
AI builder & learner | Building awesome stuff + documenting journey to first successful application| DMs open for collabs/advice. Building @Glintbase.
Joined June 2024
- Tweets1.8K
- Following154
- Followers138
- Likes2.4K
Reinforcement learning is one of the most promising paths to superintelligence. The key is making the training loop turn every experience into useful signal. With prioritised replay, process rewards, and good synthetic rollouts, even failed attempts become lasting progress instead of wasted effort.
We were expecting websites to be fighting AI companies over their content.
Instead, 93.5% of them don’t tell AI anything at all. The few that do mostly said the same sentence, and a lot of them probably never typed it.
There’s a standard called Content Signals that lets any site answer three separate questions, in one line of robots.txt or an HTTP header:
search: can you index me and link to me?
ai-input: can you use me to answer someone’s question?
ai-train: can you train your model on me?
We checked 1,121 unique sites: our leaderboard of 102 AI-agent companies, 150 popular sites, and the Tranco top 1,000. 754 responded. We wrote the method down before looking at any data, so we couldn’t move the goalposts later. 🧵
2
2
77
5/
When sites write their own rules in robots.txt, they agree on almost everything:
Search: 30 yes, 0 no
AI input: 24 yes, 2 no
Training: 16 yes, 13 no
Training is the one real argument, split almost down the middle. And 7 of the 13 sites that refuse training still let AI read them to answer questions. They’re saying yes to being cited and no to being trained on, which is exactly the distinction Content Signals was built for.
1
1
13
6/
You can check your own site in 30 seconds.
Run the two commands in the image:
the first shows what your robots.txt says, the second shows what an AI agent asking for Markdown gets told.
If the two don’t match, set a Content-Signal header on your own server. Cloudflare keeps yours instead of adding its default.
And whatever you set for ai-train, make it a choice you made, not one that got made for you.
Full report:
glintbase.dev/research/ai-us…
1
20
Connex retweeted
📢 Meet Qwen4-Max, your new AI teammate for work, creativity, and discovery!
Qwen4-Max is built on a new MoE architecture with native vision and a 1M-token context window.
Highlights:
- Outperforms Qwen3.8-Max with up to ~50% fewer tokens, while average cost per task drops over 60% in our internal evaluation.
- Advances real-world productivity through greater proficiency with professional tools, stronger long-horizon execution, and higher-quality deliverables.
- Handles real tasks, including 3D modeling, game development, manufacturing, finance, law, healthcare, and frontier scientific discovery.
- $1.5/M tokens for input, $4.5/M tokens for output, and $0.075/M tokens for implicit caching — all lower than Qwen3.8-Max, with implicit caching down by 70%.
The API is live on QwenCloud. You can also try Qwen4-Max on Qwen Studio, QwenWork and Qoder.
Start building today! 🚀
- Blog: qwen.ai/blog?id=qwen4-max
- API: qwencloud.com/models/qwen4-m…
- Token Plan: qwencloud.com/pricing/token-…
More to come! The open weights of smaller Qwen4 models will be released. Stay tuned. 👀
Connex retweeted
Introducing Hedra Model Factory: an autonomous inference and training platform for building models for the real world.
The recursive self-improvement loop behind frontier research has been locked inside a few labs. We're opening it to teams in robotics, manufacturing, biology and defense.
Our researchers come from Stanford, NVIDIA and Meta Superintelligence Labs. Now we work alongside your team to post-train and optimize your models.
Connex retweeted
OpenClaw, Hermes Agent and Prime Agent don't meaningfully learn from experience.
On Continual-ARC, our new benchmark at @letsunifyai, none of them were able to memorize anything of value and none were able to generalize to new tasks. Training a small 9B model's weights did both.
I don't know about you guys but I'm gonna be using the new fable 5.2 for extensive research. It seems to be really good at it
Replying to @claudeai
A team of 1,000 Fable 5.2 agents fully formalized Fermat’s Last Theorem in 30 hours, translating the proof into a language a computer can check step by step.
With the same setup, Opus 5.5 completed only 26% in 18 hours, and barely made any more progress after 30 hours.
1
60
Introducing Claude Fable 5.2 and Claude Mythos 5.2, designed to tackle the most daunting technical challenges.
They’re capable collaborators on ambitious projects in many fields, and show improvements in areas like coding, reasoning, and vision.
Axiom just formalised the entire Stacks Project in Lean (~16M lines) and everyone’s talking about it. Here’s why it’s a huge deal in simple terms:
The Stacks Project is a massive open textbook that researchers in advanced maths use as core reference material, thousands of definitions and results that build on each other.
Axiom used AI agents to turn the whole thing into Lean, a system where the computer checks every single step of every proof. Nothing gets a free pass.
That means a foundational chunk of working mathematical knowledge is now verified and structured in a way that both humans and AI agents can reliably build on top of. The computer rejects anything that isn’t actually correct, so the agents can help expand it without the usual trust problems.
It’s not “AI solved one hard problem.” It’s turning real research infrastructure into something machines can actually use and grow.
Most think vibecoding at scale is slop. Here’s a counterexample.
Today we release a 16M-line formalization of the entire Stacks Project, critical infra underlying modern math, with optimized code quality and its highly interconnected structure preserved.
A mathematical city 🌆
1
45
















