Winston

Singapore 🇸🇬 Medical doctor, training in a traditional zen and dzogchen lineage, Obsidian/Tana stack. Nerd out with me, I enjoy all kinds of topics :)

[Winston @syncretizm]

(1/10) In Jan, I posted how it would be amazing to have a tool that synergizes bidirectional linking and an attribute system. In June, it was like X'mas came early when I was invited to test the alpha version. Here's how Tana blew my mind🧵https://t.co/19Ftzrkwbz

[Winston @syncretizm]

Caught in a dilemma
1. Love Notion's attributes and DIY databases
2. Yet I love graph databases and their contextual backlinking - as well as the ability to thread.

I can't figure out which one I love more. I wish both were present in a single app.

[Winston @syncretizm]

@natxwang Kinda strange given they've been on the forefront of things and now wanna pull the plug when other companies are catching up

[Winston @syncretizm]

Made an AI harness that points it to codex / opencode / pi as needed. Just want to centralise my own knowledge base "here" as opposed to hosting it elsewhere. I need control over its architecture.

[Winston @syncretizm]

@natxwang Count me in

[Winston @syncretizm]

i don't think it's underrated. It's just ridiculously expensive.

[Farza 🇵🇰🇺🇸 @FarzaTV]

Watch me control my computer with just my voice. This is the future of operating systems.

No hands.

GPT-Realtime 2.0 is very, very underrated.

Demo:

[Winston @syncretizm]

Gpt 5.5 is like a precise laser. Does exactly what you tell it to do, but doesn't quite follow intuitive paths until stated. Needs broad prompting.

[Winston @syncretizm]

Wow opus 4.8 does feel like a marked improvement from 4.7. The token thing is an issue though

[Winston @syncretizm]

In the lack of data, its 'insurance' would be to cast a wide net, err on the side of horrible outcomes, etc - dropping occam's razor etc.

[Winston @syncretizm]

I've had patients coming in after consulting AI. Gemini, chatGPT, recently Claude. They sometimes tell me a long list of differentials. I nod and agree, mostly. But sometimes, AI doesn't always quite nail it.

[Winston @syncretizm]

Perhaps the next step is really to improve the model's ability to look into all of these conditions/factors that it otherwise had ignorance of, even without needing to include it within context of the prompt.

[Winston @syncretizm]

We need a way for AI to leap across domains. While AI latent spaces do discover deeper abstract concepts, next-token prediction forces the model to prioritize surface-level text statistics over those brilliant cross-domain analogies just to calculate the very next word

[Winston @syncretizm]

Embedding similarity is just proximity masquerading as understanding. Clusters local jargon but completely misses global, cross-domain analogies

[Winston @syncretizm]

@nbbaier quite a few posts i think! i might consider writing a full updated thread on it if there's enough curiosity haha

https://t.co/KoLcbsAlWg
https://t.co/al5VEKWGT0 https://t.co/i5GVrkH1x2

more of it here https://t.co/Yt29p8aqZ3

[Winston @syncretizm]

Threading in a nutshell

[Winston @syncretizm]

Built an AI harness with its own PKM style memory in my classic threading style. No markdown, no vector embeddings or search. No superficial RAG stuff. Forced LLM note by note analysis, logical retrieval by traversing paths. So far it's been an excellent conversational partner!

[Winston @syncretizm]

@horacio_ps @felixleezd It's overly proactive which sometimes works counter to what I need. Codex is way more constrained actually

[Winston @syncretizm]

@nbbaier Used codex with some coding knowledge. I now use it with codex / claude code as a hook / mcp. It's a little slower due to the traversal but I appreciate a well thought reply over a generic tasteless one any day!

[Winston @syncretizm]

Is it just me or the Twitter feed seems to be restoring it's former glory?? Starting to see accs again that disappeared from my feed for a long time