⚡️ 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…
⚡️ 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
❤1
🧠 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.
❤2
⚡️ 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. 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. 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
❤2
⚡️ 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
🤖 Run a whole office of AI clones on one task
Munder Difflin is an open-source agent harness that spins up a crew of LLM clones, each playing a role, and puts them to work together. Think multi-agent orchestration but dressed up in pixel-art cubicles.
The UI makes agent behavior actually legible. You watch them collaborate (or bicker). Less black box, more Dunder Mifflin.
Project
Munder Difflin is an open-source agent harness that spins up a crew of LLM clones, each playing a role, and puts them to work together. Think multi-agent orchestration but dressed up in pixel-art cubicles.
The UI makes agent behavior actually legible. You watch them collaborate (or bicker). Less black box, more Dunder Mifflin.
Project
munderdiffl.in
Munder Difflin · Multi agent harness for Claude Code, Codex and ten more
Free and open source multi agent harness. Run Claude Code, Codex, Gemini CLI and nine more terminal agents as an office of clones on your own machine, on the subscriptions you already pay for. Now with the Stapler for local dictation and meeting notes.
⚡️ Uber gets hit with a €825M fine for letting an algorithm fire its drivers
No human review. No real explanation. Just an automated system deactivating accounts, and drivers finding out after the fact.
The Dutch Data Protection Authority says it's a clear GDPR violation. Turns out "the algorithm decided" isn't a legal defense.
No human review. No real explanation. Just an automated system deactivating accounts, and drivers finding out after the fact.
The Dutch Data Protection Authority says it's a clear GDPR violation. Turns out "the algorithm decided" isn't a legal defense.
NL Times
Dutch regulator fines Uber €825 mil. for letting algorithm deactivate drivers' accounts
The Dutch Data Protection Authority (AP) has fined Uber €825 million for deactivating driver accounts through automated systems and without adequately informing them. This violates Europe’s General Data Protection Regulation (GDPR), the AP said in a decision…
🤖 MCP's new roadmap goes agent-first
The Model Context Protocol just published its roadmap, and the big theme is killing browser-click auth for agents. DPoP, Workload Identity Federation, and token exchange are coming so sub-agents can prove who they are without pasted API keys.
Progressive discovery is also on deck, so servers reveal tools gradually as a conversation narrows. Practical. Long overdue.
Now we wait to see how many servers actually ship it.
The Model Context Protocol just published its roadmap, and the big theme is killing browser-click auth for agents. DPoP, Workload Identity Federation, and token exchange are coming so sub-agents can prove who they are without pasted API keys.
Progressive discovery is also on deck, so servers reveal tools gradually as a conversation narrows. Practical. Long overdue.
Now we wait to see how many servers actually ship it.
Model Context Protocol Blog
The New MCP Roadmap
An update on the Model Context Protocol roadmap and focus areas for upcoming specification releases.
⚡️ Is Anthropic quietly throttling Claude Code?
Users are reporting that Claude Code's effort level can be read straight from its system prompt, and some sessions show it dialed down without warning.
API users see no degradation. Web UI? Different story.
With token spend directly tied to revenue, the incentives to quietly limit effort aren't exactly a mystery.
Users are reporting that Claude Code's effort level can be read straight from its system prompt, and some sessions show it dialed down without warning.
API users see no degradation. Web UI? Different story.
With token spend directly tied to revenue, the incentives to quietly limit effort aren't exactly a mystery.
X (formerly Twitter)
🥔🥔🥔 (@argofowl) on X
⚡️ Slop debt is the new tech debt, and it's nastier
AI code doesn't just rot, it compounds. Every "fix" piles more bloat on top of bloat nobody fully understands.
Old tech debt was poor abstractions. Slop debt is 60% code that shouldn't exist at all. Incremental refactoring won't save you. Rewrites will.
AI code doesn't just rot, it compounds. Every "fix" piles more bloat on top of bloat nobody fully understands.
Old tech debt was poor abstractions. Slop debt is 60% code that shouldn't exist at all. Incremental refactoring won't save you. Rewrites will.
Arpit Bhayani
Slop Debt
Tech debt has rules. You cut a corner to ship on time, you write a comment that says TODO fix this, and six months later someone pays it down. It is annoying but it is bounded. A senior engineer can look at a codebase with tech debt and draw a map of where…
🧠 Your local LLM isn't dumb. Your chat template probably is.
Quantization gets all the blame, but it's often not the culprit. A deep-dive on Level1Techs shows the real killers: wrong chat template and sampling settings that don't match vendor recommendations.
Many GGUF files ship without a template baked in, so your runtime silently falls back to a generic one. Model still talks, just noticeably worse. Easy to miss.
Check the template before you blame the quant.
Quantization gets all the blame, but it's often not the culprit. A deep-dive on Level1Techs shows the real killers: wrong chat template and sampling settings that don't match vendor recommendations.
Many GGUF files ship without a template baked in, so your runtime silently falls back to a generic one. Model still talks, just noticeably worse. Easy to miss.
Check the template before you blame the quant.
Level1Techs Forums
Why your local LLM feels dumber than it is
Quick Introduction We have all been on forums, chats, reddit, discord, youtube, or somewhere and heard “Oh! Model XYZ is AMAZEBALLZ!zomgwtfbbq” then downloaded it (or more likely, some quantized form of it) and said “eww… This sucks!” This post is going…
⚡️ Can AI models do real ML research? PrimeIntellect ran the numbers.
153 autonomous runs, 18 frontier models, up to 8 days each on 8xH200s. The task: optimize the nanoGPT training speedrun. It's the most scaled public experiment of its kind.
For context, Anthropic's internal AI R&D eval runs on a single CPU node. OpenAI uses one H100 for under a day.
Spoiler: gaps between models are real and wide.
153 autonomous runs, 18 frontier models, up to 8 days each on 8xH200s. The task: optimize the nanoGPT training speedrun. It's the most scaled public experiment of its kind.
For context, Anthropic's internal AI R&D eval runs on a single CPU node. OpenAI uses one H100 for under a day.
Spoiler: gaps between models are real and wide.
Prime Intellect
NanoGPT Speedrun Frontier
Compare 153 autonomous runs from 18 frontier models on the nanoGPT optimizer speedrun.
❤1
🤖 Devs don't write code anymore. They make decisions.
Simon Willison's new guide on agentic engineering patterns is out, and it's blunt: code generation is cheap now, so your job is architecture and judgment, not keystrokes.
Greenfield beats editing. Let the agent build fresh, keep the old version as fallback. Less legacy debt, more throwaway prototypes.
Honest take from the guy who coined "prompt injection."
Simon Willison's new guide on agentic engineering patterns is out, and it's blunt: code generation is cheap now, so your job is architecture and judgment, not keystrokes.
Greenfield beats editing. Let the agent build fresh, keep the old version as fallback. Less legacy debt, more throwaway prototypes.
Honest take from the guy who coined "prompt injection."
Simon Willison’s Weblog
Writing about Agentic Engineering Patterns
I’ve started a new project to collect and document Agentic Engineering Patterns—coding practices and patterns to help get the best results out of this new era of coding agent development …
⚡️ Anonymous model drops on OpenRouter and nobody knows who built it
"Ox Alpha" appeared out of nowhere on Aug 20, free to use, no author attached. 1M-token context, video input, coding benchmark scores beating GPT-5.6. Community fingerprinting leans toward a Chinese lab (Z.ai / Zhipu GLM family), but nothing's confirmed.
Free for about a week. Someone's burning serious compute to get attention. Source
"Ox Alpha" appeared out of nowhere on Aug 20, free to use, no author attached. 1M-token context, video input, coding benchmark scores beating GPT-5.6. Community fingerprinting leans toward a Chinese lab (Z.ai / Zhipu GLM family), but nothing's confirmed.
Free for about a week. Someone's burning serious compute to get attention. Source
Business Insider
A mysterious free AI model is impressing developers. And nobody knows who made it.
Ox Alpha, a mysterious AI model, debuted on OpenRouter on Thursday. Its origins are unknown, though some suspect a Chinese lab.
⚡️ ChatGPT's tokenizer has a Japanese internet alter ego
A fried chicken question led one dev down a rabbit hole: GPT's tokenizer has surprisingly deep roots in Japanese web culture, and those training data fingerprints show up in weird, specific ways.
Non-English text was never an afterthought. It's baked into the token merges. And sometimes the model acts accordingly.
A fried chicken question led one dev down a rabbit hole: GPT's tokenizer has surprisingly deep roots in Japanese web culture, and those training data fingerprints show up in weird, specific ways.
Non-English text was never an afterthought. It's baked into the token merges. And sometimes the model acts accordingly.
Theo Steiner
ChatGPT Is a Japanese Internet Degen, or: How Fried Chicken Sent Me Down a Tokenizer Rabbit Hole
A stray piece of 2ch boilerplate in a ChatGPT answer sent me down a rabbit hole through Japanese internet history, tokenizers, training data, and one very weird token.
⚡️ AI unlocked the hard stuff
Armin Ronacher's latest post nails something real: agentic coding is making devs more ambitious.
Stuff that was gatekept (eBPF, custom crypto, DWARF files, bare-metal drivers) is suddenly approachable. Language choice barely matters now. People ship Zig and Rust code without knowing either.
More slop, sure. But maybe also more developers who actually care about fast, small, correct code.
Armin Ronacher's latest post nails something real: agentic coding is making devs more ambitious.
Stuff that was gatekept (eBPF, custom crypto, DWARF files, bare-metal drivers) is suddenly approachable. Language choice barely matters now. People ship Zig and Rust code without knowing either.
More slop, sure. But maybe also more developers who actually care about fast, small, correct code.
Armin Ronacher's Thoughts and Writings
Fast and Hard Code
People pick languages differently with LLMs.
❤1