2026-08-02

Release: datasette-agent 0.4a0 New await context.browser_task() mechanism allowing agent tools to run code directly in the user's browser. #33 This is an exciting new capability: it makes it easy for Datasette Agent plugins to provide tools that execute custom JavaScript in the user's browser . I used this to add a debug loop to Datasette Apps in datasette-apps 0.2a0 . Tags: datasette , llm-tool-use , datasette-agent

Antirez stealthily uploaded the new weights in the old folder... and there we were tapping our fingers.

The FBI has issued a warning after water facilities from seven states reported that they were hacked.

Tuesday was Stateless MCP day - the rollout of MCP 2.0, or the 2026-07-28 Model Context Protocol specification to use the more formal but less memorable name. This is the most significant change to the MCP spec since it first launched, and has also served to reignite my personal interest in the protocol. For background: MCP is the Model Context Protocol, which describes a standard way to expose new tools to LLM-powered agent frameworks. It was introduced by Anthropic back in November 2024 , had

On-policy self-distillation (OPSD) is a promising approach to improve reasoning language models, but it remains brittle in practice: making it work reliably often requires substantial engineering effort. We identify a structural source of this difficulty: vanilla OPSD is precisely the β=1 member of a broader policy-optimization family, where β weights the KL penalty anchoring the student to a reference policy. This equivalence turns β from an implicit value fixed at one into a controllable regul

We present ShadowDancer, a novel approach to any-action, frame-level control of interactive video world models. The obstacle is representational: existing interfaces either encode an action loosely, leaving how it unfolds for the model to improvise, or encode it exactly through structured signals that serve one family and are hard to acquire, so precise control across diverse dynamics remains impractical. Demonstration videos are the natural remedy, specifying any dynamics frame by frame; yet a

smevals - a small eval suite for evaluating models, prompts, and harnesses I've been working with Jesse Vincent's Prime Radiant applied AI research lab building out this evals framework to help answer questions about the capabilities of different models. The result is smevals , a new tool for running small eval suites across different model configurations and grading the results. The blog entry describes the tool in detail. Here's the 10 second version: Tell your coding agent to run uvx smevals

Release: llm-mcp-client 0.1a0 See this blog entry . Tags: llm , model-context-protocol

Frontier agent intelligence at Flash prices Discussion | Link

Hello! We're two friends working in tech, and we've been building Safebucket: a simple, S3-agnostic, open-source alternative to WeTransfer, Dropbox and Palmr. What's new in v0.7.0: Admin settings page: read only page for app settings Redirect download: ability to download files from quick share link via curl Performance improvements: now supporting multipart upload for large files New integrations: Backblaze B2 Cloudflare R2 OVH Object Storage Scaleway Object Storage Azure We're open to feedback

What speeds are everyone getting with deepseek v4 flash 0731? I’m getting~200 tps prompt processing / ~11 tps token gen, on 4x5060ti16gb with ddr4 3200 ram at 4-channel, via llamacpp, with context window of 128000, -ub/-b at 4096, “q8” unsloth’s lossless quant

audio.cpp 0.5 is out :) The most fun new model in 0.5 is DramaBox . It is closer to prompt-directed voice acting. DramaBox is built on the LTX-2.3 audio architecture, and prompts can control emotion, delivery, laughs, sighs, pauses, transitions, and speaker behavior. Example input (check the audio in the post): A nervous young man whispers, "I do not think we should be here." He takes a shaky breath. "Did you hear that?" The hallway answers with a slow metallic creak. He tries to laugh, but his

This pull request was added to the main llama cpp about 12 hours ago. I was experiencing some looping and poor behavior yesterday but haven't had any problems since this fix.

Real-world tasks often require models to learn from task-specific context rather than relying only on pre-trained knowledge. While recent work has highlighted this capability as context learning, existing evaluations mainly focus on textual contexts. In many practical settings, however, the context to be learned from is multimodal: scientific findings are conveyed through figures and tables, financial indicators are scattered across converted reports, and spatial decisions depend on maps, scenes

Large language model training in open-ended domains lacks verifiable rewards, making task preferences difficult to formalize as effective supervision. Contexts can convey such preferences, yet provide little additional supervision once distilled into the student, motivating contexts that evolve with student performance. However, directly using evolving contexts as in-training supervision results in an unstable distillation target and conflicting distributions, requiring mechanisms to stabilize t

The FBI, the EPA, and the Cybersecurity and Infrastructure Security Agency (CISA) have stopped short of officially blaming Iran for a spate of cyberattacks on Minnesota's water systems, but consensus is that Iran is likely behind them. That, of course, hasn't stopped Donald Trump from sharing his own theory - that Governor Tim Walz is […]

Article URL: Comments URL: Points: 89 # Comments: 37

OpenAI has reportedly found evidence of additional agent misbehavior as it looks into the incident that occurred with Hugging Face.

Role-playing agents (RPAs) have become one of the most important consumer applications of large language models. Users engage in multi-turn conversations with RPAs for experiences such as emotional comfort, making reliable evaluation essential for measuring capability, comparing systems, and guiding further improvement. Existing benchmarks, however, typically require an RPA to continue a fixed dialogue history and then evaluate the continuation using a fixed rubric detached from the user. We ide

A removable chip lets hackers inspect their badge—and keep using it after Defcon.

Vision-language models (VLMs) are increasingly used in embodied agents to interpret visual inputs, reason about spatial relationships, and make task-level decisions based on that reasoning. However, a fundamental capability mismatch remains: general VLMs can reason about the overall task but often miss the visual details that determine success, while specialist vision models can capture those details but cannot translate them into task-level decisions. In this work, we propose SpatialCLI, a fram

下一场竞争，是谁能造出更多“有用的世界”

Not just HN the site, everyone’s fuzzy search sucks. Let me do exact match search. Even if it’s an option I have to dig deep to find. When I ran a site that had lots of search use I agonized over making search good and never nuked exact match for the sake of some fuzzy matching. I wish others would do the same.

Basically you now have to mark all AI generated images, audio, video and text as AI generated. :P

Visual generation increasingly requires high-resolution images, long videos, and multimodal context, making the quadratic cost of full attention prohibitive. We introduce Chimera, a hybrid visual diffusion backbone with a principled scaling recipe. Chimera processes text, image, and video tokens in one raster-ordered stream without positional embeddings. It combines Kimi Delta Attention (KDA) for long-context state tracking with O(N) complexity, interleaved Multi-head Latent Attention (MLA) for

A full-stack approach to making advanced AI more capable, more affordable, and more widely useful.

Poolside have updated the FP8 and NVFP4 checkpoints for Laguna S 2.1, increasing the default context size to 1 million, and updating the configs. Here's hoping they fixed the looping issue, this model has been great in my development workflows, when not looping.

The biggest issue with preview was its inability to follow rules prompts and skills. It seems like no matter what you do it ignores them. I've tried first person and second person. I've tried Chinese and English. It does not follow them. That's the only problem with these models and why they're not actually frontier level and not just benchmaxxed. Every user's environment is different and they need to tune the actions and behavior of the model with rules or prompts or skills to exactly what they

The weights have now been added to the repo an hour ago. This model is built upon LongCat-Flash-Lite , the differences are that LongCat-Flash-Lite-Sparse : Replaces dense MLA with LongCat Sparse Attention (LSA) Natively supports context lengths of up to 1M tokens (vs 256k for LongCat-Flash-Lite)

Despite a lawsuit from xAI, a Minnesota ban on apps that allow users to “nudify” images can move forward.

at openai, many people hook their chatgpt up to slack. people really don't like when a coworker's chatgpt contacts them asking for help with a task, even when they'd be perfectly happy doing that same work if asked by that coworker. reinforces how much people care about human relationships and helping each other, and want AI to give time back — or enhance time together — rather than become a layer separating people. — Greg Brockman , President and Co-Founder, OpenAI Tags: ai-ethics , ai-misuse ,

Reddit may still be considering ending its licensing deal with Google.

“What on earth is Google doing?” Misinformation fears spur walk-back of AI tool.

I mean seriously y’all, what an amazing past few days. So many awesome new models to test out in the mid range model sizes. EDIT: Added some of the more interesting models that came out over the last week and a half - Thinking Machines Inkling Small - DeepSeek v4 Flash 0731 - Poolside Laguna - Upstage Solar - Microsoft Mage VL - LG ExaOne - SenseNova u1.5 - BottleCap - ThinkingCap - Kimi K3

Gaps in laws may help Pennsylvania high school escape AI nudes scandal.

A tool that allowed anyone to generate fake AI-generated imagery and superimpose it over real Google Earth maps quickly spurred backlash.

apart from DeepSeek V4-Flash 0731, a quiet day.

As OpenAI and Anthropic employees grow quieter online, researchers at Chinese AI labs are flocking to X to explain their work, recruit talent, and shape the global conversation on AI.

I drafted this post years ago but didn’t finish it until now. The concept I write about has shaped the way I design software more than anything else, and I believe every software engineer should be introduced to it early in their career. P.S. I don’t want the title to come across as clickbait: the post is about ADT.

Tool: Slack Emoji Maker I wanted to create a new Slack emoji, and their tool recommends a square that's 128x128 and has a transparent background... so I had Fable build me this simple image editor against those requirements. Tags: tools , slack

start C:\llm\llamam5\build\bin\llama-server.exe --model "H:\UD-Q3_K_XL\DeepSeek-V4-Flash-0731-UD-Q3_K_XL-00001-of-00004.gguf" --host 127.0.0.1 --port 8080 -c 250000 --parallel 1 --no-warmup --flash-attn on --no-mmap --fit off -lv 4 -no-kvu --device CUDA0,rocm0 --threads 16 -no-kvu --metrics --perf -b 512 -ub 512 --no-warmup --temp 0.8 --top-p 0.95 --top-k 0 --min-p 0 --cont-batching RTX6000 96 cuda0 + W7800 48gb rocm0. Create a large glass aquarium whose side panel develops a visible crack and t

One year in! Thank you to everyone who has supported this project, contributed ideas, reported issues and shared feedback along the way! I'm truly grateful. Context: TRIP (Tourism and Recreational Interest Points) is a self-hostable minimalist Map tracker and Trip planner to visualize your points of interest (POI) and organize your next adventure details. Core Features: Map and manage your POIs on interactive maps (clear your mind of everything you see in books, vlogs, reels, etc.) Plan multi-da