OpenAI's Jalapeño inference accelerator moves toward deployment (9 minute read)
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 partners with Nvidia to launch Portable Computer, a fully local AI agent with zero token costs (13 minute read)
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.
|
|
Vocab Break (4 minute read)
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.
|
OpenAI and Anthropic Could Dominate Global AI Compute (48 minute read)
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.
|
Moats in the age of floods (14 minute read)
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.
|
|
Open Omnimodal World Models (GitHub Repo)
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.
|
Short-Lived Credentials for AI Agents (12 minute read)
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 LLMs: How They're Built (20 minute read)
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.
|
|
Anthropic merges Claude chat and Cowork memory, on by default (5 minute read)
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.
|
|
|
|
|
