Executive Summary
The past 24 hours in tech and AI were dominated by the explosive progress of autonomous coding agents and open-source models. OpenAI’s Codex achieved a massive milestone by porting 1 million lines of LibreOffice to Dart, while NousResearch’s Hermes Agent broke the 1 trillion tokens-per-day barrier and topped agent leaderboards. The hardware and model landscape continues to shift toward local inference, with developers pushing models like Qwen 3.6 27B on consumer RTX 3090s and running 16x parallel instances on DGX Sparks. Meanwhile, new releases like Bytedance's Seedance 2.5 (with native 4K) and Krea 2 (open weights) showed that the generative AI arms race is only accelerating.
Key Events
- OpenAI's Codex successfully ports 1 million lines of LibreOffice code to Dart, burning billions of tokens in the process → link
- NousResearch’s Hermes Agent surpasses 1 trillion tokens processed in a single day on OpenRouter and reaches 200k GitHub stars → link
- GPT-5.6 is spotted lurking in the Codex GitHub repository → link
- Bytedance releases Seedance 2.5, featuring native 4K, 30-second clips, and 3D white model support → link
- Krea 2 drops open weights for its Raw and Turbo models on Hugging Face → link
- xAI's Grok Build demonstrates advanced 3D and physics generation, building interactive worlds from single prompts → link
- OpenAI announces "Patch the Planet," using frontier AI to secure critical open-source software projects → link
- GitHub Copilot App launches BYOK (Bring Your Own Key) support, integrating with local runners like Ollama → link
- AI successfully makes Postgres multithreaded in 10 days, modifying 124k lines across 786 files → link
- OpenAI DevDay 2026 applications officially open for San Francisco event → link
Analysis
Patterns: The AI landscape is experiencing a rapid bifurcation between frontier cloud models and capable local inference. Agents like Hermes and Codex are moving from novelties to production-grade tools handling massive codebases. Meanwhile, developers are increasingly focused on hardware optimization—running models on DGX Sparks, pushing local Qwen variants on RTX 3090s, and debating the boundaries between local and cloud inference.
Escalation: The competitive tension in AI agents is escalating. Community members note OpenAI-funded accounts targeting Hermes Agent, while xAI's Grok Build is positioning itself as an accessible, frontier-grade tool directly inside a social subscription. The tooling ecosystem is also consolidating, with platforms like Firstmate and OpenCode racing to offer seamless daemon-modes and self-supervision features.
What to watch next: Keep an eye on GPT-5.6's official rollout and how it integrates into Codex. Qwen's next model checkpoints could severely disrupt the local AI hardware market if they ship a highly capable dense model. Expect further integration of AI into developer environments, particularly around AI-driven OSS security (Patch the Planet) and local-first agentic workflows.
Tweet Feed
AI Models & Capabilities
@LinusEkenstam · 2026-06-23T05:27
Seedance 2.5 🤯 → 30-second clips → Native support for 4K → Supports 50 full-modal reference materials, INSANE inc. → Supports 3D white models Simultaneously launches AI copyright commercialization platform. Bytedance is leading the way here https://t.co/Le6X0m18QL → tweet link
@victormustar · 2026-06-23T15:13
RT @krea_ai: today, we release the open weights of Krea 2. welcome Krea 2 Raw and Krea 2 Turbo, an undistilled model from mid-training mea… → tweet link
@sudoingX · 2026-06-23T16:12
grok build is cracked at 3d and physics. i don't think most people have caught up to how good it actually got. look, i sat down and described a world i had in my head, and handed it to grok build. that's it, one prompt. and it went and built the entire thing in code from nothing, no template,no me stepping in to fix things halfway through. one shot. and it's not a render or a screenshot, it's a living 3d world you can actually move through and look around in, and it just runs. i kept waiting for the part where it falls apart and i have to untangle some mess, and it never came. it just worked. sat there watching it come alive on my screen and genuinely forgot to do anything else for a while. and the part that gets me is that this is just sitting inside my x premium+. i'm not paying for some separate dev tool, it's right there, and i've been building with grok build almost every day now. → tweet link
@jezell · 2026-06-23T16:12
RT @scaling01: GPT-5.6 spotted in Codex GitHub repo https://t.co/XdH1FNCjHa → tweet link
@Ex0byt · 2026-06-22T22:50
from our friends over at MiniMax_AI. MM3 review when we have some free time. https://t.co/33M4ZNZRsT → tweet link
@Ex0byt · 2026-06-23T02:03
I guess that makes GLM-5.2 an open source rifle & MiniMax-M3 a multimodal light saber with night specs and lasers? → tweet link
@sudoingX · 2026-06-22T19:27
banger! when the ai subscription runs out, the 3090 clocks in. qwen 3.6 27b dense is getting dangerously close to "good enough." there are people spending $900 every year on phones and telling me a gpu is expensive. local ai is inevitable. → tweet link
@sudoingX · 2026-06-22T19:42
somewhere in alibaba there is an engineer sitting on the exact model i need. qwen team, a 3.7 27b checkpoint. that's all i ask. → tweet link
@Teknium · 2026-06-23T02:19
Wait we actually just broke 1T tokens in a day for the first time on OpenRouter :O Please keep contributing to the most awesome project I've ever worked on to help make Hermes Agent the best software stack on the planet! Thank you contributors🍻🍻 https://t.co/BSYK9sYGGD → tweet link
@sudoingX · 2026-06-23T07:31
Hermes agent casually sitting at the top of the leaderboard like it didn't just lap the entire field 😎 massive milestone. and somehow it still feels like day one. congrats🍻 https://t.co/EBV3lP4DOD → tweet link
AI Agents & Developer Tools
@jezell · 2026-06-23T17:27
New milestone. Codex has now ported 1M lines of LibreOffice code to dart.... still burning billions of tokens around the clock. https://t.co/DTf9ydiS6t → tweet link
@MengTo · 2026-06-23T12:04
About 30% of my Codex work now starts on my iPhone, especially when I travel. I’m in Japan right now, connected locally to my MacBook at the hotel. I can be anywhere, send the task, and Codex handles the repo on my Mac. https://t.co/vJkKdsg6Ow → tweet link
@kunchenguid · 2026-06-23T00:20
just merged a pretty awesome contribution PR to firstmate - https://t.co/NVtYmZQVoJ captain can now run "/afk" and walk away - the crew will enter a token-efficient daemon mode and self-supervise without getting stuck waiting for user messages https://t.co/zpdsq1tUKw → tweet link
@ollama · 2026-06-23T07:52
RT @_Evan_Boyle: BYOK is now live in the GitHub Copilot App! Works with @ollama, foundry, and any OAI completions or Anthropic compatible… → tweet link
@gdb · 2026-06-22T22:14
Patch the Planet: using frontier AI and working with professional security researchers to secure critical OSS projects → tweet link
@RealGeneKim · 2026-06-23T14:17
What an awesome representation of the AI SDLC -- best I've ever seen, because it shows the most important axis: time It wonderfully depicts the collapse of lead times / process times, which most of us have experienced. Literally, weeks goes down to hours or minutes. → tweet link
@sudoingX · 2026-06-22T20:07
babe wake up
openclaw feels like it's optimizing for possibility space. hermes agent feels like it's optimizing for outcomes. that distinction is about to matter a lot. → tweet link
@cooltechtipz · 2026-06-23T18:17
Guide to choosing the right AI model. https://t.co/mPM2MM7lnU → tweet link
@cooltechtipz · 2026-06-23T17:42
AI does not spend the same effort on every question. Two people can ask the same thing and use very different amounts of resources.That's quite different from traditional software. → tweet link
Hardware & Infrastructure
@alexinexxx · 2026-06-23T18:17
Sorry. I lied. I'm not new to X. I don't want to meet in Paris. The only thing I want is 8×B200s. Eight fucking B200s. I want to build a dflash drafter, not get men in my DMs. → tweet link
@badlogicgames · 2026-06-23T16:57
RT @onusoz: 16x parallel Gemma-4-26B-A4B-NVFP4 runs 🤯🤯🤯 18 output tokens/s, aggregate 300 tok/s 1 DGX Spark with 128 GB unified memory → tweet link
@sudoingX · 2026-06-23T10:45
hermes agent on step 3.7 flash building a benchmark on my dgx spark without missing a beat. most reliable agentic stack i've run. every other agent i've tried buries you in bloated tools you never use. openclaw was the worst of it. felt like coding with an anchor tied to my leg. stop running bloatware. run something that executes. → tweet link
@sudoingX · 2026-06-22T19:19
single 3090 in 2026 is... → tweet link
@badlogicgames · 2026-06-23T10:57
RT @samsja19: prime-rl can now train 1T parameters MoE blazingly fast, under 5 minutes per step, or 1k steps in ~3 days → tweet link
@ollama · 2026-06-22T21:50
RT @ttunguz: The sharpest questions in AI live at the local–cloud boundary : where should inference run, & where should your data live? → tweet link
Tech Industry & Open Source
@FinansowyUmysl · 2026-06-23T07:03
Ciekawy wykres sektora półprzewodników. Teraz i.. w czasie bańki internetowej. [Interesting chart of the semiconductor sector. Now vs.. during the internet bubble.] → tweet link
@jezell · 2026-06-23T18:14
RT @OpenAI: OpenAI DevDay 2026 applications are now open! Our biggest developer event gets even bigger. 📍 San Francisco 📅 September 29 → tweet link
@Teknium · 2026-06-22T22:09
RT @imbabybrooklyn: hey babe @NousResearch wake up - we hit 200k https://t.co/K5KMpYAYPJ → tweet link
@kunchenguid · 2026-06-23T04:29
"how to get the most out of the LLM subscriptions" was a question many people asked under my agentic engineering workflow video here's my suggestion that i think is a good default for the vast majority - 1. start with getting a $20/month subscription from OpenAI. currently at the same price point, OpenAI's plans can get a lot more done than Anthropic → tweet link
@levelsio · 2026-06-23T14:31
✨ It was really fun visiting @cursor_ai's Compile conference this week ...how we might lose ourselves with sycophantic AI coding building all these things nobody uses and nobody will ever pay for while thinking we're making groundbreaking work, so when building products with AI revenue should be the real test! → tweet link
@TheAhmadOsman · 2026-06-23T06:20
Greatest lie ever invented by tech companies was the cloud btw → tweet link