Ben Eng

«Human understanding and consensus building need to be removed as blocking inputs. These need to become non-blocking outputs.» This one guiding principle changes everything for what an AI-native enterprise would look like. 🧵

@de3ug @DanielleFong First time I've read that. Thanks for the pointer.

@jetpen @DanielleFong See Clarke’s 1st law. Much more relevant here!

danielle: "what about this crazy idea" richard swanson: "as the japanese say, very difficult!" danielle: "i see" richard: "it means impossible" danielle: "i got that"

@premium bot is no help at all, giving me references to the help documentation for search and retrieval. @Engineering is there a bug causing this search and retrieval problem with past tweets?

There is something seriously wrong with twitter's retention and ability to search for and retrieve past tweets. I'm trying to quote one of my tweets from 2026-05-06, which I know to exist based on my full data export. Its tweet id is 2052172770960670723. I'm actually wanting the whole thread. However, through both xurl and from the browser, the system does not think it exists. None of these works: xurl read 2052172770960670723 x.com/jetpen/status/… xurl /2/tweets/2052172770960670723?expansions=author_id,referenced_tweets.id Using search from the browser: from:jetpen since:2026-05-06 until:2026-05-07 None of those works. I have noticed this problem for nearly every past tweet. What the hell is going on?

@MosquitoCoasty All I will say is I used to mow the lawn every week as a child. On the farm the lawn was almost an acre surrounding the house. There were moments of excitement between long runs of monotony.

@Rage_of_the_Age Isn't he continuing in the roles of AI czar and Crypto czar?

It's funny that I'm Back to agreeing with David Sacks now that he's no longer in the Trump admin. During the Biden years I always thought he had excellent takes, during his time as crypto tzar I found myself in disagreement quite often. Funny how politics work

@DanielleFong This is a call for a moratorium on learing centers. They are a blight on our neighborhoods. They generate too much noise in the news. Not to mention the expenditures and resource consumption. Who knows how many job losses will ensue if this continues.

not like we're building up infrastructure for this

What do these superpower unlocks show? Human attention is all you need. That's the limiting factor to do anything.

I love love love that all I have to do is copy my previous slack message and paste it to start a new work session. @ hermes work in the git project /home/ben/projects/decent-registry for this session Immediately, the agent examines the git repo and the open issues, and it knows what to do next. It asks, and I can answer. Work any open issue. Choose the one that is most foundational that other issues may depend on, such that other issues will require the least refactoring next. I'm doing this while grocery shopping for dinner. The superpower unlock is having the agent do productive work, while I do other things. I can be anywhere. The other superpower is that when I have an hour to give my full attention, I can do more wayfinder or grilling tasks to surface more work to put in the backlog. An idea what used to be too big, too complex, too daunting to take the next step is now easily within reach.

The takeaway is that the cost I'm willing to incur personally is approximately one dollar an hour for up to maybe three hours a day plus my time and attention, when I'm working pro bono on behalf of humanity at large.

If nothing in this git repo is comprehensible, I get it. At this point, this would be like explaining the entirety of the universe and needing to teach cosmology as a prereq to understanding how this atom would fit into a larger whole. That's my intention on the not yet started white paper. It's all in my head (not even), and I hope to spec it out soon.

Here's all the work it did for that cost. Ticket #15 and higher. github.com/jetpen/decent-…

I've embarked on the first step of this journey. The aim is something I've always wanted to do, but getting started was always too daunting. That is, provide the foundation for decentralized applications, so that the way we interact with applications on the Internet is transformed. We need to move away from Big Tech centralization. Move toward a multitude of application replicas (maybe call it community hosting) and self-hosting with foundational platform capabilities that are scattered to the winds, so that they are unstoppable, ungovernable, and censorship resistant. Start with two projects. The first is a white paper to explain the vision. I have nothing but H1 headings in an outline today. The second is the most foundational: the registry. Yesterday, I used Hermes Agent to try Matt Pocock's /wayfinder to get this started from nothing. 4 tickets to get me to a spec. I'm using Hermes MoA in this cheap configuration: * reference models: - google/gemini-4.1-flash-lite - deepseek/deepseek-v4-flash * aggregator: - openai/gpt-5.4-nano This morning, /to-spec gave me 4 tickets for a working prototype. An hour or so later, it was finished. I was surprised how flawless the experience was, expecting the task to overwhelm the cheap models. Indeed, it was easy-peasy without even a single glitch. I was honestly expecting to have to resort to deepseek-v4-pro and claude-opus-4.8 for coding, but no need, apparently. (It really makes me wonder whether my heavy reliance on GPT 5.6 Terra medium at work is overkill.) Cheap gpt-5.4-nano performed like a champ. The /wayfinder task cost about $1. The /to-spec task finishing at a working implementation cost $1.19. Next task is to harden the implementation somewhat. This n=1 anecdote feels like a true superpower unlock.

Danger is also dependent on whether an operation happens within the workspace (e.g., under cwd of the sssion, local git repo, remote git repo) or whether it is going off-road (high-risk of doing something out of bounds).

Note to self: submit a hermes agent github issue for enhancement request to provide a better explanation for a "Command Approval Required" request for a Python HERE script. Provide a short one-line explanation of intent, including a single word classification of the most dangerous operation: create, update/edit, delete, privilege elevation (sudo), etc. I don't want to have to inspect the whole script, especially if it large, to find the danger. It's a needle in a haystack. Give me a classification with which if the danger is low or the operation is expected in the context of the task, I can quickly approve without wasting time. I only want to inspect when something looks out of the ordinary.

@MaryBowdenMD From a patient's perspective, I perceive primary care to be a well-known decision tree. Therefore, a machine that is constrained with low discretion can take care of the most typical cases, and outliers can be passed to the doctor. x.com/jetpen/status/…

For primary care services, who would you trust more?

@PensareBBall All the new arrivals from class of 2026 need roster spots. Those spots can only be opened up by players retiring or getting washed away.

@DanielleFong They have a natural affinity to higher energy regimes. Multi TeV at CERN or exa-eV cosmic rays. They can't deal with your sub-10 eV energy regime of chemical reactions.

physicists have a fear of chemistry. i'm not saying it's irrational. but you only live once

@MosquitoCoasty Note to self: when handling rope, ALWAYS wear gloves.