Solo founder building š¦ OpenSquid Persistent memory. Verified lessons. Better agents. AI-native engineering & coding agents Building in public
Joined January 2023
- Tweets2.2K
- Following737
- Followers706
- Likes5.4K
Pinned Tweet
Most sites have JSON-LD.
Almost none treat it as a system.
Multiple scripts. Duplicated entities. No shared graph.
Looks fine to humans. Breaks machine reasoning.
This is why AI answers drift.
Code is here for anyone curious: github.com/ranklabsai/ranklaā¦
GPT5.6-sol + Pi is surprisingly better than expected. I suspect Codex got bloated how similarly Claude Code got bloated.
Switched ChatGPT Pro 20x from Claude Max 20x because whatever updates Claude Code got increased number of drifts and ultimately slowed down my development on OpenSquid.
Finally got the core orchestration loop working.
The biggest lesson wasnāt making the agent smarter.
It was making the workflow deterministic.
Every task now moves through explicit stages:
scope_write ā plan ā author ā code ā deploy
with verification gates between each boundary.
Itās slower than letting an LLM freely iterate, but so far every major task has completed in one shot instead of silently drifting.
Still rebuilding deployment to follow the same pipeline (I forgot to rescope it š
), but the architecture is finally starting to look the way I originally imagined.
Building OpenSquid one bug at a time.
Sangmin Lee retweeted
šØ The Sims one shotted by GPT-5.6 Pro
this is without codex or any coding harness , one shot entire game with logic in 48 minutes, all in one .html file.
cc : @mirochill for the test
@tszzl when we can have big model taste and personality ? like after GPT 5 its so robotic
GPT 5.6 Pro continues to mogs Fable in 3d test
working on games one shot too , for comment check quote post and also share your prompts and fable output i will tag you with 5.6 pro results
credit : @mirochill
Most AI agents fail quietly.
OpenSquid tells them āno.ā
In this session the agent:
ā¢ Tried to move forward
ā¢ Was told its research was insufficient
ā¢ Had to gather evidence
ā¢ Had to prove understanding before continuing
Not because the model chose to.
Because the workflow required it.
OpenSquid š¦
AI agents donāt have a reasoning problem.
They have a discipline problem.
They often know what to do.
They just donāt reliably do it.
So I stopped trying to make them smarter.
I started building systems that can tell them āno.ā
OpenSquid š¦
Sangmin Lee retweeted
Google I/O leaks: Geminiās "Omni" push and Gemini 3.2/3.5
- A new āOmniā model is being tested, hinted via a video UI ("powered by Omni"), pointing to deeper multimodal capabilities
- Expected to go beyond current systems (like Veo), with better video generation quality and control
- Could introduce native video output inside Gemini, not just text and images
- New Gemini versions (3.2/ 3.5) are reportedly in testing, focused on faster and more efficient performance
- Gemini Ultra is expected to evolve into a more memory heavy, long context system for consistent multistep workflows
- Subscription tiers may shift to a clearer Advance / Pro / Ultra structure
- A new memory feature ("Teamfood") aims to improve how Gemini uses past chats and long-term context
- Additional visual model work (e.g., "spark Robin") suggests a strong push on image + video capabilities
Serious question for AEO / GEO specialists.
How can anyone claim to provide a complete AI visibility solution without a deep, working understanding of how AI systems actually reason?
Not prompts.
Not content playbooks.
Not āLLMs like FAQs.ā
I mean:
ā¢ retrieval vs generation
ā¢ entity resolution
ā¢ grounding
ā¢ schema graphs
ā¢ ambiguity + conflict resolution
ā¢ why hallucinations happen
Real enterprise solutions eventually expose technical surfaces:
SDKs. Specs. Tooling. System models.
So why does AEO/GEO stop at strategy decks?
If itās proprietary, whereās the technical framework?
If itās not technical, why is it sold as an AI solution?
Opinions donāt scale in AI. Systems do.
Sangmin Lee retweeted
I was re-bingeing the recent Starter Story episodes, and I noticed somethingā¦
Today was one of those days.
An enterprise brand reached out to talk about a pilot. Exciting and terrifying at the same time.
After weeks of silence and uncertainty, this felt big.
Nothing is guaranteed yet, but days like today make the grind feel worth it.
Grateful to Jesus.
Grateful for my family.
They are part of why I keep showing up, even on the hardest days.
You donāt lose until you quit.
Agentic commerce does look unpredictable right now, but itās not because agents are inherently chaotic.
Itās because most e-commerce sites publish inconsistent, human-oriented data, and agents are forced to interpolate around missing or contradictory facts.
Without machine-verifiable inputs or a clear selection layer, variance is inevitable.
I spent the last few days building an AEO auditing tool to study how ecommerce sites show up in AI answers.
I ran it across 60+ well-known ecommerce brands.
Most have excellent SEO and strong marketing.
Average score: ~58/100.
What this shows:
Sites humans understand instantly
are often still vague to AI.
AI doesnāt infer or fill in gaps.
If something isnāt explicit, it skips it.
