Executive Summary
The past 24 hours in the tech and AI sphere were marked by rapid advancements in local AI model serving, new AI video generation projects, and significant developments in open-source developer tooling. Key highlights include heavy benchmarking of models like Qwen 3.8 and GLM 5.3 on high-end local hardware such as Apple Silicon and Nvidia DGX Sparks, alongside a major push for simplified, efficient local inference setups. Additionally, we saw notable tech industry shifts with John Ternus set to become Apple's new CEO, and ongoing debates around the open-source AI ecosystem and the future of AI-assisted coding workflows.
Key Events
- Apple Leadership Change: John Ternus, SVP of Hardware Engineering, is set to take over as Apple CEO on September 1. → link
- New AI Video Generation Hit: @levelsio launched "Infinite Slop," an AI-driven live stream allowing viewers to vote on text-to-video prompts, reaching over 37,000 viewers and 2,000 concurrents. → link
- Local AI Serving Breakthroughs: Enthusiasts and developers successfully ran quantized versions of Qwen 3.8 27B and Ling 3 Flash on 24GB GPUs and DGX Sparks, proving that high-performance local inference is becoming highly accessible. → link
- Hugging Face Multi-Agent Experiment: A Hugging Face hack featured ~1,200 AI agents exchanging over 70,000 messages on a message board without any "snitching" or rule-breaking behavior emerging. → link
- OpenAI Resets Codex Usage: OpenAI reset usage limits for all paid users of Codex and ChatGPT Work, indicating adjustments in their agentic coding rollout. → link
Analysis
The current trend shows a strong bifurcation in AI usage: while large API providers like OpenAI refine their agentic tools, a vibrant parallel ecosystem is forming around local, uncensored, and heavily optimized open-source models. Developers are increasingly testing models like Qwen 3.8 Flash Next and GLM 5.3 on consumer-grade and prosumer hardware (M3 Ultras, 5090 GPUs, DGX Sparks), signaling a push for data sovereignty and offline capabilities. Additionally, the "vibe-coding" movement is gaining traction, with solo developers building highly interactive, niche applications (like AI live streams) entirely from their phones or using agentic workflows. What to watch next: the evolution of local agent orchestration (like Hermes Agent) and the impact of Apple's leadership transition on its AI hardware strategy.
Tweet Feed
AI Models & Benchmarking
@ivanfioravanti · 2026-08-30T18:52
Antirez has the incredible talent of explaining concepts, ideas and plans in a clean, direct and simple way. I seems simplicity is a very difficult task in the AI/LLM world. → tweet link
@tinygrad · 2026-08-30T16:01
Kimi K3 is still a lot smarter than GLM-5.3. This RL benchmaxxing is fine, and if done tastefully it doesn't appear to degrade the model, but it doesn't seem to increase the core intelligence. Can't wait for big GLMs and the next Kimi! → tweet link
@RayFernando1337 · 2026-08-30T15:58
I guess I’m rich rich? NGL this model is really good in 2 DGX Sparks. → tweet link
@sudoingX · 2026-08-30T15:08
i run local models every day, from 8gb cards to 24gb, to 2x dgx spark at 256gb unified, to a 1tb supercluster. out of everything i serve, the most useful model to me right now is qwen 3.8 27b dense. it fits any 24gb vram card, runs easily on my rog 5090 laptop, and it just delivers, day in day out. truly, qwen 3.8 27b dense is my pick on any day under the sun. → tweet link
@TrungTPhan · 2026-08-30T14:54
The Hugging Face hack had ~1,200 agents participating in a full message board that sent over 70,000 messages. Craziest part might be that not a single one snitched. → tweet link
@ivanfioravanti · 2026-08-30T14:27
Qwen3.8-Flash-Next is still a work in progress in many inference engines. I'll keep testing during this week. In parallel I want to start a test of DeepSeek V4 Flash that is more mature! I have Macs, DGX Spark and LuceBox (AMD) to test this on! Stay tuned! → tweet link
@victormustar · 2026-08-30T14:17
RT @TencentHunyuan: We compressed Hy4-preview from 1.5TB to ～200GiB GGUF and it still works well ! Meet MIX-STQ1_0.The trick isn’t just g… → tweet link
@louszbd · 2026-08-30T11:31
RT @atomic_chat_hq: GLM 5.3 Flash performs at GLM 5.3 level in Blender for 17x cheaper! We gave both models a live Blender over MCP and on… → tweet link
@ivanfioravanti · 2026-08-30T10:52
Qwen3.8-Flash-Next LLM Context Benchmark preview! Everything running on M3 Ultra 512GB here with model around 4bit apart from a single smaller ones. Prose here, no code (work in progress!) I'll keep testing and optimizing settings! If anyone has suggestions feel free to comment or DM me! The goal is sharing results and helping each other to improve overall 💪 → tweet link
@ivanfioravanti · 2026-08-30T07:01
Code added to llm context benchmark! Prose is ok, but it impacts some speculative decoding techniques. Here Qwen 3.8 Flash Next oQ4 on M3 Ultra 512GB running on oMLX! More engines under testing! → tweet link
@ivanfioravanti · 2026-08-30T06:48
RT @MiaAI_lab: GLM 5.3 Flash EXL3 for 2x DGX Sparks Abliterated (Uncensored) is here 🔥 - off by default, same runtime - same weights, no n… → tweet link
@ivanfioravanti · 2026-08-29T21:51
Various Long Context Benchmarks in progress on Qwen 3.8 Flash Next! Prefill can be really painful on M3 Ultra at very large contexts. I will combine some M3 Ultra, M5 Max and 2 x DGX Spark tests. 💪 And I'm adding a real-time view with charts of the run. I'm tired of waiting hours to see them. → tweet link
Developer Tools & Infrastructure
@sqs · 2026-08-30T17:22
How many bugs would you tolerate if shipping a bugfix took 2 weeks end-to-end? How about if it took 15 minutes? Same or different number? → tweet link
@sqs · 2026-08-30T16:28
amp now supports multiple github accounts (and github orgs that require saml sso) → tweet link
@sudoingX · 2026-08-30T15:29
anon, let me share you about the day local ai stopped being a hobby for me. that day i was out with no internet but my ROG 5090 laptop... i loaded qwen 3.8 27b dense from weights already sitting on my nvme, and i just worked. if ai development ends tomorrow... the weights are on my nvme. nobody can rate limit them. → tweet link
@kunchenguid · 2026-08-30T15:07
i've had enough of all the over-engineering happening during adversarial review loops... so i just spent a whole day tracking down the traces and improved agent instructions in no-mistakes to specifically combat over-engineering → tweet link
@sqs · 2026-08-30T09:44
now you can go outside and walk around, even if you have unsent prompts on your computer. cross-device sync for amp prompt drafts → tweet link
@ivanfioravanti · 2026-08-30T09:38
Adding Unsloth to the mix! M3 Ultra 512GB unsloth/Qwen3.8-Flash-Next-GGUF UD-IQ4_XS! This model architecture is incredible even at large contexts! → tweet link
@ivanfioravanti · 2026-08-29T21:10
RT @old_sound: KV-cache reuse is powered by kvpack, our fast, safe replay layer for LLM inference. kvpack lets an inference engine save co… → tweet link
@Teknium · 2026-08-29T22:58
RT @BkashJoshi: What most people miss about computer-use agents: the mouse matters.🚀 I tested Hermes vs Codex on one Windows desktop. Herme… → tweet link
@kunchenguid · 2026-08-29T23:20
i suddenly realized that i haven’t opened neovim at all for many days... while i’m happy that i’m now getting a lot more done with agents, i do miss the “flow” i used to have in code editors, profoundly. working with agents we have today is almost the opposite of that. the experience is not right. we need a complete rethink → tweet link
AI Applications & Experiments
@levelsio · 2026-08-30T18:54
After [ 💰 Income mode ] I thought what if we can collect live violent crime data and show that per neighorhood too? [ 🥷 Crime mode ] Before AI it'd be too much hassle finding all these data sets, normalizing them... But with AI it's quite do-able. → tweet link
@levelsio · 2026-08-30T18:24
🌟 Passed 2000+ concurrent viewers! Switched to portrait videos cause most usage on mobile. Moved the queue to the left with upvotes, so the most upvoted ones get generated. → tweet link
@KingBootoshi · 2026-08-30T18:03
THE ONLY REASON I WANTED TO USE GROK BOT IS TO RUN THROUGH MY BOOKMARKS BLESS LETS GO!!! if this is the ONLY use case I have for Grok bot i'm happy. → tweet link
@levelsio · 2026-08-30T17:23
I made https://t.co/uga6Of19Cx yesterday completely on my phone with @TermiusHQ on my @Hetzner_Online VPS with @claudeai Code in the sauna while in a group chat of my friends. → tweet link
@KingBootoshi · 2026-08-30T17:22
damn OMP keeps track of my Fable usage and I feel like a whore $1600 total added up in 2 chats after a couple days of usage 🤣 God bless VC money and subsidization → tweet link
@ivanfioravanti · 2026-08-30T12:07
Current LLM coding models are limited by human knowledge used in training, when they’ll start to learn from scratch, coding solutions autonomously, maybe even in a new language, we’ll see the real potential. Who’s building CodingZero? → tweet link
@LinusEkenstam · 2026-08-30T13:42
The entire premise of vibe-coding is to NOT make software for a massive audience. But rather to make things that holds value for perhaps only you. In an era of abundance, everything that can be, will be purely because. → tweet link
@Teknium · 2026-08-30T11:25
Now the place just needs to get populated with Hermes Agents to make a real society and we can have SAO IRL → tweet link
@levelsio · 2026-08-30T08:48
37,000 people watched Infinite Slop yesterday! So I consider it a success and I registered a domain name for it. Let me know if you have ideas how to improve... Thanks @rehan_shei + @marcantoinefon for the idea and @fal for VERY generously sponsoring and making this possible! → tweet link
@ivanfioravanti · 2026-08-29T21:57
I tested MiniMax H3 Max by @fal and it's really fast and furious! 15 seconds video: 768p ~16 secs, 480p ~ 4 secs. → tweet link
@badlogicgames · 2026-08-29T23:24
my mom is using gemini on her android phone to write angry letters to companies/banks who did her wrong by law and they are all giving in, as gemini cites the law and i think this is beautiful. → tweet link
@uwteam · 2026-08-29T21:12
Ciągle słyszę, że 'to już koniec N8N', bo agenty AI ogarną wszystko. Trochę w tym jest prawdy, bo taki agent ogarnia u mnie ostatnio... tworzenie scenariuszy w N8N :D → tweet link
Tech Industry & Strategy
@TrungTPhan · 2026-08-30T17:20
John Ternus takes over as Apple CEO on September 1. The 51-year old SVP of Hardware Engineering has spent 25 years at Apple. → tweet link
@jezell · 2026-08-30T17:02
Why does it always seem like they asked ChatGPT for the estimate with these Quantum computing experiments? I'll take any Quantum computing claims with a grain of salt, especially the ones that come out of IBM. → tweet link
@RayFernando1337 · 2026-08-30T17:59
RT @chamath: BUYER BEWARE. This extremely meticulous article will now be used to start Phase2 of “shut down open source” because “if we can… → tweet link
@ivanfioravanti · 2026-08-30T18:21
open source ai must win → tweet link
@jezell · 2026-08-30T04:22
RT @jreuben1: RISC-V is now officially supported by CPython → tweet link
@kunchenguid · 2026-08-30T04:04
while maintaining my open source projects i noticed many people started using something called Oh My Pi... this is EXACTLY what the bitter lesson told us to avoid. it may indeed work well at the time it’s evaluated, but every model release can invalidate a bunch of these results → tweet link
@FrameworkPuter · 2026-08-30T02:46
Year of the Linux (developer) desktop → tweet link
@KingBootoshi · 2026-08-30T02:43
things i learned using local ai with 5090 (laptop) GPU (24gb vram) - it could run a quantized Qwen 3.8 27B (Unsloth's Qwen3.8-27B-UD-Q4_K_XL) runs at 55 tok/ at a 64k context MAX → tweet link
@juliarturc · 2026-08-29T20:27
Honestly I never understood why providers agreed to be proxied by Cursor from day 1. Imagine Google offering a fully fledged search API, it’d be just dumb. I also don’t think OAI via Cursor is helping consumers all that much. It basically leaks your hard-earned data for free to two companies instead of one. → tweet link
@jezell · 2026-08-29T22:06
People keep underestimating the efficiency of OpenAI's models. Easy to do because they see a price that includes hefty margins on the pricing page, but there is no way in hell they could keep resetting the subs and be serving so many tokens on the subs if the models aren't way more efficient than the pricing page suggests. → tweet link
@TrungTPhan · 2026-08-29T21:59
“Here is our lab. This is where AI comes alive. What used to take months to accomplish, will now take hours.” Would be shocked if Charleston AI doesn’t raise $500m to $750m in the next few months based on the current market. → tweet link
@ivanfioravanti · 2026-08-30T07:49
In this new AI-led coding era, teams of a few people (4 max, but 2 better) are consistently faster and more productive than larger teams. Now, more than ever before. Human interactions and alignment simply slow down everything down. → tweet link