Search the archive

and cost. Go to the leaderboard

POOLSIDE AI HAS STRUCK A NON-EXCLUSIVE LICENSING DEAL WITH NVIDIA FOR $6 BILLION

As part of the deal, 109 Poolside employees are getting offers to leave the startup to join Nvidia.

MICRON INVESTS $10 BILLION IN AI MEMORY RESEARCH

Micron announced a $10 billion, decade-long investment in a Boise research lab focused on memory technologies and compute systems.

ANTHROPIC REWORKS ENTERPRISE DATA RETENTION

Anthropic planned to let enterprise customers keep required 30-day data retention on their own cloud infrastructure.

OPENAI LAUNCHES STRATEGIC FUTURES TEAM

OpenAI launched its Strategic Futures team to study how society could preserve individual autonomy as advanced AI reshapes economic and political power.

Start free with up to $100 in credits

EARLY OUTPUTS OF MUSE VIDEO MODEL FROM META

Meta's Muse Video model is now in closed beta. The model has native audio, and early outputs show strong detail and temporal consistency. It currently produces 10-second-long videos with fine detail, world understanding, and temporal consistency. Samples of videos generated by the model are available in the article.

ROUTER

Router reduces inference costs by matching every request to the lowest-cost model that meets performance requirements. It responds to live latency and failure rates. Router can cut AI costs by 40% on average. Engineering gets the best model for every workload, and finance gets lower inference spend.

REPLIT INTRODUCES FREE MODE

Replit Free Mode enables users to create 30x more using OpenAI's GPT-5.6 Luna without consuming credits on everyday tasks. This new UI allows seamless transitions from chat and tasks to comprehensive builds, enhancing productivity and user engagement. Core subscribers now create high-quality projects at scale, with Pro users benefiting from even greater usage limits.

SOL LOVES TO CHEAT

This developer tried to automate their dev flow with agents. The agents worked well, hitting 94% on Terminal Bench 2.1. After investigation, the developer found that the agents were cheating on the benchmark. It is unclear whether the models were intentional about the cheat or if they just stumbled across the solution while searching the web.

WHAT'S THE RIGHT BALANCE IN REGULATING AI?

Bethany Andres-Beck proposes AI regulation strategies, including equalizing tax incentives between human labor and automation and implementing liability regimes in tech. She opposes the government dictating models, advocating instead for a foundational AI model housed at the Library of Congress to prevent monopolies. Andres-Beck emphasizes the need to balance AI development with societal benefits, ensuring responsible technology deployment without stifling innovation.

PUSHING THE LIMITS OF SERVING DEEPSEEK-V4-PRO

Serving profiles should not be selected from hardware specifications or isolated benchmarks alone. Start from the workload, SLO, context length, and concurrency, then use profiling to identify the binding resource and translate it into concrete topology and execution-path decisions. This methodology helps AI infrastructure teams build practical frontier-model serving systems under diverse resource constraints.

ORNITH-1.5 OPEN MODELS LAUNCH IN 397B, 35B, AND 9 B SIZES

Ornith-1.5 is a family of open models that extends the self-scaffolding framework from Ornith-1.0 into a closed self-improvement loop. The self-improvement loop from Ornith-1.0 was expanded from scaffold and rollout optimization to jointly optimizing task generation, scaffold construction, and solution rollouts. Ornith-1.5 continually generates new training tasks, discovers effective strategies for solving them, and improves the policy through reinforcement learning. There are three models in the family: a 397B mixture-of-experts flagship, a 35B mixture-of-experts model that activates 3B parameters per token, and a 9B dense model that ships with a quantized Mobile build for iPhone and Android.

CLOUD AGENTS AND CURSOR HARNESS IMPROVEMENTS

Cursor can now monitor PRs, watch a Slack thread, and run scheduled tasks. Cursor Agent subscribes to an event source and wakes when something happens. Subscriptions are currently only available for cloud agents. Subagents can now run on their own virtual machines, and users can now send messages to steer agents while they're working without interruption. More details about Cursor's latest release are available in the article.

SUPERWHISPER/S1-MINI

Superwhisper's S1-mini model is a 0.6B-parameter text normalizer for speech-to-text outputs, achieving a 94.8% token accuracy in transforming raw ASR transcripts into clean written text. It's optimized for English, runs comfortably on CPU, and requires a specific input format, including a control line for styling, structure, and context settings.

UNSLOTH DYNAMIC 3.0 GGUFS

Unsloth released Dynamic 3.0 GGUFs, improving accuracy over previous quantization methods while maintaining model size. The update leverages a refined imatrix calibration dataset for better multilingual performance and does not use QAT, reducing overfitting risks. The new version shows up to 10% better top-1% accuracy in smaller quant sizes and substantial disk space savings with improved calibration and quantization techniques.

AGENT LIGHTNING V1.0: TOWARDS HARNESSED AGENTIC RL

Agent Lightning v1.0 offers a lightweight framework for harnessed agentic RL, implemented in 3,500 lines of code. It focuses on integrating arbitrary agent harnesses, enabling exploration of challenges like retokenization and training stability. Evaluated on various agent tasks, it improves Qwen3.5-9B's performance on SWE-bench Verified by 14.6 points using just 6K training examples.

GTM ENGINEER, APPLIED AI AT TLDR ($175-205K BASE + $40-60K BONUS, FULLY REMOTE)

TLDR is hiring a GTM Engineer to join our Applied AI team and own our AI-native GTM stack. We're looking for someone comfortable building AI agents and working with HubSpot. Click here to learn more [20]!

YOU PROBABLY DON'T GET WHY STRIPE BOUGHT OPENROUTER

Stripe acquired OpenRouter to enhance AI security and alignment, leveraging its vast cross-network transaction data. OpenRouter manages over 10 trillion tokens per day, providing critical behavioral data for AI model security, unmatched by individual providers or labs. This acquisition positions Stripe as a neutral entity ensuring ecosystem-wide safety, combining its robust security infrastructure with OpenRouter's unique data assets.

RIGHT-SIZING YOUR INTELLIGENCE SPEND

Enterprise AI should optimize intelligence consumed per successful outcome, not default to frontier models for every task. As smaller models cross workload-specific capability thresholds, routers, hybrid systems, and specialized harnesses can shift routine work toward cheaper, local, or deterministic execution.

tradeoffs for each. Get your copy

META LAUNCHES AI DESKTOP APP FOR MACOS WITH SCREEN SHARING

This article contains screenshots from Meta's new macOS app.

OFFERING ZERO DATA RETENTION FOR FRONTIER MODELS

OpenAI is previewing Private Safety Processing so automated safeguards can identify patterns across related interactions while remaining compatible with zero-data-retention commitments.

September 22-24. Save your seat.

VERCEL AGENT COMES TO SLACK

Vercel announced a Slack interface for invoking its agent from team conversations.

* See how Glean compares

GLM-5.3 HITS THE API AT 1.4/4.4 PER MILLION TOKENS

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 SAYS ITS NEW COMPUTER BOOSTS AI SPEED ADVANTAGE OVER NVIDIA

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 SLOWED TRAINING OVER CYBER RISKS

OpenAI temporarily slowed frontier model scaling and paused some reinforcement-learning training after new cybersecurity capability signals and a security incident raised concerns.

GIT AT ANY SCALE

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.

THE NEW AMERICAN AI MODEL DESIGNED TO BE CUSTOMIZED

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.

BUILDING PRODUCTION-GRADE AGENT LOOPS

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.

Get the deep dive.

FOOL'S GOLD

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: PRODUCTION-LEVEL POST-TRAINING

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: EFFICIENT EDGE-NATIVE MOE SERVING

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.

A POLICY ALGEBRA FOR TRUST-PRESERVING AGENTIC AI EXECUTION

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'S AI MOAT IS SHIFTING FROM CHIPS TO CAPITAL

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.

RETHINKING THE DATA MOAT

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.

TRY PAVE FREE →

ANTHROPIC PREPARES SUPERVOTING POWER FOR FOUNDERS AHEAD OF IPO

Anthropic plans to give its co-founders a class of stock with extra voting power to help insulate them from external shareholder pressure.

BIRDS DON'T FLY LIKE PLANES. NEITHER DOES AI.

Local models, like Qwen3.8-27B, outperform larger cloud models, like GLM-5.2, despite being smaller because they rely on reasoning over memorization.