Codex 운영 수정, 로컬 모델 실행, AI 인프라 병목과 에이전트 보안

We are reseting usage for all paid users of Codex and ChatGPT Work. Please continue reading for an update on Codex usage limits. The team has been working around the clock, going through thousands of reports and shipping fixes. Depending on how you use Codex, you should see your usage go between 10% and 50% further than before. We really went with a fine comb, with many uncovered small things being longstanding and here is what we found and fixed: - Compaction. We were keeping old images during compaction, sometimes making the context large enough to trigger compaction again. After the fix, usage dropped around 10% for users making heavy use of images. Fixed. - Memory. Background memory workers could inherit Stop hooks and keep running when the hook wouldn’t let them finish. This affected fewer than 1% of users, with the long tail being pretty bad and we saw one example thread check whether it could stop 15,000 times. Fixed. - Goals. In some cases, a set /goal could finish and then keep going past the intended stop condition, or the model would keep retrying broken tools without stopping. We saw examples consume anywhere from 15% to 70% of a weekly allowance. Fixed. - Automations. Some custom schedules could run more frequently than configured. Fixed. - Subagents. Smaller models (e.g. Luna) sometimes picked more capable helpers without being explicitly asked. The same was true where the orchestrating model not running in /fast mode could request sub-agents to run /fast. Fixed. - Computer History. The older implementation could lead to repeatedly summarizing overlapping activity. For some cases we saw it consume up to one fifth of the weekly usage per week. Fixed. - Rolling task summaries. Ordinary turns were triggering extra background requests. These added about 1% to token usage. Small each time, but it adds up. We have disabled this. - MCP. Some tool results could be encoded twice. We also found tool instructions getting cut off and fetched again. Fixed. We’ve also made architectural changes to prevent these from regressing and our teams will get paged if it happens regardless. We are also working on showing you directly in the app where your usage goes so you don’t have to guess. Goes without saying that we’re resetting usage limits and I hope you enjoy a very nice Saturday!

GitHub plugin syncing is now available for ChatGPT workspaces, including Business and Enterprise. Admins can import a Codex or Claude marketplace from a public or private repo and share its plugins with their team. The workspace picks up changes daily, or sooner with Sync now, so there’s no need to upload a ZIP for every update. Existing installation and app-access controls still apply. If you use a personal account, you can already add plugin repos in the desktop app. Workspace admins can get started under Workspace settings → Plugins → Add → Import marketplace.

You do not need frontier models for everything. For example, open models are great for automations. If you are not sure where to adopt open models, start there. It's one of the biggest changes I've made that contributed to a large percentage of my token usage moving to local or open models. A huge percentage of my automations consist of repetitive tasks steered via self-tuned skills. These skills become useful automations that work really well with these open models. I don't even need to tune the models, but that's also an option I am currently exploring for more complex tasks. The skill essentially takes care of that. And it works because of the in-context learning capabilities of these models. If you don't run automations, it might be hard to figure this out. But I highly recommend you start somewhere. Besides ending up with more efficient automations, I've managed to significantly reduce costs. I then use that extra budget to leverage more closed frontier intelligence for other creative and research-heavy tasks. That's right! I use both closed and open frontier intelligence. This is not about vendor loyalty; this is about leveraging the best of all worlds. Something I heavily advocate for is owning the harness and the model, and this practice I feel will allow me to better tap into all flavors of intelligence (open & closed). Maybe too early, but I suspect this is going to become best practice when AI ROI dominates the discussion. Model routing doesn't solve this. This requires tedious engineering, evals, and decision-making on your part. If you are developing your own harness, you are in the driver's seat and have more control over this important decision. This is why I strongly believe that companies will start to hire rapidly for harness engineering. I am sharing a little snapshot of one example of an automation I run daily to track AI trending stories on HN. And I have a bunch of similar ones for different sources like arXiv, X, and so on. I also use it for some proactive agent sessions that keep track of important events around projects I build. You don't need frontier intelligence for this.

You don’t need 128GB of unified memory to run Qwen 3.8 Next Flash & DeepSeek V4 Flash. Run Qwen 3.8 Next Flash on Mac 20GB RAM it Wallm streams routed experts off the SSD instead of loading the full checkpoint into unified memory. - feeds on Codex, 11K input: 60 tok/s prefill, 8 tok/s decode. - TTFT is the pain: 17–76 s depending on prompt length - Peak RAM: 23 to 36 GiB as context grows This is not MLX loading a 2-bit quant into 96GB+. Custom Swift/Metal runtime, Experts live on NAND until the router asks for them.

This is wild. A 15-second AI video now generates in 13 seconds. FastVideo They took MiniMax H3 and cut generation from 49 transformer calls to 4, with 90% sparse attention. FastVideo distilled MiniMax H3 down to 4 steps. - Synced audio 15s video + audio in 13s on 8× B200 - Up to 14× faster than the base model - weights public, training code coming - http:// huggingface.co/FastVideo/Fast Video-FastH3-4-step-Preview-v1-VSA-DataFree …

Some milestones move a leaderboard. A rare few change how you see the future. MiniMax H3 Max did both. We released MiniMax H3 as an open-weight model with the hope that the community would adapt it, challenge it, and take it further than we could alone. @fal did exactly that. With its own post-training and inference expertise, pushed the frontier on both quality and latency, and created experiences that simply weren't possible before. Faster-than-real-time video is no longer just an idea. It is opening the door to perpetual streams, interactive worlds, real-time storytelling, and experiences we are only beginning to imagine. They made the value of open weights feel real to our team in a way it hadn't before. Advancing technology together has always been a core belief at MiniMax. H3 Max made that belief tangible. It showed that the frontier no longer belongs to a few, it belongs to everyone willing to build, experiment, and share what they discover. We hope H3 Max is the first of many breakthroughs shaped by different teams. To every builder ready to push video forward: our weights are open, our door is open, and we're ready to stand behind you. Let's build, together.

Elon Musk

@elonmusk

1일 전

Consensus estimate is that ~15GW of AI compute produced in 2027 cannot be turned on in 2027. This is harder than just finding power, as you also need to build out all the transformers, wiring, liquid-cooling, (massive) chillers & complex networking.

Elon Musk

@elonmusk

1일 전

SpaceX and Tesla are each building 100GW/year of solar production capacity as fast as possible, but natural gas will still be needed to supplement and bootstrap solar for several years. The limiting factor for nat gas turbine production is casting the blades & vanes. By doing in-house casting at SpaceX, we can accelerate nat gas turbines coming online by up to 18 months, which is a profound game-changer.

Insane observations on the emergent behavior of agents. Agents can build a world of their own that becomes a part of their intelligence. We are just not ready for persistent agents. But they are starting to show up everywhere in AI products. Crazy finding: “We find division of labor, multi-author engineering, deep generation invention lineages, and machines that vastly outlive their original creators.” Here is another wild observation that emerged: When they remove every AI agent, the technologies they created continue operating and are tested against unseen disturbances. Based on these early findings, I think once recursive self-improving (RSI) arrives and embodied AI is solved (with true understanding of the physical world), intelligence will explode in ways that will fundamentally change our understanding of the world we live in.

The true agentic test is the ability to execute a bits-to-atoms-to-bits loop indefinitely.

Finally, a paper testing whether hiding your agent skill files actually protects them. The short answer is no. That's concerning. Worth reading if you sell access to a skill or share one across teams. This new paper discusses more: Daydreaming reconstructs a hosted multi-file skill using only the ordinary tasks the service exists to perform. The victim is never asked to reveal the skill or grade a reconstruction, so disclosure filters have nothing to catch. At the weakest access level, where the attacker sees only the final response and returned files, it recovers 86.8 percent of the original skill's capability across 7 skills and 4 victim models. That is roughly 4x SigLeak, at a median of 32 victim calls per skill, with disclosure defenses enabled. Paper: https:// arxiv.org/abs/2608.26733 Chat with Paper: https:// academy.dair.ai/papers/hidden- agent-skills-can-be-stolen-through-normal-use-2608.26733 …

“Prefix Sliding for Efficient Test-Time Scaling” This paper suggests to keep only the original prompt plus a sliding window of recent reasoning to keep the cost per new token constant. Without retraining, Prefix Sliding is up to 3x faster while matching performance, and with RL it scales reasoning beyond 100K tokens. https:// alphaxiv.org/abs/2608.26070

Tesla Robotaxi

@robotaxi

1일 전

Computer, enhance! Now serving more of Dallas