Executive Summary
The last 24 hours saw major shakeups in the AI coding tools ecosystem, highlighted by OpenAI's announcement to terminate Cursor's API access starting November 12th, which has driven developers to explore alternative harnesses like AmpCode, Hermes Agent, and Opencode2. On the hardware front, discussions around Hot Chips highlighted divergent memory architectures, such as Cerebras's SRAM-centric approach and new 3D DRAM stacking techniques aimed at solving AI's memory bandwidth bottlenecks. Additionally, local AI deployment is gaining momentum as users run powerful models like Qwen 3.8 and GLM 5.3 on consumer-grade GPUs and Apple silicon, while real-time AI video generation reached a milestone with Fal's optimized Minimax H3 model capable of generating video faster than playback speed.
Key Events
- OpenAI severs ties with Cursor: OpenAI will block Cursor from accessing its models by Nov 12th, causing a ripple effect in the dev community as Anthropic quickly reaffirms its partnership with Cursor. → link, → link
- Real-time AI video generation achieved: Fal released a fine-tuned Minimax H3 model (Max) that generates video 50x faster, allowing for continuous interactive AI livestreams. → link
- Major hardware architectural shifts revealed: At Hot Chips, companies like Cerebras and OpenAI showcased divergent memory architectures, moving towards 3D stacked DRAM and large-scale SRAM to overcome inference memory bottlenecks. → link, → link
- GLM 5.3 rolls out on Ollama: GLM 5.3 and GLM 5.3 Flash (previously Ox Alpha) are fully available on Ollama's cloud, with users praising its agentic and reverse-engineering capabilities. → link
- Local AI performance surges: Developers successfully run large models like Qwen 3.8-27B on consumer hardware (e.g., laptop 5090s) at ~55 tok/s, reinforcing the trend of self-hosted AI. → link
Analysis
The tech ecosystem is experiencing a fragmentation in AI coding tooling. OpenAI's move against Cursor is catalyzing a migration toward terminal-based and open-source harnesses (Amp, Opencode2, Hermes), signaling that developers are wary of vendor lock-in. Concurrently, the push for "owning your cognition" via local hardware is accelerating as local model performance on consumer hardware (RTX 5090, Apple M5 Max) rivals cloud offerings. On the hardware front, the industry is actively reinventing the memory hierarchy, recognizing that inference is increasingly memory-bound rather than compute-bound. Watch for further integration of AI agents into daily OS environments and increased competition among alternative model providers like Qwen, DeepSeek, and GLM.
Tweet Feed
AI Industry & Business News
@sudoingX · 2026-08-29T07:13
BREAKING: OpenAI CEO Sam Altman has terminated Cursor's access to OpenAI models effective november 12th, because he cannot be confident that a company owned by OpenAI's own co-founder will honour the terms of service. the co-founder is elon musk. the evidence cited is his sworn testimony, given in the lawsuit he filed against OpenAI for not staying open. → tweet link
@kunchenguid · 2026-08-29T04:01
many people haven't noticed the significance of this yet this is NOT just between OpenAI and Cursor if you watch closely, you'll see that Anthropic immediately committed a continued partnership with Cursor, hence SpaceXAI this indicates Elon and Dario have formed an alliance and Sam is now fighting alone against two massive competitors → tweet link
@jezell · 2026-08-29T04:35
I feel sorry for everyone using Cursor. Not because of this news though. I feel sorry for them because they are using Cursor. → tweet link
@jezell · 2026-08-28T20:37
RT @djcows: AI was supposed to give us more free time but instead it 10x'd output and 20x'd expectations → tweet link
AI Hardware & Infrastructure
@MilksandMatcha · 2026-08-29T03:55
One of the most exciting themes at Hot Chips this year was watching DRAM move into the third dimensions. D-Matrix and Cerebras both announced 3D stacked DRAM on their product roadmaps this year → tweet link
@MilksandMatcha · 2026-08-29T03:48
DRAM vs. SRAM was a big topic at this year's hot chips ...The tradeoff is capacity (DRAM) vs. bandwidth (SRAM). SRAM can sit directly next to compute and deliver enormous bandwidth... @cerebras takes the opposite extreme: putting 44 GB of SRAM directly on the WSE-3, with 21 PB/s of on-chip memory bandwidth. → tweet link
@TheAhmadOsman · 2026-08-28T19:59
M5 Ultra Studio vs RTX PRO 6000 1.2 TB/s in a single M5 Ultra Mac Studio is impressive, however, it is still NOWHERE near 2x/4x/8x RTX PRO 6000 setups... So, for 8x RTX PRO 6000s, we have 8 memory controllers processing things in parallel, which is about ~14.3 TB/s* of aggregate local memory bandwidth → tweet link
@sudoingX · 2026-08-29T07:45
just cancel your chatgpt subscription. $20 a month rents you your own thinking... a used 3090 is $900 and it runs a 27b model that outscored opus 4.6 max on livecodebench, offline, with no account attached to it. buy the card once, download weights nobody can revoke... → tweet link
@KingBootoshi · 2026-08-29T16:36
I LOVE MY FRIENDS HOLY @actualinc just added a GH200 to my computer cluster i will be sharing my journey on exploring local AI with this i am VERY interested in interacting with systems at thought speed, though tbh i have no idea what this GPU can do → tweet link
@RayFernando1337 · 2026-08-29T18:50
Let him cook. This is a fun time to be owning 2 DGX Sparks RN → tweet link
Developer Tools & Environments
@sqs · 2026-08-28T20:52
if you switched from {claude code, cursor, devin, codex, ...} to amp: (1) why? (2) what advice would you give to someone considering the same switch? → tweet link
@sqs · 2026-08-29T10:14
who wants to try some new model routing and custom model URL options in Amp this weekend? there are some that will maybe surprise you and make you laugh → tweet link
@thdxr · 2026-08-29T06:32
the opencode2 api is pretty good the obvious implication is you can build client apps against it (it's how our tooey and gooey work) but it also means someone can re-implement the server ... → tweet link
@Teknium · 2026-08-29T14:28
Thanks for the feedback on /btw - now /bg (background will behave as btw did, a fresh session in the background, response piped back to you in the session your working in, and /btw will fork your session off in the background → tweet link
@Teknium · 2026-08-29T13:31
Made a stand-alone Hermes Agent plugin for BackSearch - a wayback machine-like SaaS for agents - adds two tools, and requires their API key. → tweet link
@levelsio · 2026-08-29T14:39
When something you make get lots of irrational haters, usually you're on to something Omarchy by @DHH passes that test with flying colours 😂 → tweet link
@RydMike · 2026-08-29T10:25
Ok so Omarchy is blowing up. Fine, but how are you all developing and testing iPhone builds with it on simulators and real devices, compiling it and testing locally, when developing with native, Flutter or React Native? 🤔 Asking for a friend 😅 → tweet link
@jxnlco · 2026-08-28T23:05
RT @Rudeg: Custom sidebar sections just landed in ChatGPT/Codex desktop app 🗂 organize your tasks and projects into sections or just ask… → tweet link
@ASalvadorini · 2026-08-29T05:42
A small thread 🧵 about FutureBuilder, how its side effects break the synchronicity of your build() method, and why you shouldn't use it.
flutter #Flutterdev
@jezell · 2026-08-28T21:28
Bazel solves build problems a lot better than worktrees do... → tweet link
AI Agents & Workflows
@TheAhmadOsman · 2026-08-29T05:26
I just told GLM 5.3 Flash to put The Offer tv show on Plex for me on my living room NVIDIA Shield It literally turned my Shield on, opened to Plex, chose my profile, found the search, typed, searched, picked the show, put it on, confirmed audio & subtitles are in my preferences → tweet link
@KingBootoshi · 2026-08-28T21:51
the worst part about most agent harnesses is that skills are loaded into the harness for the agent to call themselves i think that's so ass, i hate when my agents use a skill on its own. i invoke skills for a reason. I AM THE COMBO CREATOR → tweet link
@jxnlco · 2026-08-28T19:32
research slows down in the handoffs between tools. rosalind work bench is now in research preview: guided tasks, scientific viewers, and genomics workflows that keep your question, analysis, and evidence connected. → tweet link
@iamdevloper · 2026-08-29T12:29
I'm impressed... https://t.co/RkBrlkTmnI → tweet link
AI Models, Research & Video Generation
@levelsio · 2026-08-29T09:15
Today is a very historical moment for AI video generation You can now generate AI video faster than you can watch it @fal made a post-trained Minimax H3 variant called Max which is 50x faster than the original but still maintains quality It generates 15 seconds of video in 9 seconds! → tweet link
@levelsio · 2026-08-29T17:34
Okay I built it! 🍰 Infinite Slop An infinite and interactive AI generated live stream of slop that goes on forever and ever Anything that you write in the chat is generated next and AI will try to connect it to the previous video... → tweet link
@ollama · 2026-08-29T00:13
GLM 5.3 and GLM 5.3 Flash (previously Ox Alpha) are fully rolled out on Ollama's cloud. Private. Fast. US and Europe hosted. Zero data retention. → tweet link
@ivanfioravanti · 2026-08-29T15:05
Hermes Agent + DeepSeek V4 Flash 0731 running locally on whatever hardware you can run it, it's match in heaven! I will try the new Qwen 3.8 Flash Next and GLM 5.3 Flash, but I want to say that I really like this combo! → tweet link
@KingBootoshi · 2026-08-29T03:18
i got Qwen 3.8-27B running on the laptop at ~55 tok/s LOL insane actually i just asked Fable to do it for me laptop 5090 btw → tweet link
@ivanfioravanti · 2026-08-29T16:20
Here Qwen3.8-Flash-Next-MLX-Serve-4bit running on Macbook Pro M5 Max using mlx-serve compiled from sources. Now I'll try to speed the kernel even more and then I'll plug Hermes Agent into this beast! Be ready @ddalcu 🚀 → tweet link
@kunchenguid · 2026-08-28T22:15
a few days ago @bcherny told me i can fix opus 5 with the concise output style i gave it a real try, and i really wanted it to work. But.. i'm sorry boris but it genuinely doesn't opus 5 still keep being wrong non-stop. i think it's too eager to make conclusions → tweet link
@louszbd · 2026-08-29T06:08
RT @ZixuanLi_: You can now fine-tune GLM-5.3 with @tinkerapi. It’s the first GLM model on the platform and quite possibly the strongest mod… → tweet link