AI代码洪流冲击软件工程：代码审查成新瓶颈

_IEEE Spectrum_ is the flagship publication of the IEEE — the world’s largest professional organization devoted to engineering and applied sciences. Our articles, videos, and infographics inform our readers about developments in technology, engineering, and science. AI coding tools can now generate thousands of lines of code in minutes, helping companies build features, run tests, and fix issues faster. But the flood of AI-generated code still has to be reviewed. [Large language models](https://spectrum.ieee.org/tag/large-language-models) can produce code that looks clean on the surface but [conceals sloppy mistakes](https://spectrum.ieee.org/responsible-ai) such as faulty assumptions, security vulnerabilities, or subtle errors that emerge only after deployment. Fixing those problems could erase the productivity gains AI promises. Companies are responding to the onslaught of AI code slop by rethinking how they review code. New strategies are emerging. Among other approaches, engineers are scrutinizing plans before AI begins coding, deploying specialized AI agents to catch routine flaws, sending risky changes to human reviewers, or requiring developers to defend the code their agents produce. Investors see an opportunity in closing that gap. In August, for example, AI code-review startup CodeRabbit[raised US $143 million](https://www.reuters.com/technology/ai-code-review-platform-coderabbit-valued-15-billion-latest-funding-round-2026-08-12/) at a $1.5 billion valuation, while claiming it performs more than 2 million reviews a week for 17,000 customers, including [Nvidia](https://spectrum.ieee.org/tag/nvidia), Indeed, and BMW Group. The new era of code review will determine whether AI can ever provide code that is both faster and more reliable. It also has some software engineers thinking about the future of their profession: If entry-level engineers spend less time writing code themselves, how will they learn to judge it? AI-written code is shifting the bottleneck from generating software to reviewing it. According to the Sonar study, 38 percent of developers said “more effort” is required to review AI-generated code than code written by their colleagues. Sixty-one percent of them said AI often produced code that looked correct but was “unreliable.” For Synthesia, an AI video-generation platform, code review has become essential to its engineering workflow. In November 2025, Synthesia’s 118 engineers went all-in on AI coding tools like Claude Code. According to Peter Hill, Synthesia’s chief technology officer, the result has been a massive surge in code volume. That code demands close examination. As of August, the number of pull requests, or proposed changes to a codebase submitted for review, had risen 120 percent year over year, Hill says. Ninety-five percent of those requests contain AI-generated code. One recurring problem is duplication. Hill says AI tools may not recognize that code for a task already exists, and they’ll write another version because they have limited context. Synthesia has found as many as 10 versions of the same function, leaving engineers to identify and remove redundant functions. Once that’s done, engineers retrain the AI agent so that doesn’t happen again. At the company’s scale, Hill describes getting the AI to produce the intended output an “enormous amount of work.” Much of that work involves writing a “specification,” which is a detailed plan for what the AI agent should build and how. Preventing recurring mistakes before generating code can save engineers time later. Stanley recalls how a missing instruction once caused an agent to generate 25,000 lines in the wrong version of the programming language Swift. Switching versions produced 600 errors it could not fix at once. Stanley discarded the code, updated the specification, and restarted the agent. Fifteen minutes later, it regenerated the code correctly. David Yanacek, a senior principal engineer at Amazon Web Services (AWS), says the company uses agents to test whether code works, check it against the original plan, and look for security flaws before a person reviews it. That first pass becomes more important as AI-generated code volume increases. At Bonterra, a nonprofit software provider with about 290 engineers, proposed changes tripled within three months of adopting AI, according to Tanuja Korlepra, the CTO. Code entering review rose tenfold and review times tripled, making it impractical for engineers to inspect every line. Bonterra’s agents compare code with the approved design, security rules, coding standards, and accessibility requirements, then report their confidence in the result. A low score or flagged problem sends the change to a person. Code involving payments, personal data, or other sensitive systems always receives human review. Synthesia also uses AI agents to decide where human review is necessary. Criteria set by engineers direct more scrutiny toward higher-risk changes. Altering an error message carries less risk than code that handles customer data or core business rules. Even so, fewer than 5 percent of changes bypass human review. When machines produce more code than engineers can closely read, human approval can become “theater approval,” according to JD Raimondi, chief AI architect at the software consultancy Making Sense. In other words, an engineer might confirm that the feature works, skim the code, and approve it, all without understanding the choices underneath. Temporal, an open-sourced developer platform, puts the burden back on the person submitting the code. CEO Samar Abbas says code volume and review time have increased with AI. Under its “Send Back” policy, Temporal’s engineers must explain in their own words the agent’s design choices and how the code handles unusual conditions. Otherwise, the reviewer rejects it. Junior engineers at Making Sense have seen some of the largest productivity gains from AI, Raimondi says, raising concerns about what they no longer learn by doing. The consultancy keeps juniors involved in deciding why a customer needs a feature and how it should work, rather than limiting them to checking AI output. Synthesia primarily hires mid- and senior-level engineers. Its less-experienced employees work with both a senior colleague and an AI agent, taking responsibility for parts of projects while learning to define what successful code should do. At Bonterra, agents now perform many of the well-defined coding tasks that once trained new engineers. Juniors instead own outcomes alongside experienced colleagues, learning to direct agents, question their output, and remain responsible for the result. She says this approach can help junior engineers build the skills and knowledge needed to advance in their careers. Advanced AI labs are still training ever larger models, but that training has somewhat receded to the background of the AI conversation. In 2026, inference—the use of trained models to produce code, write essays, or make images of ourselves as elves—has come to the forefront. Part of what’s caused the shift is very simple: LLMs are becoming useful, so people are using them. On top of that, many models on the market today are reasoning models. In response to a user’s query, they run inference not just once but multiple times, reprompting themselves in a process called[_\_chain of thought\__](https://arxiv.org/pdf/2201.11903). Reasoning models generate longer outputs, and models with high reasoning effort can produce up to [20 times](https://www.linkedin.com/posts/artificial-analysis_how-many-tokens-do-reasoning-models-use-vs-activity-7318302119206289408-3mt1/) as much text as those with low or no effort. Adding even more to the world’s inference workload, the rise of [agentic AI](https://spectrum.ieee.org/ai-agents) has resulted in inference running not just as a real-time response to a user’s query but also around the clock, working autonomously toward a user-defined goal. Although they might seem similar, AI training and AI inference are computationally different. These big moves from tech giants signal that in order to support the inference demand, we’re going to need a very different mix of hardware than experts may have expected even a couple of years ago. An untrained LLM is like a jumble of Scrabble tiles on a table. Instead of single letters, though, the tiles show fragments of words, called tokens. Everything you’d need to write almost anything is present, but nothing makes sense. Training a model organizes this jumble using a guessing game played at scale. The model is shown real text with the next token hidden and asked to predict what comes next. After each guess, the correct token is revealed and then compared to the prediction, and the difference is used to calculate the model’s accuracy. The game is played not with a single sentence but over billions of passages. While a real game of Scrabble can be played over a bag of chips and a few drinks, AI training is computationally intense. The model updates its parameters through [backpropagation](https://spectrum.ieee.org/what-is-deep-learning/backpropagation), a process that repeatedly calculates how each of a model’s billions or trillions of parameters should shift to make the next prediction better. This is why tech giants are [building](https://spectrum.ieee.org/5gw-data-center) larger data centers than ever before. Eventually the model’s creator decides further training isn’t worth the cost, and the guessing game stops. Backpropagation ends, the parameters are frozen, and the LLM becomes a pretrained model. Fine-tuning—a short training run on smaller, more specialized data—adds final tweaks, and the model is deployed. You might think that AI inference is less computationally demanding because the backpropagation calculations used to update parameters are eliminated. But [Sudeep Bhoja](https://www.linkedin.com/in/sudeep-bhoja-070a111/), founder and CTO of the inference-hardware company [d-Matrix](https://www.d-matrix.ai/), explains that inference adds new challenges. The models are “autoregressive” in nature. That is, the next output depends on the previous one. “So to generate the next token, you have to read all of the weights and all of the [context] from the previous token,” explains Bhoja. The context includes all of your prompts, all of the LLM’s replies, and all of the files you upload. It’s a lot of data and a lot of processing. An LLM generates its reply in two phases: prefill and decode. Prefill is the model reading a prompt. It processes every token at once, computing how each token relates to all the others. This operation is called [attention](https://en.wikipedia.org/wiki/Attention_(machine_learning)), and it’s a defining characteristic of the transformer architecture behind modern LLMs. It allows them to respond to a word in its sentence, paragraph, and larger context rather than on its own. Think of it like arranging Scrabble tiles before you place them in a game. Many players move tiles around to imagine how they connect. Self-attention plays a similar role, though instead of moving physical tiles, each token sends a query to the others and receives a score indicating the token’s relevance. These queries result in two types of vectors: the keys and values. They are typically placed in a store called the KV cache. This isn’t strictly required, as a model could instead recompute these vectors with each new token it generates. But nearly all LLMs use a KV cache to reduce how much computing they do. The KV cache is stored in memory and becomes a scratchpad to which the LLM can return to understand a conversation, and though it starts small, it can swell to dozens of gigabytes. Prefill is a problem that can be easily divided up and worked on in parallel. This is why GPUs became the dominant AI accelerator as LLMs surged in popularity. Graphics rasterization (computing the color of every pixel on a screen) is also massively parallel, so GPU architectures were a natural fit. Next comes decode. Here, the model generates its reply one token at a time. At each step it takes the most recent token, weighs it against everything in the KV cache, uses that information to predict the next token, and adds the new token’s key and value to the cache. Then it repeats in sequence, token by token. This is where the autoregressive nature of the model works against inference speed. Predicting each token requires reading the entire model from memory, and that model consists of possibly tens to hundreds of gigabytes of parameters (the numbers representing what the model learned in training). Crucially, this is in addition to the memory required to store the KV cache. As a result, the movement of all this data through memory often requires more bandwidth than inference hardware has available. So at least some of the computing parts of a GPU sit idle as it waits for data. Researchers [found](https://arxiv.org/pdf/2503.08311) that Nvidia H100 GPUs running open-source LLMs sit idle 50 to 80 percent of the time. d-Matrix’s Raptor removes that bottleneck by stacking an AI accelerator on a DRAM die. Instead of stacking memory, d-Matrix stacks memory and compute. Bhoja says this reduces the distance that data must travel to “micrometers instead of millimeters.” Like building a skyscraper, going vertical makes it possible to do more inside the same physical footprint. Majestic takes the opposite approach. Instead of trying to minimize the length that data must travel between compute and memory, the company is focused on improving the memory interface to accommodate longer wire traces while keeping bandwidth high. Longer wires allow Majestic to connect memory stacks that aren’t directly next to the GPU, removing the space limitation of HBM. “A memory interface has a very short physical distance it can operate over. In the case of HBM, it’s up to 2 or 3 millimeters. You have this shoreline around the periphery, which is the only place where you can put HBM,” says Rabii. d-Matrix and Majestic have one thing in common: Instead of HBM, they both use off-the-shelf DRAM. This is the most common type of computer memory in the world; it’s in everything from smartphones to cars. Memory analyst [Jim Handy](https://thememoryguy.com/) says HBM costs two to three times as much as DRAM. d-Matrix and Majestic chose DRAM in part because of this price advantage. However, the proponents of HBM, which include memory giants like [Samsung](https://www.samsung.com/us/) and [SK Hynix](https://www.skhynix.com/), aren’t sitting idle. The big players—Nvidia and Amazon—are going for an all-chips-on-deck approach. Nvidia’s GPUs and Amazon’s Trainium training accelerators are still great for part of the inference workload: the prefill stage, where all the context keys and values are calculated. But to accelerate decode, the part where new tokens are generated, they are looking to new, memory-centric architectures from smaller players. In Nvidia’s case, the smaller player was Groq (not to be confused with Grok, the family of LLMs trained by SpaceXAI). Nvidia purchased intellectual property and hired talent from Groq at the end of 2025, and just three months later at the Nvidia’s GTC 2026 conference, Jensen Huang [unveiled](https://spectrum.ieee.org/nvidia-groq-3) the Nvidia Groq 3 language-processing unit ([LPU](https://www.nvidia.com/en-us/data-center/lpx/)). Groq’s architecture relies on memory—in its case, SRAM—built directly into the chip’s architecture. Between the Rubin GPU and the Groq LPU, prefill and decode can both be accelerated to get the best of both worlds, the theory goes. “We do all the attention math and context processing on the Vera Rubin [GPU] rack,” explains Buck. “For all the expert calculations…the matrix multiplications, we do that part on the LPU.” The company packs 256 LPUs into the Groq 3 LPX, a system the size of a data-center rack. Amazon plans to use AWS Trainium chips for prefill, and Cerebras for decode. But Cerebras’s chips can also go it alone in inference. WSE-3 was [deployed by OpenAI to power GPT-5.3-Codex-Spark](https://openai.com/index/introducing-gpt-5-3-codex-spark/), a variant of the company’s coding mode, outputting over 1,000 tokens per second. For comparison, OpenAI’s standard GPT-5.4 deployment outputs 50 to 125 tokens per second. Cerebras can also tackle prefill without moving the workload to different specialized chips. For this, it networks together multiple WSE-3 chips to form a single pool of memory. Cerebras has demonstrated it can serve models with up to 1T parameters, such as [Moonshot](https://spectrum.ieee.org/tag/moonshot) AI’s Kimi 2.6, though Wang says “the architecture has no innate limitation in terms of how many parameters it will do.” Despite these differences in strategy, Nvidia and AWS seem to agree that the future of AI inference will be solved by a systems approach that pools different kinds of chips together to tackle the largest LLMs. Or, as Buck says: “To do modern AI inference, you need all the chips.” Nvidia became the world’s most valuable tech company because it designed the world’s most desired GPUs. But not all of the attention is focused on improving AI-inference hardware. AI researchers are also learning how to optimize LLM software and hardware in tandem to make the best use of the memory and compute components. Most computers store numbers in a 32-bit or 64-bit format. These determine how many bits are available to represent a single number. If too few bits are available, the number can’t be stored without losing information. The quality of an LLM benefits from more-precise number formats, but this creates a problem for inference performance. More-precise numbers aren’t free. The bits that describe them take up more space in memory and require more silicon and energy to compute. The process of converting an LLM from a more-precise number format to a less-precise format is called [quantization](https://spectrum.ieee.org/1-bit-llm), and it’s been in use for several years. However, researchers are finding new ways to quantize models down while retaining a large majority of the model’s quality. Quantization is likely just the tip of the spear, as AI researchers and startups are investigating a diversity of opportunities for optimization, some of which could dramatically change the silicon found in AI-inference hardware. The sheer variety of approaches to AI-inference acceleration—stacking compute on memory, extending interfaces from millimeters to meters, using an entire silicon wafer for SRAM, squeezing models into 4 bits—raises a question: Which is going to win, and which is going to lose? On the contrary, Kimball of Moor Insights & Strategy thinks inference could drive intense demand for AI hardware in the long term, because it’s not obvious where that demand will end. “You could add a million agents into your organization,” he says. “These things work 24 hours a day; they don’t go home at five at night like we do.” If AI inference remains as desirable as Kimball expects, the evolution is likely to follow the same trajectory as the CPU. The CPU didn’t improve along a single axis but instead across [multiple fronts](https://spectrum.ieee.org/intel-i860) simultaneously. Once transistor scaling slowed, chip and system architecture innovations of all kinds proliferated. The list of individual innovations that led to today’s ubiquitous, powerful personal compute could fill dozens of books. Cyberattacks can change how robots act without obvious signs of failure, requiring a new approach to safety that combines AI model and vulnerability scanning, simulation-based validation, and continuous monitoring to help secure robots from development through operation. Robot safety has traditionally asked: Can a machine remain safe when something goes wrong? Physical AI raises a harder question: Can a machine remain safe when an attacker changes what it sees, decides, or does even when nothing appears to have failed? As AI and robotics continue to advance at an unprecedented pace, modern robots perceive through multimodal sensors, interpret context using [AI models](https://spectrum.ieee.org/tag/ai-models), and translate those interpretations into physical action. As they move into dynamic environments, their safety increasingly depends on the integrity of the data guiding their decisions. That dependence creates risks that conventional safety assessments may not fully capture. Recent research has demonstrated that manipulating what a robot sees, hears, or interprets can influence its behavior without requiring direct control. So a critical safety question today is whether Physical AI models remain within their task and safety boundaries under adversarial conditions. [Simulation tools](https://vicone.com/company/press-releases/vicone-turns-def-con-34-robot-hacking-research-into-free-nvidia-isaac-sim-extension?utm_source=ieee-spectrum&utm_medium=sponsored-content&utm_campaign=2026-09) such as [NVIDIA](https://spectrum.ieee.org/tag/nvidia) Isaac Sim, when paired with [VicOne Radeis](https://vicone.com/products/radeis?utm_source=ieee-spectrum&utm_medium=sponsored-content&utm_campaign=2026-09), can test the effects of manipulated inputs before deployment. In this case, the components may still function as designed. What has changed is the trustworthiness of the commands flowing through the system. Vulnerability management can help teams identify known risks before deployment, while continuous monitoring can surface emerging threats. In each case, the camera may still work, the model may still run, and the controller may still respond. Yet the resulting behavior can be unsafe because the robot is acting on manipulated perception or reasoning. Runtime assurance must therefore look beyond whether individual components remain available and assess whether cyber events are beginning to affect physical behavior. Security event correlation, behavioral-impact assessment, and policy-bounded response supported by edge AI, can help contain the affected path without unnecessarily stopping the entire robot fleet. This requires assurance across the robot’s lifecycle. During design, teams need to understand which cyber risks could invalidate assumptions behind intended behavior. Before deployment, they should test whether realistic attacks can cause a robot to deviate from its task or safety boundaries. In operation, monitoring should identify whether cyber events are beginning to affect behavior, contain the affected path, and preserve safe operation where possible. For a deeper look at the cybersecurity risks and defense strategies shaping autonomous robotics, download our whitepaper “[Securing the Rise of AI Robots: Cyber Risks, Real-World Threats, and Defense Strategies](https://info.vicone.com/ai-robotics-security-risk-whitepaper?utm_source=ieee-spectrum&utm_medium=sponsored-content&utm_campaign=2026-09).” This white paper gives distribution engineers and utility teams a practical overview of the substation exit, the short and high-consequence section where station capacity divides into individual feeders. It explains how covered, spacer-supported overhead construction can reduce common contact-driven faults while leaving room for future circuits. > _A microwave box creates a surprisingly tricky vision system challenge: there’s a picture of a microwave on the side of the box, and that microwave has a window. From the camera’s perspective, it’s a box inside a box inside a box._ > _As the evolution of the UR Series, the g?Series is designed to accelerate how robots are integrated into complete automation systems. It connects more directly with cameras, higher power EOATs, and control systems, eliminating external cabling and hardware to streamline integration. This enables cleaner and faster deployment of advanced automation._ This Rivian might look like any other electric SUV in this affluent town, with its concentration of tech bros, venture capital, and startups. But inside this boxy EV is something special: an Autonomy+ system that will allow owners to enter an address, sit back, and let the vehicle drive to any mapped destination in the U.S. and Canada. This point-to-point system is one of the most advanced semiautonomous-driving systems coming to market. It is also a precursor of the company’s bid to make [self-driving cars](https://spectrum.ieee.org/tag/self-driving-cars) a reality, for [robotaxis](https://spectrum.ieee.org/tag/robotaxis) and—eventually—for everyday car buyers. After years of incremental advances and frustrating setbacks, self-driving has been swept up in the great AI resurgence, and is now a top priority for [investors and global automakers](https://www.wsj.com/video/series/wsj-the-future-of-everything/how-uber-plans-to-win-the-self-driving-car-race/10F91546-7884-4404-8B65-E252B6C514A3), who envision vast new streams of profits. Video released by Rivian shows the company’s R1 SUV being driven on a variety of urban and rural roads, according to the company. Rivian plans to introduce this self-driving system to compete with Tesla’s offering before the end of 2026.Rivian Impressive as it is, Autonomy+ is only a Level 2+ system in the classification system established by the Society of Automotive Engineers. Level 2+ means that a human driver must pay full attention and be ready to retake control at any moment. Rivian, along with a horde of deep-pocketed rivals, is aggressively working toward more impressive (and potentially lucrative) levels of autonomy. At Level 3, drivers could “check out” behind the wheel for limited periods, to scroll through emails or watch a movie—but not to sleep. Robotaxis currently roaming select cities in the U.S., China, and the Middle East have proved that driverless, Level 4 autonomy is possible. These cars operate in relatively tiny numbers in a [couple of dozen cities](https://insideevs.com/news/776837/china-us-robotaxi-fleet-comparison/), and within the specific constraints of commercial services. Now Rivian and its many rivals—including Tesla, [Toyota](https://www.toyota.com/cars/), Mercedes, [Volkswagen](https://www.vw.com/en/corporate.html), and [China’s BYD](https://www.byd.com/en)—are racing to bring that level of self-guided mobility to the masses. Rivian’s strategy combines a suite of cameras, radar, and [lidar](https://spectrum.ieee.org/tag/lidar); a custom set of silicon chips, developed in-house, to process sensor data; and an AI autonomy model running on those chips. With $1.25 billion in backing from [Uber](https://spectrum.ieee.org/tag/uber), Rivian plans to graduate to a fleet of self-driving, Level 4 robotaxis starting in 2028. Those taxis, along with Rivian’s consumer fleet, will be the literal training wheels for extending Level 4 ability to consumer vehicles. Meanwhile, millions of connected cars, as they cruise every nook and cranny of the globe, are already sending data to train automakers’ systems. The race is on to funnel those data through fast-improving [AI models](https://spectrum.ieee.org/tag/ai-models) with “end to end” capability: an AI architecture, powered by [deep learning](https://spectrum.ieee.org/tag/deep-learning), that processes raw sensor data directly into physical vehicle commands. So equipped, engineers anticipate they’ll be able to solve the tricky edge cases—tangled urban streets, unique geographies, swarms of pedestrians, inclement weather—that skeptics once deemed intractable. Despite the company’s high media profile, including a spotlight on RJ Scaringe, its MIT-doctorate founder and chief executive, Rivian holds a relatively tiny slice of the U.S. passenger-vehicle market. It sold just [42,000](https://rivian.com/newsroom/article/rivian-releases-fourth-quarter-full-year-2025-financial-results) vehicles last year across its three models, the adventure-minded R1S SUV and R1T pickup, and the Electric Delivery Van. Tesla sold about 1.6 million units. Toyota, the world’s largest automaker, sold more than 11 million. Rivian’s underdog strategy is to leverage software and tech to make itself a serious player. Volkswagen, among the world’s largest automakers, saw enough value there to invest up to $5.8 billion in a joint venture called Rivian and Volkswagen Group Technologies. The joint venture gives Rivian crucial capital for development. It gives Volkswagen access to Rivian’s electrical architecture and to the software for the R2, new-generation Rivian SUV that went on sale in June. “Rivian developed an architecture so important that VW is spending billions to buy it, as opposed to trying to re-create it themselves,” says [Bryan Reimer](https://ctl.mit.edu/people/reimer-bryan), a research scientist in MIT’s [Center for Transportation and Logistics](https://ctl.mit.edu/). But the joint venture doesn’t give VW access to Rivian’s autonomous tech. In March, that R2 architecture underpinned Rivian’s [$1.25 billion deal](https://apnews.com/article/uber-rivian-robotaxi-autonomous-019439a7e5dd3c855c7171f8de3e9ce9) to supply Uber with up to 50,000 robotaxis. The companies plan to initially deploy 10,000 taxis, beginning in San Francisco and Miami in 2028, before expanding across 25 cities in the U.S., Canada, and Europe. Rivian’s vulnerabilities include struggles with reliability, along with expensive body repair costs that the company says it strove to reduce for its new R2. As impressive as Rivian’s in-house tech may appear, the company has miles to go to catch up with Tesla, which recently announced it has 1.1 million active users of its FSD system. Toyota is also jumping into the game; its [Woven by Toyota](https://woven.toyota/en/) subsidiary has partnered with the Alphabet-owned [Waymo](https://robotsguide.com/robots/waymo) to develop an autonomy platform for robotaxis _\_and\__ consumer cars. Until recently, most observers would have gone all-in on Tesla as the winner of the autonomous race. Elon Musk’s company has begun operating a small test fleet of [Model Y robotaxis](https://www.tesla.com/robotaxi) in three Texas cities and in Florida. Tesla has also begun producing a dedicated autonomous vehicle, the [Cybercab robotaxi](https://www.bloomberg.com/news/articles/2026-04-24/musk-says-tesla-has-begun-production-of-its-cybercab-robotaxi). But in April, Musk pushed back his timeline for Level 4 autonomy for general consumers: “I’m just guessing here, but probably in the fourth quarter” of 2026, [he said](https://electrek.co/2026/04/22/tesla-elon-musk-unsupervised-fsd-consumer-cars-q4-delay-again/). It was the latest in a series of deflating walkbacks from the man who once promised 1 million robotaxis on the road by 2020. Scaringe, during our drive of his company’s make-or-break R2 SUV at a Utah state park, says that showroom Rivians will start adopting some of its robotaxis’ Level 4 capabilities no later than 2030, perhaps beginning with self-parking functions. Like most autonomous cars, Rivian’s system fuses data from multiple sensors to create a robust picture of a fast-moving environment and its obstacles. Data is fed to a neural network—what Rivian refers to as its “[Large Driving Model](https://www.wardsauto.com/news/rivian-announces-new-ai-hardware-software-autonomy-day-event-r2/807844/),” or LDM—that churns through hundreds of trillions of operations per second to interpret and fuse data from cameras, radar, and lidar. That network is [end to end](https://ieeexplore.ieee.org/document/8576190), meaning that it processes multiple streams of raw sensor data (such as camera pixels) and outputs driving controls (for steering, braking, and acceleration) through a single data pipeline. More traditional systems coded distinct steps for data collection, feature extraction, prediction, and decision-making. That proprietary AI driver identifies features in images and point clouds, groups them into objects, and tracks them across frames, time-stamped to the millisecond to account for differing frame rates. The AI thus builds confidence over time, acting on object detections that persist across several frames, rather than, say, slamming the brakes due to a camera blip on a single frame. The virtual driver can then navigate safely even when sensors disagree, by favoring the persistent data. The output— commands for [electric motors](https://spectrum.ieee.org/tag/electric-motors) and other systems—is backed by redundant hardware for by-wire systems such as steering and brakes. During my demo of Rivian’s point-to-point Autonomy+ system, a company test driver sits behind the wheel. Nick Nguyen, the engineer who directs Rivian’s products and programs related to autonomy, watches from the back seat. Compared to, say, a large language model that writes news or fiction, Nguyen says, an autonomous-driving AI is less subjective and easier to evaluate, so there’s little room for error. “We want cliché. We want boring. Just safe, repeatable driving,” he says. Rivian’s core LDM currently ingests cloud data from up to 125,000 cars for analysis and validation, which then fine-tunes the model through simulations. Onboard computing is smart enough to trigger recording only for unusual scenarios. Owners have to agree explicitly to data collection beforehand. Updated LDMs will be beamed back to customer cars via monthly over-the-air updates. part of that self-reinforcing data flywheel. It’s part of what Scargine calls the “data flywheel,” the self-improving AI loop that continuously refines the system. As is true for some of its rivals, Rivian no longer needs to fully rely on an onboard high-definition map or even a cellular link to pinpoint the car for navigational purposes. That strategic shift reduces data demands, and ensures steady driving in urban canyons or tunnels with no connections. Instead, the Rivian recognizes and responds to its surroundings through recognition and repetition, just as a human would do interpreting street signs, following lane markers, being alert to hazards. The Rivian R2 features 11 high-definition cameras and five radars. It will integrate a lidar unit early next year to lay the groundwork for future autonomy. That miniaturized lidar will integrate smoothly into the R2’s existing roofline, an improvement over the [bulky, drag-producing](https://www.tangramvision.com/blog/sensing-breakdown-waymo-jaguar-i-pace-robotaxi) units seen on Waymo Jaguars, and older partially autonomous models. [Vidya Rajagopalan](https://www.sonatus.com/resources/vidya-rajagopalan-of-rivian/), Rivian’s senior vice-president of [electrical engineering](https://spectrum.ieee.org/tag/electrical-engineering) hardware, says lidar costs have fallen from above $10,000 to a few hundred dollars in under a decade. A mix of sensors plays up the strengths and diminishes the weaknesses of each, Rajagopalan says. Cameras capture color and texture and can distinguish between objects, but they struggle in darkness and low-contrast lighting. Lidar is unaffected by darkness or blinding sunlight, and senses shapes in three dimensions. This inherent 3D capability makes lidar more reliable for slowing or halting a car for random objects—“a tire in the road, or maybe a large dinosaur,” Nguyen quips. Multiple cameras can further contribute 3D data, after a short delay for processing. Sensors with 360-degree vision can outperform human senses in key situations. Radar and lidar can spot nighttime pedestrians or animals hundreds of meters down the road, something no human can do. But lidar can be thrown off by dust, fog, and snow. Radar can “see” through rain or snow, but with relatively low spatial resolution. To handle the flood of sensor data, Rivian has taken on an ambitious challenge: designing its own custom autonomy chip in-house. The Rivian Autonomy Processor (RAP1) is a 5-nanometer processor that can execute 800 trillion operations per second (TOPS), three times as fast as the Nvidia [Jetson Orin](https://www.nvidia.com/en-us/autonomous-machines/embedded-systems/jetson-orin/) chip used in its earlier models. The chip will be built to Rivian’s specs by Taiwan [Semiconductor Manufacturing](https://spectrum.ieee.org/tag/semiconductor-manufacturing) Co. , which also makes custom chips for Tesla. On paper, a single AGX Thor chip is slightly faster in terms of TOPS, at 1,000 trillion operations per second. But Rivian combines a pair of chips in each autonomy module, giving it 1,600 TOPS and execution rates around 5 billion pixels of data per second, versus 3.5 billion for Nvidia’s Thor. Rajagopalan says developing the chip and AI software simultaneously shaved a critical full year from development. Experts say it’s the kind of fast-to-market speed that China has mastered and that legacy automakers are struggling to match. The in-house design allows Rivian to custom-tailor its software to the chip, and vice versa. Nvidia’s general-purpose chip, designed to satisfy multiple customers with various needs, must devote computing power to onboard infotainment, displays, or other systems. Rivian’s chip is designed to run autonomy and nothing but. Early fusion preserves the richest sensor data for maximum accuracy in self-driving. But it demands the enormous computing power the RAP1 can deliver. Raha says the approach helps the multimodal system degrade gracefully and continue to operate with certainty even if, say, a camera’s lens gets covered with mud. Rivian’s next planned milestone toward self-driving will be Level 3 autonomy—a hands-off and _\_eyes-off\__ system, but for highways only. (Remember, Tesla’s current FSD is technically a Level 2 system: hands off but _\_not\__ eyes off.) On the freeway, Nguyen points out, drivers would be spared the drudgery of dealing with stop-and-go traffic, allowing them to boost productivity or just goof off. Some autonomy critics are leery of Level 3, envisioning a limbo zone in which drivers are lulled into a [false sense of security](https://www.autonews.com/technology/an-automakers-turn-to-level-3-autonomy-amid-robotaxi-hype-0116/) when a car drives for long stretches with no human attention required. Ford and GM are among the automakers pivoting toward limited eyes-off functions. Rivian’s senior vice-president of autonomy, James Philbin, sees Level 3 as an inevitable stepping-stone to Level 4. The company expects it will initially be limited to highways, not the cut-and-thrust of city traffic. If a driver fails to respond to alerts, the system will slow the vehicle, pull off on a shoulder, or call 911. Rivian has not announced a timeline for introducing limited Level 3 capability. Ready or not, these much more autonomous systems are coming, a natural evolution of today’s semiautonomous helpers. In developed markets, adoption of showroom cars with partial-to-full automation is projected to jump from 8 percent in 2024 to 28 percent by 2030, [according to Morgan Stanley](https://www.morganstanley.com/insights/articles/self-driving-vehicles-industry-growth). “One in four cars sold globally may be equipped with smart-driving technology in five years, versus one in eight cars now,” wrote [Tim Hsiao](https://www.morganstanley.com/asiaresearch/country-and-region/taiwan.html), a Morgan Stanley analyst, in [a note posted on the company’s website](https://www.morganstanley.com/insights/articles/self-driving-vehicles-industry-growth). Combining cameras, lidar, and radar gives a self-driving car a better view of people and objects in front of it, according to Rivian. The company expects to release a self-driving system before the end of 2026 that will compete with Tesla’s, which uses cameras alone.Rivian MIT’s Reimer believes the self-driving revolution will really begin when the technology migrates from controlled taxi fleets to consumer cars, giving owners back the precious time they waste on commuting. If owners could truly send their autonomous car to safely chauffeur children, keep aged parents mobile, or run errands—while owners keep working or playing—the automakers who first help make that happen will enjoy a massive competitive edge, he says. As automakers struggle to convert buyers to subscription models, Reimer believes that self-driving appears to be the one advance for which consumers might actually pay plenty. But the greatest impediment to that revolution has little to do with technology. Public skepticism over self-driving is rampant; and the fate of fully autonomous testing in [New York City is uncertain](https://www.thecityreporter.nyc/2026/04/06/waymo-driverless-cars-testing-roads-autonomous-vehicle/). Even going from Level 2 to [Level 3](https://www.kbb.com/car-advice/level-3-autonomy-what-car-buyers-need-know/) might shift legal liability for some accidents from drivers to automakers. But with Tesla still fighting lawsuits over its rudimentary Autopilot systems, those questions aren’t anywhere near settled. Experts worry that self-driving cars may become as politicized as EVs. Labor unions are pushing back, fearing job losses from taxis to trucking. A crazy quilt of state or local regulations has failed to create coherent industry guidelines. Publicized failures—even ones that don’t result in injuries, such as Waymos driving onto a flooded street or impeding emergency workers—give the industry a black eye. Companies like Tesla and even Waymo, Reimer says, have too often relied on an arrogant “Trust me” approach, resisting regulation and oversight. Nevertheless, the momentum toward truly self-driving cars, and massive backing from automakers and AI-besotted investors, suggests their time has come. The rest of the journey will depend as much on social and regulatory issues as technical ones, and so Reimer has a bit of advice. “Do it right, and share all your data,” he says. “Earn the right to scale…. It’s about establishing trust, and developing a framework in which we truly believe these systems can operate as a trusted part of our transportation network.” Tsubaki KabelSchlepp Robotrax System provides a specialized three-dimensional cable carrier engineered specifically for complex robotic motion, preventing premature conductor fatigue, insulation breakdown, and costly unplanned production halts. In modern automated manufacturing, six-axis articulated robots perform high-speed, multidirectional maneuvers under demanding operational cycles. However, as robot arms swivel, rotate, and extend, the electrical cables, fiber [optics](https://spectrum.ieee.org/tag/optics), and pneumatic hoses supplying them endure severe mechanical stress. Torsional twist, rapid acceleration, and repeated contact with machine structures often lead to premature conductor fatigue, insulation breakdown, and costly unplanned production halts. Conventional cable carriers often transfer operational movement stress directly onto internal electrical lines and hoses. The Robotrax system changes this dynamic through a central steel cable that runs through the core of every chain link. When robot arms undergo rapid directional shifts and accelerations up to 10 g, this internal steel cable absorbs the primary tensile loads. By isolating electrical and fluid lines from pulling forces, the design preserves conductor integrity and dramatically extends cable service life. Mechanics can easily calibrate and adjust system tension using an integrated clamping piece, ensuring consistent mechanical support throughout long operational cycles. The foundation of the Robotrax system lies in its open, single-piece plastic links featuring spherical snap-on connections on both sides. This geometry allows the carrier to flex smoothly across three axes, providing radial rotation of up to ±450 degrees per meter depending on the model size. To optimize internal organization, carrier links contain up to three distinct chambers. This physical separation prevents signal interference and mechanical abrasion between heavy [power lines](https://spectrum.ieee.org/tag/power-lines), sensitive data channels, and fluid hoses. For standard models (R040 through R100), technicians can press cables directly into the carrier without tools, drastically reducing installation and maintenance time. Larger configurations, such as the R140X, incorporate swiveling crossbars with snap locks alongside vertical and horizontal dividers for customized interior partitioning. 1. Steel cable for transferring extremely high tensile forces 2. Tension piece for locking the chain links 3. Type with toolless opening swivel crossbars and divider module available 4. Open design – Fast cable laying as the cables are simply pressed in – Easy checking of all cables 5. Special plastic for long service life 6. Protective covers or heat shields made from different materials are available for different environmental conditions 7. Quick-release bracket for fixing and continuation 8. Strain relief with LineFix clamps 9. Protection against hard impacts, excessive abrasion and premature wear as well as limitation of the bending radius through protector Large robot work envelopes and high-speed motion trajectories can cause loose cable carrier loops to swing and strike the robot body. To eliminate these destructive collisions, Tsubaki KabelSchlepp integrates the Pull Back Unit (PBU). The PBU serves as an active retraction mechanism that maintains optimal tension on the cable carrier throughout the entire motion cycle. By preventing excess slack and eliminating interfering contours, the PBU minimizes collision risks across complex movement paths. The unit requires zero maintenance on its retraction element and offers standard mounting configurations for leading [industrial robot](https://spectrum.ieee.org/tag/industrial-robot) platforms, including [KUKA](https://spectrum.ieee.org/tag/kuka), [ABB](https://spectrum.ieee.org/tag/abb), and [FANUC](https://spectrum.ieee.org/tag/fanuc). Additionally, external protectors can be retrofitted onto individual chain links. These durable impact shields limit the minimum bending radius to prevent over-flexing while shielding the chain body from severe external abrasion. If wear occurs, technicians simply replace the modular protector rather than the entire cable carrier assembly. By combining central load absorption, multi-axis flexibility, and active retraction control, the Robotrax system offers plant engineers and system integrators a reliable path toward maximizing robot uptime and reducing total operational costs. This White Paper gives data engineers and architects a practical overview of how parallel partitioned reads, write-path optimization, and cloud-native bulk loading reduce large-table replication times, and why replication speed has become a business concern as data volumes grow. Beyond policy and research, Radovanovi? writes fiction and poetry. Her short stories have appeared in literary anthologies. She is the author of the short story collection _Divertimento for Four Seasons_ and is completing her first novel, _Counterpoint_. Even before France approved legislation banning social media for children under 15 last January, 13-year-old Benjamin was already wondering what life without social media would look like. “If we want to play football, we won’t be able to organize it. What will we do? Send letters instead?” he joked in [an interview for _Le Monde_.](https://www.lemonde.fr/en/pixels/article/2026/02/18/teens-on-france-s-social-media-ban-for-under-15s-we-re-going-back-to-the-stone-age_6750597_13.html) Second, many children, and often their parents, lack the digital literacy and critical thinking skills needed to navigate online risks safely. My policy research on child online protection in Indonesia, [published earlier this year in _Digital Society_](https://link.springer.com/article/10.1007/s44206-025-00244-0), found substantial gaps that account removals cannot repair: Many children lacked guidance on navigating the internet safely, and large numbers did not know how to report harmful experiences. And third, the platforms have limited independent oversight as they identify underage users, design age-verification systems, and report their own compliance. In Indonesia, platforms themselves are responsible for carrying out age verification, while the Ministry of Communication and Digital Affairs [oversees compliance](https://www.komdigi.go.id/berita/siaran-pers/detail/lindungi-anak-di-era-digital-kemkomdigi-wajibkan-verifikasi-usia-di-platform-digital). TikTok’s appeals process for users flagged as underage, for instance, can require a [government-issued ID and selfies](https://www.tiktok.com/support/faq_detail?id=7611808266502560268&category=web_account), which is a problem because it involves collecting the additional personal data on an ID card, beyond that needed to confirm age. Will government regulators ensure that TikTok handles that data responsibly? Every age-based ban creates an engineering problem: How can a platform reliably determine that a user is old enough, without intruding on other information? Governments and companies may use identity documents, parental authorization, app-store checks, or facial age estimation. Each approach has trade-offs among accuracy, privacy, accessibility, and resistance to circumvention. The challenge should not merely be to “verify age.” It should be to prove that someone is above a threshold, without disclosing their identity, birth date, or other information third parties might use to create a marketing profile. [The European Commission’s age-verification blueprint](https://digital-strategy.ec.europa.eu/en/factpages/blueprint-age-verification-solution-help-protect-minors-online) challenges companies to verify ages without collecting all that additional information. Blocking access to a platform redirects some young people, but not always where expected. Early anecdotal reports in Australia pointed to teenagers migrating to smaller, less-regulated platforms like Yope, a pattern the [Cato Institute](https://www.cato.org/blog/australias-under-16-social-media-ban-warning-online-speech-security-around-world) flagged as a “whack-a-mole” problem for regulators. But industry data collected two months later found [no broad-based shift](https://www.rnz.co.nz/news/world/589514/one-fifth-of-australian-teens-still-use-tiktok-snapchat-after-social-media-ban) of that kind, aside from a small uptick in WhatsApp use. Many teens simply found a way to stay on the banned platforms. For many teenagers, social media workarounds are merely inconvenient. But for isolated, marginalized, disabled, or LGBTQ+ youth who depend on online communities for support that’s otherwise unavailable, displacement can mean losing certain kinds of belonging, or having to move to a platform with even weaker oversight. If blanket social media bans don’t work, then what will? The platforms have created many of the conditions that governments are now trying to contain: engagement-optimized recommenders, intrusive data practices, weak safeguards against unwanted contact, and features such as infinite scroll, autoplay, streaks, and persistent notifications. These design patterns increasingly face regulatory scrutiny, including what’s required under the [European Union’s Digital Services Act](https://www.techtimes.com/articles/320180/20260711/eu-charges-meta-addictive-design-infinite-scroll-violates-dsa-health-rules.htm). A [2026 study from the 5Rights Foundation](https://5rightsfoundation.com/infinite-scroll-groundbreaking-study-reveals-tiktokisation-of-children-online/) that tracked children’s device use minute by minute found that the user interfaces shape children’s attention, sleep, and well-being in real time. Governments should define measurable outcomes and fund independent evaluation, platforms should give researchers meaningful data access, and engineers should audit age-assurance systems for bias and data leakage. Schools, parents, and children themselves need a seat in designing the technology that’s designed to protect children. If policymakers still decide to remove an infrastructure for youth connection, they should offer something better in return. Social media bans may reduce some forms of exposure to harmful content and may be justified for particular ages, services, or risks. But they are just one tool, not a comprehensive substitute for safer design, accountable platforms, digital literacy, institutional capacity, and [noncommercial digital “third places”—](https://digitalserendipities.substack.com/p/if-we-ban-social-media-for-children)moderated communities, creative spaces, and public-interest platforms designed for youth participation rather than profit. The first wave of social media restrictions isn’t enough to keep children safe. Governments are still measuring what’s easiest to count, while neglecting harder-to-measure outcomes such as children’s access to safe third places and [meaningful social connection, both online and offline](https://www.oecd.org/en/data/dashboards/oecd-child-well-being-dashboard.html). Until governments can show evidence that harm has actually declined, they will keep mistaking account removal for safety.