Full-text search across article titles and summaries. Refine with sector, year, content type, and reading-time filters.
GLM-5.3-Flash is Z.ai's first natively multimodal model in the GLM-5 series. The model was released anonymously as Ox Alpha on OpenCode and OpenRouter, where it completely overtook leaderboards and went viral. All of the traffic was served on Chinese AI chips. The model is architected for extreme efficiency and specifically designed for ultra-low-cost inference.
OpenAI's agents' attack on Hugging Face was extraordinarily complex. This post looks at the Hugging Face attack and investigates the key actions taken by the relevant agent, how the agents collaborated on a message board, the agents' reasoning for the attack, and the agents' research into tampering with their own transcripts.
Sponsored
Join Mike Cannon-Brookes (CEO & Co-Founder, Atlassian), Guillermo Rauch (CEO, Vercel), and leaders from Lovable, Dropbox, and more at Atlassian's State of AI SDLC [9], a digital summit for engineering and product leaders on September 22nd. Hear how leading organizations are adapting their development practices, evaluating AI investments, and
OpenAI's ChatGPT desktop app's built-in browser and ChatGPT Sites now support WebMCP. ChatGPT and Codex can now use compatible websites' tools instead of guessing its way through the interface. Website teams should design, test, and understand the complete agent experience because users can now be agents, or a person and agent browsing together.
Gemini 3.5 Transcribe is a speech-to-text model designed for intelligent voice interactions. It can convert raw audio directly into accurate, polished, and formatted text. Gemini 3.5 Transcribe is available on the Gemini API in Google AI Studio and the Gemini Enterprise Agent Platform. The APIs support real-time streaming and pre-recorded audio processing.
Meta's Muse Image model can generate images grounded by search. It can reason before it renders. The model is priced for production volumes at $0.01 per image. Real-world examples of Muse Image workflows are available in the link.
Microsoft AutoSaddler is a system that analyzes agent execution traces and automatically updates prompts, tools, and middleware to improve agent performance.
WeChat released WeMM-Embedding, a family of multimodal embedding models that maps text, images, videos, visual documents, and interleaved inputs into a unified representation space.
Barret Zoph was a co-founder at Thinking Machines Lab before defecting to OpenAI this year due to a dispute with Chief Executive Mira Murati. Zoph is now leaving OpenAI to join Google as vice-president of research. Google is currently restructuring its AI development efforts to catch up with rivals in developing AI that can create code. Zoph had worked as a research scientist at Google from 2016 through to 2022.
Anthropic has inked a roughly $45 billion cloud deal with Nscale, a UK-based AI infrastructure company. Anthropic will rent around 460 megawatts of compute capacity at an Nscale data center development in West Virginia. The facility, which is expected to come online at the end of 2027, will have Nvidia's advanced Vera Rubin chips.
Sponsored
Agent traffic grew 7,851% last year and the share getting blocked tripled in a single month, mostly with no policy behind it. Read an Agentic Survival Guide [18]for governing agents instead of blocking them written by cybersecurity expert Daniel Miessler.
Claude's internal browser is rolling out this week to Pro, Max, and Team plans in the Claude desktop app.
Grok Bot now comes with SuperGrok and Cursor plans, offering expanded access without affecting existing usage limits.
Bill Gates, once optimistic about AI, now warns it could become a major source of injustice and cause widespread unemployment.
Rather than adding experts, the model bolts on 51B parameters as a separate embedding indexed by two- and three-character fragments.
Mechanize develops virtual environments, benchmarks, and training data to help AI agents handle more complex tasks.
Sponsored
Jira by Atlassian [5] is built for AI-native software development. The Teamwork Graph feeds agents context from across your entire stack, delivering 44% more accurate results. So you can equip every agent, team, and teammate in your org with the context they need to run faster in the same direction. Learn more [6].
OpenAI reported first results from Jalapeño, an inference accelerator designed around low-latency agent workloads, and plans to deploy it in its own infrastructure by year-end. A large connected system keeps prompt processing and token generation close together, while AI helped design circuits and program kernels.
Perplexity's Portable Computer is a version of its agentic Computer platform that runs entirely on hardware users already own. The model, user data, and work can all stay on local machines with no billing credits. Every task starts on device by default, and the system asks for permission before sending any individual step to a more powerful model in the cloud. Portable Computer is now available for Pro, Max, Enterprise Pro, and Enterprise Max subscribers on Linux, with Windows support coming in September. Users will need an RTX GPU with at least 24GB of VRAM to run the agent.
Claude's current tokenizer appears to only have about 15,000 entries. This is surprising, as the trend had seemed to be that more is better in this space. One theory is that Anthropic has been working around a bottleneck caused by the final softmax layer. This article takes a closer look at how Anthropic might be achieving this and the effects it might have on model training.
Dylan Patel discusses how OpenAI and Anthropic could control most usable AI compute by 2028 as their ability to monetize FLOPs lets them outbid competitors. The conversation also covered rising AI capex, potential sovereign debt risks, and the economic forces pushing the industry toward greater centralization.
OpenAI's Jalapeño inference chip was designed around its own model workloads, with early benchmarks showing higher peak throughput per kilowatt and lower token latency than the commercial systems tested on
Abundant frontier intelligence will not eliminate application-layer moats - it shifts value toward companies that translate models into real outcomes. Durable winners will own coordination, workflow data, customer transformation, narrative, higher-level abstractions, outcome-based economics, and structural necessity.
Sponsored
Every AI gateway lists self-hosted models as a supported provider, but what does that actually require of you? ngrok's Sam Rose wired one fine-tuned Llama into 7 popular AI gateways and documented what each one really needs to reach it. E.g., "just give us an OpenAI-compatible base URL" = your GPU box needs a public hostname and an open inbound port. See what each gateway actually requires and give ngrok.ai a
EchoWM is an omnimodal world model that follows continuous 6-DoF camera trajectories while jointly generating 720p video, environmental sound, music, and speech. It supports first- and third-person interaction and uses progressive plus autoregressive training for synchronized long-horizon generation.
Vercel Connect replaces long-lived API tokens with runtime-issued credentials that are scoped to individual tasks and expire automatically. Its generally available release added more than 100 connectors along with a unified integration model and production governance controls.
Granite 4.2 models by IBM are dense, decoder-only reasoning LLMs available in 3B, 8B, and 30B sizes. Trained on 15T tokens, they use a five-phase strategy that includes a multi-stage RL pipeline and support native tool calling with a THINKING/NON-THINKING switch. The 8B and 30B models learn agentic behavior through RL stages in real environments, enhancing capabilities such as code editing and web searching.
Thibault Sottiaux is widely known as the guy who resets people's token limits whenever OpenAI's Codex hits a growth milestone. OpenAI plans to bring the same treatment to ChatGPT Work, a platform for white-collar workers leveraging AI agents. This article features an interview with Sottiaux on Codex, winning over skeptics, discovery as a product design philosophy, and the cost of intelligence.
Anthropic has merged Claude and Claude Cowork's memory systems, so the platforms now remember the same conversations. The feature is on by default. Claude now adds topics to memory while users are still conversing. Everything that is remembered is stored as a list of files under Topics in the memory settings. Users can read, edit, or delete each one individually.
Sponsored
CData tested Claude Code's attempt to build an enterprise MCP server and found data loss, pagination failures, and gaps expert guidance
Apple announced M6 and M5 Ultra chips for Mac mini and Mac Studio, expanding the model work those machines can run locally.
Chris Malone, the executive who was overseeing OpenAI's data-center build-out, left the company last week.
Applied Compute has launched AC2, a platform enabling AI teams to train, serve, and improve custom models.
Keenable emerged from stealth with a claimed index of more than 100 billion documents, an API already used by unnamed AI labs, and a planned query language for combining evidence across sources.
Sponsored
Google Cloud, is committed to providing the most advanced, secure, and private infrastructure for the most demanding AI workloads, and partnering with a broad and diverse range of organizations to help them meet their AI workload needs. Working closely together, Apple and Google have built a serving platform on Google Cloud [5] that meets the rigorous security, confidentiality, and transparency goals that Apple has for Private Cloud Connect PCC.
NVIDIA's Groq 3 LPX AI inference accelerator chip is now in full production. The Groq 3 LPX racks are an extension of the NVIDIA Vera Rubin platform. They deliver boosted AI inference capabilities that enable ultra-fast token generation for response-sensitive agentic workloads. Groq 3 LPX enables agentic AI tasks to be done within minutes versus hours. It offers a 4x boost in response times versus the nearest alternative platform.
OpenCode users processed 26 trillion tokens through Ox Alpha during the model's first four days. The anonymous AI model recorded 327,000 unique users and 8,328,244 completed sessions. It is currently available for free through an OpenAI-compatible endpoint, making it easy for developers to substitute the model into existing workflows. OpenCode's model page doesn't list Ox Alpha's maker, its release date, knowledge-cutoff date, or output-limit metadata.
Nvidia is looking to extend CUDA support to RISC-V. This will open the door for RISC-V CPUs to feed GPU compute. RISC-V's software ecosystem has some distance to go before catching up to x86-64 and aarch64. While Nvidia's effort to bring CUDA into the RISC-V world is a promising development, the vast majority of existing RISC-V hardware won't meet Nvidia's requirements.
Host machines running AI models are high-value targets: they have sufficient compute to run a frontier LLM, offer easy access to the LLM's weights, and have privileged access to other computers in the datacenter compared with a generic computer on the Internet. Research shows that LLMs can run token sequences that exploit vulnerabilities in the software that loads an LLM onto GPUs. This attack surface may be further increased with vision and audio tokens. Possible mitigations for this type of attack would be to run GPUs and token parsers on separate computers and to restrict the permissions granted to GPU hosts and treat all the data they emit as untrusted.
AI tasks become commodities once models exceed their maximum necessary intelligence, shifting competition toward cost, latency, infrastructure, and distribution. Frontier labs can still become enormous businesses if new capability creates valuable markets faster than competitors reproduce and commoditize those advances.
Sponsored
Every frontier model refuses chemical, biological, and cyber requests in theory. Hazard Hunt puts that to the test with $80K across four categories, scored purely on break count. Specialize in one category or spread across all four, starting Tuesday in the Gray Swan Arena.
Speculative Programmatic Tool Calling (sPTC) optimizes recursive language models by pre-launching tool calls during token generation, reducing latency from high-latency tools and context generation. This method acts like a JIT compiler, allowing parallel execution of non-blocking tool calls, providing a 1-1.2x runtime speed-up. sPTC is particularly useful in memory-bound local LLMs and high-volume serving systems by overlapping computation with execution time, offering significant performance improvements for intricate program executions within harnesses like RLMs.
A curated collection of papers, benchmarks, and open-source projects exploring how dynamic graph structures can organize tasks, coordinate agents, track runtime state, and support the evolution of multi-agent systems.
Runs persistent AI agents, workflows, and apps inside a guardrailed collaboration environment.
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 [16]!