2026-07-10

Rebranded Codex promises independent workflows that can run "for hours if needed."

Hey HN! We've spent the good part of this past year building an AI tutor that teaches kids ages 4-9 reading, math, ESL and more. Getting an AI tutor to effectively teach a child turns out to be a really hard technical challenge, this took getting the underlying architecture right. Our tutor steers the UX in real-time and makes complex decisions on the fly. Doing both at conversation speed required us to replace the standard tool-use loop. We built our own tutor harness that utilizes a streaming

TLDR: 75B-total / 9B-active MoE is the perfect shape for multi-24GB rigs, and almost nobody ships it. Qwen 27B is a great model and punches way above its weight-class, it is a frequent fallback for me. Nemotron-3-Puzzle-75B-A9B, NVFP4, vLLM 0.22.1 (the new Marlin fallbacks run FP4 on Ampere), pipeline-parallel across 3×3090 capped at 200W each. The 4th card runs a speech sidecar untouched - 3 seats × 256K ctx, fp8 KV — hybrid Mamba keeps the cache tiny - 132 t/s decode across 3 streams (~65 sing

arXiv:2607.07321v1 Announce Type: new Abstract: Tool utilization enables Large Language Model (LLM) agents to interact with the real world and resolve complex tasks. However, existing agent frameworks predominantly rely on static toolsets composed of granular atomic actions (e.g., basic file I/O or single-turn search), which forces agents to reinvent low-level logic for every recurring workflow, leading to increased reasoning overhead and failure rates. In this study, we propose that agents can

I have to admit, a lot of people we're 100% correct to make the suggestion to try this model. I am sorry I ever doubted. The 27B passed every agentic task on a neutral system prompt in 6-9 tool calls. The 75B needed a hand-tuned profile to pass at all and used 2x the turns. For agents, fewer turns beat faster tokens. The two contenders - Nemotron Puzzle-75B-A9B NVFP4, vLLM, PP=2000 across 3 cards, ~65 t/s decode. I made a post about this model. I still think its good for throughput on chatbots a

Maintainer here. OpenMed is an Apache-2.0 toolkit for clinical NLP with one hard rule: patient data never leaves your hardware. No cloud calls, no API keys, works in airplane mode. What shipped in 1.8 this week: OpenMedKit for Android (Kotlin, ONNX Runtime Mobile + ML Kit OCR): read a document, strip every name/MRN/date, entirely on the phone. iOS/Swift and React Native bridges landed too. Browser runtime : de-identification via Transformers.js / ONNX Runtime Web with wasm + WebGPU backends. Ful

Wow... Using this customized vllm provided as a docker, I'm able to run DS V4 Flash on a single RTX 6000 Pro (apparently it also works on a single 5090 - check his readme, but I haven't tried). Apparently this also works with GLM 5.2 (though you need at least two 6000 pros, which is still amazing). Setting 130K context, I needed around 150 GB of RAM to get past the safetensor sharding, but once it is fully loaded in VRAM I am able to fit it all in the GPU (If you have less than this much RAM, cr

I came across this on Hacker News and felt like I needed to share it with the dev community. The main point is simple: a lot of teams reach for extra databases, queues, search engines, caches, and services before they actually need them. This page lays out where Postgres is usually enough, and where you may actually need something else: Postgres is not perfect for everything, but it is good enough for a surprising amount of real-world work. The more I build and maintain systems, the more I appre

This post was originally written in Korean, then polished and translated into English using ChatGPT. I do run llama.cpp locally on a Tesla P40, but as someone who already pays for ChatGPT Pro, I was gradually losing the practical reason to keep running local LLMs like Qwen 3.6 27B or Gemma 4 31B. If I need access to OpenAI models through an API-like workflow, I can usually just use Codex OAuth instead. But then I realized that embedding models and reranker models are not something I can access t

arXiv:2607.07097v1 Announce Type: new Abstract: Safety evaluations of multi-agent LLM systems often compare a direct prompt with a planner-executor pipeline and report the difference as a single "pipeline effect." We argue that this aggregate is difficult to interpret because it conflates three mechanisms: harmful intent may be reframed as plausible operational work, the planner may refuse or transform the request, and the executor may act under delegation prompts implying prior approval. To sep

“token efficiency needs to drop to as much as 20% over the next 12 months, and 90% by the following year”

Hello everyone! Pangolin 1.20 is focused on how people find and reach their resources in the UI. Here's what's new: Pangolin is an open-source, identity-based remote access platform that lets you securely expose your infrastructure to your team. It supports browser based remote access and a remote access VPN in one platform with strong authentication controls. GitHub: Resource Launcher The landing page non-admins see when they sign in now is vastly more capable. Resources are grouped by site or

No real details or timescales yet, but this article has confirmation from Alexandr Wang that Meta are working on an open source variant of Muse Spark. One to keep an eye on.

MOSS-Transcribe-Diarize 0.9B is an end-to-end audio understanding model for long-form multi-speaker transcription, diarization, timestamps, and acoustic event awareness. Given an audio or video file, the model generates a compact speaker-aware transcript in one pass, including timestamps and anonymous speaker labels such as [S01] , [S02] , and beyond. Introduction MOSS-Transcribe-Diarize 0.9B turns real-world long-form audio into structured, speaker-aware transcripts in one pass. Instead of stit

The main conclusions from analysis were: The Pareto frontier for coding tasks (i.e. best quality for a given cost) includes models from OpenAI, Anthropic, and open source. This means today, only a mix of tools can provide frontier performance. Open models, and GLM 5.2 in particular, are now able to handle even the highest level of task difficulty. The token price of a model is a poor indicator of actual costs incurred on end-to-end tasks. Larger models can be far more token efficient and have lo

The feud between NightmareEclipse and Microsoft shows no signs of resolving soon.

Broadcom accuses Allstate of dodging VMware audits.

The agent-native way to ship software Discussion | Link

Windows 11 updates could soon include fixes for more security issues at once. Microsoft said in a blog post on Thursday that it's now using AI to "identify potential issues earlier," which means "customers will see a higher volume of security updates included in each security release." Hackers, even amateurs, have increasingly been using AI […]

Release: llm 0.31.1 Fix for a bug with OpenAI Chat Completion endpoints where a tool call with empty arguments could result in a JSON error from some providers. #1521 This bug came up when I was testing llm-meta-ai . Tags: llm

Hi Hacker News, I’m Yahia. I built Context.dev ( ) to make it really easy to integrate web data into your products and agents. Here’s a demo video: Since it’s an API, here are the docs: . You can send us a URL and get back clean Markdown, rendered HTML, screenshots, extracted images, etc.. You can also send us a domain and get company or brand context: name, description, logos, colors, fonts, social links, screenshots, style information, and related metadata. For more custom use cases, you can s

arXiv:2607.07229v1 Announce Type: new Abstract: Prior work has shown that chain-of-thought (CoT) reasoning is often unfaithful: a model's stated reasoning does not reliably reflect the process that produced its output. Detecting unfaithfulness, though, requires controlled experimental interventions, which cannot be applied to evaluation transcripts after the fact. We turn instead to a more tractable question that has received less attention: whether the stated reasoning is logically consistent w

Hey Guys, As promised here are the results from running MiniMax M2.7 REAP 139B Q3_K_L on llama-bench on 6x MI50's. Memory Load: Hardware: Asus X99-E-WS ( Modded BIOS to support a large number GPU's ) Intel(R) Xeon(R) CPU E5-2680 v4 @ 2.40GHz 128GB DDR4 RAM SSD 6x MI50's 96GB VRAM (Gen3 x8,x8,x8,x8,x8,x8) Results GPU Setup Model Test Result 6x MI50 / Pro VII 16GB MiniMax M2.7 REAP 139B Q3_K_L pp512 139.27 t/s 6x MI50 / Pro VII 16GB MiniMax M2.7 REAP 139B Q3_K_L tg128 24.87 t/s 6x MI50 / Pro VII 1

I made this simple 3D Geometry Wars-style game using my coding agent, Jarvis Code, with GLM 5.2. You can play it here: I was honestly surprised by the result. Most of the game came together in the first iteration, and I only needed about four small follow-up tweaks afterward. I'd be curious to hear what you think about the gameplay, the code quality, or GLM 5.2's coding ability.

arXiv:2607.06820v1 Announce Type: new Abstract: Recent advances in AI for Mathematics have focused largely on autoformalization and theorem proving, leaving the role of Computer Algebra Systems (CAS) in agentic LLM workflows underexplored. We propose a ReAct-style agentic setup that combines LLM reasoning with verifiable feedback from SageMath, together with Context7 for the up-to-date documentation. We evaluate this agentic setup across frontier models for solving research-level mathematical pr

[...] Work on web and mobile runs in the cloud. Work in the desktop app can also use local files and desktop apps with your permission. At launch, cloud Work conversations do not appear in desktop Work; desktop Work threads and local files remain on that computer. — OpenAI , trying (unsuccessfully) to clarify ChatGPT Work Tags: openai , chatgpt , ai

OpenAI is sunsetting its AI-powered browser after less than a year. But it's moving some agentic browsing features to its desktop app and a Chrome extension.

OpenAI is already shutting down ChatGPT Atlas, its browser that could do tasks for you on your behalf, less than a year after launching it. Atlas was announced in October, but as part of its wave of news about ChatGPT Work today, the company confirmed that it will be "sunsetting" Atlas and is targeting an […]

Dictate, rewrite, translate, and an agent in a single device Discussion | Link

arXiv:2607.07189v1 Announce Type: new Abstract: Vision-language models (VLMs) and agentic AI have shown strong performance on semantic visual tasks, but it remains unclear whether they can handle the physics and inverse problems that underlie computational imaging. We present ImagingBench, a benchmark of 20 computational imaging tasks spanning five categories: ray and wave optics, image signal processing, inverse reconstruction, computational sensing, and calibration. ImagingBench evaluates thre

AI cheating leads to "a failed society," professor says.

I finally decided to set up a dedicated home lab server a while ago. I priced out built-from-scratch x86 ITX configurations and low-power NAS builds, and they were easily running $500–$800 just for CPU/RAM/chassis. I hesitated. I wanted something cheap, extremely low-power, and compact. So I did what any frugal self-hoster does: I looked at old hardware I already owned. I had a 2017 Dell XPS 13 laptop (i5, 8GB RAM) sitting in my drawer. I had tried selling it locally, but since the battery was c