Monday, October 19, 2026

Pulse (last 3d) · Research 83 · Agents 56 · Industry 50 · Products 45 · Open Source 38 · Models 28 · Legal 22 · Releases 18 · Enterprise 18 · Policy 16

Trending · Anthropic 16 · OpenAI 14 · Meta 11 · Nvidia 10 · Claude 9 · Claude Code 9 · FDA 9 · Codex 7 · Hugging Face 7 · TechCrunch 7 · DeepSeek 6 · Muse Spark Open 6

---

• Harness² improves agent harnesses without verifiers, adding up to 11.5 points on frozen weights. x.com
• Perplexity Decisions returns probabilities instead of prose, at $0.04 per million input tokens with free output. producthunt.com
• shim-cli masks secrets and personal data in Claude Code tool results before the model reads them. producthunt.com
• StateAuditor is the first measured fix for agents that plan around stale state they know is outdated. arxiv.org
• A distillation attack on OpenCode drains users' API keys for Anthropic, OpenAI and Google models. reddit.com
• With lossy speculative decoding, a higher acceptance rate can come with no real speedup. arxiv.org
• Gemini 4 Argon is reported to tie GPT-6 on benchmarks. theneurondaily.com

• Krea 3 - edits images at native 2K in under 8 seconds for $0.06 each, with up to 6 reference images. producthunt.com
• Bee Frontier and Bee Eclipse - Amazon wearables that summarize conversations without keeping the audio, tied into Alexa+. producthunt.com
• Craft Agents - an open-source local desktop agent app that works with any provider, MCP tools and scheduled automations. producthunt.com
• AI Window Deck - an open-source Chrome extension that tiles many cloud coding-agent sessions across displays. producthunt.com
• Bulbul V4 Flash - Sarvam's text-to-speech starts audio in under 250 ms, with 83 voices across 11 Indian languages. producthunt.com
• Popcraft - one URL gives agents cost-previewed access to Veo 3.1, Kling 3.0 and 20+ other generators. producthunt.com
• Hybrid Reasoning Agents - Together's agents reason in latent space and act in text, with code, models and data released. arxiv.org | arxiv.org
• Nemotron Super VL and Bolt - NVIDIA adds open models built for efficiency, from data center to desktop (press release). blogs.nvidia.com
• Cajal-1 - an open 218M-parameter model that predicts whole-cortex fMRI responses to video, audio and language. reddit.com
• Connectome fly - a virtual fly runs on its real 70k-neuron wiring diagram, streamed live with no scripts. reddit.com
• CocoRNA - uses cooperative multi-agent RL to design RNA sequences. arxiv.org
• TempoPFN - linear RNNs pre-trained on synthetic data forecast time series zero-shot. i.redd.it | i.redd.it
• LagNet - lag-structured operators for long-horizon time series forecasting. arxiv.org
• Claude plugin marketplace - Anthropic's open-source skills, written in markdown and JSON, for marketing, sales, finance and brand voice. x.com
• Multi-model router, one year in - a practitioner reports seven practices that survived production and three that failed. reddit.com
• SonicBoom v0.0.3 - a GPLv3, MLIR-based runtime claiming 1.1 to 1.2x faster decode than llama.cpp (vendor benchmark). reddit.com
• ReviewBench - Laude Institute researchers claim to beat the Pareto frontier on GitHub's PR-review benchmark (single source). x.com
• Dream2Discern - predicts multi-agent runtime failures by training on hard latent variants. huggingface.co
• SciΠvot - benchmarks the judgment calls research agents make at pivotal moments, the way a principal investigator would. huggingface.co
• DiscoveryRSI - a system that recursively learns to discover valuable hypotheses. arxiv.org
• AI-driven reproducibility - on using AI to check reproducibility and analytical robustness in scientific discovery. arxiv.org
• PremiseCheck - an efficient fact-checker for the premises in human and AI peer reviews. huggingface.co
• Personalized memory gating - learns what an agent should remember from delayed, one-sided feedback. arxiv.org
• Beyond SAEs - sparse coding finds concepts more consistently than sparse autoencoders. arxiv.org
• Lexical Reference Frames - reads LM residual streams in their native lexical semantics. huggingface.co
• Belief-Trajectory Energy - measures the path a model's belief takes on the way to a prediction. huggingface.co
• Behavioral Couplings - traces how couplings form in language models and lead to emergent behaviors. arxiv.org
• Student-conditioned traces - builds distillation traces around the student model rather than teacher answers alone. arxiv.org
• LEV - looped decision models that can think either fast or deep. huggingface.co
• Multilingual embeddings - removing a language-specific residual subspace improves training-free multilingual embeddings. arxiv.org
• Embedding search in images - an embedding model can retrieve text that appears inside images. i.redd.it
• Stuck on Shuffle - studies why vision-language models fail to bind attributes to objects. huggingface.co
• Unified multimodal models - asks whether visual generation expands what unified models know. huggingface.co
• Pixel Diffusion Transformers - revisits the design space of vanilla pixel-space diffusion transformers. arxiv.org
• VocalGrad - benchmarks how well audio language models perceive acoustic properties. arxiv.org
• AVC-Reward - an event-centric reward model for audio-video consistency. huggingface.co
• ENFR - evidence networks with flow-matched references for amortized simulation-based model comparison. arxiv.org
• Spherical SO(3) CNN - builds rotational symmetry into the network so it is not relearned. arxiv.org
• Flow-box invariants - uses ML to detect singularities and near-integrability in dynamical systems. arxiv.org
• Logic-guided graph diffusion - logic constraints guide denoising diffusion for graph generation. arxiv.org
• Mesoscopic learning dynamics - theory of training dynamics in networks whose width scales with input dimension. arxiv.org
• 3D world agents - trains agents to reason and explore in 3D environments. huggingface.co
• OfficeTown - simulated organizations used as environments for evolving agent intelligence. huggingface.co
• BridgeWAM - connects video and action experts through latent bridge queries. huggingface.co
• MorphAct - robot policies that change along with the environments they act in. huggingface.co
• Warcraft III self-play - self-play RL on one desktop beats the game's built-in AI. warlock.ai
• Arrow detection - a ConvNeXt backbone pretrained with DINOv3 detects arrows. reddit.com
• GPT-2 KV cache from scratch - a PyTorch implementation with measured correctness, latency and memory. reddit.com
• amber-8 - lets 80s-style assembly programs call small local LLM decision models. reddit.com
• KillSwitch - a programming language designed to be hard for LLMs to use. blog.picheta.me
• Disney and Character.AI - Disney will hire most of Character.AI's workforce. techcrunch.com
• Cloudflare pay-per-use - Cloudflare is testing a way for AI systems to pay for the content they use. blog.cloudflare.com
• Google publisher pilot - Google updates its AI pilot that pays news publishers. blog.google
• Gemini scam screening - Google adds opt-in scam detection for text messages. theneurondaily.com
• Philadelphia delivery robot - a woman was arrested for blocking a delivery robot. theverge.com

Each result makes a fixed model perform better by changing what surrounds it: the harness, the context it is given, or the state it checks.

Harness² adds up to 11.5 points without touching weights or needing a grader Google Cloud AI Research's Harness² tunes an agent's harness for each task without ground-truth verifiers and carries what it learns across tasks, for gains of up to 11.5 points on frozen weights (preprint, unreplicated). DrugHarness applies the same evolvable-harness pattern to computational drug discovery, where gold labels are scarce and retraining a validated model means validating it again. x.com | x.com | i.redd.it | arxiv.org

The STALE gap now has its first measured fix: +5.0 points StateAuditor audits in the opposite direction from normal and raises VTA from .686 to .736 on agents that know stored state is outdated but still plan around the old value (preprint, unreplicated). This is the failure to expect from agents reading patient records that change between sessions. arxiv.org

Compressing context costs coding agents fixes Coding agents resolve more issues when they select whole files than when they read compressed context (preprint, unreplicated). That counts against the compression layer many teams added to save tokens. Separately, FC-SWE trains long-horizon software agents with RL conditioned on their own failures. huggingface.co | huggingface.co

Builders and attackers are both targeting the layer between an agent and its tools, which is where data leaks and where policy can be enforced.

Masking data before the model sees it ships as an open-source hook shim-cli is a hook and local proxy for Claude Code, Codex, Copilot CLI and VS Code that masks secrets and personal data in tool results before the model reads them, with no account and no telemetry. That is de-identification at the tool boundary, the pattern a PHI pipeline needs, and the HN case for running your own LLM proxy argues the same point for observability. producthunt.com | jrdsgl.com

Approval, policy and spec gates are becoming local, auditable code Charter wraps any API as an agent tool and blocks rule-breaking calls before they reach the API. Brocoded sends agent approval prompts to a phone with plain-language risk labels, and Pathmode checks pull requests against an approved intent.md spec. Each moves enforcement out of a vendor setting and into an artifact you can audit. producthunt.com | producthunt.com | producthunt.com

The attack surface is credentials, browsing and skill memory A reported distillation attack on OpenCode drains users' API keys for Anthropic, OpenAI and Google models (single source), and Tencent says its agents browsed in ways nobody intended. Proof Before Reuse shows agents only the skills that later sessions reproduced, which keeps a poisoned skill out of procedural memory (preprint, unreplicated). reddit.com | reddit.com | i.redd.it

Each item cuts the cost of a decision or a token, and one shows a common speed metric overstating the savings.

Perplexity now sells probabilities, not prose, at $0.04 per million input tokens Perplexity Decisions answers yes/no, multiple-choice and scoring questions over text, JSON and images with probabilities, 1 to 128 questions per request, on pplx-decider-v1-27b, with output tokens free (press release). At that price, structured triage of reports or eligibility criteria could run on every record, within a cap of 10 requests per second. producthunt.com

Speculative-decoding acceptance rates overstate the speedup Training-free lossy speculative decoding inflates output length, so a higher acceptance rate can come with no real speedup (preprint, unreplicated). DLite offers history-free speculative decoding for agent workloads, and FuTransformer runs FP8 prefill with INT4 decode on Hopper. End-to-end latency is the number that compares fairly across these methods. arxiv.org | arxiv.org | arxiv.org

Token savings now come from sending each job to the cheapest capable component Atelier has Codex find the problems and Claude only judge them, for about 71% fewer Claude tokens per review. Skisso hands an agent a 380-token UI diff instead of a 25,700-token page (vendor benchmarks). producthunt.com | producthunt.com

When top models tie on capability, how a vendor is priced and governed starts to separate them.

Gemini 4 Argon ties GPT-6, and Meta prices at the premium tier Google's Gemini 4 Argon is reported to tie OpenAI's GPT-6 on benchmarks (single source), and Meta's Muse Wonder launches priced like Opus. With capability level at the top, the choice of model for regulated work comes down to contract terms, data residency and price. theneurondaily.com | theneurondaily.com

Anthropic sets aside 5% of its shares and lets an outside watchdog in Anthropic is committing 5% of its shares to a public benefit fund and has given an outside AI watchdog employee-like access. Independent access inside a lab is the kind of assurance a compliance review can cite when vendors are otherwise equal. anthropic.com | techcrunch.com | arstechnica.com

17 items: opinion and discussion threads, a hype thread, items with no content or only a title, off-topic theory papers that state no result, and an agentic-RL thread that adds nothing new.