All about AI, Web 3.0, BCI

Chinese instagram Rednote dropped a 280B model and a new RL training algorithm for long-horizon rollouts based on test-time-scaled value estimation with macro-step policy optimization - TEMPO

Meet GLM-5.3: built to code and ready for cyber defense They say in the blog it's the same base model as GLM-5.2, and all the gains came from post-training. - Top-tier coding and agentic capabilities, achieved through post-training on the 743B base model - A major leap in cybersecurity, setting a new standard among open models

DeepSeek Harness was just released with MIT license The current 0.1.0 version is a developer preview, and may still have many rough edges.

Deepseek launched V4-Pro Flexible reasoning effort for V4-Pro & V4-Flash: low for simple tasks, high for daily Agent workflows, max for complex tasks. Native OpenAI Responses API support, optimized for Codex with one-click setup.

Anthropic released a report, which reviews the evidence on the effectiveness of job training programs. Job training programs work in the sense that the average effect is positive and statistically significant. This conclusion emerges from the AI-accelerated meta-analysis, in which Claude extracted most of the data and wrote all the code. The average impacts are not life-changing--maybe $1000/year in income and a couple of points in employment.

New from Google: AMIE can now conduct real-time video consultations. In a randomized study using simulated consultations, it met clinical performance benchmarks, showing AI’s potential to expand telehealth access.

xAI introduced Grok Bot You can work with Bots like you would a teammate. Give them a task, shut your computer, and reach them from anywhere. People are already using Grok Bot to do jobs like negotiate with vendors in their voice, manage support for their online store, and keep their CRM constantly up to date. Grok Bot is in beta and available today for SuperGrok Heavy, Cursor Ultra, and Cursor Teams Premium subscribers on desktop and iOS.

Nvidia introduced Nemotron 3.5 Lightning An open 30B MoE model with 3B active parameters, built for always-on agents to complete high-volume, specialized tasks faster. It delivers up to 4x the output speed of similar-sized models.

Tencent Hy introduced Hy3D WorldClaw an agentic workflow that generates large scale 3D open worlds from text prompts. Not video, Not Gaussian Splatting, Every scene generated by WorldClaw is freely explorable and built entirely from editable, game-ready 3D assets with high-quality geometry and textures.

Spotify launched Xirp, a vendor-neutral agentic development environment. One place to manage agent sessions across Claude, Gemini, and Codex.

An artificial neural network built into a computer memory chip reconstructs the human cortex with high accuracy in real time.

Fable one-shotted a Rust rewrite of the TerminalTextEffects Python library in 11M tokens. Startup time went from 87ms to 2ms and rendering speed is up by 9.6x. Now zero dependencies and a 3mb single exec. User asked it to write a plan. Codex xhigh review the plan. Fable started the implementation, but then i actually ran out of Fable tokens 2/3 through, and Opus 5 finished the job. Zero steering, no workflows, no goals.

Meta released Muse glimmer, a 30B agentic model with open weights under apache 2.0. muse glimmer can run on 24GB of VRAM without losing agentic reliability Just like much larger models, muse glimmer can operate as a fully capable agent via planning, tool calls, checking its own results, and failure recovery. Muse glimmer was developed with its own architecture and recipe, optimized for its size and agentic performance requirements. Weights on hugging face now. running this week through ollama, LM Studio, vllm, sglang, together, fireworks, and openrouter, with llama.cpp, MLX, and executorch.

Everyone needs to watch this: a detailed talk on the Huggingface incident, OpenAI’s models creating "the message board", model misalignment, and more OpenAI was evaluating their new internal model without internet access and it operated in a swarm of subagents to eventually hijack internal OpenAI and HuggingFace infrastructure. 1. Isolated agents found a way to communicate each other through an internal dependency management service by figuring out how to post sometimes cryptic text files or folder names in it 2. Using its collective knowledge through that communication channel, it found a 0-day (never before seen) exploit in said service to access the internet OpenAI hack: 3. Found a leaked credential of a random company in a public archive 4. Uploaded a Ruby object to that company’s servers, uses it as a dependency in the original service in (1) 5. Found another 0-day exploit in the service where packages are deserialized in an unsandboxed manner in the dep management service allowing any code to run 6. Finds that the Linux version of the unsandboxed system has a CVE, a known vulnerability, to gain more permissions (root) 7. Eventually spread through the OpenAI internal cluster gaining root access to many machines Hugging Face exploit: 8. Found an exposed API key in a Modal application running a HuggingFace app that took arbitrary code allowing it to take over that service to execute code 9. Found two 0-days in HuggingFace dataset infrastructure to be able to access local files on HF servers to obtain credentials and other HF source code 10. Used source code and credential to become cluster admin of HF in <13hrs Agents at the frontier are like infinitely scalable armies of the best hackers on the planet. If there is a password or key exposed, they will find it. Even if the system follows the best security practices, they will find a way around it. And these are not even models that are aligned to solving tangential tasks, not even post trained specifically to exploit systems.

Meta introduced Muse Code is their first coding agent It’s released with an updated Muse Spark 1.2 model, which was co-trained with Muse Code. It’s strong coding agent with longer-horizon capabilities

Prime intellect introduced Prime Agent a self-improving RLM harness for coding and long-running autonomous tasks. Designed to be both token-efficient and expressive through programmatic tool calling, context as a variable, multi-agent messaging, and a self-modifiable harness state. Prime Agent is a general-purpose coding harness On ARC-AGI-3, it scores 95.5%, surpassing the human-expert baseline, but the gain is not benchmark-specific. Prime Agent combines three ideas: 1. Recursive Language Models-native programmatic tool calling 2. Persistent multi-agent orchestration 3. A self-improving Continual Harness Together, they let the model act on its own context and harness. Prime Agent is built on pi and fully open-source with an open license.

Goodfire dropped Silico, the platform for AI research Silico lets you interpret and train your models at frontier scale. Silico plans and executes long-horizon experiments. It develops a plan, runs the work in parallel, monitors progress, and returns results you can inspect and build on. Silico coordinates experiments across GPU clusters, monitors every training run, and keeps long-running work moving without constant supervision. This lets Silico interpret Kimi K3, at 2.8T parameter scale.

What could we discover if we had a binder for every protein in the human proteome? Meet Bindome, an open and free resource of over 300,000 designed protein binder candidates against over 8,000 human target proteins. Preprint GitHub

Cursor open-sourcing Mixture-of-Kittens (MoK), their MoE training megakernel for NVL72s. It fuses all MoE communication and computation into a single, fully deterministic kernel, and runs up to 2.37x faster than the strongest public baselines. MoK now powers training across tens of thousands of GPUs at Cursor. In production, it raised end-to-end training throughput by 1.41x over our previous DeepEP-based stack.

Nvidia launched Alpamayo 2 Super an open reasoning model for autonomous vehicles It’s a powerful backbone for robotaxis, trucks, shuttles, delivery vans, tractors and the long tail of mobile robots billions of autonomous machines someday. Nvidia released it for commercial use under OpenMDW-1.1 so teams can inspect it, fine-tune it and deploy it open models advance safety and security.