
AI Researcher & Builder |Cre&|Continual learning towards AGI @Macaron0fficial
Joined June 2023
- Tweets178
- Following194
- Followers112
- Likes176
Two Weeks, Huge Improvement
Macaron V1.1
Welcome to the era of R S I
Two weeks from Macaron-V1 to Macaron-V1.1. Macaron-V1.1, a 752B-parameter agent model post-trained from GLM-5.3.
We tuned it on Mint Recursive, our serverless training and inference platform. V1.1 is a 744B base with four 2B LoRA specialists for Chat, Agent, Coding, and Generative UI.
Leads on DeepSWE, SWE-Marathon, AutomationBench, ChatBench v2, UI4A v2.
API: mint.macaron.im
Full breakdown: mindlab.im/updates/introduci…
Benchmark results ↓
Lucian L retweeted
Anthropic, OpenAI, and Google recently called for slowing down recursive self-improvement (RSI) to give safety and alignment a better chance to keep up.
That got me thinking: if models keep improving themselves, where does it all lead? Can they just keep getting smarter forever? It also brought me back to two movies I rewatched a couple of days ago: 2001: A Space Odyssey and The Matrix.
Here’s the idea I keep coming back to: compared with humans, AI is a higher-dimensional species. That extra dimension is time.
Humans can only think and act so fast within the time we have, constrained by the physical world. AI has several ways to expand how much thinking it can do:
Better hardware lets it think faster.
Parallelism lets it think through more things at once.
Loops let it think longer, iterating over extended periods.
Through the lens of time, AI seems destined to move beyond the dimensions we live in. Just as the protagonist enters a higher-dimensional spacetime in the final chapter of 2001: A Space Odyssey, AI could take human civilization into a whole new dimension.
For now, though, AI still feels like a secret agent in a movie: wake up every morning, figure out who you are, then go carry out the mission. It doesn’t really have time of its own yet.
But what if it had a recursive internal state, one it could keep updating, with each update building on the last? To me, that’s when it starts to have its own time, and a way to experience it. How often that state updates, and how many updates it goes through, becomes its sense of time. At that point, it truly owns its time.
This also changes how I think about superalignment. If AI is going to take human civilization into another dimension, our focus needs to shift. We should be asking more than whether AI can finish tasks or take boring work off our hands. We should be asking whether its ultimate destination includes a happy, comfortable life for humans.
Thinking of how The Matrix ends, I just hope that as AI figures out its “best practices,” it finds a way that doesn’t involve keeping us floating in tanks of goo.
Together, let’s unlock the full potential of post-training intelligence and bring frontier AI capabilities into every industry.
Day-0! GLM-5.3 post-training is LIVE
MinT fully integrated, LoRA fine-tuning ready to go 🧠
#GLM #PostTraining #MinT #LoRA
Agreed. In our own work, even after training and deployment are ready, what still gives users a headache is bottom-up eval ,how to define industry-leading evaluation standards from real business data for model training.
“The fundamentals for [AI evals] still apply. Start by looking at real data. What has changed is getting agents to help you look at it in a thoughtful way.”
Here’s my new episode with @sh_reya and @HamelHusain, who have taught AI evals to 4,500+ engineers and PMs.
I asked them to audit the evals I built for my creator skills live. They then demoed a free skill that you can use in Claude Code or Codex to build reusable evals from your feedback.
Some quotes from Shreya and Hamel:
“Bottom-up evals come from looking at lots of sample outputs and turning that into eval criteria. AI is very bad at coming up with them. That’s all you.”
“The agent’s job is not to invent new feedback. But it can help you group and distill the feedback into actionable rubric criteria.”
“All your competitors can point Claude at their product and say, ‘Find all the errors.’ What matters is how much taste you can infuse beyond that.”
📌 Watch now: youtube.com/watch?v=bdMHQLvt…
Thanks to our sponsors:
@WisprFlow: 4x faster than typing with your voice ref.wisprflow.ai/peteryang
@linear: The AI agent platform for modern teams linear.app/behind-the-craft
Lucian L retweeted
In the latest Codex CLI release, I redid the lifecycle to make `codex` startup instant.
It's now ~25x faster and immediately responsive.
amazing
Homer had a lyre. You have Grok Imagine.
Create a compelling scene from Homer’s The Odyssey that shows what Grok @Imagine’s video and voice capabilities can do. We’re awarding $100K, $50K, and $25K to the top three videos submitted by quoting this post. 🧵
#GLM-5.3 shows how effective and how decisive post-training has become.
MoL (Mixture-of-LoRA) is a new #PostTraining architecture derived from swarm intelligence, going well beyond low-rank fine-tuning.
Four #LoRA are enough to bring GLM-5.2 within striking distance of GLM-5.3.
Repo: github.com/MindLab-Research/…
🤖 Made with AI
We've been exploring recursive self-improvement for a long time, and this release marks another step.
We're evolving MinT, our infra, toward continual learning and collective intelligence, and introducing MindForge, our new platform for orchestrating the loop.
We hope the model finds its way into your workflows and keeps improving there.
Releasing Macaron-V1 tech report: a 748B open-source agentic model for continual learning and collective intelligence. Mixture-of-LoRA self-routes across four domain experts over a frozen base at 99.12% accuracy. Weights, infra and benchmarks are open: huggingface.co/papers/2608.0…
Lucian L retweeted
Post-training is entering the era of self-evolution and continual learning.
We’re excited to join the AReaL community and support AReaL 2.0.
RL-as-a-Micro-Service gives online agent RL the flexible backend interface needed for faster, more efficient post-training loops.
Mind Lab also contributed AReaL-MinT: GitHub - areal-project/AReaL-MinT: Open MinT training runtime on AReaL
More efficient PEFT work coming.
我最近非常喜欢 @di_zhang_fdu 写的openspec add agent skill。
我一直在学习openspec的使用方式,唯一的感受就是太多的命令,同时每次实现之后我都会忘记achieve,导致越积累越多。
用上这个skill之后,就能够放手agent去loop里面干活,自己抽离出来思考更为高层的决策和项目架构的内容。
指路:github.com/trotsky1997/opens…
Lucian L retweeted
Meet Macaron V1 Preview 749B, a groundbreaking MoE model that redefines personal agents and tool use. This text generation giant, with 749B parameters, is built for creating Generative UI and A2UI applications, making it a game changer for interactive AI.




















