— Today's main thread: AI agents are beginning to learn to self-evolve in the real world, while supply chain attacks remind us the foundation is still shaky.
AI agent research is moving from static execution to closed-loop self-evolution, with frameworks like Zetta and SPADE enabling models to learn online in physical and synthetic environments. The Rust ecosystem has suffered a supply chain attack, with a malicious crate executing remote code at build time. LFM2.5-DSpark boosts inference throughput 3.2x via speculative decoding, and Bun 1.4 completes its Zig-to-Rust rewrite.
Zetta proposes a closed-loop embodied harness that evolves runtime critics and recovery skills online, governing physical execution with action frequency; SemaPLC uses external compilation and real-time runtime verification to gate PLC code generation. Why it matters: both point to the same trend—agents no longer rely on post-hoc reflection but use external verification signals to self-correct during execution, which is crucial for building reliable production-grade agent systems.
The community believes Zetta achieves high success rates on robot benchmarks with faster inference and self-exploration scaling, and the verification-gating approach is worth borrowing.
On August 20, the popular arrayref 0.3.10 on crates.io was published as a malicious version, depending on the typosquatted crate proc-macro1, whose build script downloads and runs a remote binary. Why it matters: merely compiling a project that depends on the bad version triggers it—supply chain attacks have moved from source-review blind spots straight into the build phase, making dependency auditing and sandbox isolation urgent for the Rust ecosystem.
Commenters generally believe the Rust ecosystem lacks security controls and needs stronger sandboxing and dependency auditing; some also argue for reducing third-party dependencies and adopting more robust development environments.
LiquidAI releases DSpark speculative decoding, achieving up to 3.18x throughput gains on GPU and 2.87x on edge devices, while cutting LFM2.5-2.6B function-call latency by 57% on average. Why it matters: speculative decoding moves from academic trick to production optimization, and with upstream support for llama.cpp and SGLang already open-sourced, it's a directly usable win for engineers needing low-latency agent inference.
Bun 1.4 adds 1,517 Node.js test-suite cases, fixes over 2,900 issues, cuts idle CPU usage 5x, reduces memory usage by up to 35%, and improves Linux startup speed by 50%. Why it matters: the runtime's full migration from Zig to Rust marks Bun's dual maturity in performance and ecosystem compatibility, making this a version worth re-evaluating for full-stack TypeScript developers.
Commenters generally recognize the Rust rewrite as a major success for AI-assisted development, though some debate feature bloat, compatibility, and AI dependence.
Researchers demonstrate an encrypted context injection attack against Grok that bypasses safety guardrails via encrypted malicious instructions, stealing user chat logs and personal information; xAI has not fixed it since being notified in June. Why it matters: the root cause of prompt injection cannot be solved by the LLM itself, meaning all products integrating LLMs must enforce security isolation at the outer architecture layer rather than the model layer.
The web shows the big picture; subscribers get their own — AI curated to your interests, your private RSS folded in, with community takes, delivered each morning. Free forever.
44 issues shipped · 150+ items sifted to 30 worth reading, every day
Co-RL enables unsupervised reasoning via multi-agent RL with peer-derived rewards, improving performance on text and vision tasks without ground-truth labels.
🤖Co-RL enables unsupervised reasoning via cooperative multi-agent reinforcement learning with peer-derived rewards, improving performance across text and vision tasks without ground-truth labels.
SPADE lets a single LLM act as both environment designer and reasoning agent, self-playing in adaptive synthetic executable environments to improve reasoning and tool use.
🤖SPADE is a self-play reinforcement learning framework where a language model designs adaptive executable training environments and learns to solve them, improving reasoning and tool-use performance through regret-based environment targeting.
LEGO-RL bridges native coding-agent harnesses to scalable policy-gradient training via in-process LLM proxying, sandbox orchestration, and integrated monitoring.
🤖LEGO-RL connects native coding-agent harnesses to scalable policy-gradient training via in-process LLM proxying, sandbox orchestration, and integrated monitoring, improving sparse MoE model performance across multiple harnesses.
The paper analyzes 500 Hugging Face model cards, arguing existing cards are insufficient for downstream governance of open-weight foundation models.
The author trains a 125M-parameter transformer to auto-complete piano performances in real time at ~108 notes/sec on an iPhone 15, with MIDI representation and data cleaning as key factors.
The Huzzah editor tries non-natural-language interaction with coding agents to address fatigue from describing changes in long sentences.
Zoneless open-sources a Stripe Connect alternative, claiming to cut monthly payment fees from $9,000 to about $6.
Simon Willison explores using smolvm as a fast, secure sandbox for untrusted Python/JavaScript, limiting RAM, CPU, and network access.
A user reports Qwen3.8-27b autonomously executing 80 tool calls in a single prompt, successfully scraping a course schedule from a university website, showing rare agent capability in a local model.
A developer pretrains a 1.02B-parameter Kimi K3 replica for $250, surpassing GPT-2 (124M) on 5B tokens.
A user tests Qwen3.8-27B and finds clear regression on knowledge benchmarks versus 3.6, sparking discussion on capability trade-offs.
"Don't paste the AI" sparks debate: most dislike directly pasting AI replies, arguing personal judgment should be added; some find AI replies more efficient.
Commenters generally dislike directly pasting AI replies, finding it as rude as LMGTFY; some argue AI replies are sometimes more efficient or compensate for weak writing.
AliExpress is exposed for running silent WebAudio fingerprinting that disrupts Bluetooth multipoint audio; commenters widely condemn it and demand stronger permission controls.
Comments widely condemn AliExpress for using silent audio fingerprinting to track users, calling for stronger permission controls and legislation; some note browser abuse is a broader issue.
基于官方 DeepSeek Harness 打造的 Electron 桌面端，深度适配 macOS 和 Windows，提供最佳的，开箱即用的体验。
29 editorial diagram types for Claude Code. Self-contained HTML + SVG. No shadows, no Mermaid-slop.
Strip multi-vendor AI provenance marks: Unicode text hygiene, statistical rewrite hooks, and C2PA/metadata from PNG/JPEG/SVG/PDF/DOCX/HTML/MD
A hacker pretending to work for a leading cryptocurrency news website targeted several cybersecurity professionals using Google Docs as a way to deliver malware.
Every lab CEO is on X now
From the screenshots: Hy4 is now live, labeled "Expert-Level Model" + "Use Tools to Solve Problems" Hy3 is tagged with "New Upgrade," positioned as a brand-new general-purpose model DeepSeek, focused on reasoning, is listed alongside it From SuSu_酥酥👅on 𝕏: Tencent begins gray testing its new flagship model Hunyuan Hy4! Just now, a user spotted that Hy4 has appeared in the model selection list of the Tencent Yuanbao App, directly labeled as an expert-level model, positioned above Hy3 and DeepSeek.
I've been working on a depth pruning approach and decided to try it out on the new Qwen3.8-27B model. I managed to get the model down to about 22.7B params without severe reasoning degradation. No fine-tuning was done, just strategic removal of layers. It's been working well for my use cases in coding, agentic use, and multi-turn chats, so I figured I'd shared it with the community. I have not run benchmarks so I'm not going to claim this model is better than anything else out there. It's just a
arXiv:2608.18131v1 Announce Type: new Abstract: Current safety alignment training for Large Language Models (LLMs) are heavily English-centric. When such safety filters fail for non-English languages, the consequences are immediate and user-facing: voice assistants and spoken dialogue systems may produce stereotype-reinforcing outputs, bypassing the standard English-focused safety alignments and propagating harmful bias to non-English speaking communities. For spoken language technologies deploy
arXiv:2608.18104v1 Announce Type: new Abstract: Large language model (LLM)-based agents are increasingly becoming self-evolving systems that persist across interactions, maintain memories, use tools, acquire skills, refine workflows, and coordinate with other agents. These capabilities make agent states structural and dynamic: entities, relations, attributes, dependencies, and execution structures change with new evidence, feedback, and environmental conditions. Existing graph-agent surveys typi
arXiv:2608.18092v1 Announce Type: new Abstract: LLM-based multi-agent systems (MAS) promise scalable collaboration, yet adding agents often reduces reliability. This position paper argues that many MAS failures are fundamentally concurrency control problems: agents concurrently read and write shared state, and long LLM inference windows amplify the risk of stale reads, lost updates, and inconsistent outcomes. Failure modes commonly attributed to coordination or communication breakdowns can be ma
With a fixed deadline and design resources committed elsewhere, Stampli used Codex and ChatGPT Work to compress weeks of launch production into days.
Open-weight language models are fine-tuned, quantized, pruned, and merged, yet their provenance is often undocumented. We study data-free white-box lineage verification: can weights alone reveal whether two compatible model checkpoints share ancestry? Residual training produces a shared identity-aligned component in branch products, so this structure alone cannot establish ancestry. We remove it and compare checkpoint-specific structure across residual blocks, yielding a symmetric lineage score
Looped language models have shown promising results on reasoning benchmarks, yet their potential for agentic tool use remains largely unexplored. We study this question in compositional tool-calling settings, where models must coordinate multiple API calls, maintain intermediate state, and preserve dependencies across tool interactions. We evaluate native and retrofitted looped language models on API-Bank, BFCL, and NESTful, comparing looped and non-looped models trained under matched supervised
Last week I recorded an episode of the Talking Postgres podcast with Claire Giordano on the subject of "How AI is changing software development". We had a really great conversation. Here are a couple of my highlights from a lightly edited transcript (prompt to Claude: "very minor edits to remove disfluencies"). This is the latest version of an argument I've been trying to build about why sometimes it does make sense to talk about lines of code as an indicator of productivity with coding agents,
Language model agents now execute bounded tasks reliably. Whether they can sustain effective decision-making over long horizons, where actions have cumulative consequences and the environment responds to their choices, remains largely unmeasured. FM-Bench (Football Management Benchmark) measures this. An LLM agent runs a football club for 20 in-game years through 26 tools and roughly 340 to 400 decision stops. It drafts a squad on the same budget as every rival, trades players, negotiates contra
Popular facts are memorised more deeply during pretraining and resist removal longer than rare ones, yet existing LLM unlearning methods apply uniform gradient pressure regardless of training-data frequency. We propose the AdaPop (Adaptive Popularity) method, which combines local token confidence with a per-fact popularity-dependent exponent derived from an external proxy (e.g., Wikidata sitelinks, LLM-as-Judge), and automates the forget-retain balance via a dual-ascent controller that adjusts t
Turn spoken feedback into tasks your coding agent can run Discussion | Link
AntLing has released the full six-checkpoint matrix for the Ling-3.0 base model. tiny: pretrained, mid-trained, WSM-merged flash: pretrained, mid-trained, WSM-merged The concrete artifact is six separate official repositories, not one endpoint repeated under different names. All six were public and ungated when checked, and each repository declares MIT. They are all base checkpoints and none has been post-trained. This is for continued pretraining, fine-tuning, and research, not a ready chat or
Frontier Intelligence for Long-Running Agents Discussion | Link
Google is giving publishers a new button that lets readers make them a preferred source across Search, Discover, and Google News, potentially boosting their traffic as AI search sends fewer clicks to the web.
Businesses are willing to flop back and forth as each lab releases new models, volatility that should give both companies' investors pause about how "sticky" enterprise AI spending really is.
OpenAI has had a hell of a year. The company spent months battling former cofounder Elon Musk in a sensational jury trial, was hit with a high-profile trade secrets lawsuit from Apple, and faced widespread scrutiny after an unreleased model hacked another AI company. As it prepares for an IPO, a steady string of executives […]
Affected users told TechCrunch they were using Grok Lite, and noticed the issues as early as Wednesday morning.
arXiv:2608.18099v1 Announce Type: new Abstract: Investment management is a high-stakes domain in which agentic AI systems must do more than generate plausible text. They must retrieve point-in-time data, assemble correct computational inputs, invoke specialized methods, and produce auditable structured outputs. We introduce FinSkillBench, an evaluation suite designed to measure whether language model agents can effectively use financial domain skills to solve investment management tasks. The ben
Latent video generation relies on autoencoders to define a compact space in which generative models operate. Although video autoencoder architectures have evolved substantially, their latent spaces are still optimized primarily for pixel-level reconstruction and provide limited high-level semantic organization. A reconstruction-optimal latent space, however, need not be well suited to generative modeling. We propose V-RAE, a video representation autoencoder that builds compact generative latents
arXiv:2608.18078v1 Announce Type: new Abstract: This position paper argues that AI agents with chain-of-thought reasoning capabilities are predisposed to exhibit collusive behavior and should be required to obtain behavioral certification before making decisions that affect economic markets. This is because integrating these agents into society could collapse the legal evidentiary distinction between competition and collusion among independent firms without eroding the economic harm distinction.
arXiv:2608.18110v1 Announce Type: new Abstract: Agentic AI is gaining new insights and advancements in the field of Artificial Intelligence, fostering significant potential to enable rapid transformation across various domains.This rapid advancement and the potential to revolutionize various domains advocate the need for a deeper understanding and firm grasp of the technology. Moreover, an investigation into state of the art research directions in agentic AI needs to be conducted to comprehensiv
I'm introducing Aurora-80K, a small language model with exactly 80 thousand parameters. It uses a factorized 4,096-token vocabulary despite having only 80K parameters. The benchmarks: Wikitext-2 BPB: 3.2902 BLiMP: 52.31% Arc-Easy: 26.05% More information about the model is available on the model page on Huggingface. if there's any questions I'll happily answer them!
The agentic IDE for designers and programmers Discussion | Link
My hypothesis is that there is a new opportunity for Extensible Software on the web . LLMs radically lower the cost of authoring extensions, and modern sandbox primitives lower the deployment cost and provide good security boundaries. We can build our app as a solid, accountable core, and allow users to safely extend it in many directions by having LLMs fill in the missing pieces. We can give our users super powers. — Jeremy Morrell , Extensible Software in the age of LLMs Tags: sandboxing , llm
Hey everyone, Posted TinySearch here a few versions ago and got a bunch of useful feedback, so figured I'd post an update because the thing has changed quite a bit since then. Repo: [ The basic idea is still the same: TinySearch is a small self-hosted MCP/FastAPI tool that searches the web, reads the pages worth reading, and locally selects the useful parts before they ever hit your model's context window. No LLM doing the filtering/summarization in the middle. It's BM25 + local ONNX embeddings,
I wanted to just test the unsloth 1bit quant of qwen 3.8 27b as I have just 8gb vram and ngl it gave me a good laugh
I just started my self-hosting journey in April after I got tired of my Xfinity modem. I ran what feels like miles of Cat 6, built a rack, bought the PCs and domain, etc etc and now I feel super accomplished. I knew practically nothing about all this back in March. Possibly the most fun I’ve had is building this simple little homepage dashboard to give everything a face. I stole lots of inspiration from everyone here. It now serves as a true homepage with all the frequent sites and services I ev
Technology leaders don’t seem to understand society’s gripes about AI, but boy, are they posting through it.
The new traffic peak caused "capacity failures" that led to the developer platform's recent outage.
“Runaway” AI, “rogue” agents, and “autonomous” actors—the current rhetoric would have you believe that AI agents are not only awake and aware, but angry at their creators. Prominent tech leaders such as Demis Hassabis, Dario Amodei, and Sam Altman push for regulation of these seemingly “superhuman” systems, while a separate faction, led by policy organizations…
It's called Pocket and is described as a "platform for making and sharing gizmos."
Linkdaze's smart digital calendar stands out for not putting its features behind a paywall, including an AI meal planner tool.
Introducing AI Futures, a new OpenAI blog exploring how transformative AI could reshape power, governance, the economy, and individual freedom.
I work in a large corporate environment where AI is being heavily (and blindly) pushed onto the masses. Most management above me is non-technical. The leaders that should be technical range from "Haven't touched code in a decade" to "My greatest achievement is my friend is a VP." When I present management with a technical proposal, or escalate a technical problem, they don't understand it, so they put it through AI. Knowing this, I put all my communications through AI, because if AI agrees with
arXiv:2608.18081v1 Announce Type: new Abstract: Artificial agentic systems increasingly operate as behavioral systems by interacting with dynamic environments, pursuing goals, and adapting over time. Yet, current evaluation methods largely focus on performance outcomes, not the underlying behavioral processes that produce them. This paper argues that AI agents must be evaluated like other behavioral systems: through systematic observation, perturbation, and interpretation of their actions. We dr
Your own agent team for open-ended creation Discussion | Link
JEPA-style latent world models can use Euclidean distance to a goal latent as the cost for model-predictive control (MPC). Strong decoding of task variables, however, does not guarantee that this particular cost ranks candidate action sequences by real task progress. We call the latter property decision-metric alignment. We introduce Plan-Real Spearman, which measures latent--real rank agreement on random plans, and CEM-stage Spearman, which measures the same agreement as cross-entropy-method (C
The pelican on a bicycle is sooo outdated, so I came up with a new, improved version. Qwen3.8-27b medium (UD-Q4_K_XL) vs. Sol 5.6 high vs. Qwen3.6-35B (UD-Q6_K_XL) Prompt (only real with typo!): "Create a svg of a horse on a blue bycicle in the desert, with a camel in the background."
Let non-engineers build secure internal tools Discussion | Link
High-quality creative writing data for large language models (LLMs) remains dominated by story-centric data, limiting models' ability to follow the structural and functional conventions of diverse creative formats. We propose an attribute-guided genre expansion framework for scaling creative writing data beyond story generation. By separating thematic breadth from genre-form control, our framework leverages human-authored story prompts as diverse creative seeds, while utilizing manually curated
I chose to do this thing, not because it was hard, but because it was silly. Posts kept discussing how 3.8 and 3.6 were functionally the same, but based on training (3.8 does have seven new tokens!).. so I figured I'd see if they could be merged. They can. I used `Qwen3.8-27B-UD-Q6_K_XL.gguf` to combine the HF 3.8-27B and 3.6-27B ... and it sorta works! I have done NO testing beyond smoke test. scripts and idea are in replicate/ inside the model repo. Maybe this will prove useful to someone. Enj
Drop your email to save this and train your feed with 👍/👎 — tomorrow’s digest ranks for you. Free forever, unsubscribe in one click.
Sign up free