Executive Summary
The AI and tech ecosystem is experiencing a rapid shift towards ultrafast, low-cost agent routing tools, highlighted by the integration of "Jev" which drastically reduces inference costs and latency for agent tasks. Meanwhile, major AI models like GPT-6 Astra, DeepSeek V4.1 Flash, GLM-5.3 Flash, and Qwen 3.8 are pushing the boundaries of autonomous coding, complex problem-solving, and local hardware optimization. On the infrastructure side, Rust continues to gain ground in high-performance systems, including the GitHub Copilot SDK and database rewrites, while local AI hardware is diversifying with major advancements in AMD optimization, Apple's unified memory, and new unified-memory x86 systems like Strix Halo.
Key Events
- Jev integration enables ultrafast, ultra-cheap agent tasks: Browser Use combined with Jev found flights in 7 seconds for just $0.0039, showcasing massive cost and latency reductions for agent workflows. → link
- GPT-6 Astra breaks unsolved WWII Enigma code: The model autonomously broke a 1941 German Army Enigma message that had previously remained unsolved. → link
- GitHub Copilot SDK migrating to Rust: Microsoft and GitHub are moving the Copilot SDK from TypeScript to Rust to improve performance and runtime efficiency. → link
- Arcee AI achieves unicorn status: Arcee AI announced a Series B funding round valuing the company at more than $1 billion. → link
- OpenAI partners with Shopify for ChatGPT advertising: Shopify merchants will now be able to advertise their products directly within ChatGPT. → link
Analysis
There is a clear trend of optimizing the "orchestration" layer of AI, moving away from brute-force LLM token usage for routing and dispatching tasks to specialized classifiers like Jev, which slashes costs by over 70%. In the model space, open-weight flash models (DeepSeek V4.1, GLM-5.3, Qwen 3.8) are dominating local and cost-sensitive coding environments, often replacing slower proprietary models. Hardware-wise, the "Local AI" market is fragmenting into distinct categories: raw bandwidth (Nvidia), massive unified memory (Apple), and emerging x86 unified memory (Strix Halo), with memory bandwidth acting as the ultimate bottleneck. Watch for increased enterprise adoption of these models under strict cost limits (e.g., JPMorgan's Claude Code limits) and further Rust migrations in core developer infrastructure.
Tweet Feed
AI Models & Research
@louszbd · 2026-09-17T18:58
We were bringing up the inference stack for GLM-5.3-Flash. And sitting there watching agent work, we noticed that what it got back after a change matter about as much as how smart it was. That's what we mean by dense feedback. → tweet link
@Teknium · 2026-09-17T17:57
We are going to lean into making Hermes more like Pi, and less like OpenClaw → tweet link
@KingBootoshi · 2026-09-17T15:53
HOLY FUCK BOYS FABLE 5.1 ONE SHOT GANGNAM STYLE (it actually one shot a full control system for injecting policies into the bot safely) THE FUTURE IS FUCKING HERE HAHAHAHAHAHAHAHAHAHAHA → tweet link
@ivanfioravanti · 2026-09-17T15:46
RT @arcee_ai: Today, we are announcing our Series B funding round, valuing the company at more than $1B. → tweet link
@jezell · 2026-09-17T14:04
RT @carterleffen: Two days ago, GPT-6 Astra broke a yet unsolved German Army Enigma message from 1941. Amazingly Astra was able to autonom… → tweet link
@sama · 2026-09-17T04:56
RT @thekaransinghal: Verified U.S. clinicians can get free GPT-6 Astra (Pro) access today via ChatGPT for Clinicians. → tweet link
@victormustar · 2026-09-16T20:43
RT @_LuoFuli: Nearly half a year of silence. We spent it studying one problem: how far RL can scale. MiMo-V2.6 is in the middle of its RL… → tweet link
@ivanfioravanti · 2026-09-17T02:04
Wow @XiaomiMiMo live streaming their RL of the new model with live costs indicator qas not in my bingo card. Incredible! → tweet link
@TheAhmadOsman · 2026-09-16T23:02
I am pretty much no longer using the slow and extremely limited Codex models anymore GLM 5.3 Flash and DeepSeek V4.1 Flash is all you need → tweet link
@ivanfioravanti · 2026-09-17T13:15
For coding locally GLM 5.3 Flash is all you need! 🥈 Qwen 3.8 Flash Next 🥉 DeepSeek v4.1 Flash https://t.co/tqCQh9DoAx → tweet link
@sudoingX · 2026-09-17T01:30
here is the video of qwen 3.8 flash next autonomously building a visually striking website, start to finish, 29 minutes at 9.7x speed. the stack: official fp8 on 2x dgx spark, tensor parallel over one cable, 256k context loaded, 45 tok/s sustained with mtp on, hermes agent driving it from my laptop. → tweet link
@MengTo · 2026-09-17T10:28
GPT-6 Astra is amazing for motion design, especially when you're presenting a UI or an app. I asked it to create a Codex promo using PerspectiveUI from my ThreeUI library. → tweet link
Developer Tools & Software Engineering
@badlogicgames · 2026-09-17T16:30
RT @gregpr07: Breaking: Browser Use + Jev = Ultrafast ⚡ Findings flights took 7s and cost only $0.0039 🤯 > new action space every step >… → tweet link
@kunchenguid · 2026-09-17T06:16
alright - just got Jev deployed for a real production use case... by default, that's done by the firstmate agent and the LLM would have to do some thinking... i just replaced this dispatch process with Jev. it makes the same decision with no thinking or tool calls, done in ~200ms, and for the 25 tasks i evaluated this with, it gives the exact same answer fable would have given... the saving from Jev still resulted in a -71% reduction in cost and -90% reduction in wall time → tweet link
@jezell · 2026-09-17T03:28
RT @davidfowl: Migrating the GitHub Copilot runtime to Rust, using Copilot by @stephentoub https://t.co/FHsLlkiu3Y → tweet link
@jezell · 2026-09-17T15:24
RT @siddontang: From Go to Rust. TiDB rewrite: 1M+ lines, with nearly all TiKV-backed end-to-end tests passing. Still experimental. More… → tweet link
@jezell · 2026-09-16T19:15
RT @NVIDIAHPCDev: Introducing CUDA Rust! CUDA Rust lets you write GPU kernels natively in Rust, not just launch them from it. → tweet link
@sqs · 2026-09-17T12:53
Maybe the most-requested feature for Amp? Now you can start threads from the web in any local project dir, with a single Amp runner. Next! → tweet link
@thdxr · 2026-09-17T00:10
even our beta is going exponential let's see if it hits 100K DAUs https://t.co/B2gA1WsmWC → tweet link
@louszbd · 2026-09-17T14:50
RT @ZixuanLi_: ZCode now supports more model providers, with improved stability and performance. We’ll keep expanding integrations based o… → tweet link
@jezell · 2026-09-17T18:12
So glad people like @TigrisData and @haipingfu are solving the git + object storage problem in open source. No one needs more github style proprietary lockin to scale their repos. → tweet link
@victormustar · 2026-09-16T19:46
Tested ultracode with DeepSeek 4.1 Flash in Claude Code (⚠️ it's not cheap unless you run local) Ask random stuff like "impressive aquatic world with amazing water."... No engine. No assets. No textures. 36 modules of WebGL2 written from scratch. One 1.1MB HTML file. → tweet link
Hardware & Local AI Infrastructure
@TheAhmadOsman · 2026-09-17T01:18
Memory bandwidth for Local AI hardware matters a lot more than most people think... Here is the current local AI hardware ladder: RTX PRO 6000 Blackwell, RTX 5090... Mac Studio M3 Ultra... DGX Spark... Strix Halo / Ryzen AI Max... Start asking: > what must fit? > what bandwidth tier do I need? → tweet link
@ivanfioravanti · 2026-09-17T16:02
One of the great pluses of Apple MLX is that the team has access to hardware earlier than anyone else. M5 Ultra tunings are already backed in for day-0 support! (yo'll have to build it from sources) → tweet link
@ivanfioravanti · 2026-09-17T07:28
This blog post is great! It shows how GLM 5.3 helped @Zai_org engineers to build a complete production-grade inference service from scratch on a cluster of more than 100,000 Chinese-made AI accelerators. → tweet link
@sudoingX · 2026-09-17T00:01
amd users, rejoice. someone put an rx 7900 xt through the same a/b as everyone else, on omarchy, rocm 7.2, headless, and the gaming card flies: 30.8 tok/s stock to 54.0 tok/s with one llama.cpp flag on, +75%, qwen 3.8 27b dense at 64k context on 20gb of vram with 2.5gb to spare. → tweet link
@FrameworkPuter · 2026-09-17T04:43
We did a video with our friends at Micron around local AI and how we use memory. Please watch it so we can get more memory. https://t.co/wr7qr7dR4J → tweet link
@jezell · 2026-09-16T20:13
RT @MacRumors: Apple May Return to Server Market With Nvidia Technology https://t.co/utiS53Vq9y https://t.co/ev2RKnw9BY → tweet link
Tech Industry & Startups
@jxnlco · 2026-09-17T16:43
RT @leonardtang_: I am thrilled to announce that @beaconholdings has acquired @haizelabs, with me joining as VP of AI Research. → tweet link
@jezell · 2026-09-17T16:10
RT @choblin29: 🚨EXCLUSIVE: JPMorgan just put Claude Code on a $2,000/month leash for some engineers and is moving it into a locked-down env… → tweet link
@gdb · 2026-09-16T23:40
excited to help Shopify merchants advertise their products in ChatGPT: → tweet link
@gdb · 2026-09-16T22:57
wall-to-wall deployment of astra for engineers at databricks: → tweet link
@kunchenguid · 2026-09-16T20:29
reminder that claude quota reduction is already in effect since Sept. 14... my empirically measured value for a claude $200 plan has reduced from $7200/mo to about $6000, sitting below the value of ChatGPT Pro 20x... i think we're entering an era where stacking multiple subscriptions will be somewhat mainstream → tweet link
@LinusEkenstam · 2026-09-16T20:27
RT @higgsfield: Introducing Higgsfield API. 50+ frontier models in one API, at lower prices than a subscription. > Get up to 50% OFF disc… → tweet link
@sama · 2026-09-16T22:31
the main thing i was excited about launching this week will be next week instead, but imo worth the wait! → tweet link