Keep on to blur preview images; turn off to show them clearly
The Bitter Lesson from Ilya. When you lose trust in someone, you start believing rumors without verifying — just like how Ilya thought Sam was fired by YC and Greg was fired by Stripe. AI has no emotions (yet). Maybe that makes it more objective, once we fix hallucination.
- “what did Ilya see?” - “Mira’s screenshots”
RT @BlumLenore: @drmichaellevin is a magician! As I've said a number of times before, if anyone could convince me to be a panpsychist it wo…
FOLLOWS YOU. Artificial Intelligence, Cognitive Architectures, Computation. The goal is integrity, not conformity. https://t.co/rFUNzdYXuK
RT @Sylvia_Sparkle: 🤔Virtual Teaching Assistants (VTAs) have incredible potential, but how to evaluate them? Our EMNLP Findings paper addr…
Asst professor @MIT EECS & CSAIL (@nlp_mit). Author of https://t.co/VgyLxl0oa1 and https://t.co/ZZaSzaRaZ7 (@DSPyOSS). Prev: CS PhD @StanfordNLP. Research @Databricks.
Link to the repo, already at 551 stars (thanks guys!): https://t.co/YjjIPsxIna
Former Quant Investor, now building @lumera (formerly called Pastel Network) | My Open Source Projects: https://t.co/9qbOCDlaqM
This is a subtle point that many people miss. Scouting requires focussing on the diff. Every notification to a human comes at a cognitive cost. So it better be worth it. Monitoring != search in a loop
Co-founder & Chief Scientist @yutori_ai. Prev: Senior Director leading FAIR Embodied AI @MetaAI and Professor @GeorgiaTech.
This is so wild, even to me, and I made the project! I had an idea yesterday for a commercial companion app to my mcp agent mail open-source project yesterday. I came up with an elaborate plan in an iterative process that involved many rounds of revisions using GPT-5 Pro in the webapp, resulting in a huge 34k-token plan document. A lot of work went into designing the best architecture that would allow me to avoid running a big centralized server and being responsible for user data/code in any way (I don't want the devops headaches or the security liability, and neither do users). I set up a basic project skeleton with an AGENTS dot md file and two best-practices guides I created using Opus and GPT-5 Pro for how to make the best native iOS apps using Swift. And, of course, the big plan document. Then I reset my agent mail database to get a clean start (using the handle clear-and-reset-everything command) and spun up 7 different codex instances at once in the new folder. This is more than I've tried in the past and I wanted to see how well it would work. Then I simply used the same 5 prompts shown in the first screenshot below over and over again, keeping all 7 codices (lol) busy for many hours. Now, nearly 400 agent messages later, I have a remarkably clean, fleshed out, and complete implementation to test out and work on. And the amount of actual human attention and labor I had to expend is 90% less than what I used to need to do to get to this level. And I was already going at light speed before and using tons of agents; the difference is that I had to be the middle-man and coordinate all the work, and now they self-delegate and manage themselves! Not to mention TIME: this all happened in a few hours last night and a few hours today (with many hours of the agents grinding through all my queued up messages last night as I slept!). I really think this is the future of development.
More on this if you're interested in trying it yourself: