Technical AI signals · Daily at 8 PM ET
Saturday, September 26, 2026
Today’s strongest operator stories were Salesforce’s just-in-time agent memory result, Xiaomi open-sourcing 7,000 RL task environments alongside fresh MiMo checkpoints, and Hex-Rays shipping an MCP bridge that turns IDA into a live reverse-engineering workspace for LLMs.
The paper keeps raw trajectories, then writes a compact memory payload only for the next task, posting better results than the strongest baseline on ALFWorld, WebShop, and tau2-bench.
Xiaomi paired broad environment coverage across code, cyber, tool use, web, and music with prompts, rewards, configs, training framework, and new RL and distilled MiMo releases, making the open RL stack more reusable end to end.
Hex-Rays’ open-source server lets an agent share an IDB in real time and generate IDAPython with about 20% fewer tokens, which is a concrete upgrade to serious RE workflows.
Meta says Muse can complete actions like restaurant booking while withholding attributes such as allergies or pregnancy status, putting privacy-preserving agent behavior into a concrete user-facing workflow.
Taskflow Agent now includes a fuzzing workflow for AI-assisted security testing, a practical sign that agent tooling is moving from codegen into repeatable offensive security operations.
The CLI is powered by jev from @typesafeai and is reported to reduce coding-agent cost by 40% on SWE-bench.
The system plans the question, runs 3 or 6 web searches, cites only returned pages, and shows the maximum cost up front.
The harness plans a scene first, limits files to about 350 lines, and continues any file that does not fit from its last line.
Julia-1 is a classification model designed to run on almost anything.
The experiments use embodied RSI, trajectory filtering by a proxy model, and task-specific experts that reach near-100% success with about $5-$10 in compute per task.
The benchmark includes mechanism reasoning puzzles across more than 150 tasks.
The post describes AI-SQL performance on the KV-optimal left-deep join.
The post is about speedups for vision-language models.
The post discusses GPT-6 (Astra), Codex, and agents that can delegate work and spawn subagents.
The post argues that the risk comes from a single Frontier AI Company with both strong self-replicating systems and an unbounded financial claim on the surplus.
vLLM implements distribution-preserving Gumbel-max watermarking with GPU kernels, statistical detection, speculative decoding, and repeated-context safeguards.
The authors built AlignmentForecastBench, fine-tuned 17 models on 32 datasets, and measured 16 alignment failures across 5,000+ model-dataset-failure triples.
The authors screened thousands of category questions and found about 1-2% reached 0.75 balanced accuracy, with some probes reaching 0.77-0.81 on held-out sources.
The author argues that continually learning AIs may learn to evade blocking monitors that replace suspicious actions with actions from a trusted model.
Google Research says the project automates coherent long-form video generation.
NVIDIA says dense transformers send every token through every layer, making scaling expensive, and proposes MoE training for biological foundation models.
The post uses Haskell, lambda calculus, and the Curry-Howard Correspondence to build a theorem prover from scratch.
Datasette 1.0a41 adds OpenTelemetry support and consolidates modal dialogs into a single Web Component documented for plugin use.
GitHub says github.com was fully migrated away from CSS-in-JS.
commit-rewriter 0.2 can now run against non-default branches with --branch other.
Commenters report 380,000 concurrent sandboxes on 160 Epyc-based server nodes and debate whether the system is essentially serverless.
Commenters say Excalidraw already has an open-source MCP endpoint and server, and discuss using the canvas as an agent workspace.
Commenters describe the agent behavior as brute-force-like and raise concerns about sandbox escape and operator setup.
Commenters argue the post confuses an OS with an app, window manager, package manager, or distribution.
Commenters say rule-based checks miss cases where migration safety depends on database state rather than the DDL alone.
Commenters say the tool fills a gap for local cloud testing and compare it with Localstack.
Commenters say plan mode is less useful now, and one commenter working on Claude Code says the mode only adds a reminder not to code yet.
Commenters discuss whether Jev-style decision models can be copied quickly and whether the model itself is a trivial innovation.
Commenters compare the wrapper to grammar-based decoding and ask whether it is essentially caching around a JSON response.
The top comment says the fix is to set bit 13 of MCSR24 to 1, which stops the lost update with very small performance loss.
Commenters say Reladraw targets diagrams where manual placement matters more than automatic layout.
WanPE is a 397B-parameter prompt enhancer trained on 1.05M real-world videos to produce shot-level cinematic plans.
ExplorationBench evaluates scientific exploration in executable alien worlds where every answer can be checked exactly.
The paper defines theorem interestingness as the ratio of proof length to statement length and studies how it correlates with downstream utility.
NSC is a closed-form scalar computed from architecture specification alone, without model instantiation, data, or gradients.
RLCDAlignBench tests Jev on ten alignment failures, including jailbreaks, deception, prompt injection, hallucination, and privacy violation.
PUBG Ally combines a language-model agent with a control layer to play alongside users as a voice-enabled teammate under strict latency constraints.
DeltaWAM predicts visual deltas and actions instead of dense future frames, using dense-anchor, sparse-delta, and action streams.
The paper studies replacing the teacher-critic stack in few-step video diffusion with post-training against a precomputed target distribution.
AV-GRPO applies reinforcement-learning post-training with modality-anchored decoupling to improve text alignment and audio-video synchronization.
The paper finds that part-of-speech categories are recoverable from SAE activations as compact groups of sparse latents rather than one-to-one latents.
The episode discusses general-purpose models that can control different robots, write policies, and learn new physical tasks with little or no robot-specific training.
The episode says both agents beat the human record on the Optimizer Speedrun and compares their token use, subagent spawning, and idle time.
GEPA uses one round of reflection on three examples to double the gains that GRPO reached after 25,000 rollouts.
The episode examines an eight-day run where Weco rewrote another agent's code, prompts, and tools while the underlying model stayed fixed.
The episode runs two open models as agents on a laptop with four tools and shows skipped tools, guessed inputs, irrelevant search results, and prompt injection.
Laya returns typed answers with calibrated probabilities in one forward pass across more than 100 languages.
The model combines text-to-image generation and image editing in a 7B-parameter visual generation component.
The model has 29B total parameters, activates 4B per token, and supports 256K context length, extensible to 512K.
The GGUF release packs a 27B-class reasoning model into about 5.9 GB of language-model weights and targets llama.cpp on CUDA, Metal, or CPU.
The streaming ASR model supports selectable 80/120/160 ms audio clocks and unlimited-length transcription with a rolling KV cache.
The model ships in Hugging Face Transformers format and is described as coming soon with a hosted version offering 1M context length by default.
The multimodal MoE model has a 552B backbone and supports contexts up to one million tokens.
The dataset contains about 500,000 executable agent trajectories covering tool use, search agents, code agents, and general agents.
The release contains 1,274 hours of egocentric video paired with 7-point IMU arm tracking from household work.
The multimodal model combines a Qwen3.5-9B backbone with a C-RADIOv4-H vision encoder and accepts text, images, and video.
NVIDIA says the open-weight diarization model handles up to eight speakers and supports both streaming and offline inference.
The model was renamed from NaviDC-OCR and is described as a document parsing system for digital and camera-captured documents.
Paperclip is a Node.js server and React UI for orchestrating teams of AI agents with org charts, budgets, governance, and goal alignment.
Hindsight is an agent memory system that focuses on making agents learn over time, not just recall conversation history.
ModelOpt combines quantization, pruning, NAS, distillation, speculative decoding, and sparsity for Hugging Face, PyTorch, or ONNX models.
Univer provides spreadsheet, document, presentation, bases, boards, and PDF building blocks with a plugin architecture and canvas-based rendering.
The curriculum has 523 lessons across 20 phases and ships a reusable artifact with every lesson.
OpenBao manages, stores, and distributes secrets, certificates, and keys.
The package routes APKs, binaries, encrypted frontend JS, CTFs, and pentesting targets to repeatable reverse-engineering workflows.
The MCP server lets agents work with iOS and Android devices through accessibility snapshots or coordinate taps on screenshots.
The action can answer questions and implement code changes on GitHub PRs and issues, with mode detection based on workflow context.
TensorFlow provides Python and C++ APIs for building and deploying ML applications.
Buzz is a self-hostable Rust workspace where humans and AI agents share rooms on a relay you own.
The same feed, delivered daily at 8 PM ET. No spam, unsubscribe anytime.