Nvidiallama.cpp releases b eleven seven one nine with CUDA thirteen support
MetaOnlyFans Creators Are Pissed That Meta Lets Muse Talk DirtyvLLM releasesvLLM zero point three two adds structured endpoints, FP4 KV cache expansion
OpenAIOpenAI Python SDK three point three five point one fixes Realtime byte reads
OllamaOllama zero point four zero seven adds Responses effort, PPLX Decider support
AnthropicReplit switches default AI provider from Anthropic to OpenAIWIREDWhy AI Isn’t Ready to Tackle Our Convoluted GenomesVercelVercel AI Gateway adds session affinity view to request logsWIREDData Centers Have a New PR Plan: Your Kid’s Field TripMetaMeta removes clause allowing scraping of business account domainsNvidiaNVIDIA RTX Spark laptops launch starting at two thousand three ninety nine dollarsweb-edit (previously) is my tiny web app for browsing and editing files in a folder. I mainly built it for editing files on different machines via Tailscale.
Daily briefing on generative AI
Updated · 12 stories · 9 sources · 3 min read
vLLM zero point three two brings structured endpoints and FP4 KV cache to more GPUs, while Replit switches from Anthropic to OpenAI as its default model provider.
Step 1 · 30 seconds
Step 2 · 4 min
Voiced by ElevenLabs from today's briefing.
Good morning. This is dailybrief, and here are today's top stories in generative AI, for Sunday, October eighteenth. vLLM has released version zero point three two with eight hundred fourteen commits from three hundred sixty three contributors. The update adds a new POST /v1/systemone endpoint that answers choice, noul and score questions over structured state, according to the vLLM project. The release expands low-precision KV cache support, with NVFP4 KV cache now running on SM8x, SM12x and SM90 GPUs. For GLM-5.x models, FP8 KV cache on SM100 cuts KV memory per rank in half while delivering two point three to five point five percent more throughput, though with a seven to eleven percent time-to-first-token cost. The release drops Python three point ten support and requires transformers version five point sixteen point one or higher. Turning to model providers, Replit has made OpenAI its default AI provider over Anthropic, CEO Amjad Masad said Thursday at TechCrunch Disrupt. Masad cited cost, partnership and performance as reasons for the change, saying ChatGPT now performs better on the company's internal benchmarks. Replit had been a top-five customer of Anthropic until recently, according to Masad. Anthropic models remain available to Replit customers, but OpenAI is now the default, Masad clarified after the event. The shift could significantly reduce Replit's spend with Anthropic since most customers stick with defaults. In developer tools, Vercel has added a Sessions view to AI Gateway logs that groups requests into sessions and shows how session affinity routed each one. Each session displays the coding agent or app that ran it, along with cost, models, providers and cache hit rates across every turn, according to Vercel. Session affinity routes each turn to the provider that served the previous one, so the provider already has most of the prompt in cache. Developers can group requests by sending the same session ID in provider options or headers. Meanwhile, llama.cpp has released version b eleven seven one nine with CUDA twelve point eight and thirteen point four library support. The release includes builds for Ubuntu and Windows on x64 and arm64, along with ROCm ten point zero support for Ubuntu and Windows. Ollama released version zero point four zero seven, adding support for OpenAI Responses effort on non-thinking models. The update also adds PPLX Decider v one point two support to the MLX backend and lets llama-server infer the draft speculative type automatically. Also today, NVIDIA launched RTX Spark laptops with Blackwell RTX graphics and up to one hundred twenty eight gigabytes of unified memory. Laptops from ASUS, Dell, HP, Lenovo, Microsoft and MSI are available now starting at two thousand three hundred ninety nine dollars, according to NVIDIA. The company celebrated the launch with events in Berlin, Beijing, Istanbul, Seoul, San Francisco and Taipei. And finally, Meta removed language from its commercial terms that appeared to permit scraping websites of businesses with Meta accounts for AI training. Legal experts told Nieman Lab the provision looked like permission to scrape publisher websites and use content to train AI models. Meta said the terms only applied to ads and business products and clarified the language in the terms themselves. That's dailybrief for today. Every story is linked to its original source on our website.
Podcast Two-host conversation with Jessica & Eric · 5 min
Jessica: Okay, so Replit just switched their default AI from Anthropic to OpenAI.
Eric: Wait, Replit? They were like a top-five customer for Anthropic.
Jessica: Exactly. The CEO said it was about cost and performance.
Eric: That's a massive shift.
Jessica: This is dailybrief, I'm Jessica.
Eric: And I'm Eric. It's Sunday, October eighteenth, twenty twenty six.
Jessica: So yeah, according to TechCrunch, Replit's CEO Amjad Masad confirmed OpenAI is now the default model provider. Users can still pick Anthropic if they want, but—
Eric: But defaults drive everything.
Jessica: Right. And he mentioned they've been building these vibe-coded applications, where you basically describe what you want and the AI builds it from scratch.
Eric: I mean, that's interesting, but this really feels like it's about money more than anything.
Jessica: Yeah, he did cite cost. And apparently their enterprise business grew from two percent to fifteen percent of revenue.
Eric: Okay, that's real traction.
Jessica: Speaking of enterprise spending, NVIDIA just launched RTX Spark laptops with their new Blackwell graphics.
Eric: How much?
Jessica: Starting at two thousand three ninety nine dollars.
Eric: Of course.
Jessica: But you get up to one hundred twenty eight gigs of unified memory, which is actually pretty wild for a laptop.
Eric: Wait, unified memory meaning shared between CPU and GPU?
Jessica: Exactly. So you can run way bigger models locally without shuffling data around. ASUS, Dell, HP, Lenovo, Microsoft, MSI—they're all shipping them now.
Eric: Okay, for someone who actually needs to run models on the go, that's compelling. Expensive, but compelling.
Jessica: On the open-source side, llama dot cpp released version b eleven seven one nine with CUDA thirteen support.
Eric: That's the tool for running models locally, right?
Jessica: Yep. So now you get CUDA twelve point eight and thirteen point four library support on Ubuntu and Windows, both x64 and arm64. Plus ROCm ten for AMD GPUs.
Eric: What does CUDA thirteen actually get you?
Jessica: Newer GPU features, better performance hooks. It's mostly about keeping up with NVIDIA's latest drivers and libraries.
Eric: Cool. Makes sense.
Jessica: And vLLM put out version zero point three two with a bunch of upgrades. They added structured decision endpoints—
Eric: Okay, what does that mean?
Jessica: So there's a new endpoint at slash v one slash systemone that handles choice, noul, and score questions. Like, instead of parsing a blob of text, you just get structured answers back.
Eric: That's actually really useful.
Jessica: Yeah. And they expanded FP4 KV cache to more GPU types—SM8x, SM12x, SM90.
Eric: That's the four-bit key-value cache, right? Cuts memory use in half?
Jessica: Exactly. Though heads up, they dropped Python three point ten support.
Eric: Always something.
Jessica: Ollama also updated—version zero point four zero seven. They added support for OpenAI's Responses effort parameter on non-thinking models.
Eric: Wait, I thought effort was just for the reasoning models.
Jessica: It was, but now Ollama lets you use it more broadly. They also added PPLX Decider v one point two support on the MLX backend, and the llama-server now infers draft speculative type automatically.
Eric: That last one's nice—less config, fewer knobs to turn.
Jessica: Quick fix from OpenAI: Python SDK three point three five point one patches a bug in the Realtime API.
Eric: What broke?
Jessica: Byte reads over websockets. So if you're using the streaming voice API, this fixes how it reads data off the wire.
Eric: Okay, good to know.
Jessica: Vercel added a sessions view to their AI Gateway logs. It groups requests and shows you which provider handled each turn.
Eric: So you can see cache hit rates?
Jessica: Yep. It's all about session affinity—routing each request to the same provider so you reuse the cached prompts. You get cost, models, providers, cache hit rates, all in one view.
Eric: That's genuinely useful for debugging your spend.
Jessica: Meanwhile, Meta quietly removed some language from their commercial terms after legal experts raised red flags.
Eric: What'd they say?
Jessica: According to TechCrunch, the terms appeared to let Meta scrape the websites of any business with a Meta account for AI training. Experts at Nieman Lab flagged it, and Meta pulled it.
Eric: Did Meta say it was a mistake?
Jessica: They said the terms only applied to ads and business products, not web crawling. But... the fact it was in there at all is kind of telling.
Eric: Yeah, that's not great.
Jessica: Simon Willison updated his web-edit tool—tiny web app for browsing and editing files over Tailscale. He added copy-to-clipboard for text and images.
Eric: That's handy. I use Tailscale for that kind of thing all the time.
Jessica: Alright, that's it for today. Every story we talked about is linked to its source on the dailybrief website.
Eric: See you tomorrow.
Step 3 · 2 minutes
web-edit (previously) is my tiny web app for browsing and editing files in a folder. I mainly built it for editing files on different machines via Tailscale.
llama.cpp released version b eleven seven one nine with CUDA twelve point eight and thirteen point four library support across multiple platforms.
Explain it simply
In plain words llama.cpp, a tool for running large language models locally, released a new version with support for NVIDIA's CUDA thirteen libraries on Linux and Windows.
Term to know CUDA — NVIDIA's parallel computing platform that lets programs run on GPUs
What it means for developers: CUDA thirteen support enables developers to use newer NVIDIA GPU features and libraries
On top of that, Vanniall struggles to build an audience on popular social media platforms, like Instagram, without getting in trouble.
Today's newsmakers
The companies that feature most in today's stories, ranked by how often and how prominently they appear.
3 stories today
Steady2 stories today
▲ 1+1 more story
2 stories today
New today+1 more story
1 story today
Steady1 story today
▼ 3Company names and logos are trademarks of their owners, shown only to identify them.
Open source · for developers
New repositories gaining stars fastest over the last 30 days. Source: GitHub.
AIA plugin that lets you use AI models including DeepSeek and Kimi without login, registration, or API keys for free.
AIA menu bar tool that provides access to various AI coding models including Codex on DeepSeek and Claude Code on Kimi.
AIAn open-source parametric CAD application written in Rust that runs on desktop, web, and supports automation through MCP.
AIA version control system designed specifically for AI agents to manage code.
AIConnects an AI vision model to control a real iPhone using a cheap ESP32 via AirPlay and USB HID.
AITools and skills that enable Claude Code to reverse engineer and modify PC games you own.
AIA website framework that automatically finds trending topics and writes daily reports based on your configured sources and criteria.
AIAn open-source personal finance application with AI advisor for tracking net worth, investments, ETFs, cash, and debts locally.
Optional · if you have more time
9 more stories
vLLM released version zero point three two with eight hundred fourteen commits, adding structured decision endpoints and low-precision KV cache across more GPU types.
Explain it simply
In plain words vLLM, an inference server for running large language models, released version zero point three two. The update adds endpoints for structured decisions and expands support for four-bit KV cache to more GPU types.
Term to know KV cache — stored key-value pairs that speed up text generation by reusing computations
What it means for developers: FP4 KV cache halves memory use per GPU while structured endpoints simplify decision tasks
OpenAI released Python SDK version three point three five point one, fixing beta Realtime byte reads on websockets.
Explain it simply
In plain words OpenAI released a bug-fix update to its Python SDK that fixes how the Realtime API reads bytes over websocket connections.
Term to know websockets — protocol for two-way persistent connections between client and server
What it means for developers: Fixes byte-reading bug for developers using OpenAI's streaming Realtime voice API
Ollama released version zero point four zero seven, allowing Responses effort on non-thinking models and adding PPLX Decider v one point two support.
Explain it simply
In plain words Ollama, a tool for running language models locally, released version zero point four zero seven with support for OpenAI's Responses effort parameter and PPLX Decider model.
Term to know speculative type — method for predicting next tokens to speed up text generation
What it means for developers: Responses effort support and automatic inference simplify model configuration for developers
Replit CEO Amjad Masad said OpenAI is now the default model provider, citing cost, performance and partnership, though Anthropic models remain accessible.
Explain it simply
In plain words Replit, an AI coding platform, changed its default AI model provider from Anthropic to OpenAI. The CEO cited better performance and lower cost, though users can still choose Anthropic.
Term to know vibe-coded applications — apps built from scratch using AI coding agents based on prompts
What it means for developers: Default changes drive most usage, potentially shifting significant revenue between AI providers
In fact, the human genome is less a script than a puzzle that gets harder the closer they look.
Vercel AI Gateway logs now group requests into sessions and show how session affinity routed each turn to maximize prompt cache hits.
Explain it simply
In plain words Vercel's AI Gateway now shows which AI provider served each part of a conversation and how much of the prompt was read from cache, helping developers see costs and routing.
Term to know session affinity — routing each request to the same provider to reuse cached prompts
What it means for developers: Visibility into cache hit rates and provider routing helps optimize costs and latency
The nonprofit 7x24 Exchange International wants to promote data centers, which they believe are "mission critical” infrastructure.
Meta removed commercial terms language that appeared to permit scraping any business account's website for AI training after legal experts raised concerns.
Explain it simply
In plain words Meta changed its commercial terms after legal experts said language appeared to let Meta scrape websites of any business with a Meta account for AI training.
Term to know web scraping — automated extraction of data from websites for other uses
Why it matters: Clarifies that Meta business accounts don't grant website scraping or AI training rights
NVIDIA RTX Spark laptops with Blackwell RTX graphics and up to one hundred twenty eight GB unified memory are now available from six manufacturers.
Explain it simply
In plain words NVIDIA launched RTX Spark laptops with its new Blackwell graphics chips and up to one hundred twenty eight gigabytes of memory, starting at two thousand three ninety nine dollars.
Term to know unified memory — RAM shared between CPU and GPU for faster data access
Why it matters: High unified memory and Blackwell GPUs enable running larger AI models locally on laptops
Last step · 1 minute
5 questions on today's stories
Browse other days
How we work
Lab announcements, release notes and SDK changelogs come before commentary. Every story links to the original.
Figures, model names and prices are taken from the published text. Where a detail is not stated, we leave it out.
Each story states what changed, by how much, and what it means in practice. No hype, no speculation.
vLLM zero point three two brings structured endpoints and FP4 KV cache to more GPUs, while Replit switches from Anthropic to OpenAI as its default model provider. 12 stories from 9 sources are summarised on this page.
1. web-edit 0.1a2. 2. llama.cpp b11719 released. 3. OnlyFans Creators Are Pissed That Meta Lets Muse Talk Dirty.
New Models (3), Developer Tools (6), Research (1), Chips & Infrastructure (2).
Headlines and short summaries come from the public feeds of 9 publishers, including NVIDIA Blog, Ollama releases, OpenAI Python SDK releases, Simon Willison's Weblog, TechCrunch, Vercel. Every story links to the original article.
Yes. A voice recording covering every story is at the top of this page and is published each day.
Every headline links to the original. How it works
