menagerie (https://github.com/google-deepmind/mujoco_menagerie).mujoco_menagerie offers a wealth of open-source robot models in .xml format, simplifying the simulation of complex systems. These models encompass diverse designs, as illustrated in the image above.In dm_control, (installable also with pip: pip install dm_control with its own version of MujoCo), a very useful code base is provided for creating Reinforcement Learning pipelines from MujoCo models as environments classes with suitable .step(), .reward() methods. These pipelines are available via its suite subpackage, and are intended to serve as benchmarks, on which different proposed Reinforcement learning methods can be evaluated and compared. Therefore, it should not be alerted when used for that purpose.These benchmarks can be shown by running the following:# Control Suitefrom dm_control import suitefor domain, task in suite.BENCHMARKING: print(f'{domain:<{max_len}} {task}')which will give the following domains and tasks among others:Additionally dm_control allow the manipulation of the MJCF models of the entities from within the running script, utilizing its PyMJCF subpackage. Therefore, the user doesn't need to change the xml files in order to add new joints, or replicate a certain structure for example.MJCF MuJoCoMuJoCo XML Configuration File Format MJCFTo show a working example of an MJCF file, we will review the car.xml source code available with MujoCo github repository here. It basically exhibits a triple-wheel toy vehicle, with two front lights, where it has two main degrees of freedom DoFs: Forward-Backward and Left-Right movements.By taking a look at the first part of the code we note that the code always is between <mujoco> ..</mujoco> tags. We also note the compiler tag defines what <compiler> is used (Euler by default) and allow setting its options.<mujoco> <compiler autolimits="true"/>Next, as some objects in the model may need its own customized texture and geometric shape unlike standard shapes such as spheres and boxes, the <texture>, <material> and <mesh> tags can be utilized as follows. We note also in the <mesh> tag that the exact points coordinates are provided in the vertex option, where each row represent a point in the surface.<asset> <texture name="grid" type="2d" builtin="checker" width="512" height="512" rgb1=".1 .2 .3" rgb2=".2 .3 .4"/> <material name="grid" texture="grid" texrepeat="1 1" texuniform="true" reflectance=".2"/> <mesh name="chasis" scale=".01 .006 .0015" vertex=" 9 2 0/> -10 10 10 9 -2 0 10 3 -10 10 -3 -10 -8 10 -10 -10 -10 10 -8 -10 -10 -5 0 20" </asset>The <default> tag is helpful to set some default values for certain classes, like the wheelclass which will be always of certain shape, size and color (defined with type, size, and rgba respectively)<default> <joint damping=".03" actuatorfrcrange="-0.5 0.5"/> <default class="wheel"> <geom type="cylinder" size=".03 .01" rgba=".5 .5 1 1"/> </default> <default class="decor"> <site type="box" rgba=".5 1 .5 1"/> </default></default>The first body in Mujoco models is always is the <worldbody> with the order of 0, as a parent object for all the other bodies in the model. Since we have only one car, the only children body of it should be car.Within each body we can define its children of other bodies, geometries, joint or lights, with <body>, <geom>, <joint>, <light> tags respectively.This is shown in the next snippet, in which we note the options of name, class and pos among others, which define the unique name, the defined class in default and the initial position of the parent tag respectively.<worldbody> <geom type="plane" size="3 3 .01" material="grid"/> <body name="car" pos="0 0 .03"> <freejoint/> <light name="top light" pos="0 0 2" mode="trackcom" diffuse=".4 .4 .4"/> <geom name="chasis" type="mesh" mesh="chasis" rgba="0 .8 0 1"/> <geom name="front wheel" pos=".08 0 -.015" type="sphere" size=".015" condim="1" priority="1"/> <light name="front light" pos=".1 0 .02" dir="2 0 -1" diffuse="1 1 1"/> <body name="left wheel" pos="-.07 .06 0" zaxis="0 1 0"> <joint name="left"/> <geom class="wheel"/> <site class="decor" size=".006 .025 .012"/> <site class="decor" size=".025 .006 .012"/> </body> <body name="right wheel" pos="-.07 -.06 0" zaxis="0 1 0"> <joint name="right"/> <geom class="wheel"/> <site class="decor" size=".006 .025 .012"/> <site class="decor" size=".025 .006 .012"/> </body> </body></worldbody>As the car can move in any direction including jumping and flipping with respect to the ground floor, it gets <freejoint/> tag with 6 DoFs. While each of its wheels: right and left wheels, get one DoF, along its previously defined axis with the zaxis="0 1 0"option the yaxis.The active control handles in MujoCo are defined with the <tendon> tag, defining group of joints as the <fixed> tag, and then with the <actuator> tag, to define the exact name and control range of the motor <tag>. As in the following code.<tendon> <fixed name="forward"> <joint joint="left" coef=".5"/> <joint joint="right" coef=".5"/> </fixed> <fixed name="turn"> <joint joint="left" coef="-.5"/> <joint joint="right" coef=".5"/> </fixed></tendon><actuator> <motor name="forward" tendon="forward" ctrlrange="-1 1"/> <motor name="turn" tendon="turn" ctrlrange="-1 1"/></actuator>This system of tendons conveniently control the car, as we can control the linear movement of the car, with the "forward" tendon, having forward displacement of 0.5 for both wheels, and the turning movement with the "turn" tendon, having opposite directions of displacement for each of the wheels, which physically will make the car turn.The degree of displacement is controlled with both of the defined motors, by multiplying their values with the coef values of the tendons.Lastly, the <sensor> tag defines the joints that should read, as generalized displacements value on its DoF.<sensor> <jointactuatorfrc name="right" joint="right"/> <jointactuatorfrc name="left" joint="left"/> </sensor></mujoco>The TaskIn order to train and run the reinforcement learning agent to control the car, we must set a clear purpose of the intended behavior. For example we may aim to make the car take a circular path or drive towards a fixed, but unknown position.For this example, we will define a reward so that the car drive from its initial position A=(0,0,0) towards B=(-1,4,0). This point is somehow to the left of the car, so it has to turn as well as drive in straight line, as shown below.Made by authorFor this task, we must define a reward function in relation to the euclidean distance between the current position of the car and the target position. We choose to take the exponent of the negative distance np.exp(-np.linalg.norm(A,B)) to represent this reward so that the values are always in the range [0,1].Continuous Proximal Policy OptimizationAs we noted in the XML file, the range of the actuators values is continuous, from -1 to 1. This means that the action space is continuous too; therefore, the training algorithm should be suitable to address these scenarios.This means that algorithms like DQN will not be suitable, as it can only be applied to discrete action spaces. However, actor critic methods, like PPO, can still be used to train models of continuous action space.The PPO code used here for this task is based on CleanRL single-file implementation for the continuous PPO, but with modifying some parameters and replacing the environment with our newly written environment enveloping the previous MujoCo model.Practically we train for 2e6 steps, with 2500 steps per episode. As the default update rate in MujoCo is 2ms, then 2500 steps translates to 5 seconds.It is worth noting that the discrete PPO update formulas are the same for the continuous case, expect for the type of the output distribution in the policy model, where it will be categorical Categorical in the discrete case, and Gaussian Normal, or any other continuous distribution in the continuous case.Next we will show the used environment for stepping and simulating the MujuCo model, which will be used for the training program of PPO.Training EnvironmentAs we will be using the main MujoCo package (not dm_control), we will do the following imports:import mujocoimport mujoco.viewerimport numpy as npimport timeimport torchWe then define the init method of the environment class, in which:The XML file of the model is loaded with the command: mujoco.MjModel.from_xml_path(), which will result in the model structure containing the geometries and constants such as time-steps, and gravity constant in model.opt.The data structure are loaded from the model structure with the command data = mujoco.MjData(model). In this structure, the current state values (like generalized velocity data.qvel, generalized position data.qpos, as well as actuator values data.ctrl), can be read and set.Duration is 5 seconds, which can be mapped to the simulation time by delaying it in specific amount, as the simulation is usually much faster. For example 5 seconds maybe simulated in 0.5 seconds.Rendering: if the render variable is set to True. A viewer GUI will be initialized with mujoco.viewer.launch_passive(model,data). The passive feature is needed so that the GUI doesn't block the code execution. However, it will be updated to the recent values in data when viewer.sync() is called, and it should be closed with viewer.close()class Cars(): def __init__(self, max_steps = 3*500, seed=0,render=False): self.model = mujoco.MjModel.from_xml_path('./car.xml') self.data = mujoco.MjData(self.model) self.duration = int(max_steps//500) self.single_action_space = (2,) self.single_observation_space = (13,) self.viewer = None self.reset() if render: self.viewer = mujoco.viewer.launch_passive(self.model, self.data)In reset() method, the data structure will be rested based on the original model using mujoco.mj_resetData.Here we can choose the shape of the state we will be using to solve our problem. We note as the task is only moving in 2D, therefore we need the current Cartesianposition of the car data.body('car').xpos, in addition to its orientation data.body('car').xquat, and lastly the velocity data.body('car').cvel also maybe helpful to judge if we want to accelerate of decelerate.Note that data.body() or data.geom() allow named access to these objects as defined in the XML file, or even by their index number , where 0 always indicate the worldbody.def reset(self): mujoco.mj_resetData(self.model, self.data) self.episodic_return = 0 state = np.hstack((self.data.body('car').xpos[:3], self.data.body('car').cvel, self.data.body('car').xquat)) if self.viewer is not None: self.viewer.close() self.viewer = mujoco.viewer.launch_passive(self.model, self.data) return stateAs our task is to reach the point [-1,4], then our reward can be as simple as the distance between the current position and the destination. However, taking exp(-distance) seems more suitable since it will restrict the rewards values to the range [0,1], which can lead to better stability in learning.As mentioned previously all we have to do to synchronize the change to the viewer window is to invoke the command self.viewer.sync().def reward(self, state, action): car_dist = (np.linalg.norm(np.array([-1,4]-state[:2]))) return np.exp(-((car_dist)))def render(self): if self.viewer.is_running(): self.viewer.sync()def close(self): self.viewer.close()In the step() routine, the actual model will be updated. First by setting the current action of the forward and turning movements in the data.ctrl. But it is noted that the action is transformed with the np.tanh() which has the output range of [-1,1]. This will allow the neural network of the policy to be trained on the full range [-Inf, Inf] for its output vector, which is easier to represent, as smaller values may get rounded during training.We additionally keep count of the episodic return, and handle the terminal case, by resting the environment.def step(self, action): self.data.ctrl = np.tanh(action) mujoco.mj_step(self.model, self.data) state = np.hstack((self.data.body('car').xpos[:3], self.data.body('car').cvel, self.data.body('car').xquat)) reward = self.reward(state, np.tanh(action)) self.episodic_return += reward done = False info = {} if self.data.time>=self.duration: done = True info.update({'episode':{'r':self.episodic_return,'l':self.data.time}}) info.update({"terminal_observation":state.copy()}) state = self.reset() return state, reward, done, infoThis finished the main environment class of the car model. It is not that complicated or hard to write. However, in dm_control a customized environments and pipelines with various tools are already available and ready to be used for training RL agents. An extensive topic, that is left for explorations in future posts.Training ResultsAfter training the PPO program with the previous environment and using a suitable agent network, we record the following training curve for the episodic return.Made by authorWe can see that the model is clearly learning, albeit slowly. There you have it. Your first simulated and controlled RL agent with MujoCo.However, we still need to see it in action: does the robot really move towards point [-1,4]? To do that we need to run the following testing program with the render variable set to True.def main(): duration = 5 env = Cars(max_steps=duration*500,render=True) #2000000 is the training iterations policy = torch.load(f'ppo_agent_cars_{2000000}_mlp.pth') state = env.reset() start = time.time() while time.time() - start < duration: with torch.no_grad(): action = policy.actor(torch.Tensor(state).to('cuda')).cpu().numpy()[:2] state, reward, done, info = env.step(action) if done: break time.sleep(0.003) env.render() env.close()After initializing the environment and loading the trained model with pytorch. We get the initial state by resetting the environment. Inside the while loop, we alternate between inferring the action from the actor model, and stepping the environment. Lastly we keep rendering each frame with env.render().If we run the program without any delay, we will get a very fast simulation that we may not be able to observe and depending on our while condition, it may get repeated many times before the program finishes.To avoid that, we delay the execution by some amount with time.sleep(). The program may still run several times (before duration seconds has passed), but it will be observable.In my case, this code shows the car moving exactly as shown in the image above (in The Task section), but as the speed is limited and the episode length is only 5 seconds, the simulation ends before reaching the point [-1,4], because reaching the point will be physically impossible in that case, no matter how long the model is trained.ConclusionWhile this tutorial merely scratches the surface of MuJoCos vast API capabilities, it equips you with the foundational knowledge to embark on your robotic simulation journey. MuJoCos C++ foundation enables lightning-fast performance, making it ideal for training intricate robots of diverse configurations.This versatility positions MuJoCo as a valuable tool in both research and industry:Research: Researchers can rigorously test and compare novel reinforcement learning algorithms within challenging, realistic scenarios without the logistical complexities and costs of physical prototyping.Industry: Manufacturers can thoroughly evaluate robot designs and models in environments mirroring real-world conditions, ensuring optimal performance before deployment.This Reinforcement and Imitation Learning series will delve deeper into specific, popular algorithms, exploring their intricacies and applications. Subscribe or follow along to stay informed and explore the full potential of these powerful techniques!Join thousands of data leaders on the AI newsletter. Join over 80,000 subscribers and keep up to date with the latest developments in AI. From research to projects and ideas. If you are building an AI startup, an AI-related product, or a service, we invite you to consider becoming asponsor. Published via Towards AI

You don't have to wait for the new 2025 LG TVs to drop to score a great deal on an older generation model. Right now as part of Amazon's Presidents' Day Sale, you can pick up a 2023 65" LG Evo C3 4K OLED TV for just $1,196.99 with free shipping on Amazon. This is a better deal than what I saw during Black Friday. The LG Evo C-series of TVs have consistently been our favorite high-end 4K TVs, especially for 4K HDR movies and gaming.65" LG Evo C3 4K OLED Smart TV for $119765" LG Evo C3 4K OLED Smart TVThe LG Evo C3 is a 2023 model, just one year behind the 2024 LG Evo C4. So what are the differences? Aside from the $650 price difference, not much. The most important trait -- picture quality -- is more or less identical. Where the C4 improves on the C3 are an upgraded processor, newer WebOS smart interface, and a higher 144Hz refresh rate.OLED TVs are considered the best TVs you can buy right now. Compared to traditional LED LCD TVs, they offer better image quality, deeper blacks, better contrast ratio, wider color gamut, and super fast response times. Because of these advantages, the OLED TV are easily the best type of TV for viewing 4K HDR content in all of its intended glory. LG OLED TVs particular have been out for years and benefit from several generations of optimizations.The LG C3 also has all the features you'd want in a gaming TV as well. It has a native 120Hz panel and all four HDMI 2.1 ports for running 4K at 120Hz on a PS5 or Xbox Series X. It also supports variable refresh rate (VRR), auto low latency mode (ALLM), and DTS audio, which is good for people who still watch Blu-ray discs. The C3 is also much easier to set up than its predecessors; the rear cabinet housing is made of a composite fiber that drops the weight to a mere 36 pounds. The LG Evo C-series TV is our favorite high-end 4K TV of 2025 because of the brilliance of its OLED display along with a host of quality features that don't quite push it to the point of an excessively high price. This model brings better contrast and clarity than the previous year's already luminous LG C2. Its a sight to behold, especially when you add in the deep blacks and well-balanced colors on the crisp 4K screen. Once you choose OLED, it's hard to go back to anything else.How Does This Compare to the Upcoming LG Evo C5?The LG Evo C5 hasn't yet been released, but it was showcased during CES 2025. At least on paper, the improvements of the C5 seem to be incremental, with no major ground breaking updates that would make you want to hold off until its launch. More importantly, the C5 will probably release at a very high retail price which will take months to get down to a level that's competitive with other TV deals.Looking for more options? Check out all of the best TVs of 2025.Why Should You Trust IGN's Deals Team?IGN's deals team has a combined 30+ years of experience finding the best discounts in gaming, tech, and just about every other category. We don't try to trick our readers into buying things they don't need at prices that aren't worth buying something at. Our ultimate goal is to surface the best possible deals from brands we trust and our editorial team has personal experience with. You can check out our deals standards here for more information on our process, or keep up with the latest deals we find on IGN's Deals account on Twitter.Eric Song is the IGN commerce manager in charge of finding the best gaming and tech deals every day. When Eric isn't hunting for deals for other people at work, he's hunting for deals for himself during his free time.

Secretlab has already kicked off its Presidents' Day Sale, which starts now and runs until February 17. Save up to $139 off Secretlab's popular Titan line of gaming chairs, Magnus gaming desks (including the Magnus Pro electric standing desk model), and accessories like the Secretlab Skins upholstery covers, desk mats, cable management, and more. Unfortunately, newer releases like the Titan Evo Nanogen chair and the recliner add-on are exempt from this sale. It's no secret that we love our Secretlab gaming chairs. Two of the six chairs in our best gaming chair roundup are Secretlab models. Of all the gaming chairs we covered in our "Budget to Best" roundup video earlier this year, my colleague Akeem Lawanson considered the Secretlab Titan Evo to be the most comfortable. No good chair comes cheap and Secretlab chairs definitely cost a premium, but we think the craftsmanship, materials, and customizability are worth it.TL;DR - The 7 Best Secretlab Deals Secretlab Titan EvoSecretlab Titan 2020Secretlab Titan Evo LitePreorder (Out November 24)Secretlab Titan Evo Nanogen Edition$799.99 at SecretlabNot Yet Available (Out 2025)Secretlab Titan Recliner Add-On$199.00 at SecretlabFixed Gaming DeskSecretlab MagnusElectric Standing DeskSecretlab Magnus ProYou can quickly browse through all of the listed products on sale above. For more information on each product and why they are worth your consideration, read through below.Secretlab Titan EvoSecretlab Titan EvoThe Titan Evo starts at $519 during the sale. This is Secretlab's flagship chair and it's available in small, medium, and large sizes. Upholstery options include Neo Hybrid leatherette, SoftWeave Plus fabric, or premium Napa leather. The chair features cold-cure foam upholstery for the seat, a supportive four-way lumbar system, full length backrest with 165 degrees of recline, full metal 4D armrests with magnetically attached PU cushions, and a memory foam headrest pillow.Aside from the build quality, the Titan Evo also stands out thanks to the sheer number of officially licensed designs from popular video games, TV shows, and more. Some of the more popular examples include The Witcher, Overwatch, Attack on Titan, League of Legends, World of Warcraft, and Game of Thrones. They generally cost more than the standard colors, but they're worth it if you're looking for that extra personal touch.In our Secretlab Titan Evo review, Chris Coke wrote that "after two years of daily use, the Secretlab Titan Evo has proven that it can stand the test of time and still be one of the best gaming chairs you can buy. Meaningful ergonomics paired with Secretlabs wide selection of designs, it remains a fantastic option, especially for fans of bright colors or designs."Secretlab Titan 2020Secretlab Titan 2020The prior model Titan 2020 gaming chair is available for $474, which is $45 less than the base model Titan Evo. The Titan 2020 is still an excellent chair and not much different than the current Evo model. In fact, outside of an upholstery change (the PU leather has been updated with Neo Hybrid Leatherette), the changes are mostly cosmetic. You are limited to fewer design options, so if you want to build out something that's truly unique, you might want to splurge a bit extra for the current generation Titan Evo model.Secretlab Titan Evo LiteSecretlab Titan Evo LiteAmong the Titan chairs, the Evo Lite is definitely the best value with its starting price tag of $419, or a full $100 less than the base model Titan Evo. It's built upon the same frame as the Titan Evo and has the same core features like the cold-cure foam cushioning, lumbar, 165 degrees of recline, and 4D armrests. What it compromises on is customization, with "only" two upholstery options, two sizes, and five colors, a non-adjustable lumbar system, simpler arm rests, and no included head rest. If none of these tradeoffs bother you, then you'll be saving quite a bit of money.Secretlab Titan Evo Nanogen EditionPreorder (Out November 24)Secretlab Titan Evo Nanogen Edition$799.99 at SecretlabAlthough the Titan Evo Nanogen Edition isn't on sale, it deserves mention simply because this is our top pick for the best gaming chair. In our Titan Evo Nanogen Edition review, Chris Coke wrote that "the Secretlab Titan Evo Nanogen Edition deserves every bit of the overwhelming praise Ive given. Granted, at $799 its significantly more expensive than the original and not far off from an entry-level Herman-Miller. But the return it offers in comfortable, supportive gaming is well worth the extra cost thanks to dramatically improved materials in both the fabric and multi-layered padding. The Titan Evo Nanogen Edition is class-leading, and is hands-down the most comfortable gaming chair Ive ever used."Secretlab Titan Recliner Add-OnNot Yet Available (Out 2025)Secretlab Titan Recliner Add-On$199.00 at SecretlabSecretlab also announced a new recliner add-on to anyone who already owns the Titan Evo chair. It's so new that not only will this recliner ship out sometime next year, it's not even available for preorder yet. We have received a unit for testing, however, and it has turned out to be a very practical addition. In our recliner add-on review, Chris Coke wrote that "while both comfort and value are subjective things, the recliner is able to take the Titan Evo and transform it from one of the best racing style gaming chairs to standing head and shoulders above the competition at its price point. Its novel enough that I wouldnt be surprised to see other brands following suit in the near future. If you dont mind paying for it, its an absolutely killer upgrade for your gaming chair."Secretlab Magnus and Magnus ProFixed Gaming DeskSecretlab MagnusElectric Standing DeskSecretlab Magnus ProThe Magnus and Magnus Pro are also on sale during Cyber Monday. The Magnus is a traditional fixed-frame gaming desk while the Magnus Pro ups the ante with a custom designed electric standing desk frame for an additional $250. Both desks feature an all-metal desktop surface, solid steel legs and cleverly thought out areas for cable management, but the Magnus Pro has some really unique features including a power cable that runs internally inside one of the telescoping legs and an in-line control panel that you won't bump into.In our Secretlab Magnus Pro review. Mark Knapp writes that "the Secretlab Magnus Pro is a fantastic desk, bringing the brilliant cable management solutions of the original Magnus to a fast, quiet, and wide-ranging motorized standing desk. The desk is built well and proves an excellent platform for work and play alike. Its an expensive desk though, and for the money, it would have been nice to see a smarter safety mechanism for the motors and the desk mat included. Still, the overall quality you get is a big step up from cheaper standing desks, and the optional accessories truly enhance the experience. Anyone whos not committed to a standing desk should save their money and go for the standard Magnus if everything else about this model sounds good, but for gamers who love a tidy desk and want the flexibility of a standing desk, the Magnus Pro should be the first they consider."PlayWhy Should You Trust IGN's Deals Team?IGN's deals team has a combined 30+ years of experience finding the best discounts in gaming, tech, and just about every other category. We don't try to trick our readers into buying things they don't need at prices that aren't worth buying something at. Our ultimate goal is to surface the best possible deals from brands we trust and our editorial team has personal experience with. You can check out our deals standards here for more information on our process, or keep up with the latest deals we find on IGN's Deals account on Twitter.Eric Song is the IGN commerce manager in charge of finding the best gaming and tech deals every day. When Eric isn't hunting for deals for other people at work, he's hunting for deals for himself during his free time.

Anybody missing Luther could do worse than tune into new Bradford-set crime drama Virdee. DCI Harry Virdee may not have quite as brilliant an intellect as DCI John Luthers, but he operates in a similarly grey area between criminality and the law, is just as handy with his fists, and faces a villain in series one whose methods are so grotesque and lurid that they could have stepped straight out of the Idris Elba-led drama. You thought Luthers twins were deranged? Stay tuned.What sets Virdee apart from Luther, aside from their differing British cultural contexts (Harry is Sikh with Indian heritage and lives in Bradford, John is a Black Londoner), are their marriages. Unlike estranged and then tragically widowed Luther, Harry is very much a team with his wife Saima, even if he keeps things from her. Their love and young son are the bedrock of both their lives, and hopefully what will keep Harry Virdee from going too far down a dark pathMeet the Virdee characters and see where you might recognise the actors from below.Staz Nair as Harry VirdeeDCI Hardeep Harry Virdee is a 39-year-old Bradford cop with conflicted loyalties to his beloved wife Saima and young son Aaron, his brother-in-law Riaz, his job, and his city. Hes the lead character in AA Dhands five-strong Harry Virdee book series (Streets of Darkness, Girl Zero, City of Sinners, One Way Out, The Blood Divide) and is played in this BBC One series by actor and singer Staz Nair. TV fans will recognize British actor Nair for his roles in US dramas Game of Thrones, in which he played Dothraki general Qhono; Krypton, in which he appeared as Dax-Baron, and for his regular role in Supergirl as William Dey, and in Zack Snyders Rebel Moon. He also appeared as Rocky in a 2016 TV version of The Rocky Horror Picture Show, and was formerly a member of The X Factor band Times Red.Vikash Bhai as Riaz HyattRiaz is a Bradford crime kingpin, and Harrys brother-in-law. The two grew up together, and despite going in different directions after Riaz was sentenced to prison, the bond between them remains unbeknownst to Harrys wife and Riazs sister Saima. Hes played by Vikash Bhai, an actor familiar to fans of sci-fi series Pandora, as well as BBC thriller Crossfire, US sci-fi series Hanna, and many more. Bhais voice might also be recognised by listeners to Big Finishs audio Doctor Who adventures, of which hes recorded many.Aysha Kala as Saima HyattSaima Hyatt is a nurse, mother of young son Aaron, married to Harry, and the sister of drug kingpin Riaz not that she knows that her brothers operating an organised crime group out of his cash-and-carry warehouse. Clever and independent, Saima is a proud Muslim of Pakistani heritage whos not prepared to compromise her faith to placate her bigoted father-in-law. Shes played by screen and stage actor Aysha Kala, seen recently in Apple TV+ crime drama Criminal Record and known previously on TV for ITVs Indian Summers, as well as recent National Theatre roles in The Motive and the Cue, and The Father and the Assassin.Kulvinder Ghir as Ranjit VirdeeRanjit is Harrys father, though he hasnt seen him or met his grandson Aaron for eight years by the time that Virdee begins. An Indian Sikh who holds a strong prejudice against Pakistani Muslims, he disowned his son when he married Saima, and refuses to acknowledge him to this day. Hes played by Kulvinder Ghir, a very familiar face on British TV for his time as a castmember on beloved comedy series Goodness Gracious Me, as well as Beecham House, Still Open All Hours, and recently, Apple TV+ sci-fi Foundation.Tomi May as Enzo TobinEnzo is Riazs right-hand man in Bradford West, and the one who gets his hands dirty when violence is called for in the fight against Vasil Sharmas rival gang. Line of Duty fans will recognise actor Tomi May as having played Miroslav Minkovicz, a member of the organised crime group being hunted by AC-12s police officers. May has also appeared in Killing Eve, The Trouble With Maggie Cole, The Man Who Fell to Earth, Headhunters and an episode of the videogame-to-TV adaptation Halo.Danyal Ismail as DS Khalil AminNo TV crime drama would be complete without a new DS for our lead to show the ropes and explain things to (and by extension, also to us) along the way. In Virdee, thats Khalil, a new recruit to Harrys Bradford team. Khalil quickly understood that his new boss who doesnt do desks also doesnt quite follow the rules, and the question is, will he support or report Harry for it? This is Ismails fourth crime drama TV role, following parts in ITVs Vera, Madonald & Dobbs and Ridley.Elizabeth Berrington as DS Clare ConwayDS Conway is Harrys police colleague (and are we sure that shes only a DS as shes credited? She acts more like DCI Virdees boss). Shes a supporting character about whom little is known, and shes played by Elizabeth Berrington. Where have you seen Elizabeth Berrington before? Everywhere. From The Office to Waterloo Road to Stella to The Responder to Good Omens, via basically every British TV show made in the last 20 years, Berringtons been in it.Join our mailing listGet the best of Den of Geek delivered right to your inbox!Elaine Tan as Rebecca ArmitageRebecca Armitage is part of the UK Crime Agency, a fictional organisation in Virdees world. The UKCA is called in to take over a high-profile investigation into a ritualistic killer. Shes played by Elaine Tan, who recently appeared inRed Eye and Sky sci-fi drama The Lazarus Project, and before that Tom Clancys Jack Ryan, Acquitted and multiple episodes of British soap EastEnders in the role of Li Chong. Nicola Burley as Sophie BrodenhamSophie is a mystery at the start of Virdee, but is soon established as Riazs confidante. Her backstory will be unravelled as the series approaches its finale. Shes played by Nichola Burley, who was recently in ITVX crime drama Protection, as well as playing Brenda in BBC true-crime drama The Gold, appearing in Netflix fantasy thriller Behind her Eyes, and many more.Ramon Tikaram as Jai PawaNo spoilers here for anybody who hasnt yet binged all episodes of Virdee series one. Jai Pawa is a powerful figure from Virdees past who returns to Bradford set on vengeance. Hes played by another very familiar face on British TV: Ramon Tikaram, seen recently in Netflix fantasy KAOS, but also Brassic, Pennyworth, multiple Doctor Who audio adventures, Stella, EastEnders, and many others, including, of course, for the role of Ferdy in 1990s favourite This Life. ALSO APPEARING The Lazarus Project and Waterloo Roads Nina Singh as Harrys niece Tara Virdee-Duggal. We Are Lady Parts, Mary Poppins Returns and theatre actor Sudha Bhuchar as Harrys mother Jyoti Virdee Gangs of London and The Gentlemens Andi Jashy as Vasil Sharla, the leader of a rival drug operation going up against Bradford West. Newcomer Charlie Mann as Paul King, a local Bradford thief. All episodes of Virdee are streaming now on BBC iPlayer.

Warning: contains Virdee finale spoilers.A rooftop. A banging view over Bradford city centre. Two compostable cardboard cups of chai, and a steely nod. Thats how BBC One crime thriller Virdee, based on the five-book series by AA Dhand, came to an end. Lead DCI Harry Virdee (Staz Nair) had survived a deranged campaign of murder from a past enemy, but at what cost? His wife Saima (Aysha Kala) no longer trusts him, and hes now even more indebted to his criminal kingpin brother-in-law Riaz (Vikash Bhai) than before.One positive for Harry was a tentative reunion with his estranged father, an Indian Sikh whod cast Harry out from the family eight years ago for marrying Saima, a Muslim woman with Pakistani heritage. After refusing to ever acknowledge his young grandson Aaron, Ranjit Virdee bent to the pleas of their other family members, and eventually swallowed both his pride and his prejudice to meet with Harry and to clasp his hand.With major final spoilers, lets dig in to the Virdee ending and see where it leaves Harry, Saima, Riaz and co. for a potential second series.Harry Now Officially Part of Bradford West?Harry had been indebted to Riaz since hed had taken the blame for a murder Harry committed in self-defence in 2001. During that years Bradford Riots, Riazs mother was knocked unconscious by racist thief Paul King, whom she and Riaz had caught stealing from the till of their family shop. Paul attacked Riaz, who froze, and Harry arrived and protected him, accidentally stabbing King in the neck with a pair of scissors during the fight. To maintain his honour, Riaz insisted on lying that he was the killer.Riaz went to prison and came out a criminal who worked for his drug-dealing former cellmate Jai Pawa (Ramon Tikaram). He, Pawas wife Sophie (Nicola Burley) and Harry now a police officer informed on Pawa, who was imprisoned in India. Unbeknownst to Harry, during the operation to arrest Pawa, Riaz and Sophie had stolen 30 millions worth of his heroin to set up their own drug-dealing operation.In the years since, Harry and Riaz have maintained an uneasy truce, communicating via secret burner phones and using one another for information. In series one though, Riazs right-hand-man Enzo insisted that if the crew were to help Harry save his kidnapped wife Saima, they needed him to come over to their side and work for them fully (a bit rum, considering that Saima is also Riazs sister). Harry agreed, which means that in a potential second series, hell be a fully corrupt officer working for Riazs organised criminal gang.DS Khalil Amin: Spy?Harry being a Bonafide member of Bradford West in future is a problem because in the final episode of series one, Harrys new partner DS Khalil Amin was recruited by the United Kingdom Crime Agency to inform on corrupt activity in Bradford Police. Harry is therefore working with somebody tasked with rooting out corrupt officers in the pay of organised crime groups. Khalil already knows that Harry has a secret source of criminal intel because in episode one, he realised that Harry didnt get Ateeqs location from the low-level drug dealer he pretended to have heard it from. The question is: will Khalil shop his partner to his new boss?What Did Jai Pawa Want?Revenge against everybody who had wronged him, and to use his intel on Harrys murder of Paul King (see above) to destroy the career of the officer who arrested him. When Riaz was inside for Kings murder, he shared a cell with Pawa and told him the truth that Harry was actually Kings killer and that he had taken the blame. Pawa used this to try to coerce Harry into confessing to murder as part of his vengeance plan.Pawa first bribed a prison officer in the Punjab to help him escape his cell, and then travelled to England to commit a campaign of terror against the people who had betrayed him and orchestrated his imprisonment: the accountant whod given evidence against him in court; Harry, the police officer whod made his name by arresting him; Bradfords Chief Constable Jonthan Boardman; Riaz, the lieutenant who stole from him and usurped his position in the criminal world; and Sophie, his ex-wife who was not only part of Riazs scheme to get him out of the picture, but also Riazs secret lover.Pawa killed his former accountant and set up shop in the basement of his house, breeding parasitic wasps he then used ritualistically in a series of murders. He killed Priti Parmer, a young drug dealer whod recently switched sides from Vasil Sharmas organisation to work for Riaz and Bradford West, then he released a toxic gas into an LGBTQ+ nightclub, kidnapped the Chief Constables son Alistair, and kidnapped Harrys reporter niece Tara, and left a message to Harry and Riaz on her skin reading 420 the address of the old corner shop where Harry had killed Paul King. Harry saved Tara and Alistair, but Pawa had also kidnapped Harrys wife Saima, and threatened to kill her unless Harry confessed to Kings murder.Join our mailing listGet the best of Den of Geek delivered right to your inbox!Saima escaped, and Riaz shot Pawa dead, then placed his body next to that of Vasil Sharma, whom Riaz had also shot dead, to stage the murders as drug-on-drug violence. Now Riaz and Sophie are the only kingpins in Bradford left standing.Will Riaz Regret Sparing Ateeq Farooqis Life?The teenage drug-dealer whom Harry saved from the Vasil Sharmy crew in episode one is a tricky character whose loyalties are unclear. After his kidnap, he petitioned to work for Bradford West, earning a face-to-face meeting with Riaz by helping Enzo (Tomi May) to steal over 1 million from the rival gang. Riaz took a chance on him, but the whole plan was a set-up so that Vasil Sharmys gang could take out Riaz at a time and place set up by Ateeq (who clearly had misgivings about it, but went along with it anyway). Ateeq arranged the hit, Sharmys assassin shot at Riaz, but didnt kill him. Riaz shot Sharmy in retaliation but spared the life of Ateeq because of his moral rule about not involving underage kids in criminal activity. Now that Riaz is the biggest fish in Bradfords criminal pond, will Ateeq prove a loyal lieutenant or a dangerous rival?What Next for Harry and Saima?What sets DCI Harry Virdee apart from DCI John Luther (his predecessor in the handsome-detectives-who-operate-outside-the-law genre) is his functional, loving and supportive marriage with Saima. Intelligent and independent, Saima did her own sleuthing in Virdee series one, and discovered that her brother Riaz wasnt just a successful cash-and-carry owner but a drug-dealer and murderer. She also discovered that her husband Harry knew all about Riazs criminal operation, and the two had been secretly swapping intel.At the end of series one, Saima and Aaron are living separately from Harry, and she tells her husband I dont know who you are anymore, I dont think I ever did. Saima doesnt yet know that Riaz took the blame for Paul Kings murder, but Harrys niece Tara heard him and Riaz talking about King, so is close to discovering the truth.Harry Reunited With his FatherAfter being petitioned by Saimas father, Harrys dad Ranjit finally conceded to reunite with the son hed banished from the family years earlier for marrying a Muslim woman of Pakistani heritage. In Virdees final moments, Harry and his dad sat together on a park bench. You made me a stranger, Harry told his father, and the pair clasped hands, suggesting a brighter future for the father and son. A second series of Virdee hasnt been officially announced at the time of writing, but well keep you posted if that happens.All episodes of Virdee are streaming now on BBC iPlayer.

At a glance, the most notable change to Siri in iOS 18 is its new design. The floating orb has been replaced with a rainbow glow around the edge of the device. This signals that Siri is now powered by Apple Intelligence, but the biggest AI features for Siri havent rolled out yet. Thats about to change in the next few iOS 18 updates.Apple lists 13 new Siri features in iOS 18. Most of these have already shipped, but it may not feel like the big iOS 18 update for Siri has arrived.I think thats because the biggest AI-related Siri feature so far is handing off a request to ChatGPT. Im a big ChatGPT user, but using Siri isnt one of my top three ways of interacting with ChatGPT. These are the 10 new Siri features that are included in iOS 18 so far:More Resilient Request HandlingIf you stumble over words or change your mind mid-sentence, Siri will still follow along.For example, you might say:Siri, set an alarmwait no, sorry, I meant a timer for 10 minutesactually, lets make that 15.Siri will adjust accordingly without needing a new command.Maintains Conversational ContextSiri remembers recent requests to make follow-up interactions easier.For example, if you ask, When are the Warriors playing next? and then say, Add that to my calendar, Siri will understand what that refers to.Glowing Edge LightA new glowing edge light wraps around the screen when you speak to Siri. This responsive animation follows your voice and allows you to keep scrolling, typing, or navigating your device while using Siri.Predictive Text in Siri KeyboardAs you type a request to Siri, predictive text will help complete your request faster. This model is designed specifically to speed up Siri interactions.More Natural VoiceSiri now sounds more natural, expressive, and clear. The voice is synthesized entirely on-device using Apple Intelligences new language models.Product KnowledgeSiri can answer thousands of questions about Apple product features and settings by referencing Apple support documentation and large language models.Type to SiriYou can now type to Siri at any time. All voice commands can also be input via text, and you can switch between text and voice seamlessly.Suggestions in Siri KeyboardSiri now provides suggested requests above the keyboard, offering ultrafast access to common actions.More Visually Rich ResponsesSiri responses now match the look and feel of the app they reference. For example, when asking about the weather, the response will appear as if it were directly from the Weather app.ChatGPT IntegrationSiri can tap into ChatGPT to assist with certain requests when it determines ChatGPT may provide a more helpful answer.Additionally, you can ask about photos, documents (including PDFs and presentations), and more. Youll always be asked before any information is sent to ChatGPT. Siri will then present the response in a seamless manner.If it feels like Siri is mostly just the same in iOS 18, thats because the three biggest upgrades are still being developed. While weve expected these Siri features to arrive in iOS 18.4 beta as soon as this week, we now expect some to be held back until iOS 18.5. At any rate, these are the three Siri upgrades remaining for iOS 18:Personal Context UnderstandingWith the on-device semantic index powered by Apple Intelligence, Siri can understand emails, messages, photos, calendar events, files, and more. This allows Siri to provide answers tailored to you.For example, when you ask, What was that movie that Jamie told me I should check out? Siri will retrieve the relevant message from last week. Similarly, if you say, Add my passport number here, Siri can extract the number from a saved photo of your passport and insert it for you.Onscreen AwarenessSiri can now understand and interact with content on your screen. If a friend texts you about a new coffee shop in your neighborhood, you can simply ask, How long would it take me to walk there from home? and Siri will process your request based on the message.In-App ActionsSiri can perform hundreds of new actions across both first- and third-party apps. While editing a photo, you can say, Make this photo warmer, and Siri will adjust the settings in the Photos app. Siri can also operate across appsafter editing, you can say, Add this to my ratatouille recipe note, and Siri will seamlessly move from Photos to Notes to complete the task.9to5Macs TakeWhich of these features will arrive in iOS 18.4 beta? We should find out soon. But how well they work is another question.It feels like a make-or-break moment for Siri, and these three remaining iOS 18 features are key indicators of Apples ability to leverage AI with Siri.These three Siri features are AI uses that only Apple can provide not because of Apples AI talent, but because of Apples position as the platform operator.Apple needs to finalize these ambitious features before making any big promises for Siri and Apple Intelligence with iOS 19. WWDC 2025 is less than four months away.Siri should be the best voice assistant available on the iPhone.Its the only one with deep device and operating system integration. Thats even more reason for Apple to be a strong player in AI and not just a platform for innovative technologies. If Siri cant keep up, a strong case will mount for Apple to support changing the default voice assistant and allowing competing voice assistants to hook into iOS. Otherwise, standalone AI devices will actually have a reason to exist.Limited integration with temporary ChatGPT requests isnt enough. In time, the iPhone could look less appealing to someone who values having the best AI-powered voice assistant more than blue iMessage bubbles.Best Apple accessoriesFollow Zac Hall on X, Threads, and Bluesky, and listen to Runtime with co-host Sophia Tung on Apple Podcasts and YouTube.Add 9to5Mac to your Google News feed. FTC: We use income earning auto affiliate links. More.Youre reading 9to5Mac experts who break news about Apple and its surrounding ecosystem, day after day. Be sure to check out our homepage for all the latest news, and follow 9to5Mac on Twitter, Facebook, and LinkedIn to stay in the loop. Dont know where to start? Check out our exclusive stories, reviews, how-tos, and subscribe to our YouTube channel