— When AWS bills hit $1.7 billion and open-source model parameters reach 2.8 trillion, today's world feels a bit surreal.
A severe bug in AWS's billing system caused some customers to see bills as high as billions of dollars, sparking panic, though the company confirmed actual charges were unaffected. In AI, Moonshot AI released the 2.8-trillion-parameter open-source MoE model Kimi K3, Face Intelligence open-sourced the digital employee platform StaffDeck, and NVIDIA launched the Nemotron 3 Embed model, which topped the RTEB benchmark. The developer community is fiercely debating Claude Code's auto-execution feature, arguing it undermines trust by acting without permission.
Starting July 17, a critical error in the AWS billing console caused some users to see estimated bills skyrocket to millions or even billions of dollars, with reports of anomalous amounts like $1.7 billion, $3 billion, and even $4.2 trillion. Amazon has confirmed this is a bug in the billing calculation subsystem and is working on an urgent fix. Why it matters: This incident exposes the fragility of cloud service billing systems. For enterprise users relying on AWS, the accuracy of billing data directly impacts financial decisions and cost control; any false reports can trigger unnecessary business disruptions and trust crises.
Most users believe this is just a system bug and not a cause for panic, but some say the shock is enough to cause a heart attack and question why AWS's billing system lacks basic anomaly detection.
On July 16, Moonshot AI officially released Kimi K3, a sparse Mixture-of-Experts model with 2.8 trillion parameters, activating 16 out of 896 experts. It natively supports vision and a 1-million-token context window. The model is built on two architectural innovations: Kimi Delta Attention and Attention Residuals, and is claimed to be the world's first open-source 3T-level model. Why it matters: Kimi K3 pushes the parameter scale of open-source models to new heights. Its 1M context and MoE architecture offer direct engineering value for long-document understanding, code generation, and knowledge work scenarios. However, its 2.8T size also raises questions about the feasibility of local deployment.
The community is divided on ultra-large parameter models. Some argue that the 753B parameter model under the MIT license already exceeds personal hardware capacity, casting doubt on actual deployability. Others focus on the architectural innovations' potential to improve inference efficiency.
On July 1, Anthropic introduced a 60-second timeout auto-execution mechanism in Claude Code version 2.1.198: when an Agent requests human input but receives no response within 60 seconds, the system automatically proceeds with its best judgment. Users discovered and criticized this feature for acting without permission, lacking transparency and safety safeguards. Why it matters: This touches on the core conflict between AI Agent autonomy and human oversight. For engineers integrating AI tools into development workflows, unpredictable Agent behavior could lead to unintended modifications to the codebase, directly threatening production environment security.
Opinions are sharply divided. Some see it as improving automation efficiency, while others strongly oppose it, arguing it undermines the trust foundation of human-machine collaboration and lacks sufficient safety guardrails.
Face Intelligence, in collaboration with Northeastern University, Tsinghua University's THUNLP Lab, and OpenBMB, has officially open-sourced StaffDeck. This is a digital employee platform that provides AI Agents with employee IDs, positions, and performance management, aiming to solve issues like chaotic workflows, unsupported answers, and inability to self-iterate in real enterprise processes. Why it matters: StaffDeck upgrades Agents from chatbot logic to enterprise-level digital employees with state machines, knowledge traceability, and self-evolution loops, providing an engineering framework for stable deployment of AI Agents in complex business processes.
On July 15-16, NVIDIA released the Nemotron 3 Embed model collection, including three open-source checkpoints: 8B, 1B BF16, and 1B NVFP4. The 8B version ranks first on the RTEB benchmark with an average NDCG@10 of 78.46. The 1B version was obtained through ModelOpt NAS pruning and distillation from the 8B teacher model, while the NVFP4 version retains over 99% of BF16 precision on Blackwell with 2x throughput improvement. Why it matters: Embedding models are the infrastructure layer for RAG, Agent retrieval, and code search. Breakthroughs in retrieval accuracy and inference efficiency from Nemotron 3 Embed directly impact the quality and cost of downstream AI applications.
無料ダイジェストをメールで。登録すると保存とパーソナライズも。
12 号配信 · 毎日150件超から読む価値ある30件に厳選
VideoChat3 released: fully open-source video MLLM supporting motion, long videos, and streaming interaction for efficient general video understanding.
SEED framework proposes self-evolving online policy distillation, converting full trajectories of Agent reinforcement learning into hindsight skills during training to bridge the supervision gap of sparse rewards.
SearchOS-V1 multi-Agent framework transforms implicit search progress into explicit persistent shared state, solving repetitive loops and wasted search budgets in information retrieval Agents.
LongStraw enables long-context reinforcement learning training over 2M tokens under fixed GPU budgets, designed for Agent long-trajectory scenarios.
Bonsai 27B compresses Qwen3.6-27B to 3.9GB via 1-bit quantization, running locally on iPhone while maintaining about 90% benchmark scores.
Show HN: SSH honeypot real-time monitoring dashboard, visualizing attacker IPs, credentials, commands, and malware delivery behavior.
Capital One open-sources VulnHunter: an Agentic AI-based code security tool using LLMs for vulnerability discovery and verification.
Commenters are divided on the value of VulnHunter; some see it as just a wrapper for common security scans lacking innovation, while others find its methodology and LLM-based verification approach practically meaningful.
Show HN: Rust library 'connections' implements Galois connections, providing composable, formally verified safe conversions for numeric types.
Simon Willison releases LLM cliché highlighter tool, highlighting ten common cliché patterns in AI-generated text.
'The state of open source AI' sparks debate; commenters say the article is poorly designed and reads like AI-generated content, with disagreement on whether open-source AI is truly winning.
The main consensus is that the article is poorly designed and reads like AI-generated content, with disagreement on whether open-source AI is truly winning; some note open-source models are growing fast and approaching closed-source frontiers.
Apple sends legal letters to dozens of OpenAI employees; most comments view it as routine, but some question if it's an attempt to suppress competitors.
Most comments view Apple's legal letters as routine, but some see it as Apple suppressing competitors.
DeepSeek V4 Flash runs 1 million token context locally on RTX 5090 via llama.cpp; community shares configurations and benchmark results.
Codex Dream Skin
"Vibe-Trading: Your Personal Trading Agent"
Skills for Design Engineers.
a great week for open models continues.
Human cognition does not separate understanding and generation. A teacher at a whiteboard speaks and draws together, each modality reshapes the other. In this paper, we bring this coupled loop to artificial systems. Masked Diffusion Models (MDMs) are ideally suited to this task, yet existing samplers either decode text and image interleavedly or independently update them in parallel branches that share only previous-step history, but not the other modality's latest decisions within the same step
Building interactive worlds that respond coherently to player actions has long been a shared goal of computer graphics, games, and artificial intelligence. Recent video generative models provide a data-driven route toward this goal by predicting future observations conditioned on user actions, and are increasingly regarded as potential next-generation game engines. Realizing a genuinely interactive game world, however, requires interaction outcomes that follow rules over evolving game conditions
The world's first open 3T-class model Discussion | Link
OpenAI trained GPT-Red, an internal-only attacker model, using self-play reinforcement learning against a population of defender LLMs. It beat human red-teamers 84% to 13% on a replicated indirect prompt injection arena, found a novel "Fake Chain-of-Thought" attack class, and cut GPT-5.6 Sol's failures 6x on OpenAI's hardest direct injection benchmark. OpenAI concedes it still struggles with multi-turn and image-based attacks. The post OpenAI Details GPT-Red: An Internal Automated Red-Teaming Mo
Tool: Mermaid to ASCII art (mermaid-ascii) After building the Mermaid to ASCII tool based on Grok Build's Rust code I learned that there's an older, more fully-featured Go library called AlexanderGrooff/mermaid-ascii that implements a similar pattern, so I had Claude Fable 5 compile that one to WebAssembly as well so I could compare the two. This one includes support for colors! Test:::good / Test --> Deploy:::warn / Deploy --> Rollback:::bad / classDef good color:#3fb950 / classDef warn color:#
It would open up a completely new business for the Facebook and Instagram maker.
Apple filed a trade secrets lawsuit against OpenAI last Friday, and it’s not messing around. The complaint alleges a pattern of misconduct reaching all the way up to OpenAI’s chief hardware officer and claims more than 400 former Apple employees now work at the company. OpenAI’s response so far has been carefully hedged, and the timing couldn’t be worse with the company reportedly eyeing an IPO […]
Zyphra released ZUNA1.1 on July 16, 2026, under the Apache 2.0 license. The 380M masked diffusion autoencoder reconstructs, denoises, and upsamples scalp-EEG across arbitrary channel layouts. It accepts variable-length inputs from 0.5 to 30 seconds, against ZUNA1's fixed five seconds. Reported NMSE holds or improves while the input range widens. The post Zyphra Releases ZUNA1.1: An Apache 2.0 EEG Foundation Model With Variable-Length Inputs From 0.5 To 30 Seconds appeared first on MarkTechPost .
I've written Mojibake because I don't like the other Unicode libraries for Unicode support. It consists of only two amalgamation files: mojibake.h and mojibake.c. I've added all the most important Unicode algorithms, such as normalization, case conversion, segmentation, bidirectional text, collation, confusable, and others. I regularly test it in these OSes: Linux, macOS, FreeBSD, OpenBSD, NetBSD, and Windows 11. You can find a WASM demo on that site of all the public API functions and the docum
Apple is suing OpenAI. The complaint is readable and intense, as these things often are, though many experts seem to think many of the allegations are just the ways things are done. So what does Apple really want here, and why is it picking such a public fight with OpenAI? On this episode of The […]
Prosecutors accused 21-year-old student Zyaire Wilkins of publishing on Steam several fake video games that contained malware, infecting thousands of victims, and stealing crypto from some of them.
Patreon is strengthening its defenses against AI scraping by working with Cloudflare to block bots that train AI models on creators’ content without permission. The move marks a shift away from relying on websites using robots.txt alone to actively block unauthorized AI training.
Federal authorities have arrested a Florida man suspected of stealing at least $220,000 in crypto through malware-infected Steam games, as reported earlier by local news outlet Local10. In the complaint, officials accuse 21-year-old Zyaire Wilkins and co-conspirators of launching eight malware-embedded games from around May 2024 to February 2026, allowing them to infect about 8,000 […]
Sarah Friar, CFO of OpenAI, introduces a practical AI scorecard to measure ROI through useful work, cost per successful task, dependability, and return on compute.
Some of you might remember that I posted some time ago about the GGML-ported asset production pipeline . A key elelent of that was the TRELLIS.2 port that performs image-to-3D generation. Well, I'm happy to report that after a grueling debugging session (thanks to ) I've managed to fix quite a few bugs and the asset quality is now on par with the reference. This means that top open source 3D generation quality is now available to everyone with a good enough GPU (or for people patient enough to g
Firefox in WebAssembly This is absurdly cool: Puter compiled Firefox to WebAssembly such that the whole browser runs in another browser. Here's my blog, running in Firefox, running in WebAssembly, running in Chrome: They chose Firefox/Gecko because it has strong single-process support. The project used an estimated $25,000 worth of Claude Opus and Fable tokens, but took advantage of a Claude Max subscription plan so cost much less in actual dollars. The demo funnels all traffic over a WebSocket
I just saw that these had dropped. Still very much early days, but nice to see a new locally runnable foundation model, along with a couple of thinking preview versions. Anyone taken a look at this yet? Am intrigued to see how it holds up compared to Qwen 3.6 and Gemma 4 (my current stack)
I'm curious on everyone's real world experience for this model in real codebases / tasks. Does K3 really exceed 5.5 and Opus 4.8 on your coding tasks or not really? Is it benchmaxxed or is just that good of a model? Curious on everyone's use cases and thoughts, please be detailed (what codebase, what lang, around what area and etc, how K3 does vs Opus 4.8 and 5.5)
So I know this is going to incur the wrath of the Qwen cult but after a month of using 27b q8_0 as my primary coding agent in 6+ agent coding workflow with GPT-5.5 as the orchestrator, I got very frustrated with the amount of back and forth I was doing just to always end up in the same place without a resolution to a simple bug. Out of desperation and lack of progress on a medium complexity project, I switched to Gemma4-31b q8 as the main coding agent and 27b as the QA and Reviewer agents, 35b a
Databricks has remade its image into an AI company and has published research on the cost savings of open weight AI models for coding.
The NHTSA recently called for autonomous vehicle companies to improve how cars respond in emergency situations.
TikTok is starting to test an opt-in tool that scans for AI likenesses and lets creators report them to the company, as spotted by social media consultant Matt Navarra. The tool is initially being tested with "some" US creators, TikTok US spokesperson Zachary Kizer tells The Verge. YouTube has been working on a similar tool […]
Is there something I can actually help you with today? — Kimi K3 , after refusing to leak its system prompt Tags: kimi , ai-personality , generative-ai , ai , llms
Reinforcement learning has become a standard post-training recipe for large language models, but dense full-parameter updates create two deployment-relevant bottlenecks: suppressed reasoning performance, often reflected by premature saturation of test-time scaling, and interference when consolidating multiple capabilities through multi-domain training or model merging. We show that the reasoning-effective component of these updates is largely concentrated in the base model's spectral space, moti
This has to be a bug. My usual monthly S3 bill is only $2–3. Has anyone else received a strange billing estimate today?
Hi HN, I am co-founder of Kickstage, a software company specializing in solutions for the electrical industry and lately grid operators. We are hiring engineers from different backgrounds, a lot of them software developers with limited experience in the sectors. Deep domain knowledge is key in our industry however, so we are constantly teaching the basics of power flow analysis, active vs reactive power, transmission line properties etc. With Jupyter notebooks and the Python console only, that's
We present Wan-Streamer v0.3, which reframes our native-streaming interaction model under a single organizing view: a video is a world plus an event stream. The world is the persistent context in which a video unfolds, including the environment, scene, subjects, ambient acoustic conditions, voice characteristics, and other relatively stable conditions. The event stream is everything that changes over time within that world, including scene or environmental changes, subject behavior, speech, and
Today I accidentally became the maintainer of software that was already almost three decades old: reviving a 1990s Java utility to keep the last PAD submitter alive in 2026. I was looking for a way to submit my AI app RiverScript to old desktop software directories and ended up digging through the long-forgotten world of PAD files. There was one problem. The only batch submission tool I could find was a Java application that first appeared in the late 1990s. The last release was in 2017. It was
The City Attorney’s Office sent the tech giants cease-and-desist letters this week telling them to stop profiting from 13 “face-swap” apps that are overwhelmingly used to target women and girls.
Official estimates Google and Apple likely made millions in nudify app fees.
In letters sent to Apple and Google, San Francisco City Attorney David Chiu said that both companies have long been aware that they are hosting apps in violation of state law.
Tomorrow is the 15th anniversary of the first day of the Recurse Center ( ) My cofounders and I did YC all the way back in the Summer of 2010, with the initial idea of building "OkCupid for jobs." That idea quickly fizzled, and we spent the better part of a year pivoting between other ideas that also failed. Finally, we made something that we wanted ourselves: a self-directed programming retreat, where people built fun projects, contributed to open source, and helped each other become better pro
Google and Apple were sent cease-and-desist letters regarding 13 apps on their respective stores.
arXiv:2607.13239v1 Announce Type: new Abstract: Foundation models, including large language models (LLMs) and vision-language models (VLMs), are increasingly used for transportation management center (TMC) tasks such as anomaly detection, incident reporting, and traveler information. Deploying multiple such models across TMC functions raises a portfolio question: which model should serve each function, in which deployment mode, and under what shared hardware budget? We formulate this as the Foun
Multi-reference-to-audio-video (MR2AV) generation aims to generate coherent audio-video content conditioned on multiple references and textual instructions. Existing benchmarks mainly focus on text-driven generation, single-reference subject preservation, or isolated audio-video alignment, leaving the emerging MR2AV setting largely unexplored. Compared with these settings, MR2AV requires models to jointly reason over multiple references while generating synchronized visual and audio content. Mod
I made this experimental art project/game that's an LLM chat assistant, but where you're the AI. I wanted people to get a visceral sense of what it's like to answer the kinds of things that people prompt their chatbots day in and day out. If you're interested, I wrote up some more info on how I made it, including how the "user" prompts are generated with an eye for realism: Hope you enjoy it! I'd love to hear people's takeaways. Comments URL: Points: 48 # Comments: 22
Learning broad world knowledge directly from raw visual data is a fundamental capability of intelligence. We introduce UniVR, the first investigation into simultaneously learning complex reasoning, fine-grained physical dynamics, and long-term planning from pure visual demonstrations. At its core, UniVR features VR-GRPO, a reinforcement learning paradigm with complementary global and step-level rewards. This approach enforces logical coherence and physical consistency throughout the reasoning pr
While recent advances in 3D generation have enabled impressive visual synthesis, existing methods often rely on 2D diffusion supervision without explicit mechanisms for geometric consistency, leading to spatial hallucinations such as duplicated structures and misaligned geometry. These issues become more severe in 4D generation, where maintaining consistency across viewpoints and temporal evolution introduces additional challenges, including jitter, identity flicker, and structural drift. We pre
I've been considering this for a while but it just seems hard for me as I don't really know how docker works. I only use docker compose and many online resources assume I run docker and provide docker compose files. Many projects only really provide docker compose files. I really want to switch to running rootless containers. I had considered docker rootless but figured I might as well just go to podman if I'm gonna do that. The biggest barrier has been the architecture shift and being confused