💻 Building products & remote teams. 🤖 Exploring LLMs, AI & ML. 🎯 Hacker mindset, builder energy.
Joined October 2012
- Tweets1.6K
- Following1.2K
- Followers999
- Likes8.8K
Pinned Tweet
Playing around with Grok Imagine. Trying out some hard cases for TTV.
A few things that consistently trip it up:
-Doors: opening and closing doors seems to break things.
-Being near doors: walking near doors or walking out of doors often causes unnatural walking or turning behavior.
-Walking through doors: what’s on the other side tends to be rendered late, almost like a fog-of-war effect.
Doors might really be the final boss for text-to-video. 🚪🤖
Happy 4th of July.
My family moved to the U.S. in the early 1990s from a Soviet communist/socialist country. I was very young, but I still remember waiting in bread lines with paper tickets, and ice cream being something special we only had a handful of times a year when it was available.
When we moved to America, one of the things I vividly remember was walking into a store and seeing so many options: different kinds of bread, and even a bucket of ice cream with three different flavors in it.
It may sound small, but those memories stayed with me. I am deeply grateful to live in this country.
Happy Independence Day. Please find a way to celebrate responsibly.
Cheers.
How to increase revenue?
Make /init generate a larger AGENTS file. Easy.
Stop letting LLMs write your AGENTS file. It becomes a big token hog.
It should be very small. Written by a human. Only include the things the LLM consistently gets wrong.
The LLM doesn’t need your entire project structure on every inference. The rest it can figure out through tool calls and context.
The way people talk about tokens: always counting, always measuring, never enough, always running out. The global currency isn’t USD. It’s tokens.
What if AI doesn’t make everything more effective? What if organizations drown in AI slop that becomes a DDoS attack on decision makers?
I predict a significant number of organizations will become more efficient at the individual node level, and less effective as an organization.
Significant amount of data especially Enterprise data is behind a "Talk to Sales" firewall. This limits Agents, LLMs knowledge, access.
This can go into two directions
A) make X have more crypto spam
B) build the next global COIN or HOOD
I genuinely want crypto to proliferate on X, but applications that create incentives to spam, raid, and harass random users is not the way.
It meaningfully degrades the experience for millions of people — only to enrich a few people.
And yes, we are launching a number of features in a couple weeks, including Smart Cashtags that will enable you to trade stocks and crypto directly from timeline.
LLMs are big, complicated programs that eat data and electricity. How much of the data have they eaten that they shouldn’t have eaten?
