Executive Summary
The dominant story of the day is Google's release of Gemma 4 12B, a unified multimodal model (text, image, audio, video) that was rapidly ported to MLX, Ollama, and quantized to run locally on just 8GB RAM. Open-source AI agent tooling also saw a major launch with Nous Research's Hermes Desktop client and Portal gateway, while MiniMax's M3 model entered the top-10 on benchmarks with sparse attention and 1M context. Meanwhile, Codex (OpenAI) faced widespread user backlash over dramatically reduced rate limits and outages, pushing developers back to alternatives like Claude. On the infrastructure side, a developer demonstrated running 42 parallel AI agents on a local cluster of 14 RTX 3090s—signaling that local-first agent deployment is becoming viable.
Key Events
- Google releases Gemma 4 12B: New dense unified multimodal model supporting audio, image, video, and text. Immediately optimized for local inference via MLX, Ollama, and Dynamic GGUFs (8GB RAM). → link
- Hermes Desktop launched by Nous Research: Cross-platform native GUI for the Hermes agent, supporting local and remote/cloud model connections via Tailscale. → link
- MiniMax-M3 enters top rankings: Sparse attention + 1M context + multimodality; inference quickly scaled after launch-day traffic surge, stabilizing at 50–70 TPS. → link
- Ideogram 4.0 released: Claimed best open image model, with downloadable weights for fine-tuning. → link
- Codex limits severely nerfed: OpenAI's coding agent hit with major rate limit reductions after a promotional 2x period ended; $200/month Pro users report near-unusable quotas. Widespread 403/429 outages reported. → link
- Codex + GPT-5.5 ports 210k lines of LibreOffice to Dart: Demonstrated sustained multi-day agentic coding workflow at massive scale. → link
- Local 42-agent cluster demonstrated: 14× RTX 3090s running Qwen 3.6 27B with exl3 quantization and Aphrodite engine at full 256k context. → link
- Sam Altman endorses new US AI Executive Order: States it "gets the balance right" between AI leadership, safety, and cyber defense. → link
- ACE-Step open-source AI music generation reaches quality parity: Marked as a breakthrough moment for open-source music models. → link
Analysis
Patterns & Trends: - The local inference wave is accelerating. Every new model release this cycle (Gemma 4 12B, Hermes, MiniMax-M3) was immediately optimized for consumer hardware—8GB RAM, MLX, Ollama, exl3 quantization. The ecosystem is converging on "run anything anywhere." - Agent tooling is commoditizing fast. The Hermes Desktop + Portal stack, Codex, and various agent-agnostic workflows (as @kunchenguid advocates) signal that multi-agent orchestration is moving from lab experiment to daily developer tool. - Provider lock-in fatigue. Multiple developers expressed frustration with Codex rate limits and outages, highlighting the advantage of agent-agnostic and on-premise strategies. @TheAhmadOsman explicitly offers consulting to migrate enterprises off Claude Code/OpenAI to self-hosted LLMs.
Escalation: - Open AI and Anthropic are both tightening usage limits on their coding agents, which may accelerate adoption of open-source alternatives (Hermes, local deployments). - Chinese labs (Moonshot, DeepSeek) are cited as "S-Tier" by practitioners, signaling intensifying competitive pressure on Western model providers.
What to Watch: - Whether OpenAI walks back Codex limits given community backlash, or if it permanently pushes users toward alternatives. - MiniMax-M3's trajectory on benchmarks now that inference is stabilized. - The "vibe-coded slop" vs. serious agent software debate—expect a market correction for AI-wrapped products shipping low-quality output.
Tweet Feed
AI Model Releases & Updates
@victormustar · 2026-06-03T16:10
🚨 New open-source Google model: Gemma-4-12B → tweet
@victormustar · 2026-06-03T16:14
RT @osanseviero: Super excited to introduce Gemma 4 12B! 💎 - Multimodal: audio, image, video, and text input - Novel architecture: we remo… → tweet
@badlogicgames · 2026-06-03T18:26
RT @Prince_Canuma: 🚀 Gemma 4 12B is here! We partnered with @GoogleDeepMind to bring and optimize their new dense and unifed multimodal mo… → tweet
@gospaceport · 2026-06-03T17:08
RT @UnslothAI: Gemma 4 12B can now run locally on just 8GB RAM via Dynamic GGUFs. Google's new model, Gemma 4 12B Unified supports image,… → tweet
@SkylerMiao7 · 2026-06-03T04:37
M3 brings sparse attention + 1M context + multimodality, and Together did the hard serving work to make it fast. Great collaboration with the Together team. → tweet
@SkylerMiao7 · 2026-06-03T03:13
M3 traffic got wild, so we shipped overnight. Inference serving upgraded at 22:00 Beijing / 7:00 AM PT. TPS much smoother now. Most users should be seeing 50–70 TPS. → tweet
@SkylerMiao7 · 2026-06-03T07:28
RT @RyanLeeMiniMax: MiniMax-M3 is rank #6 on https://t.co/myhk0Z6tSp → tweet
@victormustar · 2026-06-03T16:19
RT @ideogram_ai: Introducing Ideogram 4.0: the best open image model in the world. Think it. Make it. Own it. Download the weights, fine-… → tweet
@victormustar · 2026-06-03T08:38
Don't understand anything but it's a banger 🔥 thanks to ACE-Step open source AI music has caught up! → tweet
@TheAhmadOsman · 2026-06-03T17:45
S-Tier Chinese Labs: Moonshot and DeepSeek. These 2 are levels above everyone else → tweet
AI Agents & Developer Tools
@Teknium · 2026-06-02T23:38
RT @Teknium: It's finally here. The official Hermes Desktop app. Available on all platforms. → tweet
@Teknium · 2026-06-03T01:32
Just pushed an update to help remote connecting with the Hermes Agent GUI over tailscale to function! Please update if you had any issues! → tweet
@Teknium · 2026-06-03T01:48
RT @NousResearch: The next evolution of Hermes Agent is here! Introducing Hermes Desktop: everything you love about Hermes, now native on… → tweet
@Teknium · 2026-06-03T01:18
RT @NousResearch: Nous Portal is the simplest way to power your Hermes Agent. Run 'hermes portal' to switch today. → tweet
@Teknium · 2026-06-03T04:27
RT @cobi_bean: Hermes Desktop dropped today, so i hooked my cloud agent into it. → tweet
@Teknium · 2026-06-02T19:43
The GUI can connect to remote instances like so: → tweet
@Teknium · 2026-06-02T23:35
PSA: we have no official mobile app. The official desktop app is only found here: https://t.co/fXwDEwJyz5 → tweet
@ollama · 2026-06-03T03:20
You can use Hermes Desktop with Ollama using local or cloud models. Get started 👇👇👇 → tweet
@gdb · 2026-06-03T01:48
Build and launch apps to your team, using Codex: → tweet
@gdb · 2026-06-03T07:55
codex for computer work is growing very fast → tweet
@jezell · 2026-06-03T04:55
After 4 days with /goal, Codex + GPT5.5 have ported 210k lines of LibreOffice to dart with the associated tests. Still humming along. I think it's gonna be a while still. → tweet
@MengTo · 2026-06-03T06:56
I've been using Codex since day 1. I built 3 massive web apps, 1 mac app and 1 iOS app with it. It's honestly the best app I've used this year. Seriously thinking about creating a course on this. Anyone interested? → tweet
@RydMike · 2026-06-03T07:04
Hmm @OpenAI pulling an @AnthropicAI with the Codex limits? 🤔 Certainly may feel like it when the 2x promo ended. Btw the $20 plan did not have 2x and is only good for about 2 prompts before the 5h limit is done, so it is almost useless even as a trial. → tweet
@RydMike · 2026-06-03T07:00
RT @bridgemindai: Codex limits are NERFED. With my $200 ChatGPT Pro plan I used to never worry about my limits using Codex. → tweet
@kunchenguid · 2026-06-03T05:02
codex is completely down. 403 / 429 on every request. back to claude. having my workflow being fully agent-agnostic pays off almost every day → tweet
@jezell · 2026-06-03T05:40
Codex is taking the night off folks → tweet
@Teknium · 2026-06-03T14:29
RT @anildelphi: Honestly feel like this will be the ChatGPT moment for agents. Helped two friends who have barely used AI set up their Herm… → tweet
@nummanali · 2026-06-03T06:46
Codex / ChatGPT design system is become very inviting. It makes me feel like the future is mind with the subtle glows and gel like icons. Would love to know how they did the design research → tweet
@TrungTPhan · 2026-06-02T23:36
RT @bearlyai: FYI. If you've been using agentic AI coding agents, try OpenADE (our free open-source AI coding tool for easy collaboration a… → tweet
ML Infrastructure & Compute
@TheAhmadOsman · 2026-06-03T02:52
14x RTX 3090s + Qwen 3.6 27B. Running 42 agents IN PARALLEL at full 256k context. - exl3 6bpw - fp8 KV Cache - Aphrodite Inference Engine w/ tp=2, pp=7. The world of agents will run locally btw → tweet
@TheAhmadOsman · 2026-06-03T18:53
Evals. Data. Compute. In that order → tweet
@tinygrad · 2026-06-03T17:30
Because it's the full stack from Tensors to MMIO, the ceiling on speed in tinygrad is higher than in any other framework. → tweet
@gospaceport · 2026-06-03T17:06
This poat conforms the 3090 is still pretty damn kino. Also why I only dabbled with a pair of 5060ti's when I read about "future support" like those 2 words send chills. → tweet
@sudoingX · 2026-06-03T18:31
this is what a telescope actually sees. not a planet, not even a clean star, just a smudge of light smeared across eight pixels. that bright blob on the left is the entire star, KIC 6922244, as raw as it gets. you add up those pixels every single frame and you get one number, brightness… somewhere in that flat line, a fraction of a percent deep, is a real world crossing its star. you pull it out by understanding the geometry… that's a kind of spatial intelligence no textbook hands you. → tweet
@sudoingX · 2026-06-03T18:13
this is the first chart i ever made on this project. that sharp spike dropping out of the noise is a real planet crossing its star… then i built a neural net from scratch, and taught it to find these in the noise the way the real pipelines do. you've got the most capable machine ever built within arm's reach and you're using it to reword emails. i pointed mine at the actual sky. → tweet
Industry Commentary & Business Strategy
@TheAhmadOsman · 2026-06-02T23:22
Grifters shipping vibe-coded slop are everywhere now and it is getting exhausting ngl. The issue is not that they are vibe coding… The issue is pretending the output is serious software when it belongs in the trashbin. A lot of these people are not building products - They are producing screenshots, GitHub activity, Fake momentum. Very happy for them though: The graph is green, the software is not → tweet
@TheAhmadOsman · 2026-06-03T16:14
If you're business / enterprise is trying to migrate off Claude Code / OpenAI and want to host your LLMs on-premise I do consulting for that kind of stuff btw. Bonus: you'll be setup with a path forward to training models on your tasks / workflows and save so much $$$ long term → tweet
@kunchenguid · 2026-06-03T17:18
more and more people started to talk about "AI makes us busier than ever" and the discussion is becoming more and more misguided… we're busy because many people choose to be busy, and we've structured our society to reward business. this has nothing to do with AI. → tweet
@kunchenguid · 2026-06-02T20:53
finally made this video to fully unpack the tokenmaxxing game, and the economy behind the AI industry… how did tokenmaxxing start, why companies ask employees to burn tokens, the capital games being played, what to do as leaders, what to do as employees → tweet
@ASalvadorini · 2026-06-03T14:21
The number of tokens you consume is not a metric. It matters: (1) what you're building with it (2) which problem solves (3) when you are shipping it. Otherwise it's just a waste of money 💰💰💰💰💰 → tweet
@mipsytipsy · 2026-06-02T23:11
i wrote a post about the growing divide between AI skeptics and AI enthusiasts. wins and costs are both real, but too often fall on different groups of people. when you're only experiencing half the story, it's too easy to write each other off. → tweet
@thdxr · 2026-06-03T15:27
every hit product in the past few years could have been made by an established company. shopify could have made cursor, airbnb could have made claude code, stripe could have made lovable. everyone will have good reasons as to why but remember amazon made aws → tweet
@thdxr · 2026-06-03T15:17
RT @alan__rice: AI prices are getting ridiculous → tweet
@TheAhmadOsman · 2026-06-03T14:49
Using Windows in the age of AI is a permanent underclass move btw → tweet
@TrungTPhan · 2026-06-03T15:52
when the agent on your AI PC sees you open an incognito browser tab, lock the door and plug headphones into the laptop → tweet
@TheAhmadOsman · 2026-06-03T06:09
This will be true by Summer 2027 → tweet
@badlogicgames · 2026-06-03T18:08
fell into a youtube ai slop hole of fan films. and it's hilarious. kal-el and his love interest keep shaking hands. it's also very obvious what the training data composition is. beautiful slop. → tweet
Startups & Product Strategy
@MengTo · 2026-06-03T10:12
Best decision I made this year: switching to paid trials. Free users were 50% of our token costs, brought abuse and spam, and rarely converted. Paid trials converted better, lowered noise, and increased customers. As a bootstrapped product, I'd rather serve serious users than compete with VC-funded free plans. → tweet
@levelsio · 2026-06-03T17:58
This is how you do a great startup video. Humble and chill and a real cool backstory → tweet
@levelsio · 2026-06-02T22:44
RT @eostudi0: Meet @yasser_elsaid_, founder of @chatbase. He hit $10M ARR without raising a dollar. We asked him 15 questions to break dow… → tweet
@victormustar · 2026-06-02T20:38
RT @ClementDelangue: Arcee needs more attention that it gets! There aren't a lot of great American open-source AI model companies and they'… → tweet
Open Source & Developer Projects
@kunchenguid · 2026-06-03T05:53
just merged 2026 open source PRs in 2026, close to 5k stars across my projects since i quit my job and started building - you can tell from the chart when that was :) → tweet
@kunchenguid · 2026-06-03T03:38
baby-menu v0.1.12 - bunch of bug fixes and reliability improvements since initial release. this is my baby-menu after talking to it for a bit - what would you want as your personal menu bar app? i'm now getting addicted to this new paradigm of hyper-personal software → tweet
@jsuarez · 2026-06-03T18:03
Reinforcement learning research with Joseph Suarez → tweet
@hnasr · 2026-06-03T03:43
Root Cause paperback now available on Amazon India. Grab your copy → tweet
@ASalvadorini · 2026-06-03T12:22
RT @dedene: POV: you're still using GitHub Copilot after June 1st, 2026 → tweet
@jezell · 2026-06-02T21:37
RT @craigaloewen: WSL now has built in Linux container support with both a CLI and an API, announced today and coming soon by the end of th… → tweet
@steipete · 2026-06-03T12:11
RT @_lopopolo: This is pretty rad. coreutils on Windows, based on uutils → tweet
Hardware
@FrameworkPuter · 2026-06-03T00:30
We spent the last few months overhauling our logistics infrastructure around refurbishment, and today we launched a broad set of refurbished products into the Framework Outlet. We'll be able to turn around customer returns into refurbs faster now too! → tweet
@FrameworkPuter · 2026-06-03T00:08
LPCAMM2 memory is one of the biggest improvements we've brought into Framework Laptop 13 Pro, enabling higher throughput and better power efficiency without sacrificing upgradeability. We published a blog post today deep diving into how we designed for it. → tweet
Policy
@sama · 2026-06-03T00:48
theUSshould lead on AI by continuing to develop the very best models, making sure they're safe, and getting cyber tools into the hands of trusted defenders. the new EO gets the balance right. → tweet