Michael Sinanian retweetet
Michael Sinanian
2.6K posts
Michael Sinanian
@sinansays
currently @GenCogAI & teaching AI for PM @UCLA prev: @revenue_io | @lynxdxlabs | @dealpath | @belkin | @gyroscope_app | @eff | @ucberkeley | @umsi | @umichlaw
./US/CA/SF Beigetreten Ocak 2009
1.1K Folgt349 Follower
As an aside: this is why I love information visualization as a field. B/c it brings into sight (literally!) what is simply imperceptible to human senses. It's pure magic. There's a reason the point of an academic paper is 'the figure'. (and why Tufte is GOAT'd)
English
But some of us noticed fossil sequences, continental fit, tidal periodicity, infection patterns, nonlinearity, etc. Our present reality is strong shaped by notions that are hardly 'real' to us. Inferring latent structure from observable proxies is *it*. AI will accelerate this...
English
So much of our present civilization rests on a handful of people noticing phenomena (through human-scale proxies/clues) that we can't otherwise observe with our senses. You cant "see" evolution, plate tectonics, gravitation, microbial disease, market dynamics, network fx, etc.
English
Simple yet powerful demo of using Claude Cowork to auto-post to LinkedIn, etc. We rapidly need solutions to rebuild trust ('proof of work'/'proof of thought'/'proof of credibility') etc. for the digital commons. The value of online comments has rapidly diminished in just 6-12 mos
Tigran@tigranbs
The dead internet theory is real! I did an experiment to prove that to myself! Using basic Claude Cowork, without any great technical skills or other coding work, you literally can endlessly spam people 🤷♂️ It is kinda sad that most of LinkedIn, Twitter, or Instagram is literally a bunch of AIs talking to each other, and now spammers have the most powerful tools to amplify content on social media with little effort. After doing this experiment, I understood that my feed and probably yours as well, is hitting the most refined AI slop that you can imagine. We have been trashing Moltbook for being a social media of AI-generated content, but at least they were honest from the beginning.
English
Solid insights from @Stammy's YIR. my take-aways:
- forgetfulness in human memory is a feature, not a bug
- real moat in AI is personalization & memory
- product thinking now more than ever requires restraint and saying 'no' a lot
- design moat is curiosity n habitual curation
Stammy@Stammy
Better late than never. I wrote 9,313 words for my 2025 year in review. Sold my house, left my last startup Limitless, took a few months off, quit caffeine, had my 2nd year of no alcohol, became obsessed with Claude Code, and started fresh at Sesame. paulstamatiou.com/2025-year-in-r…
English
serious corollary: continue to be skeptical about sweeping AI claims (esp ones with benchmarks). real impact is felt when daily frictions (like tab count!) are meaningfully removed. did a new tool/workflow truly reduce cognitive load? improve decisions? just make things easier?
English
idea: national debt clock but it's a countdown of my decreasing browser tabs as we approach ASI
English
no, we are not in the Singularity. i am still drowning in browser tabs!!
English
step by step. slowly, quietly. then suddenly.
Awni Hannun@awnihannun
Distributed inference in MLX on Apple silicon will be much faster in Tahoe 26.2
English
@ChrisGPotts Yes exactly! Timeline's progressed much faster. I would've agreed with ur estimate in '24. But end of '25 now and the avg user query triggers web search, maybe includes retrieval from gDrive/Box/etc. The pace of consumerization has been far faster than was reasonably guessed
English
@sinansays Oh, I had forgotten about that part! Yes, I say, "I think OpenAI is not going to expose more of this. I think they regard these as trade secrets." Maybe they did but innovation and the market forced them to be forthcoming!
English
Remarkable to watch @ChrisGPotts's talk "Shift from Models to Compound AI Systems" from Dec'24 where he argues OpenAI likely wont expose inference-time reasoning models like o1, treating them as trade secrets with most exploration staying in research literature. But in reality...
English
@ChrisGPotts Yes, the Q&A ~42:50. But it's why I’m interested in UX for compound systems: making legible what lives in weights of base model vs tools/retrieval/APIs. Ppl dont realize they’re talking to a system; e.g. my mom won't learn DSPy, so how can the system view be best shown to users?
English
@sinansays Oh, thanks so much for this – very incisive. Are you referring to the section starting at 11:19? I suppose my mistake was in underestimating the speed with which other groups would figure all of this out and bake it into their models, creating a pressure for openness.
English
It adds a little friction that biases you to think that what’s behind it must be worth it. Feels rooted in the same psychological trick that a process 'taking time to think' must mean it's delivering a better result
English
Definitely have mixed feelings about how the open web is changing, but will admit a strangely cool feeling when a site stops you for a Cloudflare robo-human “security check". Like an x-ray pat-down before entering a speakeasy where u might learn something new or do something cool
English
However you feel about iOS 26’s “Liquid Glass,” one upside is the revitalized icon design language. The whole OS suddenly feels fresh again. Makes me realize how flat everything was for so long. It was fine in its time, but we were quite overdue for a refresh
English
Years with text-only LLMs taught me to describe the world to an entity that can only "see in words". Visuospatial orientation becomes a first-order filter when describing things: shape, orientation, sequence, etc. Multimodality may obviate that, but notable cognitive side effect
English
We're at peak substack. It's only downhill from here. Completely oversaturated. Doesn't have a "where the cool kids hang" or "where the smart takes are" feel anymore. Has become Medium circa ~2018. there will always be a virtual place that's 'cool'. substack aint it. not anymore.
English