Founder @ Causeway - Building AI infrastructure for autonomous operations - blog: https://xitter.cf/t.co/t8RrWP2uTX
📨 Newsletter →
Joined November 2021
- Tweets5K
- Following1.3K
- Followers6.4K
- Likes5K
Pinned Tweet
Introducing an enhanced version of the prompt turning ChatGPT into AutoGPT!
Adding:
• A project manager to oversee tasks and coordinate efforts.
• Expert advisors
• Regular status for progress updates, goal tracking, and next steps.
Experience AutoGPT in ChatGPT:
What if RAG was differentiable?
I ran an experiment: separate memory content (database) from memory access (neural network). Make retrieval trainable.
Key finding: embedding entity names separately from facts gives +27% accuracy. The model learned on its own that entity matching is 84% of the signal.
Results: → 85% retrieval on completely unseen entities → Instant updates → No retraining
Bounded compute, unbounded knowledge.
alexander.space/#proto-agi
Vibe coding exposed a truth: most moats were just friction in disguise.
Now building is nearly free.
So where does value actually accumulate?
Fastest way to vibe code a presentation?
I found using slidev worked awesome!
Why? Because its code, you can get claude code to make it for you and it actually looks professional! Unlike what you get if you ask it straight up.
claude code can make pptx but i burned hours on stuff a good template would've solved instantly
prompt template repos + vibe coding starter kits feel inevitable
the awesome-claude-prompts repo is coming, someone just has to be first
Using Neovim and Claude Code feels like the ultimate dichotomy.
Reinventing the keyboard with feedback, I really wonder what that feels like.
This quoted post is unavailable.
I got sick and tired of telling Claude Code the same things over and over.
"Use ALTER TABLE, not DROP TABLE"
"Use uv, not pip"
"Never force push"
It learns. Session ends. Next session? Amnesia.
Every correction lost to the context window void.
So I built Causeway . run
→ Watches your corrections
→ Extracts rules automatically
→ Enforces via hooks OUTSIDE the context window
Tell Claude once. Never repeat yourself.
Link to open source repo🔽
More details:
Landing: causeway.run
Source: github.com/CausewayRun/cause…
• MIT licensed — fully open source
• RAG-based memory — corrections are embedded and retrieved contextually
• Zero tokens consumed — rules live outside Claude's context window
Star it, fork it, break it. PRs welcome.
Gemini was NOT happy with my attempt at clean up poorly converted png to svg code.
Describing it as a nightmare! 😂
A bit surprised by how little tooling exists for browser use AI.
Providers really only provide the browser with actions, the LLM you have to bring yourself.
Currently prototyping human in the loop using browserbase and modified operator repo.
Google just announced their vibecoding IDE antigravity: url: antigravity dot google!
I tried it, it takes the task based LLM work to the next level. Check out the screenshot.
Rather than esoteric graphs. I added no-code traversal tool. Makes it easy for the user to iterate on the knowledge graph, and for the LLM to use the same when doing tool calling.
Allows the user and AI to collaborate and understand what happens.
The advantage is that the LLM can do MANY steps and find the end result in a grounded, understandable way.
Have you ever tried running 7 Claude Code terminals at once?
If vibe coding works for one, why not N? What's stopping you?
Not the model—YOU are. You're the coordination layer. The bottleneck isn't the agent, it's the conductor.
causeway.run - orchestrate N agents simultaneously.
First 10 gets $100 credits.
DM if you're already running N>3.
Giving LLMs tool to construct their own context is magic.
I created a knowledge graph tool, where the LLM can query it. I mean, feels like context problem is solved.
The main issue with RAG enthusiast, is that you pay an upfront cost of maintaining the graph. But I mean, reasoning becomes childs play.
Like: "Ahh, LLM is updating a line of code, the graph shows ALL dependencies, both code, data, busines logic etc".
This is AGI infra.
