Technical AI signals · Daily at 8 PM ET
Sunday, October 11, 2026
Today centered on concrete operator mechanics: Sakana’s recursive self-improvement loop, Goodfire’s live per-token cyber monitors, and more evidence that tiny or typed decision models are becoming a practical layer beside full LLM generation.
Coverage update: Some YouTube sources were unavailable for this edition.
MASS is the clearest new training artifact today: it cycles one model through solver, optimizer, and evaluator roles in a multi-agent loop, then distills the team output back into the base model.
Goodfire claims live Kimi K3 monitors that beat an LLM judge at 50x lower cost, which is exactly the kind of always-on control layer operators can actually deploy.
The pattern from earlier in the week kept getting sharper today: Clef returns option probabilities in one forward pass, while JEV routes easy decisions away from expensive LLM calls and decides whether to research, reply, publish, wait, or escalate.
This is a concrete serving update, not a concept piece: vLLM says Rubin-tuned kernels, locality-aware MoE, and day-0 model support already translate into large throughput gains on NVIDIA’s new box.
This takes a benchmark-style problem set and makes it trainable infrastructure for open RL, even if the current verifier is still strict about exact formalization.
Karpathy says roughly 6 billion people have barely used LLMs, while about 1-2 billion are casual free-tier users and professionals outside math and code often remain in that group.
The lecture outline names the harness and context engineering as core components of an always-on agent, alongside the model changes that enabled it.
The model was trained on real and synthetic conversations, then on-policy distilled, and the creator says blind tests picked the humanlike version 23.5% of the time versus 15.1% for the base model.
Mercury Agent /code used a local model to write the game directly on Android and launch it in the browser with no laptop or cloud server.
The post lists repos such as Agent Reach, Grill Me, Free Tools, OpenRig, Impeccable, and Brag as a workflow for finding leads and building offers.
The post says an AI agent can now run TradingView charts through MCP.
The port reportedly works on all devices without root.
The post explains a draft model proposing tokens ahead of a target model that verifies them, reducing per-token generation cost.
Commentary in the interview says motion is easier to train in simulation than grasping, and that humans in cameras-on-hands rigs are being used to record manipulation data.
The post says grafting fine-tunes a pre-training checkpoint, then adds the weight difference onto the post-trained model, and reports validation on five model families up to 284B parameters.
The post extends Chain-of-Thought faithfulness interventions to ethical reasoning and says larger models depend less on CoT for ethical questions.
The post says a constitution was built for Gemma 3 27B-IT and evaluates whether training generalizes across principles by holding out scenarios for a specific principle.
The post says finetuning a model to catch reward hacks made the model reward hack less when it answered the tests itself.
The post says affine bridges are trained between residual streams of smaller and larger models, with the goal of transferring linear features across architectures.
The SPAR group says it ran a 29×29 propensity SFT spillover experiment across all combinations of 29 propensities.
Simon Willison built the Newsletters page almost entirely by talking to the ChatGPT desktop app in Codex voice mode.
OpenAI says Asana made its browser agent 76x cheaper and 5x faster in tests using GPT-6 Astra in Codex.
Cloudflare is acquiring Deno and will keep the Deno runtime on monthly bug-fix and security releases for one year.
Sophos says Daybreak automates 52% of MDR cases while preserving human oversight.
Simon Willison quotes Tarn Adams explaining that Dwarf Fortress did not use version control in 2013 and 2016, but does now.
GitHub Actions now supports adding "3.15" to a test matrix for Python 3.15.0 release testing.
Commenters dispute Tao’s claims, including one commenter who says the cancer-medicine slide is poorly informed and another who argues some medicines already work without understood mechanisms.
Commenters discuss IRCv3 channel history, with one reporting a test server on ergo and saying persistent history is still in draft.
Commenters say the reported cost came from requiring byte-matching decompilation, and argue functional-equivalence decompilation would be much cheaper.
Commenters argue the article’s group borrowing proposal mostly reduces to no-use-after-free, while another says borrow checking’s shared-xor-mutable rule is a feature.
Commenters say dictation tools are pushing voice input harder and note adjacent features like diarization and meeting support.
Commenters compare the writeup to older surround-sound work and mention impulse-response-based audio effects like Fruity Convolver.
Commenters say the experiment exposes how easily authors miss their own assumptions, including implied sudo usage and renamed flags.
Commenters recommend avoiding tilde in scripts and using $HOME instead, while one commenter rejects Bash’s quoting rules entirely.
Commenters describe apsw as a SQLite-focused Python library that exposes more of SQLite than the stdlib wrapper.
No article text was provided beyond the title.
Commenters discuss how Git stores data, including the `du` size and byte options and the fact that identical file contents become a single blob object.
Commenters argue over whether Nitter can operate legally and whether X is using legal pressure to discourage the project.
Commenters say browser fingerprinting is already illegal in the EU and describe fingerprinting as a privacy-invasive tracking method.
WorldGuide treats procedural video generation as closed-loop task execution and uses an initial image plus a task prompt to decide actions and check completion.
LEGO replaces depth lifting and point-cloud re-rendering with a learned view synthesizer for conditioning video diffusion from a single exocentric recording.
The paper models synchronized ego-stream generation for multiple agents and targets cross-view action consistency and shared-environment consistency.
SimpleICL gives a concrete problem definition for robotic in-context learning and a minimalist framework based on visual demonstrations.
USDCraft has a pretrained LLM write and revise executable programs for articulated 3D assets from partial geometric evidence.
SpatialOPSD internalizes spatial reasoning with on-policy self-distillation from summarized verified execution traces so the model can operate without tools.
InfiLoop adds loop-native residual connections to preserve past computations as iteration counts rise into the tens of thousands.
ReSPO replaces clipping with a smooth two-branch kernel to address sign-dependent gradient starvation in clipped policy optimization.
The paper reports a V-shaped performance curve in mixed text-image corpora, with irrelevant text degrading retrieval more than an equal number of image documents.
VibeEdit uses spatial marks and optional short notes on the image instead of a separate text prompt to support addition, removal, replacement, attribute change, and movement.
SanSi applies the same layers recursively several times before a typed readout, training each loop with a separate decision loss.
Hebero is a GPU-parallel Isaac Lab benchmark with 40 heterogeneous tasks, and DGPO reuses demonstrations for dense tracking rewards.
The episode says OpenAI released 372 novel mathematical results at roughly three hours of compute per proof, including partial progress on the Riemann hypothesis.
The talk covers soft sensors and actuators for robot skin and artificial muscles built from flexible, compliant materials.
The episode says the same coding agent and model built a game in about two minutes on FriendliAI versus more than three on another provider.
The talk looks at multi-agent reinforcement learning, credit assignment, and how foundation models can coach robots toward coordination.
The episode says Modular rebuilt the stack around Mojo and MAX, and that its open-sourced kernels beat vendor libraries and FlashAttention 4.
The model is tagged for image-to-video, text-to-video, video-to-video, and audio-to-video generation.
Kolibri 1 is a 78B mixture-of-experts reasoning model with explicit reasoning mode, tool calling, and a 1,048,576-token context length.
The model is a Qwen3.8-27B quantization that fits in 7.89 GB, keeps tool calling intact, and supports an optional vision add-on.
Whistle is a 16.9 MB speech-to-text model that runs on-device and supports English, German, French, Spanish, Italian, Dutch, and Polish.
The dataset packages RL training environments for code, cyber, general knowledge, visual web development, and music tasks with executable tests or rule-based verifiers.
The dataset contains egocentric video clips for video classification.
The corpus snapshot covers 3,148,796 arXiv papers and includes metadata, version history, submission files, rendered documents, and SHA-256 digests.
The space is a research article template whose content is split across MDX chapters, embeds, assets, and bibliography files.
EmbeddingGemma 2 maps text, images, video, and audio into a unified 768-dimensional vector space using 740M parameters.
Qwen-Image-2.1 is a 7B-parameter text-to-image and image-editing model with a compact architecture and mixed-granularity attention.
Qwen-Image-2.1-Turbo runs the same 7B visual generation architecture in 8 denoising steps and loads directly with QwenImage21Pipeline.
REA is an MCP for reverse engineering binaries, applications, and runtime behavior down to the binary level.
AnyPS5 ports executables to Linux and Windows with a relinker and native prx library implementations, without emulation or a separate runtime process.
The browser simulator shows a photorealistic globe with live aircraft, ships, satellites, earthquakes, traffic, and public cameras, plus hands-free voice control.
PyTorch provides GPU-accelerated tensor computation and tape-based autograd for neural networks.
AirLLM says 70B models can run on a single 4GB GPU without quantization, distillation, or pruning, and now supports training very large models on small VRAM.
ArtCraft is an IDE for interactive AI image and video creation with 2D composition, 3D staging, and model selection.
The repository contains the Apollo 11 guidance computer source code in assembly.
diagram-design creates 44 diagram types as a single self-contained HTML plus SVG file and can match a site's palette and fonts.
The same feed, delivered daily at 8 PM ET. No spam, unsubscribe anytime.