⚡️ Google's DiffusionGemma generates text like an image diffuser. And it's fast.
DiffusionGemma is Google DeepMind's open-weight text diffusion model built on a clever trick: it's a 25.2B-parameter MoE with only ~3.8B active params per forward pass.
No retraining from scratch. They converted an existing Gemma 4 checkpoint. One tester hit ~670 tok/s on consumer hardware.
Paper
DiffusionGemma is Google DeepMind's open-weight text diffusion model built on a clever trick: it's a 25.2B-parameter MoE with only ~3.8B active params per forward pass.
No retraining from scratch. They converted an existing Gemma 4 checkpoint. One tester hit ~670 tok/s on consumer hardware.
Paper
arXiv.org
DiffusionGemma Technical Report
We introduce DiffusionGemma, an experimental open-weight language model that uses discrete diffusion to generate text at exceptionally high speed. Rather than decoding one token at a time,...
⚡️ Aaron Swartz faced 35 years. Meta scrapes the whole web for AI and gets a shrug.
Swartz was prosecuted into a corner for downloading academic papers. Meta scrapes at planetary scale to train its models, with no criminal exposure in sight.
Same act. Different consequences. Turns out the CFAA bites hard when you're 26 and broke, and barely grazes you when you're a trillion-dollar company with aligned government interests.
Swartz was prosecuted into a corner for downloading academic papers. Meta scrapes at planetary scale to train its models, with no criminal exposure in sight.
Same act. Different consequences. Turns out the CFAA bites hard when you're 26 and broke, and barely grazes you when you're a trillion-dollar company with aligned government interests.
quailblog
I'm Upset Again About a Co-Creator of RSS Being Prosecuted For Something Meta Is Doing With Little Consequence
Also here's a cool unrelated photo of a chipmunk because I need some chill
⚡️ Tired of babysitting coding agents? This dev built a pseudocode editor instead.
Write rough pseudocode, hit save, get real code. Intent stays stored alongside the output so the "why" doesn't vanish.
It's a PoC, but the instinct is right: full-sentence prompting gets old fast, and agents start losing the plot on big codebases.
Write rough pseudocode, hit save, get real code. Intent stays stored alongside the output so the "why" doesn't vanish.
It's a PoC, but the instinct is right: full-sentence prompting gets old fast, and agents start losing the plot on big codebases.
www.danielvaughn.dev
Huzzah
My personal portfolio site and blog.
⚡️ Amazon buys rare books, scans them, destroys them. That's "fair use" now.
A 404 Media investigation confirmed it with a $29 AirTag: bulk book orders go to Amazon's LAS8 warehouse, where spines get cut and pages scanned for AI training data.
Meanwhile, the Internet Archive is being sued into the ground for digitizing books to preserve them.
Same act. Very different courtrooms.
A 404 Media investigation confirmed it with a $29 AirTag: bulk book orders go to Amazon's LAS8 warehouse, where spines get cut and pages scanned for AI training data.
Meanwhile, the Internet Archive is being sued into the ground for digitizing books to preserve them.
Same act. Very different courtrooms.
Blogspot
It is a sign of the times that Amazon gets to call this fair use while huge corporations try to sue the Internet Archive out of…
This 404 report justly been getting considerable coverage. Amazon is buying massive quantities of books, scanning them for AI training dat...
⚡️ Codex on AWS Bedrock is billing users ~10x too much
A caching bug in OpenAI's Codex on Bedrock is causing cache writes to fire constantly while reads barely register (under 5% hit rate). GPT-5.6 charges 1.25x for writes and 0.10x for reads, so a near-zero hit rate turns into a brutal bill.
Workaround: set `web_search = "disabled"` in your config.
Source
A caching bug in OpenAI's Codex on Bedrock is causing cache writes to fire constantly while reads barely register (under 5% hit rate). GPT-5.6 charges 1.25x for writes and 0.10x for reads, so a near-zero hit rate turns into a brutal bill.
Workaround: set `web_search = "disabled"` in your config.
Source
GitHub
Native Bedrock Codex GPT-5.6 Sol lacks explicit cache controls, producing high cache-write spend · Issue #37674 · openai/codex
Summary Native Codex CLI requests to Amazon Bedrock Mantle cannot opt into GPT-5.6 Sol explicit prompt caching. On an agentic coding workload, this has produced a large volume of cache-write tokens...
🤖 Apple Music is making AI song labels mandatory
Optional tags weren't cutting it. Apple introduced "Transparency Tags" back in March, but disclosure was left to labels and distributors' discretion. Now it's going compulsory.
Apple says it's built internal tools to identify AI-made tracks, fight spam, and catch impersonation by monitoring streams. It can even auto-pull AI tracks farming fake plays.
Labeling is the softest possible move. But it's a signal that synthetic content has officially scaled past the "niche problem" stage.
Optional tags weren't cutting it. Apple introduced "Transparency Tags" back in March, but disclosure was left to labels and distributors' discretion. Now it's going compulsory.
Apple says it's built internal tools to identify AI-made tracks, fight spam, and catch impersonation by monitoring streams. It can even auto-pull AI tracks farming fake plays.
Labeling is the softest possible move. But it's a signal that synthetic content has officially scaled past the "niche problem" stage.
MacRumors
Apple Music to Label AI-Generated Songs
Apple sent emails to its music industry partners about the upcoming launch of a Made With AI label that Apple will apply to AI-generated content (via The Hollywood Reporter). Content tagged as being "materially generated using AI" will have a label visible…
🤖 AI companies are scan-and-shredding books. Copyright is why.
Licensing e-books for AI training? Publishers won't play ball. So companies buy physical copies in bulk, gut them through scanners, and trash the rest. Cheaper and legally murkier.
Anna's Archive is calling for emergency digitization of rare and out-of-print titles before they disappear into a training run.
Perverse incentives, all the way down.
Licensing e-books for AI training? Publishers won't play ball. So companies buy physical copies in bulk, gut them through scanners, and trash the rest. Cheaper and legally murkier.
Anna's Archive is calling for emergency digitization of rare and out-of-print titles before they disappear into a training run.
Perverse incentives, all the way down.
annas-archive.gl
AI companies destroy physical books — let’s scan rare books before it’s too late
AI companies are secretly buying, scanning, and destroying millions of physical books to train their models, permanently locking human knowledge inside private corporate servers. Anna’s Archive is urgently calling on volunteers worldwide to scan and upload…
🤖 Mystery model just dropped on OpenRouter. Nobody knows who made it.
"Ox Alpha" is a frontier reasoning model, free to use, with a 1M token context window, coding + agentic focus, and text/image/video input.
The provider is staying anonymous "during this preview." Your prompts are retained. Not for training, they say, but for what exactly? Nobody's saying.
People are already poking its guardrails trying to fingerprint the lab. Try it if you're brave.
"Ox Alpha" is a frontier reasoning model, free to use, with a 1M token context window, coding + agentic focus, and text/image/video input.
The provider is staying anonymous "during this preview." Your prompts are retained. Not for training, they say, but for what exactly? Nobody's saying.
People are already poking its guardrails trying to fingerprint the lab. Try it if you're brave.
openrouter.ai
Ox Alpha - API Pricing & Providers
Ox Alpha is a reasoning model designed for coding, sustained agentic work, and production workloads. This model is free to use. 1,048,576 token context window, maximum output of 131,072 tokens.
⚡️ DeepSeek drops V4-Flash-Vision-Exp on their API
Multimodal vision is now live via `model='deepseek-v4-flash-vision-exp'`. On agent benchmarks requiring visual understanding, it delivers a significant leap over V4-Flash and brings multimodal agent capabilities close to Opus-4.8.
Still experimental. But shipping fast. Source
Multimodal vision is now live via `model='deepseek-v4-flash-vision-exp'`. On agent benchmarks requiring visual understanding, it delivers a significant leap over V4-Flash and brings multimodal agent capabilities close to Opus-4.8.
Still experimental. But shipping fast. Source
X (formerly Twitter)
DeepSeek (@deepseek_ai) on X
DeepSeek-V4-Flash-Vision-Exp is now live on the DeepSeek API Platform! 🚀
🔹 This experimental multimodal model matches DeepSeek-V4-Flash on text capabilities—including agents, reasoning, and world knowledge.
🔹 On multimodal agent benchmarks, V4-Flash-Vision…
🔹 This experimental multimodal model matches DeepSeek-V4-Flash on text capabilities—including agents, reasoning, and world knowledge.
🔹 On multimodal agent benchmarks, V4-Flash-Vision…
⚡️ DeepSeek V4 Flash just got eyes
DeepSeek quietly shipped vision support for V4 Flash. One model name: deepseek-v4-flash-vision-exp. This matters because V4 0731 was notoriously bad at knowing it was blind, and kept hallucinating fake image-reading tools.
384 tokens per image, ~2,500 images per dollar. Cheap. But capped at 800×800 px, which kills OCR on anything A4-sized.
Open weights? Unclear. Watch this space.
DeepSeek quietly shipped vision support for V4 Flash. One model name: deepseek-v4-flash-vision-exp. This matters because V4 0731 was notoriously bad at knowing it was blind, and kept hallucinating fake image-reading tools.
384 tokens per image, ~2,500 images per dollar. Cheap. But capped at 800×800 px, which kills OCR on anything A4-sized.
Open weights? Unclear. Watch this space.
Deepseek
Vision | DeepSeek API Docs
The deepseek-v4-flash-vision-exp model accepts images alongside text, so you can ask the model to describe pictures, read text from screenshots, analyze charts, and more.
⚡️ Intelligence just got 100x cheaper. Again.
A fixed capability level that cost $20/M tokens in late 2022 now runs under $0.10. Not a rounding error. A structural collapse in the price of thinking.
And it's still accelerating. The apps that were "too expensive to build" a year ago are cheap today.
A fixed capability level that cost $20/M tokens in late 2022 now runs under $0.10. Not a rounding error. A structural collapse in the price of thinking.
And it's still accelerating. The apps that were "too expensive to build" a year ago are cheap today.
CatalystNeuro
What Happens When the Cost of Intelligence Drops 100x — CatalystNeuro
A given level of LLM intelligence costs about a thirtieth of what it did a year ago, and the rate of decline is accelerating. Using Artificial Analysis's measured cost per task, this post traces the falling price of each capability level, explains why cheaper…
⚡️ Devs are quietly switching from Claude to Codex
One dev's week with Codex stuck out: it's faster, writes cleaner Ruby, and doesn't drown your codebase in comment blocks that read like an LLM's inner monologue.
Claude's verbosity is becoming a real complaint. Code comments shouldn't explain the AI's reasoning process to future devs.
Codex isn't perfect (it overengineers). But "terse and fast" is winning.
One dev's week with Codex stuck out: it's faster, writes cleaner Ruby, and doesn't drown your codebase in comment blocks that read like an LLM's inner monologue.
Claude's verbosity is becoming a real complaint. Code comments shouldn't explain the AI's reasoning process to future devs.
Codex isn't perfect (it overengineers). But "terse and fast" is winning.
All about coding
Quick impressions: A week of using Codex more than Claude
Claude goes above and beyond what is asked and guesses what you might want. Codex does what you tell it and stops at the first sign that it might be done. Ten impressions from a week of using Codex more.
⚡️ Someone built a tool to stop Claude from writing like it's pitching a listicle
"Claudette" (nobuzz) pipes Claude's output through Gemini to strip the corporate fluff and superlatives in real time.
It's a skill you drop into Claude Code. One command, previous reply detoxed.
Anthropic chose to anthropomorphize the model hard. The open-source community is now billing them for the cleanup. Source
"Claudette" (nobuzz) pipes Claude's output through Gemini to strip the corporate fluff and superlatives in real time.
It's a skill you drop into Claude Code. One command, previous reply detoxed.
Anthropic chose to anthropomorphize the model hard. The open-source community is now billing them for the cleanup. Source
GitHub
nobuzz/README.md at main · adnanakil/nobuzz
A Claude Code skill (/debuzz) that pipes Claude's answers through Gemini to remove the BuzzFeed voice - adnanakil/nobuzz
🧠 NVIDIA won't tell you how its memory works, so someone ran the experiments
Doubleword's deep-dive into GPU memory access reverse-engineers NVIDIA's undocumented memory subsystem through timing benchmarks because the official docs just... don't go there.
Turns out a lot of AI kernel optimization is flying blind. Worth the read.
Doubleword's deep-dive into GPU memory access reverse-engineers NVIDIA's undocumented memory subsystem through timing benchmarks because the official docs just... don't go there.
Turns out a lot of AI kernel optimization is flying blind. Worth the read.
Doubleword
What happens when a GPU reads memory | Doubleword
Following an LDG.E SASS instruction through the hardware units in an RTX 4090.
⚡️ Open-source TTS hits 34 ms latency on one H100
Nari Labs squeezed Qwen3-TTS down to 34 ms p95 time-to-first-audio at 10 req/s on a single H100. No quality loss. Code is open-source.
Standard vLLM/SGLang setups were hitting 200 ms+. That gap is the whole story.
Real-time voice agents are an infra problem as much as a model problem.
Nari Labs squeezed Qwen3-TTS down to 34 ms p95 time-to-first-audio at 10 req/s on a single H100. No quality loss. Code is open-source.
Standard vLLM/SGLang setups were hitting 200 ms+. That gap is the whole story.
Real-time voice agents are an infra problem as much as a model problem.
Nari Labs
Pushing the Speed-Cost Frontier for Qwen3-TTS
10 RPS with p95 TTFA under 50 ms on a single H100: how we optimized Qwen3-TTS serving.
⚡️ OpenAI cuts GPT-5.6 Sol prices 20%+ for devs
Input drops from $5 to $4 per million tokens. Valid for three months, with Anthropic and Chinese models breathing down their neck.
Frontier AI is getting cheap fast. The EV price war just found a new industry.
Input drops from $5 to $4 per million tokens. Valid for three months, with Anthropic and Chinese models breathing down their neck.
Frontier AI is getting cheap fast. The EV price war just found a new industry.
⚡️ OpenAI cuts GPT-5.6 Sol pricing 20%
It's a 3-month promotional drop. Also: Luna (the lightweight one) gets slashed 80%, and Terra 20%.
The 20% efficiency gain was partly driven by Sol itself, which autonomously rewrote production code kernels to cut serving costs. Paying models to reduce their own bill. Sure.
Source
It's a 3-month promotional drop. Also: Luna (the lightweight one) gets slashed 80%, and Terra 20%.
The 20% efficiency gain was partly driven by Sol itself, which autonomously rewrote production code kernels to cut serving costs. Paying models to reduce their own bill. Sure.
Source
OpenAI Developers
GPT-5.6 Sol Model | OpenAI API
❤1
⚡️ OpenAI acquihires the Instant team
The folks behind InstantDB, a real-time "Firebase for the modern stack," are joining OpenAI. Cloud app shutdown hits August 31, 2027 (so, a year out, not a fire drill).
Classic acquihire: the tech and the people go in, the product sunsets. Seventeen OpenAI acquihires in three years and counting.
The folks behind InstantDB, a real-time "Firebase for the modern stack," are joining OpenAI. Cloud app shutdown hits August 31, 2027 (so, a year out, not a fire drill).
Classic acquihire: the tech and the people go in, the product sunsets. Seventeen OpenAI acquihires in three years and counting.
Instantdb
The Instant team joins OpenAI
The Instant team is joining OpenAI. Instant Cloud runs until August 31st, 2027.
⚡️ New Rust LSP claims 100x lower RAM than rust-analyzer
rust-glancer is a fresh, open-source language server for Rust built around one idea: don't hold the whole workspace in memory. Skips incremental analysis in favor of a leaner, on-demand model.
Still early (v0.1), but if you've ever watched rust-analyzer eat 4GB on a big workspace... you get the appeal.
Source
rust-glancer is a fresh, open-source language server for Rust built around one idea: don't hold the whole workspace in memory. Skips incremental analysis in favor of a leaner, on-demand model.
Still early (v0.1), but if you've ever watched rust-analyzer eat 4GB on a big workspace... you get the appeal.
Source
rust-glancer.github.io
Hello, world! · Rust Glancer
Rust LSP that doesn't eat memory for breakfast