Parse notes
Sponsored
AI work gets expensive when models have to search across fragmented systems and burn tokens rebuilding context. Glean uses enterprise context, intelligent routing, and efficient retrieval to get more done with less. In a benchmark against Claude Cowork, Glean averaged $0.45 per task versus $1.84—a 4x cost advantage. See how a context-first approach can improve AI performance while keeping spend under control.
The API for GPM-5.3 is now available. Z.ai plans to make the model's weights openly available, but it has yet to set a release date. API pricing remains unchanged from GLM-5.2, so developers gain substantially stronger coding and long-horizon agent performance without paying more.
Cerebras' new computer, the CS-4, is multiple times faster than its predecessor, which Cerebras already claims is more responsive than systems built with Nvidia's processors. The machine is currently being sampled by a small group of customers. It will be more widely available in the third quarter. The product is a significant step in the company's bid to challenge Nvidia in the market for AI data center hardware.
OpenAI temporarily slowed frontier model scaling and paused some reinforcement-learning training after new cybersecurity capability signals and a security incident raised concerns.
Cursor explained why Git's packfile-centric, distributed architecture becomes difficult to operate as a centralized service at large scale, outlining approaches that distribute the filesystem, packfiles, or Git itself.
Thinking Machines released a model called Inkling in July. Inkling is the company's first model trained from scratch. Its weights are available on Hugging Face under an Apache 2.0 license. This article walks through the various choices Thinking Machines made while building Inkling. It covers the model architecture, position encoding, how images and audio enter the model without a separately pretrained encoder in front of them, the thinking effort setting, and more.
Liquid AI used autonomous coding agents to build toktoktok, a production BPE tokenizer trainer requiring both ML and systems expertise. The experiment highlighted concrete specifications, multi-domain tasks, and external verification as key ingredients for reliable long-running agent workflows.
Sponsored
Stale, badly chunked, or ungoverned data produces confidently wrong answers. This AWS workshop [12] covers ingestion, chunking, incremental hash-based refresh, source-to-response lineage, and access control at retrieval, with Amazon Bedrock Knowledge Bases and AWS Glue. Sept 1 | 60 min | Technical demos
Safety alignment in open-weight language models is trivially removable. Abliteration projects a refusal-mediating direction out of the weights in minutes, and no release-time defense can prevent it durably. Decoy hardening is a technique that concedes the refusal strip and poisons its payoff. Once refusal is stripped, most answers to hazardous operational requests are confident, fluent decoys whose critical elements are falsified. The defense is inert against in-context jailbreaks by design and applies to first-release models only.
Miles v0.1 is an open system for improving AI agents through reinforcement learning after their initial training. For example, a team training a coding agent could let many copies attempt tasks inside isolated environments, score which attempts worked, feed those results back into training, and distribute the updated model to the workers without stopping the whole pipeline. Miles packages the rollout, sandboxing, asynchronous training, replay, model-update, and multi-hardware pieces needed to run that loop at scale.
FreeToken continuously remaps experts, model state, CPU/GPU work, and agent state reuse to the bandwidth and memory actually available on a personal machine. The authors report support for more than 20 MoE models, from 35B models on an 8GB laptop GPU to a 753B GLM model on one workstation GPU.
The paper proposes a system for enforcing an AI agent's permissions throughout an entire task, not just when the task begins. For example, a refund agent could read the correct customer's record, calculate a refund, use the payment tool only below its spending limit, request human approval when necessary, and leave an audit trail, all under one combined set of rules. The authors report that their runtime stopped or corrected 94.8% of rule-breaking actions while still completing 86.9% of legitimate tasks.
Nvidia is leveraging its capital strength to maintain AI leadership, moving beyond chips as competitors like AMD and Google gain ground. Through a $105 billion investment in an Ohio data center supporting OpenAI, Nvidia aims to sustain growth amid rising competition by expanding its AI infrastructure. By forming partnerships with Wall Street financiers for GPU investments, Nvidia seeks to widen its market influence and diversify revenue streams.
Dwarkesh Patel discussed AI research automation with Ryan Greenblatt, who emphasized algorithmic progress as a major driver of AI advancements, suggesting improvements even without human expert data. Shuchao Bi, co-founder of YouTube Shorts, echoed similar sentiments by highlighting the need for refined data distributions to enhance AI models. Both experts argue that AI research could accelerate with better data processes, challenging traditional notions about human data's significance in model improvements.
Sponsored
Stop being the single point of failure. Describe the process, and Pave builds it into real software.
Anthropic plans to give its co-founders a class of stock with extra voting power to help insulate them from external shareholder pressure.
Local models, like Qwen3.8-27B, outperform larger cloud models, like GLM-5.2, despite being smaller because they rely on reasoning over memorization.
Harvey II enhances legal work efficiency by enabling AI agents to start with context, memory, and preferences from prior tasks, reducing repetitive setup.
Vercel launched a two-week security challenge offering up to $1 million to researchers who could escape its Firecracker-based Sandbox.
Etched shipped its first hardware rack to Jane Street and raised $700M at a $21B valuation.
Warp Factories is an out-of-the-box system for streamlining AI software development using a software factory model.