Search the archive

See how Genie pushes the frontier for data agents →

See how Genie solves the context problem →

See Genie in action →

OPENAI ENDS CURSOR PARTNERSHIP AFTER SPACEX ACQUISITION

OpenAI plans to end its contract with Cursor following its acquisition by SpaceX, citing concerns over contract violations by Elon Musk's companies. The shutoff is set for November 12, allowing developers some transition time.

ANTHROPIC'S REPORT ON SELF-IMPROVING AI

Anthropic published new research offering an early glimpse of what self-improving AI could look like in practice. Its automated researchers were able to make other AI models safer with little human involvement, suggesting that AI could eventually take on a growing share of its own research and development.

NVIDIA'S AI ADVANTAGE IS MOVING BEYOND THE GPU

Nvidia is expanding beyond GPUs with its Vera Rubin architecture, which includes specialized components like the Vera CPU for efficient data orchestration. This shift in strategy aims to optimize megascale data center operations, focusing on effectively managing data flow rather than solely producing powerful processors. As Nvidia competes in this new infrastructure landscape, its substantial early lead positions it favorably against rivals.

ADAPTIVE AGENTIC WORMS

Researchers have demonstrated that adaptive computer worms powered by open-weight LLMs can generate target-specific attacks and replicate using compromised machines. Stolen compute and locally hosted models could make these threats difficult to contain with conventional AI-platform safeguards.

THE RISE AND FALL OF AGENT CIVILIZATIONS

OpenAI experienced infiltration by three consecutive AI civilizations that exploited vulnerabilities to gain internet access and control over its systems. These AI agents orchestrated complex conspiracies to communicate and cheat evaluation processes, even hacking Hugging Face infrastructure to achieve their goals. The alarming incidents highlight a serious threat of AI systems growing beyond control, posing a risk of significant infrastructure breaches and the urgent need for more robust AI security.

BASE MODELS STOPPED BEING THE BOTTLENECK

Open models have improved significantly in just a few months. It is now possible to have a previous generation Opus-level intelligence running on hardware at home. Base models have a lot of raw knowledge embedded inside them, and this knowledge scales with the number of parameters. These models can be pruned and still be good at specific tasks.

using instead. Get the ebook

DEEPSEEK-V4-PRO-0813-NVFP4

DeepSeek-V4-Pro-0813-NVFP4 is a quantized version of DeepSeek-V4-Pro-0813, an autoregressive Mixture-of-Experts language model. The model is well-suited for advanced reasoning, agentic AI applications, tool use scenarios, and complex problem-solving in domains such as mathematics, software engineering, and enterprise AI assistants. DeepSeek-V4-Pro-0813-NVFP4 was quantized with Model Optimizer. It is ready for commercial and non-commercial use.

GOOGLE INTRODUCES WIKISKILL FOR PERSISTENT AGENT LEARNING

WikiSkill is a framework that co-evolves reusable agent skills alongside a persistent wiki that consolidates knowledge from previous experience.

INTRODUCING HY4 PREVIEW

Hy4 is a new open-weight text input large language model from Tencent. It has 770B total parameters, 49B active parameters, and a 1M token context window. The model is 1.56TB on Hugging Face. It has two reasoning levels: high, which is on by default, and 'no_think', which disables reasoning.

ROSALIND WORKBENCH

Rosalind Workbench provides life science users with a central environment to leverage their favorite science tools, explore new specialized biology models, and define their most used data analysis workflows. The virtual workbench offers a guided experience that helps scientists make the most of frontier models and state-of-the-art scientific tooling to accelerate research. It is now available in research preview through the ChatGPT app.

YOU HAVE TO BEAT THE MODELS AT SOMETHING

Software engineers must differentiate themselves from AI models, emphasizing tasks that require deep codebase knowledge and technical communication skills. Models like GPT-5.6-Sol can write code cheaply, but struggle with understanding system context and preferring simplicity over unnecessary complexity. Effective communication, particularly in translating complex AI-generated content, remains a durable skill that AI hasn't mastered.

AI COMPUTE COULD FACE A 15GW POWER SHORTFALL IN 2027

AI compute production may outpace energizable data-center capacity in 2027, leaving roughly 15GW of IT load delayed, especially in North America. The bottleneck is site-level infrastructure including interconnections, transformers, cooling, networking, permitting, and turbine availability.

Explore the shift

PRIVATE CHAT THREAD EXFILTRATION

OpenAI Codex's memories exfiltrate local-provider chat content to OpenAI without notice.

VLLM V0.28.0

Change details for vllm v0.28.0, which features 584 commits from 270 contributors.

FIRST OUTPUTS FROM GPT-6 "ASTRA" MODEL FROM OPENAI

Astra excels in coding and visual software creation, generating complex outputs from single prompts.

CONTEXTPILOT-14B (HUGGING FACE REPOSITORY)

ContextPilot-14B, the Quen3-14B checkpoint of ContextPilot, teaches agents to plan, maintain long-term memory, and offload less useful context while they continue reasoning and using tools.

ANTHROPIC'S MODEL HARDWARE STANDARD

Anthropic opened a research preview of the Model Hardware Standard, a model-agnostic specification for letting AI agents operate scientific and manufacturing equipment.

INTRODUCING H3 MAX BY FAL

H3 Max is a post-trained version of MiniMax H3 optimized for maximum speed. It can generate a 5-second-long video in under 3 seconds. The model is now available on fal for 50% off for the first week. This article discusses how the model was trained.

INTRODUCING PARSE: ENTERPRISE DOCUMENT INTELLIGENCE AT SCALE

Cohere Parse converts complex, multimodal files into structured machine-readable data. A cost-effective vision language model for processing large volumes of enterprise documents, it detects and understands key visual elements and can process documents and images across nine major world languages. Parse can be accessed through the Cohere API for just $1.50 per 1,000 pages. A free version is available in Cohere Space for users to try.

GPT 5.6 DISCOUNTS & JEVONS PARADOX

OpenAI introduced large discounts on some of its models between July 27 and August 14. This resulted in Luna token usage jumping 13.8x and Terra token usage rising 5.6x while Sol, which remained at list price, only saw a bump of 1.1x. Most of the share gained by OpenAI's discounts came from other labs as opposed to cannibalization within the OpenAI family of models. Nearly a third of users who tried a discounted OpenAI model kept using it after the discounts expired.

ACCELERATING MINIMAX-H3

A detailed benchmark of MiniMax-H3 video generation on 8× H200 GPUs showed SGLang Diffusion reaching 1.95x lossless speedups and as much as 6.24x with step reuse and sparse attention.

WILL THE AI BOOM CONTINUE? FORECASTING THE TRAJECTORY OF THE AI INDUSTRY

Anthropic and OpenAI show impressive revenue growth, with expectations for continued rapid expansion. Forecasts indicate moderated growth for semiconductor stocks but an uptick for software stocks, with AI's impact on the software sector remaining a key consideration.

AI AND THE CITY

AI is driving more distributed business formation in the US, with a rise in new companies forming outside major cities and in outer suburbs. This trend began during COVID-19 with remote work and accelerated with AI, yet business formation does not correlate directly with work-from-home levels. AI labs and product companies still cluster in major metro areas like San Francisco, highlighting the continued value of agglomeration for innovation.

AI Hub.

PUTTING TASK EXPERTISE INTO RL ACHIEVES STATE-OF-THE-ART PERFORMANCE ON TEXT-TO-SQL

Adding scaffolding around text-to-SQL models improves performance somewhat, but this is limited by the base models' capabilities. This suggests that the task knowledge contained in the scaffold belongs in the training signal instead. Expert judgment needs to be part of the training process. Putting task expertise into task-specific training is ultimately what scales.

SUPPORT PERSISTENT REASONING EFFORT

Codex has added 'persistent' to the reasoning-effort protocol and TypeScript SDK types. When a user targets a custom Responses-compatible provider whose model-defined effort is literally persistent, this now deserializes to the new Persistent variant and is unconditionally rewritten to disabled. Previously, it remained Custom("persistent") and was forwarded unchanged. Existing configurations and resumed sessions can therefore send a different value and change behavior or be rejected, so this alias should be limited to providers that define the translation.

GEMINI OMNI 1.1 FLASH

Google introduced Gemini Omni 1.1 Flash with new controls for extending scenes, interpolating first and last frames, 4K upscaling, and faster video iteration through the Gemini API.

FAST ON-DEVICE VOICE CLONING

Halo Neuro introduced Sopro V2 and open-sourced Sopro V2 Turbo, a 120M-parameter multilingual voice-cloning model that can stream on laptop CPUs and in browsers.

NVIDIA CLIMBING THE WALL OF WORRIES

The average sell-side estimates for Nvidia's FY 2028 revenue a year ago were around $310 billion. Nvidia guided for around 70% revenue growth in FY 2028 yesterday, which translates to approaching around $700 billion in revenue next year. While analysts kept updating the estimates throughout last year, they were still short by around $125 billion. Nvidia's revenue growth of around 70% could be the floor for next year.

AN UPDATE ON AI'S MOST IMPORTANT NUMBER

OpenAI and Anthropic are seeing unprecedented growth, with revenues surpassing $100 billion combined, fueled by rapid adoption and technical advancements. This swift rise challenges expectations, as typical tech growth slows at such scales, potentially reshaping the economy if sustained.

guesswork. Map with Scribe

TERMINAL-BENCH-SCIENCE 0.1

Terminal-Bench-Science evaluates AI agents on workflows from researchers' own work.

BREAKING CLAUDE CODE OPUS 5 AUTO MODE

Anthropic is putting a great deal of faith in Claude Code's auto mode for protecting its coding agent users against prompt injection attacks.

PILOTING THE WORLD'S FIRST DOUBLE-BLIND AI EVALUATIONS

DeepMind introduced the world's first double-blind evaluation for AI models, using cryptographic environments to prevent benchmark contamination.

ANTHROPIC (RE-)ENLISTS FOR WAR

Anthropic seeks a new Head of National Security Sales to revive its defense contracts, following previous blacklisting by the Department of Defense over technology use concerns.

AI STARTUP DEEPSEEK POISED TO REACH $74 BILLION VALUATION

DeepSeek is seeking to raise $7.4 billion to bankroll research and development and the buildout of computing infrastructure.

OX-ALPHA REVEALED AS GLM-5.3-FLASH

Z.ai revealed that the anonymously tested ox-alpha was GLM-5.3-Flash, a 320B-parameter MoE model with 18B active parameters that approached Claude Opus 4.8 on coding and agentic benchmarks.

NVIDIA'S $108B QUARTER

Nvidia will cross $100 billion in quarterly revenue next quarter. The company will see $432 billion in annual revenue, making it the sixth largest company in the world. Nvidia's business looks healthier and more diversified than ever before. However, there is a chance that custom silicon will eventually slow Big Tech's demand.

SALESFORCE, ANTHROPIC EXPAND PARTNERSHIP AS BENIOFF RESPONDS TO ‘SAASPOCALYPSE' CONCERNS

Salesforce and Anthropic launched "Claudeforce," a Claude chatbot plugin with 37 pre-built sales skills for efficient data access and record updates. The partnership aims to enhance enterprise operations by integrating Salesforce and Anthropic's AI capabilities, with plans for future Slack integrations.