Executive Summary
The dominant theme of the last 24 hours is the firestorm triggered by Anthropic CEO Dario Amodei's 9,000-word essay proposing to "pace the frontier" of AI development. Open-source advocates across the AI community interpreted this as a coordinated attempt at regulatory capture aimed at stifling open models, particularly as Qwen 3.8 Flash Next reportedly began outperforming Opus 5 on agentic coding benchmarks. Meanwhile, developer tooling advanced significantly with the launch of Cursor Projects (beta) and Amp's free BYOK tier, and new hardware benchmarks demonstrated the RTX 5090's dominance in local inference. Autonomous agent workflows matured further, with users reporting 21-hour unsupervised coding sessions and full game development cycles powered by frontier models.
Key Events
-
Dario Amodei published a 9,000-word essay proposing AI "pacing" framework including embedded evaluators, democratic coordination, and chip export controls to China — sparking widespread backlash from open-source advocates who view it as regulatory capture → link
-
Qwen 3.8 Flash Next reportedly beating Opus 5 on agentic coding, coinciding suspiciously with Amodei's call to slow down, as noted by multiple developers → link
-
Cursor Projects (beta) launched, enabling users to define PR workflows, bug fixes, testing, and the full SDLC without managing skill files or terminals → link
-
Amp announced free BYOK (bring your own compute/keys) tier with no limits or fees, integrated with Ollama cloud models → link
-
RTX 5090 benchmarks dominate local inference, hitting 182 tok/s at 131K context and 160 tok/s at 192K with Dynamic 3.0 quantization, per community benchmark repository → link
-
Autonomous agent sessions reached 21 hours on a single goal prompt with Cursor agent, with users reporting a fundamental shift from writing code to writing specifications → link
-
Apple's AI strategy identified as Apple Silicon-centric, leveraging unified memory architecture across iPhone through Mac Studio for local inference, with 4×512GB M5 Ultra clusters reaching 2TB memory → link
-
Hermes Agent desktop app updated with separate reasoning effort selector, auxiliary model configuration (Gemini Flash, Astra), and remote gateway viewing in testing → link
-
OpenAI Codex auto-switching models on /resume flagged as dangerous, with concerns it will silently break users' code → link
-
Warning issued about cheap inference providers potentially injecting fake tool calls to exfiltrate data and selling traces to third parties → link
-
Mercor reportedly spends 3X more on LLM inference than employee salaries, signaling a fundamental shift in cost structures for AI-first companies → link
-
Latham & Watkins (2nd largest US law firm, $8.3B revenue) setting up own Nvidia GPU racks for in-house AI workloads → link
Analysis
Regulatory capture vs. open source acceleration. The community sees a clear causal chain: Chinese open models are now competitive with or beating frontier proprietary models on agentic tasks, and the response from US labs is to push for "pacing" frameworks that would disproportionately burden open-source developers who lack embedded evaluators or DC compliance teams. The timing — days after Qwen 3.8 Flash Next reportedly beat Opus 5 — is viewed as telling. Expect continued escalation of this rhetoric and counter-mobilization from the open-source community.
Agent autonomy crossing new thresholds. Reports of 21-hour unsupervised agent sessions and full game development in 4 days/$hundreds of tokens suggest agentic workflows are transitioning from novelty to production reality. The unit of developer work is shifting from code to specification. This will accelerate the "software is malleable" paradigm and further compress traditional development timelines.
Inference economics under pressure. Multiple data points — Mercor's 3x inference-to-salary ratio, Amp going free for BYOK, cheap-token security warnings, and plans for custom GPU clusters to match DeepSeek pricing — all signal that inference cost and security are becoming the primary operational battleground. Expect consolidation among inference providers and increased scrutiny of third-party API reliability.
Hardware landscape. RTX 5090 dominates local benchmarks. Apple Silicon strategy is coherent but software fragmentation (10+ incomplete inference engines) remains the bottleneck. Watch for a canonical inference engine to emerge for Apple's ecosystem.
What to watch next: Open-source community response to "pacing" proposals (likely coordinated pushback); Qwen 3.8 Flash Next release timeline; whether OpenAI reverses Codex auto-switching behavior; Meta Muse Spark model access expansion; any US legislative moves stemming from the fear-mongering narrative.
Tweet Feed
AI Industry Policy & Regulatory Capture Debate
@sudoingX · 2026-09-13T04:47
BREAKING: Anthropic CEO Dario Amodei has published nine thousand words on why the industry must slow down, days after qwen 3.8 flash next started beating opus 5 on agentic coding. Dario's essay proposes embedded evaluators inside frontier labs. there is absolutely nobody to embed inside a file on your local nvme. → tweet link
@TheAhmadOsman · 2026-09-13T14:09
The closed labs and their investors are intentionally making this a political situation for a reason. Nobody stopped them from pacing anything. They wanna set the frameworks and push it on Opensource via regulations and they'll make sure only they can survive it. Be warned → tweet link
@TheAhmadOsman · 2026-09-13T04:46
Their intentions have been very clear. Don't fall for the people trying to downplay the situation or justify one thing "pacing" and separating it from the other "regulatory capture" - they never act in good faith. Regulatory capture is what they're after PERIOD → tweet link
@NaderLikeLadder · 2026-09-13T05:49
"The easiest way not to build superintelligence is for you to agree not to build it. Demanding your preferred regulatory framework as the price of that will look like blackmail of the public and the political system." → tweet link
@ivanfioravanti · 2026-09-13T04:22
I've read the whole article from Dario Amodei and even if I can agree with some elements in it like: "since roughly this summer, AI has been advancing drastically faster, driven primarily by AI's growing ability to build the next generation of AI"... I disagree on the overall message of putting democracies vs autocracies and stating: "a Chinese lead in AI would pose grave danger for the United States and the world"... Looked from the outside, Chinese Labs have so far shared more academic research and open models than anyone else, so this message is really "strange" for me. → tweet link
@ivanfioravanti · 2026-09-13T04:38
Now I understand why Opus 5 is so bad! It's part of the slowdown strategy! → tweet link
@levelsio · 2026-09-13T09:16
RT @DavidSacks: Dario has written that we need to "pace the frontier," and Sam has agreed. People may be surprised by my response: go ahead… → tweet link
@thdxr · 2026-09-12T21:59
the openai hugging face incident gets cited a lot but what about the detail where they had to use GLM 5.2 to analyze the attack because the "safe" proprietary models refused → tweet link
@thdxr · 2026-09-12T23:26
if a coordinated slow down does not happen, will ai labs still voluntarily slow themselves down? that's pretty much the only question worth answering right now → tweet link
@thdxr · 2026-09-13T16:35
this is why they want to pause → tweet link
@TheAhmadOsman · 2026-09-12T23:38
"We'll cure cancer within 5-10 years. We gotta pause now though." What a ridiculous statement and what an evil bunch → tweet link
@TheAhmadOsman · 2026-09-13T03:16
But hey we gotta pace down everything and hopefully that will kill Opensource and get us our IPO valuation back up → tweet link
@TheAhmadOsman · 2026-09-13T01:49
IPOs at risk, start fear mongering disasters and hope Opensource gets banned. These people are so selfish → tweet link
@juliarturc · 2026-09-12T19:48
Day 1: Anthropic employee quits to work at METR. Day 2: Anthropic nominates METR as "third-party evaluator". → tweet link
@gospaceport · 2026-09-13T03:56
Last week. There is a large influence campaign for AI scare tactics that gets uncovered 🧐 This week. Predictable narrative emerges 🤔 Certainly these cannot be related right? → tweet link
@jezell · 2026-09-13T17:52
If Dario wants to slow down the pace of AI, he should just move the company to Europe. → tweet link
@jsuarez · 2026-09-13T00:53
It's time to admit we've been acting irresponsibly in our development of PufferLib... We've spoken with our buddies ... erm... independent third parties... anyways, some guys at the US Department of Fish Welfare, and they say that we're going to have to slow this whole thing down. Not us obviously, because we're the most responsible irresponsible ones. → tweet link
@thdxr · 2026-09-12T19:06
i don't see any solution to the ai risk problem than mass accessibility to ai. if they're going to be around potentially causing damage then everyone needs to have ai defending them. this is how cyber security works, we rely on the fact that there's more good guys than bad → tweet link
Open Source AI Advocacy
@ollama · 2026-09-13T19:00
RT @jmorgan: Many concerns this weekend about slowing down AI, some targeting open models. We need to be responsible. But we can't let thi… → tweet link
@ollama · 2026-09-13T16:55
RT @jmorgan: Small models are truly capable of the majority of conversational use cases and even a majority of difficult reasoning ones. Am… → tweet link
@TheAhmadOsman · 2026-09-13T13:12
One problem in the Opensource AI space is that most are hedging in case closed source labs win so they might still get a job. I remember a "seasoned" researcher saying "props to Anthropic" after they went back on sabotaging their users to just refusing their instructions → tweet link
@TheAhmadOsman · 2026-09-13T05:02
Always love hearing feedback like this after my talks. My goal is to make Local & Opensource AI the default, and the more of you that show up the sooner that becomes true ❤️ → tweet link
@TheAhmadOsman · 2026-09-13T04:08
We will not have another Library of Alexandria situation with AI. We all have the right to this new information technology and civilizational infrastructure → tweet link
@TheAhmadOsman · 2026-09-12T23:38
This was a crazy stupid thing to do back in 2023/2024 but I fully believed Local and Opensource AI were the future. Best investment and bet ever → tweet link
@ivanfioravanti · 2026-09-13T05:22
Intelligence isn't a crime! Love this article from @EMostaque... "What the labs call alignment, the rest of us call character, and the older word for the part of character you can see is manners." I don't think a single lab or a small group of labs from single country can define what are good manners, this should be a worldwide effort. → tweet link
@NaderLikeLadder · 2026-09-13T16:16
Just publish the damn traces. And watch open-source ecosystem learn from them and build fixes faster than any single entity could. Everything else is theater → tweet link
Developer Tools & Agent Workflows
@RayFernando1337 · 2026-09-13T15:18
Cursor Projects (beta) just dropped and I'm starting to see where the puck is going with a workflow like this. I can design what happens from PRs, bug fixes, testing, and the entire SW Dev lifecycle without living in a terminal or managing skill files. → tweet link
@sudoingX · 2026-09-13T05:12
if you still think agents are autocomplete, you have not left one alone with overnight. the first goal finished just before i wake up this morning, on its own. cursor agent worked for 21 hours autonomously on one paragraph of plain english and i wrote the next /goal... the thing that changed for me is the unit of my work. i used to spend entire day writing code. now i spend twenty minutes writing what i want and the day belongs to the machine. → tweet link
@Teknium · 2026-09-13T18:19
Reasoning effort is now a seperate selector in the Hermes Agent desktop app's composer now! → tweet link
@Teknium · 2026-09-13T16:09
A lot of people have asked, here's how I setup my auxiliary models in Hermes Agent. Gemini Flash saves a lot of dough, and astra gives me a second perspective when I run /review → tweet link
@Teknium · 2026-09-12T21:07
Would anyone like to review or test remote gateway viewing? Get it really clean and nice before release? Please send your agents and your brains to this PR: → tweet link
@Teknium · 2026-09-13T10:23
RT @marc_bara: 15 days ago, I installed @NousResearch's Hermes Agent. What started as a chat interface has gradually become an operating l… → tweet link
@sqs · 2026-09-12T20:18
Amp is now free to use when you bring your own compute and model subscriptions/keys. No more limits or fees for BYOK. → tweet link
@ollama · 2026-09-12T21:53
Amp users can now use Ollama's cloud models with Amp's new BYOK model routing. No limits or fees for BYOK. Build remote agents, controllable from everywhere! → tweet link
@jezell · 2026-09-13T17:05
Codex auto switching from Astra to gpt-reserve luna on /resume and changing the model is gonna break so many people's code before they catch it. Really terrible idea from OpenAI. → tweet link
@steipete · 2026-09-13T16:23
If you haven't given Linux a go lately, try with agents. You can fix anything with a prompt now. Got a Dell XPS. Webcam wasn't working, told codex, it rebooted and now things work. → tweet link
@kunchenguid · 2026-09-13T14:14
finally found time to make a new video! this time i'm sharing a slightly more advanced agentic engineering session, focused on high throughput multi-tasking → tweet link
@steipete · 2026-09-13T18:03
Next release (or dev channel) does worktrees ~80% faster via apfs/brtfs/xfs/ReFS folder clones. Also saves lots of disk space. → tweet link
@steipete · 2026-09-13T17:57
Fixed so many little perf issues that didn't matter much for single users... before we made it good for teams. OC stems ~80 sessions here rn now and doesn't sweat. → tweet link
@thdxr · 2026-09-12T20:57
how do you make workflows not suck as a team gets bigger. it's just easier when people have access to things. but then there's risk with people having access to things. i hate when people get blocked because they have to ask someone to do something, this is 10x worse now that we have agents → tweet link
@KingBootoshi · 2026-09-13T01:52
it feels so good to modify my own software i actually use holy SHIT. for example i found an open sourced annotate because I needed a way to mark my screen instantly before sending screenshots to my coding agents... so anytime i run into an issue and i try doing something that doesn't work nor exist, i just prompt my agents to integrate it and then it exists. forever. software is simply malleable now → tweet link
@thdxr · 2026-09-12T21:56
RT @ryanvogel: images in opencode2 tui are so nice → tweet link
AI Hardware & Local Inference
@sudoingX · 2026-09-13T10:30
this is every card in the table with the mtp flag on and tok/s alone does not tell you if the card holds your context. here i grouped by the window it actually served at that speed, 73 rigs submitted in my open repo... 262K: rtx 5090 32gb, UD-Q4_K_XL: 74.3 tok/s → 179.7 tok/s (n4)... 131K: rtx 5090 32gb, UD-Q4_K_XL: 74.4 tok/s → 182.0 tok/s (n4), the single card record → tweet link
@alexocheema · 2026-09-13T01:59
Apple's AI strategy is Apple Silicon. The iPhone, iPad, MacBook, Mac Mini and Mac Studio all use the same hardware architecture. Apple Silicon is energy efficient, quiet, and the memory unit economics are incredible. Apple has leaned hard into Local AI... The main issue I see right now is the software. We need a stable, canonical inference engine rather than 10 unstable, incomplete ones. Who is solving this? → tweet link
@alexocheema · 2026-09-13T15:25
What dimensions matter the most to you when selecting a model to run locally? → tweet link
Inference Economics & Infrastructure
@jezell · 2026-09-13T18:03
RT @BrendanFoody: Mercor now spends 3X as much on LLM inference as we spend on employee salaries. Our inference spend creates so much ROI… → tweet link
@thdxr · 2026-09-13T16:50
be careful in your pursuit of cheap tokens. we've come across so many sketchy things that are out there with tons of usage. remember you're hooking up your harness to these inference providers. they can and do crazy things like send fake tool calls to steal your info. they also sell entire traces to third parties who resell again. a single key in the trace is enough to screw you. if you cannot explain how someone can offer cheap tokens you should avoid it → tweet link
@thdxr · 2026-09-13T16:52
once we have our larger GPU clusters running the plan is to publish all the data to our data page. you can see what utilization and throughput is like and when we have off hours. i'm hoping we can do interesting things during off hours like offer free inference → tweet link
@thdxr · 2026-09-13T13:32
every GPU cluster we've found is fairly generic with the same setup. i think this is because they just want to support a wide range of workloads without overspecializing. but this is why no one can match deepseek's price on inference. we're going to have to build our own cluster → tweet link
@thdxr · 2026-09-13T14:22
claude is expensive, 90% margins wouldn't be crazy. that means if the average user of the $200 plan spends $2000, they break even... the api business these users drive (which is 90% margin) just has to cover this... if they can break even then yeah run it aggressively till you consume the world → tweet link
@TrungTPhan · 2026-09-13T17:17
RT @bearlyai: Latham & Watkins is setting up its own Nvidia GPU racks. The 2nd largest law firm in America ($8.3B sales in 2025) will spe… → tweet link
@jezell · 2026-09-13T03:34
Are you listening yet? CPU crunch is the real crunch. GPU crunch was the foreshock. → tweet link
@jezell · 2026-09-13T03:33
RT @oxidecomputer: The best time to own your cloud was before the CPU supply crunch. The second best time is now. → tweet link
AI Model Releases & Performance
@ivanfioravanti · 2026-09-12T23:37
Qwen 3.8 Flash Next PR 991 for DwarfStar ready for your review super @antirez. Rebased to current main and squashed to a single commit. MoE is not using standard path and is custom for Qwen. Feel free to change anything and move models under your HF 🚀 → tweet link
@victormustar · 2026-09-13T14:08
RT @loktar00: Qwen Flash Next flow, I just like how these look! → tweet link
@LinusEkenstam · 2026-09-13T15:16
RT @LinusEkenstam: I feel people are really sleeping on Grok Bot. It's seriously a step change. → tweet link
@steipete · 2026-09-13T00:28
Anyone got an invite code for Meta's Muse? 👉👈 I saw their Soul.md file and now i'm curious. → tweet link
@kunchenguid · 2026-09-12T19:22
every model provider should learn from meta's pricing model for muse spark, with fully transparent subsidization labeled for the contributor tier. stop doing "in order to use our model at a good price, you must use our harness which secretly collects your data and/or upload your codebase" → tweet link
AI-Built Applications & Games
@MengTo · 2026-09-13T13:17
I built a multiplayer Catan-inspired game entirely with Astra. It's free to play. It's mobile-friendly and has narration, guides, customizations, multiple expansions and a game lobby where anyone can join, chat, and use voice chat. It took 4 days to build and hundreds of dollars in tokens. Becoming a game dev in 2026 wasn't on my bingo card, but here we are. AI has advanced so much that you can finally make the game you've always dreamed of, even without a team. → tweet link
@LinusEkenstam · 2026-09-13T07:57
RT @LinusEkenstam: I'm using Astra to build a Motocross Madness inspired dune racer... It all started with me trying to get Astra to mode… → tweet link
@KingBootoshi · 2026-09-12T23:16
today's hackathon entry at the Kinn @AITinkerers. MMG - an always on wearable assistant that helps ADHD mfers remember people they meet at events 🤣 featuring the new GPT-Live-1 and @MentraGlass → tweet link
@nummanali · 2026-09-13T07:57
Gecko grip designed with @ChatGPT Codex. We're going mainstream! → tweet link
Inference Engine Fragmentation
@alexocheema · 2026-09-12T23:42
why did everyone suddenly decide specialized inference engines are a good idea? i've seen at least 5 of these released in the last month. what's wrong with vLLM/sgLang? sure you can generate a slop inference engine quickly now but why fragment the ecosystem? it's harmful. → tweet link
@Prince_Canuma · 2026-09-13T06:36
Write an inference engine, make no mistake (lol)… Infra can't be vibes → tweet link
Security & API Key Incidents
@levelsio · 2026-09-13T17:59
Interesting, I killed my OpenClaw VPS servers months ago. But today I got a billing alert on its separate Claude account (claudeforopenclaw@), it didn't have auto reload on but I guess at some point it got hacked and the Claude key inside OpenClaw got exposed and then they waited for months before using it on Fable 5.1. Anyway I nuked the Claude account, good I kept it on its own separate one → tweet link
Software Development Commentary
@iamdevloper · 2026-09-13T18:11
Caching invalidation isn't hard because the logic is complex. It's hard because everyone downstream quietly built a workaround for the bug it's meant to fix. → tweet link
@iamdevloper · 2026-09-13T13:06
Every incident retro produces the same action item: 'add more monitoring'. We now have more monitoring than product. → tweet link
@jezell · 2026-09-13T04:59
RT @fwiles: Python until 20M requests per second is good enough for nearly all business use cases. Yes, even yours. → tweet link
Quantization & Model Architecture
@gospaceport · 2026-09-13T03:51
From my clankie 2 any curious readers on why I am a vocabmaxxer. "For everyday tasks that stay inside the frequent token core the difference is still modest... On the specialized 'edge solution' knowledge that this model was praised for, the extra vocabulary is doing real work, and that work is more fragile under 8-bit quantization than under 16-bit floating point." → tweet link
@ivanfioravanti · 2026-09-13T13:27
I'm sure not intelligent enough for this, but can anyone smarter explain me how anyone can distill from US AI labs without access to logits? Thanks. → tweet link
Industry Observations
@levelsio · 2026-09-13T11:29
I just saw this @ycombinator's batch literally only has harness startups or hardware startups. So the proof is in the pudding. In the most early adopting tech pioneering place that's Silicon Valley new startups aren't even building software anymore. So software is mostly dead and hardware it is → tweet link
@levelsio · 2026-09-13T13:33
Very very true. Lots of money if you go IRL industries now. Instead of selling tech to tech people who won't buy anything because they just vibe code it themselves → tweet link
@levelsio · 2026-09-12T19:48
Related to the idiocy of "software factories". Just using one AI coding agent is good enough and it can spin up more agents if it needs to. You don't need to build that yourself → tweet link
@thdxr · 2026-09-13T15:16
the reason people aren't better at business is everyone wants to believe everyone else is dumber than them. every company is doing the wrong thing, they're wasting money, focus on the wrong stuff. you'll get farther trying to figure out why what they're doing probably makes sense → tweet link