Uncategorized

There is a quiet failure mode that lives at the center of every AI-assisted coding workflow. You ask Claude Code, Cursor, or Windsurf to modify a function. The agent does it confidently, cleanly, and incorrectly — because it had no idea that 47 other functions depended on the return type it just changed. Breaking changes ship. The test suite screams. And you spend the next two hours untangling what the model should have known before it touched a single line. An Indian Computer Science student built GitNexus to fix that. The open-source project, now sitting at 28,000+ stars and 3,000+ forks on GitHub with 45 contributors, describes itself as ‘the nervous system for agent context.’ That description undersells what it actually does. What Actually is GitNexus? GitNexus is a code intelligence layer, not a documentation tool. It indexes an entire repository into a structured knowledge graph — mapping every function call, import, class inheritance, interface implementation, and execution flow — and then exposes that graph to AI agents through a Model Context Protocol (MCP) server. The agents stop guessing. They query. To understand why this is significant, you need to understand what AI coding agents currently operate on. Most tools like Cursor, Claude Code, and Windsurf rely on either file-based context windows (they read the files nearby and hope for the best) or traditional Graph RAG approaches (they query a graph with a series of prompts, hoping to discover what matters). Neither approach gives an agent a structural map of the repository before it acts. GitNexus pre-computes the entire dependency structure at index time. When an agent asks ‘what depends on this function?’, it gets a complete, confidence-scored answer in one query, instead of chaining 10 successive queries that each risk missing something. The Indexing Pipeline Running npx gitnexus analyze from the root of a repository kicks off a multi-phase indexing pipeline that does the following: First, it walks the file tree and maps folder and file relationships (the Structure phase). Then it parses every function, class, method, and interface using Tree-sitter ASTs (Abstract Syntax Trees). Tree-sitter is a high-performance, incremental parser originally developed at GitHub that produces concrete syntax trees for any supported language. GitNexus uses it to extract symbols with precision that regex or simple text search cannot match. After parsing, GitNexus performs cross-file resolution: it resolves imports, function calls, class heritage, constructor inference, and self/this receiver types across the whole codebase. This is the step where it learns that UserController in src/controllers/user.ts calls into UserService, which authRouter imports, which handleLogin depends on. Next comes clustering — GitNexus groups related symbols into functional communities using Leiden community detection on the call graph, assigning each cluster a cohesion score. Then it traces execution flows from entry points through full call chains to build what it calls ‘processes.’ Finally it indexes everything for hybrid search using BM25 (a keyword ranking algorithm), semantic vector embeddings, and RRF (Reciprocal Rank Fusion) to merge results. The graph is stored in LadybugDB, an embedded graph database with native vector support formerly known as KuzuDB. This entire pipeline runs locally — no code leaves your machine. A particularly useful flag for teams: gitnexus analyze –skills takes the Leiden community detection one step further. Instead of only grouping symbols internally, it generates a custom SKILL.md file for each detected functional area of your codebase under .claude/skills/generated/. Each skill file describes that module’s key files, entry points, execution flows, and cross-area connections — so an AI agent working in the authentication module gets targeted architectural context for that specific area, not a generic overview of the entire repo. Skills are regenerated on each –skills run to stay current. https://github.com/abhigyanpatwari/GitNexus Seven Tools and Two Prompts Your Agent Gets Once indexed, GitNexus registers an MCP server that exposes seven tools and two guided prompts to your AI agent. impact runs blast radius analysis. Given a target symbol, it returns every upstream caller grouped by depth with confidence scores — handleLogin [CALLS 90%], UserController [CALLS 85%] — so the agent knows what it risks breaking before it touches anything. context gives a 360-degree view of any symbol: its callers, its callees, every process it participates in, and which step of each process it occupies. query runs process-grouped hybrid search across the codebase, returning matching symbols alongside the execution flows they belong to. detect_changes performs git-diff impact analysis — it maps changed lines to affected processes and assigns a risk level before you commit. rename executes coordinated multi-file symbol renames using the graph for high-confidence edits and text search for the rest, with a dry-run mode to preview changes before applying them. cypher exposes raw Cypher graph queries for engineers who want to write custom traversals against the knowledge graph directly. list_repos handles the multi-repo case — GitNexus uses a global registry at ~/.gitnexus/ so one MCP server can serve multiple indexed repositories simultaneously. Beyond the tools, GitNexus also exposes two MCP prompts for guided workflows. detect_impact runs a pre-commit change analysis that surfaces scope, affected processes, and an overall risk level — think of it as a structured checklist before any significant edit. generate_map produces architecture documentation directly from the knowledge graph, complete with Mermaid diagrams, making it useful for onboarding engineers or documenting a codebase that has grown faster than its docs. Editor Support and Deepest Integration with Claude Code GitNexus supports Claude Code, Cursor, Codex, OpenCode, and Windsurf. Editor support varies by tier. Windsurf gets MCP only. Cursor, Codex, and OpenCode get MCP plus agent skills. Claude Code gets the full stack: MCP tools, agent skills (Exploring, Debugging, Impact Analysis, Refactoring), PreToolUse hooks that enrich every search with graph context before Claude acts, and PostToolUse hooks that auto-reindex after commits. For Claude Code users, GitNexus installs itself completely — hooks, skills, and an AGENTS.md / CLAUDE.md context file — in a single npx gitnexus analyze command. The Model Democratization Angle One of the less obvious implications of this architecture is what it does for smaller models. Because

For years, the computer vision community has operated on two separate tracks: generative models (which produce images) and discriminative models (which understand them). The assumption was straightforward — models good at making pictures aren’t necessarily good at reading them. A new paper from Google, titled “Image Generators are Generalist Vision Learners” (arXiv:2604.20329), published April 22, 2026, blows that assumption apart. A team of Google DeepMind researchers introduced Vision Banana, a single unified model that surpasses or matches state-of-the-art specialist systems across a wide range of visual understanding tasks — including semantic segmentation, instance segmentation, monocular metric depth estimation, and surface normal estimation — while simultaneously retaining the original image generation capabilities of its base model. https://arxiv.org/pdf/2604.20329 The LLM Analogy That Changes Everything If you’ve worked with large language models, you already understand the two-phase playbook: first, pretrain a base model on massive text data using a generative objective, then apply instruction-tuning to align it for downstream tasks. The pretraining phase is where the model develops a rich internal representation of language that can be repurposed for almost anything. The Google team’s core claim is that image generation training plays the exact same foundational role for vision. Their base model, Nano Banana Pro (NBP), is Google’s state-of-the-art image generator. By performing a lightweight instruction-tuning pass — mixing a small proportion of computer vision task data at a very low ratio into NBP’s original training mixture — they created Vision Banana. The key insight: generating photorealistic images implicitly requires a model to understand geometry, semantics, depth, and object relationships. Vision Banana learns to express that latent knowledge in measurable, decodable formats. Critically, no training data from any of the evaluation benchmarks is included in the instruction-tuning mixture — ensuring that all results reflect true generalist capability rather than in-domain memorization. How It Works: Perception as Image Generation Rather than adding specialized decoder heads or regression modules for each task, all vision task outputs are parameterized as RGB images. The model is instruction-tuned to produce visualizations that follow precise, invertible color schemes — meaning the generated images can be decoded back into quantitative outputs for benchmark evaluation. The research team identified three key advantages of this strategy. First, it supports a wide variety of tasks with a single unified model — after instruction-tuning, only the prompt changes, not the weights. Second, it requires relatively little new training data, since instruction-tuning is solely teaching the model how to format computer vision outputs as RGB. Third, it helps the model retain its original image generation capabilities, since the outputs are simply new RGB images. For semantic segmentation, the model is prompted with instructions such as: “Generate a segmentation visualization of this image, using the color mapping: {‘cat’: ‘red’, ‘background’: ‘yellow’}.” Each pixel is colored by its predicted class, and because color assignments are specified in the prompt, no fixed label vocabulary is needed. For instance segmentation, since the number of instances is unknown in advance, Vision Banana uses a per-class inference strategy — running a separate pass per class and dynamically assigning unique colors to each instance. Masks are recovered by clustering pixels with similar colors using a threshold. Metric depth estimation uses a bijective mapping between unbounded metric depth values in [0, ∞) and bounded RGB values in [0, 1]³. A power transform (shape parameter λ = −3, scale parameter c = 10/3) first “curves” metric depth values, which are then encoded as a false-color visualization that traverses the edges of the RGB cube, following the structure of a 3D Hilbert curve. This transform is strictly invertible, so the generated depth image decodes cleanly back to physical metric distances. Crucially, no camera parameters — neither intrinsics nor extrinsics — are required at training or inference time. The model infers absolute scale purely from visual cues and world knowledge embedded during pretraining. The depth training data is also entirely synthetic, generated from simulation rendering engines, with zero real-world depth data used. For surface normal estimation, the mapping is more direct: surface normals are unit vectors (x, y, z) ranging from −1.0 to 1.0, which map naturally to RGB channels. Facing-left normals encode as pinkish-red; facing-up normals encode as light green; normals pointing toward the camera encode as light blue/purple. The Numbers: Beating Specialists at Their Own Game Vision Banana’s results across benchmarks — all in zero-shot transfer settings, where the model has never seen any training data from the evaluated datasets — are significant: Semantic segmentation on Cityscapes val: mIoU of 0.699, compared to SAM 3’s 0.652 — a 4.7-point gain. Referring expression segmentation on RefCOCOg UMD val: cIoU of 0.738, edging out SAM 3 Agent’s 0.734. Reasoning segmentation on ReasonSeg val: gIoU of 0.793, beating SAM 3 Agent’s 0.770 — and notably surpassing even non-zero-shot methods trained on in-domain data, including X-SAM. Instance segmentation on SA-Co/Gold: pmF1 of 0.540, on par with DINO-X (0.552), and ahead of Gemini 2.5 (0.461), APE-D (0.369), and OWLv2 (0.420) under zero-shot transfer. Metric depth estimation: average δ1 of 0.882 across six major benchmarks; on the four datasets where Depth Anything V3 was evaluated (NYU, ETH3D, DIODE-Indoor, KITTI), Vision Banana scores 0.929 versus Depth Anything V3’s 0.918 — while using zero real-world training data and no camera parameters. Surface normal estimation: average mean angle error of 18.928° across four datasets, compared to Lotus-2’s 19.642°. On indoor datasets specifically, Vision Banana achieves the lowest mean angle error (15.549°) and lowest median angle error (9.300°) among all compared methods. On generative benchmarks, Vision Banana holds its own against its base model: it achieves a 53.5% win rate against Nano Banana Pro on GenAI-Bench (text-to-image), and a 47.8% win rate on ImgEdit (image editing), where Nano Banana Pro scores 52.2%. Overall, the results confirm that lightweight instruction-tuning does not degrade the model’s generative capabilities. Key Takeaways Image generation pretraining is a generalist vision learner: Just as LLM pretraining unlocks emergent language understanding, Google’s research shows that training on image generation naturally develops powerful internal visual representations that transfer to perception tasks

Training frontier AI models is, at its core, a coordination problem. Thousands of chips must communicate with each other continuously, synchronizing every gradient update across the network. When one chip fails or even slows down, the entire training run can stall. As models scale toward hundreds of billions of parameters, that fragility becomes increasingly untenable. Google DeepMind is now proposing a different model entirely. Google DeepMind researchers introduced Decoupled DiLoCo (Distributed Low-Communication), a distributed training architecture that decouples compute into asynchronous, fault-isolated ‘islands,’ enabling large language model pre-training across geographically distant data centers without requiring the tight synchronization that makes conventional approaches brittle at scale. The Problem with Traditional Distributed Training To understand why Decoupled DiLoCo is important, it helps to understand how distributed training typically works. Standard Data-Parallel training replicates a model across many accelerators (GPUs or TPUs), each processing a different mini-batch of data. After each forward and backward pass, gradients must be averaged across every device — a process called AllReduce — before the next training step can begin. This blocking synchronization step means every device must wait for the slowest one. Across thousands of chips spanning multiple data centers, that bottleneck is not just inconvenient; it makes global-scale training effectively impractical. Bandwidth is another hard constraint. Conventional Data-Parallel training requires approximately 198 Gbps of inter-datacenter bandwidth across eight data centers — far beyond what standard wide-area networking (WAN) can support between geographically distributed facilities. How Decoupled DiLoCo Works Decoupled DiLoCo builds on two prior systems from Google. The first is Pathways, which introduced a distributed AI system based on asynchronous data flow, allowing different compute resources to work at their own pace without blocking on one another. The second is DiLoCo, which dramatically reduced the inter-datacenter bandwidth required for distributed training by having each worker perform many local gradient steps before communicating with peers — dramatically reducing how much data needs to flow between data centers. Decoupled DiLoCo brings both ideas together. Built on top of Pathways, training is divided across separate clusters of accelerators called learner units — the ‘islands’ of compute. Each learner unit trains semi-independently, performing many local steps, before sharing a compressed gradient signal with an outer optimizer that aggregates updates across all learner units. Because this outer synchronization step is asynchronous, a chip failure or slow learner unit in one island does not block the others from continuing to train. The bandwidth savings are dramatic. Decoupled DiLoCo reduces required inter-datacenter bandwidth from 198 Gbps to just 0.84 Gbps across eight data centers — multiple orders of magnitude lower — making it compatible with standard internet-scale connectivity between datacenter facilities rather than requiring custom high-speed network infrastructure. Self-Healing Through Chaos Engineering One of the most technically significant properties of Decoupled DiLoCo is its fault tolerance. The research team used chaos engineering, a method that deliberately introduces artificial hardware failures into a running system to test its robustness during training runs. The system continued training after the loss of entire learner units, and then seamlessly reintegrated those units when they came back online. This behavior is what the research team describes as ‘self-healing’. In simulations involving 1.2 million chips under high failure rates, Decoupled DiLoCo maintained a goodput (the fraction of time the system is performing useful training) of 88%, compared to just 27% for standard Data-Parallel methods. Goodput is the practical metric that matters here: a training run with high nominal compute but low goodput wastes significant resources. https://deepmind.google/blog/decoupled-diloco/? Critically, these resilience gains come with minimal degradation in model quality. In real-world experiments using Gemma 4 models, Decoupled DiLoCo achieved an average ML benchmark accuracy of 64.1%, compared to 64.4% for the conventional baseline — a difference well within the noise of typical evaluation variance. Training a 12B Model Across Four U.S. Regions The research team validated Decoupled DiLoCo at production scale by successfully training a 12 billion parameter model across four separate U.S. regions using just 2–5 Gbps of wide-area networking, a bandwidth level achievable with existing commercial internet infrastructure between data center facilities. The system accomplished this more than 20 times faster than conventional synchronization methods. The key reason: rather than forcing compute to pause and wait for communication to complete, Decoupled DiLoCo incorporates required communication into longer periods of computation, eliminating the “blocking” bottlenecks that make conventional distributed training slow at global scale. Mixing Hardware Generations An underappreciated implication of the architecture is its support for heterogeneous hardware. Because learner units operate asynchronously, they do not need to run on identical hardware at the same clock speed. The research team demonstrated training runs that mixed TPU v6e and TPU v5p chips — different hardware generations with different performance characteristics — in a single training job, without degrading ML performance relative to homogeneous runs. This has two practical consequences worth noting. First, it extends the useful life of existing hardware, allowing older accelerators to continue contributing meaningfully to large-scale training. Second, because new hardware generations do not arrive everywhere at once, being able to train across generations can alleviate the recurring logistical and capacity bottlenecks that arise during hardware transition periods — a real operational challenge at organizations running large training infrastructure. Key Takeaways Decoupled DiLoCo eliminates the single-point-of-failure problem in large-scale AI training by dividing training across asynchronous, fault-isolated “islands” of compute called learner units — so a chip or cluster failure in one island does not stall the rest of the training run. The architecture reduces inter-datacenter bandwidth requirements by orders of magnitude — from 198 Gbps down to 0.84 Gbps across eight data centers — making globally distributed pre-training feasible over standard wide-area networking rather than requiring custom high-speed infrastructure. Decoupled DiLoCo is self-healing: using chaos engineering to simulate real hardware failures, the system maintained 88% goodput compared to just 27% for standard Data-Parallel training under high failure rates, and seamlessly reintegrated offline learner units when they came back online. The approach was validated at production scale, successfully training a 12 billion parameter model across four

This is today’s edition of The Download, our weekday newsletter that provides a daily dose of what’s going on in the world of technology. We’re in a new era of AI-driven scams When ChatGPT was released in late 2022, it showed how easily generative AI could create human-like text. This quickly caught the eye of cybercriminals, who began using LLMs to compose malicious emails. Since then, they’ve adopted AI for everything from turbocharged phishing and hyperrealistic deepfakes to automated vulnerability scans. Many organizations are now struggling to cope with the sheer volume of cyberattacks. AI is making them faster, cheaper, and easier to carry out, a problem set to worsen as more cybercriminals adopt these tools—and their capabilities improve. Read the full story on how AI is reshaping cybercrime. —Rhiannon Williams “Supercharged scams” is one of the 10 Things That Matter in AI Right Now, our essential guide to what’s really worth your attention in the field. Subscribers can watch an exclusive roundtable unveiling the technologies and trends on the list, with analysis from MIT Technology Review’s AI reporter Grace Huckins and executive editors Amy Nordrum and Niall Firth. Healthcare AI is here. We don’t know if it actually helps patients. Doctors are using AI to help them with notetaking. AI-based tools are trawling through patient records, flagging people who may require certain support or treatments. They are also used to interpret medical exam results and X-rays. A growing number of studies suggest that many of these tools can deliver accurate results. But there’s a bigger question here: Does using them actually translate into better health outcomes for patients? We don’t yet have a good answer—here’s why. —Jessica Hamzelou The story is from The Checkup, our weekly newsletter that gives you the latest from the worlds of health and biotech. Sign up to receive it in your inbox every Thursday. The must-reads I’ve combed the internet to find you today’s most fun/important/scary/fascinating stories about technology. 1 DeepSeek has unveiled its long-awaited new AI modelThe Chinese company has just launched preview versions of DeepSeek-V4. (CNN)+It says V4 is the most powerful open-source platform. (Bloomberg $) + And rivals top closed-source models from OpenAI and DeepMind. (SCMP)+ The model is adapted for Huawei chip technology. (Reuters $) 2 More countries are curbing children’s social media accessNorway is set to enforce the latest ban. (Reuters $)+ The Philippines could follow soon. (Bloomberg $)+ Americans are pushing to get AI out of schools. (The New Yorker) 3 The US has accused China of mass AI theft as tensions riseA White House memo claims Chinese firms are exploiting American models. (BBC)+ Beijing calls the accusations “slander.” (Ars Technica) 4 OpenAI set itself apart from Anthropic by widely releasing its new modelIt’s releasing GPT-5.5 to all ChatGPT users, despite cybersecurity concerns. (NYT $)+ OpenAI says the new model is better at coding and more efficient. (The Verge) 5 Meta is cutting 10% of jobs to offset AI spendingRoughly 8,000 layoffs are set to be announced on May 20. (QZ)+ Anti-AI protests are growing. (MIT Technology Review) 6 Palantir is facing a backlash from employeesThanks to its work with ICE and the Trump administration. (Wired $)+ Surveillance tech is reshaping the fight for privacy. (MIT Technology Review) 7 The era of free access to advanced AI is coming to an endAI labs are under mounting pressure to start turning profits. (The Verge) 8 Elon Musk’s feud with Sam Altman is heading to court The case has already revealed several unflattering secrets. (WP $) 9 A new movement is encouraging people to ditch their smartphones for a month“Month Offline” is like a Dry January for smartphones. (The Atlantic) 10 Spotify has revealed its most-streamed music of the last 20 yearsFeaturing Taylor Swift, Bad Bunny, and The Weeknd. (Gizmodo)  Quote of the day “We want a childhood where children get to be children. Play, friendships, and everyday life must not be taken over by algorithms and screens.”  —Norwegian Prime Minister Jonas Gahr Store announces age restrictions for social media. One More Thing NASA/JPL-CALTECH VIA WIKIMEDIA COMMONS; CRAFT NASA/JPL-CALTECH/SWRI/MSSS; IMAGE PROCESSING: KEVIN M. GILL The search for extraterrestrial life is targeting Jupiter’s icy moon Europa As astronomers have discovered more about Europa over the past few decades, Jupiter’s fourth-largest moon has excited planetary scientists interested in the geophysics of alien worlds.  All that water and energy—and hints of elements essential for building organic molecules —point to an extraordinary possibility. In the depths of its ocean, or perhaps crowded in subsurface lakes or below icy surface vents, Jupiter’s big, bright moon could host life.  To find further evidence, NASA is now searching for signs of alien existence on Europa. Read the full story on the mission. —Stephen Ornes We can still have nice things A place for comfort, fun and distraction to brighten up your day. (Got any ideas? Drop me a line.) + Here’s a fun look at the secret collaborations of pop history.+ Meet the mannequins showing how the “ideal” body has evolved.+ A photographer has cataloged all 12,795 objects in her home into an archive of a life.+ Slime molds are unexpectedly beautiful when viewed through these high-detail macro shots.