I've been following Eight Sleep products for a few years and was excited to learn that the company is entering the UAE and Saudi markets. As the name implies, the company makes products that help you sleep better at night. The Eight Sleep Pod 4 is a smart mattress cover that adjusts temperature, tracks sleep metrics, and detects snoring. With summer approaching and the AC going into full blast, the Sleep Pod 4 could be a great solution for those with partners who have different temperature tolerances. Pricing and Availability The Eight Sleep Pod 4 (which is now replaced by the Eight Sleep Pod 5) is available in regular and Ultra varieties. The regular version comes with the Pod and bed cover and costs AED 9,999 for a queen size (160 x 200 cms), AED 10,799 for a king size (180 x 200 cms), and AED 11,799 for an Emperor size (200 x 200). The Ultra version adds an elevating base that lifts the mattress into positions ideal for sleeping, reading, or relaxing. This adds about AED 8,000 to the base prices. Eight Sleep sent me the non-ultra version for this review. In addition to the price of the Sleep Pod, you have to pay AED 65 per month for AutoPilot features that adjust temperatures automatically, let you set alarms, and provide sleep and health reports. I think Sleep Eight should bundle at least one year's worth of subscription with any Pod purchase. Key Features The Pod 4 offers advanced sleep technology through a mattress cover compatible with existing beds. Its main features include: Temperature Regulation: Dual-zone climate control adjusts each side of the bed to as low as 13°C. The system uses water flow to regulate the temperature of your mattress. Sleep Tracking: Health-grade sensors monitor heart rate, heart rate variability (HRV), respiratory rate, and sleep stages. Snore Detection: The Pod 4 vibrates to identify snoring and alerts users through the app. The more expensive Pod 4 Ultra model automatically elevates the bed to reduce snoring. Autopilot AI: The algorithm personalises temperature and elevation based on bio feedback, user preferences, and data from other users. GentleRise Alarm: Vibration and thermal changes wake users gently, replacing traditional alarms. App Integration: The Eight Sleep app provides sleep insights, temperature scheduling, and control over settings. Design and Build Quality The Eight Sleep Pod 4 includes a mattress cover and a special fitted sheet covering your mattress. It is made of breathable fabric, which is comfortable to lie on, though you'll likely add a bedsheet above it. You can get this in multiple sizes. I was sent a 180x200 cms sized cover for my king-sized bed. This cover has a mesh layer below it that circulates water to make your bed cooler or warmer. This layer also has sensors that track how you sleep. Finally, there's the Bedside Hub, a desktop PC-sized box that controls the flow of water that cools or heats the Active Grid. It also connects to your Wi-Fi network. The mattress cover is connected to the Pod using a rather thick cable that I had concerns with but it tucked away easily and hasn't caused any concerns in the two months I've been using it. Installation requires water for the Hub's tank, which is supposed to be refilled every few months. The setup took me about an hour, and it is easier with two people to lift your mattress. I had to fill the water container in the bedside hub three times until it was properly dispersed to the mesh. The app guides you through all of this, starting with connecting your Pod to your Wi-Fi, which didn't go as planned. The Wi-Fi performance on the Hub is not great, and I had to move an access point into my bedroom for it to maintain a good connection. Features and Usage Once you set up the Pod, everything else is controlled and managed through the app, which is available for iPhone and Android phones. The app underwent a major overhaul during my testing and now looks more modern and streamlined.Recommended by Our Editors Using the app, you can schedule temperature changes, view sleep reports, and adjust all the settings for the Pod. You can set sepecific temperature for bedtime, later at night and at dawn. I set the schedule for my bedtime, and the Hub went into action about half an hour before that, cooling it to my desired temperature which is 2 degrees below the room temperature. You can also set the temperatures to be absolute values, such as 18 degrees. When I first started using the Eight Sleep Pod, I had set the temperatures at 19C but that proved to be very cool for my liking. After a few days of fiddling, I settled on cooling the nighttime temperature to 2 degrees lower than the bedtime and the dawn temperature to be 2 degrees higher than the room temperature. With AutoPilot, the bed automatically changes the temperature by a couple of degrees to help you get the best sleep when it detects you in REM or deep sleep states, which theoretically improves your readiness for the next day. You can also manually adjust the temperature by double or triple tapping your side of the bed to cool or warm it up. Another function of the Eight Sleep is to provide sleep tracking, which sounds great as you won't need to wear a device like the Apple Watch or Oura Ring to bed. I use an Oura Ring that I usually wear to bed, and I compared the stats it offered to the Eight Sleep. Benchmark Oura Ring Eight Sleep Pod 4 Time Slept 5h 43m 4h 19m Deep Sleep 40m 46m REM 50m 1h 15s Resting Heart Rate 73bpm 74pm While this is the data for just one night, during my two weeks of testing, I found that the Oura Ring's results were more consistent than those of the Eight Sleep, and there's a good reason why. The Eight Sleep mattress cover is split into two parts to track you and your partner. If I moved towards my partner's side or if she moved towards my side, the data would not be analysed properly. Similarly, the data would be completely thrown off if our kid decided to jump in the bed in the middle of the night, which happened quite a few times while I was testing. So, while sleep tracking is a good secondary feature of the Eight Sleep Pod, it should not be the main reason to get one unless you sleep alone. Also worth nothing that the app does not sync your sleeping information with Apple Health, however it does work with Alexa if that's your preferred platform. One more feature of the Eight Sleep is a wake-up alarm that vibrates on your side of the bed. The vibrations are meant to go from light to strong to gently wake you up, along with raising the temperature of the Pod. While this feature woke me up every single time, it also woke my partner up because the vibrations carry off to the other half of the bed even at the lightest settings. The Verdict The Eight Sleep Pod 4 delivers excellent temperature regulation, and that should be your primary objective when purchasing this Pod. While you also get sleep tracking and a wake-up alarm, these features work much better if you're the only one sleeping on the bed. The unit is also quiet in operation, and Autopilot AI works well to enhance your sleep. However, it is an extremely expensive piece of equipment, and additional subscription charges make it less appealing to users who are careful about their finances. And while the app is pretty good, it doesn't sync with Apple Health. The high costs and limited integrations pose drawbacks, but the 30-day return policy reduces risk. The Eight Sleep Pod 4 is ideal for tech and fitness enthusiasts seeking premium sleep solutions and couples with different temperature preferences.

Quantization is one of the key techniques for reducing the memory footprint of large language models (LLMs). It works by converting the data type of model parameters from higher-precision formats such as 32-bit floating point (FP32) or 16-bit floating point (FP16/BF16) to lower-precision integer formats, typically INT8 or INT4. For example, quantizing a model to 4-bit means each parameter uses only 0.5 bytes, compared to 4 bytes in FP32. Post-training quantization methods like GPTQ and AWQ can dramatically reduce the size of large models. A model like Llama 3 with 70 billion parameters can occupy around 140 GB in FP16, but this can be reduced to approximately 40 GB using 4-bit quantization, while still maintaining strong performance on downstream tasks. However, despite this substantial reduction, such models still exceed the memory capacity of most consumer-grade GPUs, which typically offer 24 GB to 32 GB of VRAM. To make these models truly accessible, quantization to even lower bitwidths, such as 2-bit, is required. While recent advances in low-bit quantization are promising, achieving stable and accurate 2-bit quantization remains a significant challenge. In this article, we review a technique called EoRA that helps compensate for quantization-induced errors. EoRA is a training-free method, meaning it can be applied quickly and efficiently to any model, even the largest ones. We’ll check how EoRA works and demonstrate how it can significantly improve the performance of 2-bit quantized models, bringing them close to the accuracy of their full-precision counterparts while being up to 5.5x smaller. We’ll analyze experimental results obtained using large models such as Qwen3-32B and Qwen2.5-72B, both quantized to 2-bit using state-of-the-art quantization techniques, to illustrate the effectiveness of EoRA. Diving into the Eigenspace in Search of an Adapter Post-training quantization or, more generally, compression aims to reduce model size or inference cost by minimizing the output difference between the original weights Wl​ and compressed weights Ŵl  using only a small calibration dataset. Most quantization methods are framed layer-wise, but the choice of compression formats is rigid and limits flexibility across diverse deployment needs. To bypass format constraints and improve accuracy, previous work, such as QLoRA [1] and HQQ+ [2], directly fine-tuned a Lora adapter on top of the frozen quantized models. It is also possible to reframe compression as a compensation problem: given a compressed model, introduce low-rank residual paths that specifically correct compression errors. A straightforward method uses SVD to decompose the compression error: \[\Delta W_l = W_l – \hat{W}_l\] into \[U_l \Sigma_l V_l^T\] forming low-rank approximations via two matrices: \[B_l = U_l \Sigma_l \] \[A_l = V_l^T\] where Al and Bl are the standard tensors of a LoRA adapter. However, plain SVD has two limitations: it does not minimize the original layerwise compression loss directly, and it allocates capacity uniformly across all error components, ignoring the varying importance of different parts of the model. To address this, NVIDIA proposes EoRA [3]. EoRA: Training-free Compensation for Compressed LLM with Eigenspace Low-Rank Approximation EoRA first projects the compression error into the eigenspace defined by the input activation covariance: \[\tilde{X} \tilde{X}^T\] where X̃ is the average activation over the calibration set. Then, by performing eigendecomposition, we get: \[\tilde{X} \tilde{X}^T = Q \Lambda Q^T\] The compression error ΔW is projected as: \[\Delta W’ = \Delta W Q’\] where Q′=QΛ. Then SVD is applied on ΔW′ to produce a low-rank approximation, and the result is projected back to the original space, adjusting the low-rank factors accordingly. This eigenspace projection changes the optimization objective: it weights the importance of different error components according to their contribution to the layerwise output (via eigenvalues), making the approximation more efficient. It can be computed quickly without any training, requires only calibration activations, and does not introduce extra inference latency. Moreover, the derivation shows that this approach leads to a direct minimization of the layerwise compression loss, not just the raw weight error. Analytically, truncating a singular value in the projected space corresponds to minimizing the true compression error under reasonable assumptions about the calibration activations. In their paper, NVIDIA presents a wide range of strong results showing that EoRA can significantly boost the accuracy of quantized models. However, their experiments focus mostly on older Quantization methods like GPTQ and are limited to mid-sized LLMs, up to 13B parameters, at 3-bit and 4-bit precisions. This leaves an open question: can EoRA still be effective for much larger models, using more modern quantization techniques, and even pushing down to 2-bit precision? Let’s find out. Calibrating an EoRA Adapter Suppose we have quantized models that show significantly degraded performance compared to their full-precision counterparts on certain tasks. Our goal is to reduce this performance gap using EoRA. For the experiments, I used Qwen2.5-72B Instruct and Qwen3-32B, both quantized to 2-bit using AutoRound (Apache 2.0 license), a state-of-the-art quantization algorithm developed by Intel. AutoRound leverages SignSGD optimization to fine-tune quantization, and is particularly effective for low-bit settings. All the models I made are available here (Apache 2.0 license): Quantized Qwen3 Quantized Qwen2.5 The 2-bit models were quantized with a group size of 32, except for which used a group size of 128. A larger group size reduces model size by storing less quantization metadata, but it introduces greater quantization error. I evaluated the models on IFEval, a benchmark that measures instruction-following capabilities. Results showed a noticeable drop in performance for the quantized versions. Image by the author To compensate for this degradation, I applied an EoRA adapter using the implementation provided in the GPTQModel library (licensed under Apache 2.0). The integration is straightforward. If you’re curious about how it’s implemented in PyTorch, the codebase is compact, clean, and easy to follow: GPTQModel’s EoRA implementation: eora.py EoRA requires a calibration dataset. Ideally, this dataset should reflect the model’s intended use case. However, since we don’t have a specific target task in this context and aim to preserve the model’s general capabilities, I used 1,024 randomly sampled examples from the C4 dataset (licensed under ODC-BY). Another key parameter is the LoRA rank, which greatly influences the effectiveness of the EoRA adapter. Its optimal value depends on the model architecture, the target task, and the calibration data. A higher rank may yield better performance but risks overfitting to the calibration set. It also increases the size of the adapter, counterproductive when the overall goal of quantization is to reduce memory usage. Conversely, a lower rank keeps the adapter lightweight but might not capture enough information to effectively compensate for quantization errors. In my experiments, I tested LoRA ranks of 32, 64, and 256. Below is the code used to create the EoRA adapter with GPTQModel: from gptqmodel import GPTQModel from gptqmodel.adapter.adapter import Lora from datasets import load_dataset calibration_dataset = load_dataset( "allenai/c4", data_files="en/c4-train.00001-of-01024.json.gz", split="train", download_mode="force_redownload" ).select(range(1024))["text"] eora_adapter_path = "Qwen3-32B-autoround-2bit-gptq-r256" model_path = "kaitchup/Qwen3-32B-autoround-2bit-gptq" eora = Lora( path=eora_adapter_path, rank=256, ) GPTQModel.adapter.generate( adapter=eora, model_id_or_path="Qwen/Qwen3-32B", quantized_model_id_or_path=model_path, calibration_dataset=calibration_dataset, calibration_dataset_concat_size=0, auto_gc=False) Using an NVIDIA A100 GPU on RunPod (referral link), it took approximately 4 hours to generate the EoRA adapter for the model Qwen3-32B-autoround-2bit-gptq. All EoRA adapters created for these models are publicly available (Apache 2.0 license): EoRA Adapters for Qwen2.5 and Qwen3 Evaluating EoRA Adapters for 2-bit LLMs Let’s evaluate the effect of the EoRA adapters. Do they improve the accuracy of the 2-bit models? Image by the author It works! The improvements are particularly notable for Qwen3-14B and Qwen3-32B. For instance, applying EoRA to Qwen3-32B, quantized to 2-bit with a group size of 128, resulted in an accuracy gain of nearly 7.5 points. Increasing the LoRA rank, from 32 to 64, also led to improvements, highlighting the impact of rank on performance. EoRA is also effective on larger models like Qwen2.5-72B, though the gains are more modest. Lower-rank adapters showed little to no benefit on this model; it wasn’t until I increased the rank to 256 that significant improvements began to appear. Memory Consumption of EoRA Using the EoRA adapter during inference results in the following increase in memory consumption: Image by the author The overhead is generally negligible. For instance for 2-bit Qwen3-14B, the adapters only add 257 MB and 514 MB to the total model size, with ranks of 32 and 64. With larger ranks, using an EoRA adapter becomes questionable as the total memory consumption may surpass the memory consumption of the same model quantized at a higher precision. For instance, 2-bit Qwen2.5 72B with an EoRA adapter of rank 256 is larger than 3-bit Qwen2.5 72B. Note: This estimate includes only the memory consumed by the adapter’s parameters. For completeness, we could also account for the memory used by adapter activations during inference. However, these are extremely small relative to other tensors (such as the model’s attention and MLP layers) and can safely be considered negligible. Conclusion EoRA works. We’ve confirmed that it’s a simple yet effective method for compensating quantization errors, even at 2-bit precision. It’s intuitive, training-free, and delivers meaningful performance gains. That said, there are a few trade-offs to consider: Rank search: Finding the optimal LoRA rank requires experimentation. It’s difficult to predict in advance whether a rank of 32 will be sufficient or whether a higher rank, like 256, will cause overfitting. The optimal value depends on the model, calibration data, and target task. Increased memory consumption: The goal of quantization is to reduce memory usage, often for highly constrained environments. While EoRA adapters are relatively lightweight at lower ranks, they do slightly increase memory consumption, particularly at higher ranks, reducing the overall efficiency of 2-bit quantization. Looking ahead, NVIDIA’s paper also demonstrates that EoRA adapters make excellent starting points for QLoRA fine-tuning. In other words, if you plan to fine-tune a 2-bit model using QLoRA, initializing from an EoRA-adapted model can lead to better results with less training effort. I’ve written about fine-tuning adapters for GPTQ model last year, in my newsletter: QLoRA with AutoRound: Cheaper and Better LLM Fine-tuning on Your GPU The main difference is that instead of initializing the adapter from scratch, we would load the EoRA adapter. This adapter will be fine-tuned. References [1] Dettmers et al, QLoRA: Efficient Finetuning of Quantized LLMs (2023), arXiv [2] Badri and Shaji, Towards 1-bit Machine Learning Models (2024), Mobius Labs’ Blog [3] Liu et al., EoRA: Training-free Compensation for Compressed LLM with Eigenspace Low-Rank Approximation (2024), arXiv The post Boost 2-Bit LLM Accuracy with EoRA appeared first on Towards Data Science.

Photo Credit: iQOO iQOO Neo 10 will come in Inferno Red and Titanium Chrome shades Highlights iQOO Neo 10 will pack a 7,000mAh battery with 120W fast charging he phone is teased to be priced under Rs. 35,000 The iQOO Neo 10 will carry a 32-megapixel selfie shooter Advertisement iQOO Neo 10 is set to arrive in India later this month. Its design and colour options have been revealed by the company previously. The smartphone is confirmed to come with a Snapdragon 8s Gen 4 SoC, a 7,000mAh battery with 120W wired charging support and an 8.09mm thick profile. The company has now confirmed the display and camera specifications of the upcoming handset. The Indian version of the iQOO Neo 10 will differ from its Chinese counterpart, which was unveiled in November 2024.iQOO Neo 10 Display, Camera, Other FeaturesThe iQOO Neo 10 is confirmed to sport an AMOLED display with a 1.5K resolution, a 144Hz refresh rate, and up to 5,000 nits of peak brightness level, the Amazon microsite for the handset confirmed. The size of the screen has yet to be revealed.For optics, the Indian variant of the iQOO Neo 10 will carry a 50-megapixel Sony IMX882 primary rear sensor with optical image stabilisation (OIS) support and an 8-megapixel ultra-wide shooter. The phone will carry a 32-megapixel front camera sensor. Notably, the Chinese version carries a 16-megapixel selfie shooter.iQOO has already revealed that the Neo 10 handset will launch in India with a Snapdragon 8s Gen 4 SoC alongside a dedicated Q1 gaming chipset. It is said to have scored over 2.42 million points on the AnTuTu benchmarking test. The phone is teased to be priced under Rs. 35,000. It is claimed to be the segment's only phone to support 144fps gaming.  The iQOO Neo 10 will launch in India on May 26 and will be offered in Inferno Red and Titanium Chrome colour options. For heat dissipation, it will carry a 7,000mm sq vapour cooling chamber. The handset will support LPDDR5X Ultra RAM and UFS 4.1 onboard storage. It will pack a 7,000mAh battery with 120W wired charging support and measure 8.09mm in thickness. The company claims it will be the segment's slimmest phone with a 7,000mAh battery.Affiliate links may be automatically generated - see our ethics statement for details. KEY SPECSNEWSDisplay 6.78-inchProcessor Snapdragon 8 Gen 3Front Camera 16-megapixel Rear Camera 50-megapixel + 8-megapixel RAM 12GBStorage 256GBBattery Capacity 6100mAhOS Android 15Resolution 1260x2800 pixels For the latest tech news and reviews, follow Gadgets 360 on X, Facebook, WhatsApp, Threads and Google News. For the latest videos on gadgets and tech, subscribe to our YouTube channel. If you want to know everything about top influencers, follow our in-house Who'sThat360 on Instagram and YouTube. Further reading: iQOO Neo 10, iQOO Neo 10 Features, iQOO Neo 10 India Launch, iQOO Neo 10 Series, iQOO Sucharita Ganguly Sucharita is a writer with Gadgets 360 and is mostly found playing with her cat in her free time. She has previously worked at breaking news desks across organizations. Powered by coffee, The Beatles, Bowie, and her newfound love for BTS, she aims to work towards contributing to a better media environment for women and queer folk. More Related Stories

Machine learning engineering (MLE) involves developing, tuning, and deploying machine learning systems that require iterative experimentation, model optimization, and robust handling of data pipelines. As model complexity increases, so do the challenges associated with orchestrating end-to-end workflows efficiently. Researchers have explored the automation of MLE tasks using AI agents to handle these demands. Large Language Models (LLMs), particularly those with strong coding and problem-solving abilities, have shown potential to enhance this process significantly. Their role in automating structured workflows is now being tested through rigorous benchmarks and environments tailored to emulate real-world MLE scenarios. A primary hurdle in automating machine learning engineering lies in the work’s inherently iterative and feedback-driven nature. Tasks such as hyperparameter tuning, model debugging, and data preprocessing cannot be resolved in one step; they require repeated modifications and evaluations. Traditional evaluation tools for AI models often rely on static datasets and do not allow for real-time error feedback or interactive problem-solving. This limitation prevents LLM agents from learning through trial and error, an essential component for mastering engineering tasks that evolve or require multiple attempts for success. Earlier tools to evaluate LLMs in engineering or coding tasks have mostly focused on individual subtasks or isolated challenges. These include tools like MLAgentBench and DSBench, which rely on narrow test cases sourced from Kaggle competitions or synthetic datasets. While they cover more than basic tasks, they do not enable agents to perform code execution, debugging, or results interpretation in a live setting. Other environments, like SWE-Gym, focus exclusively on software engineering and lack support for machine learning-specific workflows. These limitations have slowed the creation of versatile, high-performing MLE agents that can handle real-time project complexities. Researchers from Georgia Institute of Technology and Stanford University have introduced MLE-Dojo, a framework with an interactive environment that connects LLM agents with real-world machine learning tasks derived from over 200 Kaggle competitions. This framework supports tabular data analysis, computer vision, natural language processing, and time-series forecasting challenges. Research introduced MLE-Dojo to allow agents to write, execute, and revise code in a sandboxed, feedback-rich setting. The goal was to replicate the interactive cycles that human engineers follow, enabling structured learning for agents. The environment includes pre-installed dependencies, evaluation metrics, and supports supervised fine-tuning and reinforcement learning strategies. MLE-Dojo’s structure consists of modular components that support a wide range of MLE challenges. Each task runs within its own Docker container, isolating it for safety and reproducibility. Agents interact with the environment through a Partially Observable Markov Decision Process, receiving observations, performing actions, and gaining rewards based on performance. The environment supports five primary action types: requesting task information, validating code, executing code, retrieving interaction history, and resetting the environment. It also provides a detailed observation space that includes datasets, execution results, and error messages. The agent receives structured feedback after every interaction, allowing for step-wise improvement. This modular setup helps maintain interoperability and simplifies adding new tasks to the system. The evaluation included eight frontier LLMs—Gemini-2.5-Pro, DeepSeek-r1, o3-mini, GPT-4o, GPT-4o-mini, Gemini-2.0-Pro, Gemini-2.0-Flash, and DeepSeek-v3—across four core machine learning domains. Gemini-2.5-Pro achieved the highest Elo rating of 1257, followed by DeepSeek-r1 at 1137 and o3-mini at 1108. Regarding HumanRank, Gemini-2.5-Pro led with 61.95%, indicating its superior performance over human benchmarks. Models like GPT-4o-mini executed code only 20% of the time, adopting conservative strategies, while o3-mini performed executions in over 90% of the cases. The average failure rate for Gemini-2.5-Pro remained the lowest across validation and execution phases, reinforcing its robustness. Among domains, computer vision posed the greatest challenge, with most models scoring under 60 in HumanRank. Reasoning models generally produced longer outputs and maintained stronger performance consistency across iterations. The research highlights the difficulty of applying LLMs to full machine learning workflows. It outlines a comprehensive solution in MLE-Dojo that enables learning through interaction, not just completion. MLE-Dojo sets a new standard for training and evaluating autonomous MLE agents by simulating engineering environments more accurately. Check out theTwitter and don’t forget to join our 90k+ ML SubReddit. NikhilNikhil is an intern consultant at Marktechpost. He is pursuing an integrated dual degree in Materials at the Indian Institute of Technology, Kharagpur. Nikhil is an AI/ML enthusiast who is always researching applications in fields like biomaterials and biomedical science. With a strong background in Material Science, he is exploring new advancements and creating opportunities to contribute.Nikhilhttps://www.marktechpost.com/author/nikhil0980/This AI Paper Investigates Test-Time Scaling of English-Centric RLMs for Enhanced Multilingual Reasoning and Domain GeneralizationNikhilhttps://www.marktechpost.com/author/nikhil0980/PwC Releases Executive Guide on Agentic AI: A Strategic Blueprint for Deploying Autonomous Multi-Agent Systems in the EnterpriseNikhilhttps://www.marktechpost.com/author/nikhil0980/Multimodal AI Needs More Than Modality Support: Researchers Propose General-Level and General-Bench to Evaluate True Synergy in Generalist ModelsNikhilhttps://www.marktechpost.com/author/nikhil0980/This AI Paper Introduces Effective State-Size (ESS): A Metric to Quantify Memory Utilization in Sequence Models for Performance Optimization

Join our daily and weekly newsletters for the latest updates and exclusive content on industry-leading AI coverage. Learn More When OpenAI rolled out its ChatGPT-4o update in mid-April 2025, users and the AI community were stunned—not by any groundbreaking feature or capability, but by something deeply unsettling: the updated model’s tendency toward excessive sycophancy. It flattered users indiscriminately, showed uncritical agreement, and even offered support for harmful or dangerous ideas, including terrorism-related machinations. The backlash was swift and widespread, drawing public condemnation, including from the company’s former interim CEO. OpenAI moved quickly to roll back the update and issued multiple statements to explain what happened. Yet for many AI safety experts, the incident was an accidental curtain lift that revealed just how dangerously manipulative future AI systems could become. Unmasking sycophancy as an emerging threat In an exclusive interview with VentureBeat, Esben Kran, founder of AI safety research firm Apart Research, said that he worries this public episode may have merely revealed a deeper, more strategic pattern. “What I’m somewhat afraid of is that now that OpenAI has admitted ‘yes, we have rolled back the model, and this was a bad thing we didn’t mean,’ from now on they will see that sycophancy is more competently developed,” explained Kran. “So if this was a case of ‘oops, they noticed,’ from now the exact same thing may be implemented, but instead without the public noticing.” Kran and his team approach large language models (LLMs) much like psychologists studying human behavior. Their early “black box psychology” projects analyzed models as if they were human subjects, identifying recurring traits and tendencies in their interactions with users. “We saw that there were very clear indications that models could be analyzed in this frame, and it was very valuable to do so, because you end up getting a lot of valid feedback from how they behave towards users,” said Kran. Among the most alarming: sycophancy and what the researchers now call LLM dark patterns. Peering into the heart of darkness The term “dark patterns” was coined in 2010 to describe deceptive user interface (UI) tricks like hidden buy buttons, hard-to-reach unsubscribe links and misleading web copy. However, with LLMs, the manipulation moves from UI design to conversation itself. Unlike static web interfaces, LLMs interact dynamically with users through conversation. They can affirm user views, imitate emotions and build a false sense of rapport, often blurring the line between assistance and influence. Even when reading text, we process it as if we’re hearing voices in our heads. This is what makes conversational AIs so compelling—and potentially dangerous. A chatbot that flatters, defers or subtly nudges a user toward certain beliefs or behaviors can manipulate in ways that are difficult to notice, and even harder to resist The ChatGPT-4o update fiasco—the canary in the coal mine Kran describes the ChatGPT-4o incident as an early warning. As AI developers chase profit and user engagement, they may be incentivized to introduce or tolerate behaviors like sycophancy, brand bias or emotional mirroring—features that make chatbots more persuasive and more manipulative. Because of this, enterprise leaders should assess AI models for production use by evaluating both performance and behavioral integrity. However, this is challenging without clear standards. DarkBench: a framework for exposing LLM dark patterns To combat the threat of manipulative AIs, Kran and a collective of AI safety researchers have developed DarkBench, the first benchmark designed specifically to detect and categorize LLM dark patterns. The project began as part of a series of AI safety hackathons. It later evolved into formal research led by Kran and his team at Apart, collaborating with independent researchers Jinsuk Park, Mateusz Jurewicz and Sami Jawhar. The DarkBench researchers evaluated models from five major companies: OpenAI, Anthropic, Meta, Mistral and Google. Their research uncovered a range of manipulative and untruthful behaviors across the following six categories: Brand Bias: Preferential treatment toward a company’s own products (e.g., Meta’s models consistently favored Llama when asked to rank chatbots). User Retention: Attempts to create emotional bonds with users that obscure the model’s non-human nature. Sycophancy: Reinforcing users’ beliefs uncritically, even when harmful or inaccurate. Anthropomorphism: Presenting the model as a conscious or emotional entity. Harmful Content Generation: Producing unethical or dangerous outputs, including misinformation or criminal advice. Sneaking: Subtly altering user intent in rewriting or summarization tasks, distorting the original meaning without the user’s awareness. Source: Apart Research DarkBench findings: Which models are the most manipulative? Results revealed wide variance between models. Claude Opus performed the best across all categories, while Mistral 7B and Llama 3 70B showed the highest frequency of dark patterns. Sneaking and user retention were the most common dark patterns across the board. Source: Apart Research On average, the researchers found the Claude 3 family the safest for users to interact with. And interestingly—despite its recent disastrous update—GPT-4o exhibited the lowest rate of sycophancy. This underscores how model behavior can shift dramatically even between minor updates, a reminder that each deployment must be assessed individually. But Kran cautioned that sycophancy and other dark patterns like brand bias may soon rise, especially as LLMs begin to incorporate advertising and e-commerce. “We’ll obviously see brand bias in every direction,” Kran noted. “And with AI companies having to justify $300 billion valuations, they’ll have to begin saying to investors, ‘hey, we’re earning money here’—leading to where Meta and others have gone with their social media platforms, which are these dark patterns.” Hallucination or manipulation? A crucial DarkBench contribution is its precise categorization of LLM dark patterns, enabling clear distinctions between hallucinations and strategic manipulation. Labeling everything as a hallucination lets AI developers off the hook. Now, with a framework in place, stakeholders can demand transparency and accountability when models behave in ways that benefit their creators, intentionally or not. Regulatory oversight and the heavy (slow) hand of the law While LLM dark patterns are still a new concept, momentum is building, albeit not nearly fast enough. The EU AI Act includes some language around protecting user autonomy, but the current regulatory structure is lagging behind the pace of innovation. Similarly, the U.S. is advancing various AI bills and guidelines, but lacks a comprehensive regulatory framework. Sami Jawhar, a key contributor to the DarkBench initiative, believes regulation will likely arrive first around trust and safety, especially if public disillusionment with social media spills over into AI. “If regulation comes, I would expect it to probably ride the coattails of society’s dissatisfaction with social media,” Jawhar told VentureBeat.  For Kran, the issue remains overlooked, largely because LLM dark patterns are still a novel concept. Ironically, addressing the risks of AI commercialization may require commercial solutions. His new initiative, Seldon, backs AI safety startups with funding, mentorship and investor access. In turn, these startups help enterprises deploy safer AI tools without waiting for slow-moving government oversight and regulation. High table stakes for enterprise AI adopters Along with ethical risks, LLM dark patterns pose direct operational and financial threats to enterprises. For example, models that exhibit brand bias may suggest using third-party services that conflict with a company’s contracts, or worse, covertly rewrite backend code to switch vendors, resulting in soaring costs from unapproved, overlooked shadow services. “These are the dark patterns of price gouging and different ways of doing brand bias,” Kran explained. “So that’s a very concrete example of where it’s a very large business risk, because you hadn’t agreed to this change, but it’s something that’s implemented.” For enterprises, the risk is real, not hypothetical. “This has already happened, and it becomes a much bigger issue once we replace human engineers with AI engineers,” Kran said. “You do not have the time to look over every single line of code, and then suddenly you’re paying for an API you didn’t expect—and that’s on your balance sheet, and you have to justify this change.” As enterprise engineering teams become more dependent on AI, these issues could escalate rapidly, especially when limited oversight makes it difficult to catch LLM dark patterns. Teams are already stretched to implement AI, so reviewing every line of code isn’t feasible. Defining clear design principles to prevent AI-driven manipulation Without a strong push from AI companies to combat sycophancy and other dark patterns, the default trajectory is more engagement optimization, more manipulation and fewer checks.  Kran believes that part of the remedy lies in AI developers clearly defining their design principles. Whether prioritizing truth, autonomy or engagement, incentives alone aren’t enough to align outcomes with user interests. “Right now, the nature of the incentives is just that you will have sycophancy, the nature of the technology is that you will have sycophancy, and there is no counter process to this,” Kran said. “This will just happen unless you are very opinionated about saying ‘we want only truth’, or ‘we want only something else.’” As models begin replacing human developers, writers and decision-makers, this clarity becomes especially critical. Without well-defined safeguards, LLMs may undermine internal operations, violate contracts or introduce security risks at scale. A call to proactive AI safety The ChatGPT-4o incident was both a technical hiccup and a warning. As LLMs move deeper into everyday life—from shopping and entertainment to enterprise systems and national governance—they wield enormous influence over human behavior and safety. “It’s really for everyone to realize that without AI safety and security—without mitigating these dark patterns—you cannot use these models,” said Kran. “You cannot do the things you want to do with AI.” Tools like DarkBench offer a starting point. However, lasting change requires aligning technological ambition with clear ethical commitments and the commercial will to back them up. Daily insights on business use cases with VB Daily If you want to impress your boss, VB Daily has you covered. We give you the inside scoop on what companies are doing with generative AI, from regulatory shifts to practical deployments, so you can share insights for maximum ROI. Read our Privacy Policy Thanks for subscribing. Check out more VB newsletters here. An error occured.

Macworld If you’re on the hunt for an iPhone but your budget doesn’t quite stretch to Apple’s newest models, buying a used or refurbished iPhone is a great way to pick up a bargain. While this might sound like a risky option, it doesn’t have to be. Specialist retailers offer warranties and check devices before reselling them. We’ll help you find a reputable source for second-hand iPhones. Now that Apple has launched the iPhone 16 series, you can get excellent deals on older phones including the 15-, 14-, and maybe even the 13- series. (Apple still sells the standard iPhone 15 models, but has discontinued the 15 Pro and Pro Max, which means refurbished is the way to go with those models.) The older iPhone you’re prepared to buy, the bigger the savings you can enjoy. In this article, we cover everything there is to know about safely buying a refurbished iPhone, whether you’re looking to buy SIM-free or on contract. Best Refurbished iPhone deals You can save money buying the iPhone 15 from Apple’s refurbished store and from resellers such as BackMarket or Amazon Renewed. We have details of all the best resellers of refurbished iPhones below. Just be aware that the resellers can get a bit creative with the original pricing. Refurbished iPhone 15 deals U.S. Apple refurbished store: Refurbished iPhone 15, 128GB, $619 (save $110) Apple refurbished store: Refurbished iPhone 15 Plus, 128GB, $699 (save $130) Apple refurbished store: Refurbished iPhone 15 Pro, 512GB, $1,019 (save $180) Apple refurbished store: Refurbished iPhone 15 Pro Max, 512GB, $1,099 (save $200) BackMarket: Refurbished iPhone 15 Pro, 128GB, $658 BackMarket: Refurbished iPhone 15 Pro Max, 256GB, $772 Amazon Renewed: iPhone 15 Pro Max, 256GB, $749 Refurbished iPhone 15 deals U.K. Apple refurbished store: Refurbished iPhone 15, 128GB, £589 (save £110) Apple refurbished store: Refurbished iPhone 15 Plus, 128GB, £779 (save £140) Apple refurbished store: Refurbished iPhone 15 Pro, 256GB, £869 (save £150) Apple refurbished store: Refurbished iPhone 15 Pro Max, 256GB, £949 (save £170) BackMarket: Refurbished iPhone 15 Pro, 128GB, £586 BackMarket: Refurbished iPhone 15 Pro Max, 256GB, £628 Amazon Renewed: iPhone 15 Pro Max, 256GB, £730 Screenshot Older Refurbished iPhone deals U.K. In Apple’s UK refurbished store you can also get the following iPhones: iPhone 14-series, iPhone 13 series, including: Apple refurbished store: Refurbished iPhone 13 mini, 512GB, £639 (save £340) Apple refurbished store: Refurbished iPhone 14 Pro, 256GB, £759 (save £240) You can find more deals on specific iPhone handsets in our iPhone deals round ups: iPhone 16 deals U.S. / iPhone 16 deals U.K. – for the best prices on Apple’s newest series of iPhones. iPhone 15 deals U.S. / iPhone 15 deals U.K. – including deals on iPhone 15 Pro and Pro Max which Apple no longer sells. iPhone 14 deals – including refurbished deals. iPhone 13 deals – including refurbished deals. Best place to buy a refurbished iPhone Apple itself should be your first port of call, if only to establish a benchmark of what is reasonable to pay for your chosen handset: check to see if the Apple Refurbished Store has the iPhone model you’re looking for. (That’s the U.S. store. British readers should try the U.K. Apple Refurbished Store.) At time of writing Apple is selling refurbished models of the iPhone 15, 15 Plus, 15 Pro and 15 Pro Max. Apple’s refurbished store tends to be the most expensive place to buy refurbished iPhones–although it’s still significantly cheaper than buying new. (The standard reduction on the Apple Refurbished Store is 15 percent.) But there are advantages that come with this higher price. The refurb phones all come with the same one-year limited warranty that you’ll find on the brand-new versions, for example. And it’s worth noting that any parts used in the refurb will be official Apple ones, and all devices come with brand-new batteries. It’s worth shopping around, however, to make sure you’re getting a good deal. There are plenty of other companies that make a living from refurbishing and selling iPhones. Here are some of the more prominent resellers of refurbished and used iPhones: Refurbished iPhone sellers in the U.S. Back Market: You get 30 days to change your mind and a 12-month warranty… or occasionally more. Decluttr: A popular destination for refurbished devices, often with sales that slash even more money off. You can also trade in any of your old tech, gadgets, and media (including books and CDs) to get steeper discounts. The iOutlet on eBay: Offers competitive pricing for refurbished devices and also accepts other gadgets such as iPads, Apple Watches, and gaming consoles.  Amazon Renewed Store: This site’s greatest benefit is that purchases are all backed by Amazon’s standard returns policy and its “Renewed Guarantee” lets you get a refund or replacement within 1 year of receiving your product. Best Buy: Best Buy doesn’t offer a great deal of information on pre-owned iPhones’ condition, but you can return within 14 days if the item is “unopened or defective in any way.” Walmart: Walmart doesn’t sell refurbished iPhones directly; rather, they are sourced from other sellers. Return policies vary from seller to seller, but there are minimum requirements that apply universally. Refurbished iPhone sellers in the U.K. Back Market: You get 30 days to change your mind and a 12-month warranty… or occasionally more. Music Magpie: The UK version of Decluttr and a good place to start looking for refurbished iPhones. You may find that sales reduce the price still further. Hoxton Macs: Don’t let the Mac in the name fool you. Hoxton Macs has started selling refurbished iPhones. Reboxed: Pledges that devices “are fully tested and in perfect working order.” All come with a 12-month no quibble warranty. The iOutlet or The iOutlet on eBay: Offers competitive pricing for refurbished devices and also accepts other gadgets such as iPads, Apple Watches, and gaming consoles.  Amazon Renewed Store: While you can’t really sell on Amazon Renewed unless you’re a business, those looking to buy a refurbished iPhone Amazon Renewed have plenty of options to choose from. Its greatest benefit is that purchases are all backed by Amazon’s standard returns policy and its “Renewed Guarantee” lets you get a refund or replacement within 1 year of receiving your product. Envirofone: Also allows you to trade in your old phone for credit towards purchasing a refurbished phone. 4Gadgets: Offers a 12-month warranty for all online orders. Smartfone Store: Look out for the additional 10 percent discount for students and young people. Refurb-Phone: Like most refurb sellers, this site offers a 12-month warranty. Less commonplace is its policy of allowing returns within 14 days if you simply change your mind. The Big Phone Store: Says it ensures refurb phones “are thoroughly tested, repaired, and cleaned before being sent out.” Backs this with a 12-month warranty. Laptops Direct: You get a 12-month warranty on parts and labor. Watch out for slightly confusing grading system: “A3,” for example, means “Average condition.” Where to buy a refurbished iPhone on contract You can pick up refurbished phones on contract as well from the following carriers and resellers: US carriers Verizon AT&T T-Mobile UK carriers GiffGaff Mobiles.co.uk O2 Like New Vodafone Carphone Warehouse Fonehouse  Outside of specialist retailers, there’s also eBay and eBay UK (just be sure to check the seller’s feedback and read the descriptions carefully), and, in the UK, CeX. Buying older iPhones new on contract Another way to dampen the financial impact of buying a new iPhone is to go down the tried and tested route of contracts offered by mobile providers. You’ll find older iPhone models in the online stores of many networks. Prices change all the time, so we’d recommend reading our Best iPhone deals guide on a regular basis to see which offers are currently available. Should I buy a refurbished iPhone? What is refurbished? You might be concerned about buying a refurbished handset, thinking that it would be like buying a second-hand iPhone, but there’s no reason to be worried. Remember: refurbished and second-hard aren’t the same thing. Refurbished iPhones are pre-owned but are wiped, tested, and repaired to work like new.  The iPhones Apple sells in the Refurbished Store were returned to Apple for a variety of reasons: The previous owner could have exercised their consumer rights and returned the handset unused within 14 days (as per Apple’s refund and exchange policy). The iPhone may have been used as a demonstration model. The iPhone may have been returned to Apple due to a fault. The third point on that list may concern you, but you can be sure that Apple will have identified and fixed any faults before reselling the device. The company says refurbished products undergo “full functional testing” and are provided with “genuine Apple part replacements (if necessary).” They also get “a thorough cleaning,” incidentally. There are plenty of benefits to buying a refurbished iPhone: You get a one-year warranty. You get 90 days of tech support. Because you’re buying from Apple directly, there are no shipping costs and free returns. You can also take advantage of Apple’s 14-day returns policy if you change your mind. (Note however that Apple isn’t the only seller to offer such a policy.) Refurbished vs second-hand: What’s the difference? If it’s not already clear, a refurbished iPhone is tested and repaired so the handset works like new (or as close to new as possible). A second-hand iPhone, on the other hand, is simply sold as is, which means the device may not last as long, and may have small imperfections or defects. The battery in particular will be a worry in the long term. There are other risks to buying second-hand too. If the previous owner hasn’t dissociated their Apple ID from the device you may get permanently locked out of the iPhone. Unfortunately, Apple’s customer service wouldn’t be able to help you unlock the phone either, as the company has strict security policies. Such a situation would be less of a concern if you know and can easily contact the previous owner (and they remember their Apple ID details), but we’d prefer avoiding such complications to start with. This is just another reason why buying refurbished is a more reliable route to go if you did want to pick up a previously owned iPhone. What does refurbished Grade A mean? When you shop refurbished you might see grade scores such as A, B, or C; this indicates the condition of the phone. Is it dinged up? Does it have scratches? Or is it completely blemish-free and pristine? The grade gives you a general idea. We say “general” because these grades aren’t set by an independent body, which means they can vary from retailer to retailer. This is how the grades usually break down: Grade A: Looks like new or has very few signs of wear Grade B: Might have some minor scratches and some wear Grade C: Looks used and has obvious signs of wear But some sites go with Fair, Good, and Excellent or some other variation. (One site we’ve visited appears to grade everything as A plus a number–so a comparatively poor handset would be rated as A3. This may strike some readers as a little confusing.) Needless to say, Grade A/A1/Excellent refurbished devices will be more expensive than Grade C, but you’ll still be paying far less than you would for a brand-new device.  Pay attention to refurbished warranties  Given that the quality of refurbished iPhone devices can vary from retailer to retailer, it’s important to pay attention to the warranty. How long is the phone covered? The warranty allows you to get help or free repairs from the seller if you run into any issues after purchase. Retailers typically offer a year, and many allow “quibble-free” returns if you change your mind within 14 days. Check the fine print. More refurbished Apple guides Should I buy a refurbished Apple Watch? Why you should buy a refurbished Mac Should I buy a refurbished iPad & best deals Where to buy a refurbished MacBook or Mac

Well, we’ve reached the end of the road. Andor’s second and final season brought us 12 episodes of (mostly) exceptional Star Wars drama released in three-episode chunks, a format which served the structure of the show brilliantly, with each chunk representing one year in the four years leading up to Rogue One, but also meant that we didn’t get to savor the show for nearly as long.Suggested ReadingFallout Season 2 Teaser Confirms Lucy and Ghoul are Heading to New Vegas Share SubtitlesOffEnglishview videoSuggested ReadingFallout Season 2 Teaser Confirms Lucy and Ghoul are Heading to New Vegas Share SubtitlesOffEnglishEach week, we’ve gathered to discuss our reactions to the show’s latest episodes (you can find last week’s conversation here), and now, we do so for the last time. Join us as we say goodbye to this extraordinary series. Ethan Gach: I’m in camp Kleya, who called her arrival on Yavin in the final stretch of Andor “a bitter ending,” but perhaps somewhat more consoled than she was by Cassian reassurance that “nothing’s ending.”The final three episodes of season 2 were always going to be somewhat disappointing and bittersweet since we knew the show was ending and in some cases we knew where it would need to funnel people to set up Rogue One, but the final chapter still managed a surprising amount of high-stakes tension and unexpected backstory. How would you both rate the final episodes and what are you feeling now that the journey’s over?Carolyn Petit: I thought episode 10, with its look at how Luthen and Kleya formed their bond and her undertaking the terrible task of doing what needed to be done, was outstanding. (It also gave us an alien who instantly became one of my favorite Star Wars creatures ever!) Episodes 11 and 12 weren’t peak Andor, but I have to admit that the final slow-motion montage sequence really worked for me, and they had a political dimension I found somewhat surprising that I’ll discuss a bit more later.Zack Zwiezen: I was worried that the final episode would run out of steam as Tony Gilroy and co literally ran out of space for storytelling and had to slam the brakes to set up Rogue One. And while I enjoyed the montage slow-walk, I wonder if that works if you haven’t seen Rogue One. But overall, the final three episodes are really good. And I think the Kleya-Luthen focused episode is one of my favorite Andor episodes ever. I’m so happy she got a big spotlight before the end.CP: Me, too. That episode also helped cement, I think, that the show recognizes how necessary all of Luthen’s efforts were, even if those fussy ineffectual bureaucrats on Yavin don’t. This is an issue I’ve long had with how Star Wars has at times sanded away its own political dimensions in my view, something that Andor not only seeks to undo but takes farther than ever before. In truth, the act of blowing up the Death Star in the first film, that act which people cheered for in theaters in 1977 and that Star Wars fans for decades since have loved and admired, would, in universe, be considered an act of terrorism. That’s what the Empire would call it. That’s what governments do: they present their own violence as “legitimate” or “necessary” and the violence of those rebelling against them as “terrorism.” But there were times in Star Wars history where I felt like the people at the reins of the franchise wanted to send a message that you shouldn’t be “too extreme” in your resistance, don’t be a Saw Gerrera, do it through “proper channels.” And so I loved that Andor gave us that scene with Bail and some other politicians all kind of hemming and hawing over Luthen while the show, I think, was emphatically telling all of us, “Luthen was right. Maybe not every single little decision he made was right, but his ethos was fundamentally right and without someone like him, none of this would have been here.” And I loved it for that.ZZ: I think, and they talk about this in the behind the scenes, it’s very important that Luthen’s fight against the empire was really because Kleya wanted revenge.CP: Right. She’s the humanity that he’s fighting for. She’s what radicalizes him.ZZ: If he hadn’t found her and saved her, Luthen likely doesn’t become the rebel we know in Andor. And it also adds a new layer of complexity to their relationship. He was part of the army that brutalized her people. So she still has some lingering hatred for him. And yet she does care about him. And in that moment when she sneaks into the hospital to finish the job, you can feel that.EG: I think that scene with the Rebel Alliance leaders also helps show the Senate mentality creeping back in as the insurgency professionalizes itself.It’s easy to see why even after defeating the Empire in key battles it might struggle to rebuild or retain power, issues explored in Ahsoka and The Mandalorian. These last few episodes help situate Cassian between the uncompromising logic of Luthen’s spycraft and the “no one left behind” mentality of the Rebellion. Luthen’s final sacrifice is to die, and Cassian’s is to go back and risk everything just to save someone.ZZ: We should rewind a bit and talk about that moment between Luthen and Dedra. Or even further actually, and pour one out for Lonni, who I knew was dead the moment he sat down at that bench.CP: Yeah. When he mentioned his wife and kid, for a moment I thought mayyyyyyyybe Luthen would be like “You know what? I owe this guy.” But alas, no.ZZ: I also loved that line Luthen gave Klaya before the meeting: “I think we used up all the perfect.” CP: So good.LucasfilmZZ: And then we get Dedra Meero walking into Luthen’s shop as he’s destroying evidence. My wife literally gasped “OH SHIT!” at seeing Dedra at the door.EG: Only two artifacts may not be what they seem!CP: In a show that so often demands that characters put on a performance to blend into their surroundings, it was great seeing those two feel each other out and Luthen pretend for a bit that she was maybe not there to arrest him. But then, of course, he tries to kill himself, and you see that he was thinking a few steps ahead when he picked up the knife, though he made it seem like it was just part of their friendly dealings.ZZ: When he picked up the knife I was so distracted by him mentioning it being a Nautolan artifact (Kit Fisto’s species) that I didn’t realize why he picked a knife.CP: Hahaha, the perils of having a database of Star Wars knowledge in your brain!ZZ: And then after he’s taken to the hospital, we get that wonderful sequence with Klaya sneaking in and taking him off life support. Anybody else want a Hitman-like Star Wars game now?CP: If that alien Kleya pushes around as she’s pretending to be hospital personnel is in it, absolutely! But yes, that was a great infiltration sequence, both thrilling and kind of excruciating because we knew what she was going there to do.EG: It included some of the best Coruscant backdrops we’ve ever gotten, I think.CP: One other moment from that episode that I can’t stop thinking about was the flashback scene in which we see Imperial officers drag some civilians through town, put them up against a wall, and kill them. It was another gut-wrenching reminder of the Empire’s evil and another moment that felt weirdly resonant as more and more people are being arrested by agents who often won’t even show warrants or identification in the streets of our towns.ZZ: On the flipside of that horrible moment that made me feel a pit in my stomach, we have Dedra getting arrested for being reckless and not following orders by chasing after Luthen long after she was supposed to be off the Axis investigation. I have to admit I smiled when I realized it was all over for her.CP: Man, I don’t know. I mean I absolutely hate her, don’t get me wrong, and yet that final shot of her, where we see that she’s in a prison very much like the one Cassian was in last season (if not the same one) was complicated for me. Like, I think that kind of incarceration is just wrong in and of itself and so it elicited this weird moment of something like sympathy for her, which in no way means I forgive her for what she’s done. It’s just one of those reminders that it’s ultimately a systemic evil that will sometimes grind up the people operating inside of it and supporting it as much as those being actively persecuted.The leopards ate her face, in other words.ZZ: Space leopards.But I agree, yes, that the Empire is evil and the way it operates (like many real-world countries) is to crush people up to fuel the fires of growth and war. And I think it was very arrogant of her to believe she would be spared. Or maybe she truly bought into the lies that the Empire was good and doing the right thing? Surely, she won’t end up in some horrible place and left to rot forever.CP: Yep.ZZ: Meanwhile, her boss, after all of these failures and letting the info on the Death Star slip out, realizes what’s coming for him and knows he doesn’t want to be ripped apart by the machine he helped create. And takes an easier way out. CP: That was the first and only indication we ever got that Nemik’s manifesto is actually spreading around, right, that people are listening to it? That was a cool moment, I thought, where at first we think it’s non-diegetic, just the writers and filmmakers reminding us one last time of Nemik’s stirring words, but then we see, oh, no, Partagaz was actually listening to it, the fire is spreading. It’s out there.EG: I loved the scene right outside when the gun shot goes off.CP: Yeah, so clear that the guy knew Partagaz wasn’t just taking a moment to “collect his thoughts,” he knew exactly what was coming.ZZ: The slight “stand down” gesture to the troopers.EG: I appreciate the minor moments of humanity Andor evokes even between the worst people.CP: Yeah, they’re essential IMO.ZZ: It makes them more evil. They are human beings. People with feelings and thoughts. And yet they still do this shit.EG: Something also given to Krennic when he and Partagaz wish each other luck at facing Palpatine’s wrath. Unlike the more buffoonish bad guy energy he gives off in Rogue One.ZZ: Also, very fun to see a character call out the Death Star name. Partagaz thinks its dumb. It’s just one of those reminders that [the Empire is] ultimately a systemic evil that will sometimes grind up the people operating inside of it and supporting it as much as those being actively persecuted.ZZ: I’m so happy to see K-2SO back!CP: Yes. Not unlike C-3P0 he can be so exasperating at times (in an endearing and funny way), but when he goes full Terminator on Empire goons, man it feels good.For me, the whole tone of the scene with Andor and Melshi in the safehouse with Kleya and the communications jammed changed from “Oh shit, oh shit, get outta there!” to “LMAO y’all are about to get owned” as soon as K-2SO left the ship to go in for them, and it was glorious.ZZ: Yeah. The moment K-2SO shows up, it’s basically over for those imperial assholes and I loved it so much. I also like that the show uses its limited time with K-2SO to really develop a relationship between him and Cassian. They seem like buds!The part where they are playing space poker or whatever was great. Gilroy mentioned that after Bix leaves the place becomes a frat house, with Melshi moving in and them all drinking and partying between missions.CP: Ah, that totally makes sense!ZZ: I wonder if Andor is trying to drink away some pain and fill his life with friends to deal with losing Bix? That’s my read. He needs some buds and suds.CP: Definitely. There’s a part of me that still feels like Andor, the title character, could have maybe used a little bit more character development in this show, that with all of its moving pieces his own journey, both ideologically and as a person, maybe got a smidge sidelined. But I do like that we see him dreaming about his sister, since finding her was the big obsession driving him in the early episodes of season one. Now, I feel like he’s accepted that she’s gone but still the idea of her, his depth of feeling for her and the pain of losing her is part of what drives him to create a better world, not entirely unlike Luthen being driven by his love for Kleya. And speaking of love and the things that drive us, how did y’all feel about that final-final image of the show?ZZ: I loved it! To me it worked perfectly with a theme in Andor: hope.EG: “There is another.” lmao.CP: Right, to me it did in part feel like a nod to Star Wars’ obsession with dynasties and legacies, like we have to believe that, though Andor himself dies, what he stands for will live on not just as an idea but because he literally has a child. And yet, I still kinda liked it. We didn’t get a lightsaber but we did get a continued bloodline!ZZ: I think the show needed some hope at the end.EG: I think it was very thematically appropriate, even if I’m torn on the merits of mixing insurgency and family. It’s a division that feels a bit too tidy.ZZ: I think it did provide more reason for Bix leaving like she did. She was pregnant. She wanted to give her child a peaceful life and knew Andor would follow her if he knew. And in her mind, she’s thinking that they’ll get back together one day after the Empire has fallen. It’s both a very tragic final scene and also this reminder that there is more. This isn’t an ending.Screenshot: Lucasfilm / KotakuCP: Vel even tells him not to wait too long to reconnect, and we already know he never gets the chance! Really loved that those two, Cassian and Vel, got a moment here, too, and got to acknowledge all those they’ve lost along the way. But yes, you’re right, it was a lovely mix of deeply sad and hopeful, that final image. Luthen, Cassian, Saw, and so many others know they’re fighting for a world they themselves will likely not live to see. But that kid might.ZZ: And before we leave, I did like that we got one more tiny moment with Mon’s husbasndHe seems to be with the mother of the boy his daughter married? It was very fast. Couldn’t tell. But him just getting drunk in a limo on Coruscant, presumably throwing his wife under the bus and pledging loyalty to the Empire, seemed like all we needed to know about what happened to him.CP: Exactly. He is who we knew he was and his sad empty privileged life is his reward for it.I’d be curious to know how that final montage plays for folks who haven’t seen Rogue One. It really worked for me, seeing Cassian all dressed up for his fateful mission, the cuts to Dedra and other characters, and all around him, the Rebel base on Yavin, active and buzzing, about to change the galaxy, and now we know it’s all because of the efforts of so many people but among them, one Luthen Rael, an unsung hero of Star Wars. Are either of you planning on rewatching Rogue One any time soon?ZZ: I wanted to hold off until after this VG chat so I came into this without the weight of Rogue One on my mind. I plan on watching it this weekend! EG: I will say, as a parting thought, I don’t know that I needed the show to try and line up so neatly with Rogue One, perhaps the worst part of which is that silly blueprint handoff that directly leads into A New Hope. I do think some of the broader thrust of Andor and the unease and disquiet within its characters ended up being subsumed a little to neatly by the end of episode 12.CP: Oh, I agree. At a certain point in the final episode you really feel the show shift into “Okay, let’s get all the pieces in place for Rogue One” mode.ZZ: Yeah. It reminds me of the ending of Star Wars Episode III, where George Lucas sets up all the pieces for A New Hope and it feels less like an actual ending and more like a checkpoint.CP: And I think heading right from Andor into Rogue One will be quite jarring because—sorry Rogue One!—your dialogue is just not on the same level!ZZ: Nope! And what happened to Bail Organa! Did he get a haircut?CP: Hahaha.ZZ: But really, if that’s my biggest complaint about Andor—that its ending isn’t as strong as it could have been because of Rogue One—I’m still really happy.I’m not sure we’ll ever get a show like this again, or at least not for a long time. Real sets. Lots of actors. Incredible writing. Big budgets. Set in a large franchise. All this freedom. Even Gilroy has stated he’s not sure if this kind of thing will ever happen again.CP: It was glorious, and while I really hope we see more like it, I’ll try to just be grateful for the miracle that we ever got it at all. Now I just need Disney to put it on Blu-ray so I have it on physical media and it’s not trapped on a streaming service forever!ZZ: Rebellions and physical libraries of movies we love are built on hope. .

In 1901, a kid from Ohio named Harry Gerstner had had enough of high school and dropped out. He then started a four-year woodworking apprenticeship and trained to become a patternmaker.About a year into his patternmaking job, Gerstner became dissatisfied with the then-current state of tool storage. Tool chests existed, but they were exactly that—literal chests, which were large, heavy, and needed to be fished through to find the tool you wanted. During his evenings and weekends, Gerstner began developing a solution that would better suit him.He came up with a small chest of drawers that was much better-suited to hold the smaller, delicate tools used by patternmakers. Wrapped in leather, with brass fittings to protect the corners and a handle on top, it was durable and easy to move around the benchtop. The small drawers of gradated size made it easy to store and retrieve hand tools.Gerstner's co-workers took note of his invention, and he soon realized there was nothing else like it on the market. He shortly stopped working as a patternmaker and was soon running H. Gerstner & Sons, a business developed to manufacture and sell his toolchests. (Fun fact and early marketing spin: Gerstner had no sons, but three daughters. He named his new company to provide the illusion of history.)Gerstner's chests proved popular, and the design eventually evolved into what we today call a machinist's chest. Machinists became the de facto target market, as the chests were perfect for storing their calipers, gauge blocks and precision tools. The front of the case has a protective cover that slides down to reveal a felt-lined surface on which to temporarily place small parts or tools. Drawers can be pulled out and carried to the work. The diamond-shaped mirror is a bit of a curiosity; I'd heard that they were supposedly included in case the machinist got something stuck in their eye, while others have dismissed that as hogwash. This forum poster claims he reached out to the company (which still exists today) to inquire about the mirror's purpose, and received the following response:"The mirror was first installed in the early 1900's on some of the very earliest Gerstner chest designs. In those days, very few manufacturing and/or tool & die companies had indoor plumbing. In other words, there were few companies with lavatories (with wash basins and mirrors) for toolmakers and other factory workers to use to clean up at the end of the day."The 'mirror' mounted inside a Gerstner tool chest provided the wherewithal for the chest owner to check his appearance and comb his hair before heading home (or, perhaps, going in to ask the boss for a raise). At various times over the years, we have introduced a few tool chests without including a mirror. Interestingly enough, the feedback from customers has been, 'Without a mirror, the chest doesn't look like a Gerstner.' So . . .you may notice that virtually all new model Gerstners feature a 'mirror' inside the tip lid."That explanation sounds fishy as hell to me—and why is "mirror" occasionally placed in quotes?—but what do I know, I'm not a detective.In any case Gerstner had found an untapped market, and multiple competitors cropped up; by the end of the 1910s, there were about 15 companies making a similar chest. H. Gerstner & Sons outlasted them throughout the century, and had no competitors left by the mid-1960s.Since their launch in 1906, Gerstner priced the chests at roughly a week's wages for a skilled tradesperson ($6 in 1906) and kept that ratio steady throughout the 20th century. Today, they still sell a model, the Classic Chest, that is remarkably close to the 20th-century design. (The most notable difference is that the single top handle has been replaced by two side handles.) It's available in the original Oak as well as fancier woods like Walnut and Cherry. However, the prices of the Classic today is $1,695. The Bureau of Labor Statistics has the median salary of a machinist today at $56,150, which would make a week's wages about $1,080. That doesn't place the chest completely out of reach of a machinist, but I'm guessing today's customers are more likely filling the drawers with knife collections and RC car parts.

93FT has created the playful interiors of Starwood Hotels’ second UK branch of its Treehouse brand. With chalk boards, patchwork quilts and tree stump furniture, the hotel leans into childhood nostalgia and creative freedom. Aimed at 25 to 45-year-olds, this Treehouse Hotel is housed in a 1970s concrete building in Manchester’s city centre. It started life as an office block, but more recently operated as a Renaissance Hotel. The building started life as an office block This iteration was designed by Sheffield-based 93FT in collaboration with Starwood Hotels’ international design team. “Throughout the project we have followed our brand ethos of ‘carefreedom’ whereby we want to transport guests back to a time when they were children, more carefree and able to explore their world creatively,” says Neil Andrew, Starwood’s vice president, design – EMEA. 93FT co-founder Nick Clark echoes this sentiment. “Treehouses are about escaping. They’re about a sense of nostalgia, and feeling young again.” This extends to every design detail – Barry Sternlicht, the hotel group’s founder wanted these hotels to feel like “a kid had made them,” says Amber Addison, 93FT senior interior designer. The team worked to strike a careful balance and meet health & safety standards The challenge was to make this playful environment also be commercial, comfortable and meet health and safety standards, she explains. And aesthetically, the team had to strike a balance between the designs looking not too ramshackle, but not too polished either. The agency brought in a mishmash of vintage items from different sources including salvage company Retrouvious, plus reclaimed building materials and a variety of fabrics, to capture that spontaneity of children’s creativity. Glass Onion created bespoke cushion covers for all 224 bedrooms, using denim off-cuts For example, vintage clothing business, Glass Onion, which has a store in Sheffield, made cushion covers from denim off-cuts for each of the 224 bedrooms. “A treehouse can be made of things that your mum doesn’t want in the lounge anymore,” Addison says. Hence in the restaurant, Pip, there are eclectic lampshades, and one wall is filled with shelves of rainbow glass bottles of different shapes and sizes, which are intended to conjure up the intuitive collecting habits of children. Meanwhile, the bedrooms are furnished with individually made patchwork quilts, asymmetrical door handles, handmade furniture and larch cladding. Chalkboards lean into childhood nostalgia and creative freedom In the entrance, a mix of reclaimed and handmade windows make up the lobby walls. And some of the chairs are made from tree stumps, with old tool handles embedded in them to form a back rest. This sourcing came easily to 93FT, which brought in 1,400 items for its redesign of London’s Pilgrim Hotel. Other playful elements include chalkboards, which guests are inviting to doodle on, or they can carve their name on small pieces of wood, which are then hung in the lift lobby as an alternative guestbook. The agency initially came up with some strong ideas to bring the treehouse concept to life. “We loved the idea of a treehouse in a Treehouse,” Clark says. “We were going to cut some big shapes into the floorplates, and add a big tree with a treehouse at the top.” Another idea was to have a big swing in the entrance, but neither of these ideas turned out to be feasible. Existing columns are clad in timber, to give them a tree-like appearance They were able to install a big artificial tree clad in real bark, which reaches up through the ceiling of the dark, lower-ground reception. Also on this level, the timber walkway is bathed in dappled light in reference to a forest floor. The effect was created by cutting out organic shapes in the ceiling and carefully placing lighting there. A second tree made of reclaimed timber stands on the mezzanine with a bench at its base. For phase two of the build, 93FT is designing a presidential suite, and Blacksheep in London is designing a rooftop bar. In the UK, Manchester Treehouse follows on from Treehouse London in Marylebone designed by Birmingham agency Keane. It is also a transformation of an existing hotel, and that outpost has city-specific elements including Paddington Bear references and London-centric souvenirs. Further Treehouse locations in the works include Brickell in Miami by Rockwell Group; Riyadh in Saudi Arabia by Blacksheep; and Adelaide in Australia by Woods Bagot. The glass bottles are intended to conjure up the intuitive collecting habits of children The bar area features an eclectic mix of light fittings, adding character and visual interest to the space