ANR adds AI-powered frame generation to games, letting supported GPUs render extra frames for smoother motion. It works via a simple installer that hooks into the game's executable and can be compared against DLSS, FSR, XeSS and others.
Why it matters. Game developers and modders who want a free alternative to proprietary upscalers can test ANR in their titles.
spotted Nov 13
Today in one paragraph. Today's wave of open-source AI is eroding the monopoly of big-tech, with projects like apex-compute's ANR delivering frame-generation quality that rivals Nvidia's DLSS on both Windows and Linux. From cheaper speech-to-text on Spot instances to a 300 M LLM running on an iPad, the toolbox keeps expanding.
directory-opus is a ready-to-run desktop package that includes every feature of the Directory Opus file manager out of the box. You just download it and launch, no configuration needed.
It installs a ready-to-use set of cheat modules and preset configurations for the Delta Force series on Windows. After a single click the tools run from the desktop with no extra setup.
Manus is an autonomous AI agent that can write and run code and control a web browser on your behalf. You describe a task and it figures out the steps, handling everything from form filling to small script testing.
Manus is an autonomous AI agent that can write, execute scripts and control a web browser based on natural language instructions. It runs tasks like form-filling, data gathering or small code jobs without you typing each step.
The model rewrites nine internal weight matrices without any new training data, using its own activations. It scores higher on GPQA Diamond and LiveCodeBench while needing 27-54% fewer thinking tokens.
Edge0-145B-NVFP4 is a large language model for generating text. It runs on Nvidia GPUs using a 4-bit quantized format to fit the 145-billion parameters.
Qwen3.5-4B-3D-Coder Studio lets you type a description and watches a fine-tuned LLM write a complete Three.js script that runs instantly in your browser. The interface streams the code as it’s generated, shows the live scene in a sandboxed iframe, and lets you download the result.
43 likes
From the labs
Straight from the companies building them, top 9 of 14
Amazon’s new AgentCore runtime V2 snapshots the initialized environment and restores it for each new session, so startup time no longer scales with image size. The post walks through measuring startup latency and presents four architecture patterns for fast sessions.
OpenFold3-preview is now available as a SageMaker JumpStart model, enabling you to run asynchronous inference on proteins, DNA, RNA and ligand complexes directly in your AWS account. It supports multi-GPU parallelism, custom multiple sequence alignments for higher accuracy, and keeps your data private.
AssemblyAI demonstrates a pipeline that runs large-scale speech-to-text inference on AWS using ECS and Spot GPU instances. An SQS-based retry loop keeps latency in-line with SLAs while delivering about 70 % cost savings versus on-demand servers.
Amazon Bedrock now automatically tags each inference request with the IAM principal that issued it, so you can see exactly who is spending on AI models. No code changes or extra instrumentation are required.
Amazon SageMaker HyperPod Data Capture automatically stores the inputs and outputs of a running inference endpoint into an Amazon S3 bucket. The stored traffic can then be queried with Athena to spot blind spots and used to fine-tune a new model version.
The walkthrough shows how to set up a UR17G arm with a Robotiq 2F-140 gripper and a beverage can in Isaac Sim using pre-validated SimReady assets. The assets include collision, mass and joint-drive data so the simulated robot can actually pick up and move the can with agent-generated controls.
XYBench is a benchmark of over 8,000 real questions that test whether AI assistants can spot the XY problem, when users ask for the wrong solution. It measures a model’s ability to redirect users to the real issue instead of just answering the misguided request.
ChatGPT Wallet adds a built-in payment flow for agents in ChatGPT Work, Codex and related tools. Users approve each transaction and the feature uses existing card protections to keep payment data safe.
llama.cpp v1.0.0 offers a portable C/C++ library that runs large language models on everything from a laptop to a server, using quantization and optimized backends. It gives developers a lightweight way to embed private AI inference without relying on the cloud.
Research
The papers researchers upvoted most today, top 9 of 33
BirdDreams is a neural model that creates both the driver’s camera view and a top-down bird’s-eye view of a road scene as it evolves under different vehicle actions. By learning motion separately from appearance it produces more plausible traffic behavior.
PhysLoop uses a large language model’s physics knowledge to evaluate and correct videos generated by AI models. The feedback loop creates fine-grained supervision without human labelling, helping the video model learn more realistic dynamics.
G-STAR lets image generators pack more visual detail into fewer tokens, cutting generation time. It groups tokens and applies product quantization to keep quality while compressing the sequence.
Duo is a full-duplex world model that watches video and listens to audio while simultaneously generating spoken replies. It runs continuously, treating the AI’s mind and the physical scene as one shared world.
The paper classifies decision models into three roles, actor, judge, and teacher, based on whether they pick actions, evaluate past actions, or assign scores for learning. It then measures how each role performs on tasks that require planning many steps ahead.
It provides an open-source implementation that turns vision problems into conditional image-editing using a pretrained diffusion model. Users can run the model to perform tasks like segmentation or object detection by editing images rather than training a separate network.
The study dissects the hyper-connection blocks used in cutting-edge language models, identifying which pieces are essential and which waste compute. It then proposes a leaner version that retains accuracy while lowering training overhead.
Ropedia Xperience 2 is a large-scale collection of physical AI environments and recorded robot behaviors. It lets researchers train and evaluate models on long-horizon, contact-rich tasks using visual inputs.
TTB pre-computes the test-time training layers of large view synthesis models so they can render new camera angles faster. It keeps the quality of KV-cached decoder-only models while cutting inference cost.
19 upvotes
The conversation
What builders are arguing about on Hacker News, top 7 of 14
The analysis scans publicly released model checkpoints for numerical patterns that only arise when training on NVIDIA’s Blackwell GPUs. It shows that DeepSeek, Moonshot AI, MiniMax and Ant Group have been using the restricted hardware despite export bans.
Edge0-145B is an open-source AI model with 145 billion parameters that can be run locally on a personal computer. It offers coding, reasoning, and agent capabilities comparable to cloud-based frontier models without any usage fees.
A machine-learning system was trained to read Gödel’s 1,009 pages of German Gabelsberger shorthand and translate them into modern mathematical notation. The output shows dozens of previously unknown proved theorems in logic and set theory.
The Foundation Models framework gives Swift developers a built-in way to call large language models, either the on-device Apple model or any remote model you point it at. It also supports multimodal prompts, tool calling and image understanding, so you can add summarisation, chat agents or visual analysis to an iOS or macOS app with a few lines of code.
Developers and autonomous AI agents can call the Brave Search API and pay per query with USDC, using the open x402 payment protocol. The API returns a 402 response with price details when no payment is attached, and the client retries with a signed payment.
The author ran 600 trials on LeWM, a joint-embedding predictive architecture that plans actions in a compact embedding space instead of pixels. The study shows that a tiny visual detail in the goal image can impact success more than swapping the predictor for a physics engine.