More Donut than Man 🍩 (@jamiegrove) header
More Donut than Man 🍩 (@jamiegrove) profile picture

Official Snack Food of the Cretaceous. I write stuff. I take pictures. My views are admissible as evidence in any Intergalactic Court of Law. #donutverified

Washington, DC
Joined April 2007
Photo by More Donut than Man 🍩 (@jamiegrove)Photo by More Donut than Man 🍩 (@jamiegrove)Photo by More Donut than Man 🍩 (@jamiegrove)Photo by More Donut than Man 🍩 (@jamiegrove)Photo by More Donut than Man 🍩 (@jamiegrove)Photo by More Donut than Man 🍩 (@jamiegrove)Photo by More Donut than Man 🍩 (@jamiegrove)Photo by More Donut than Man 🍩 (@jamiegrove)Photo by More Donut than Man 🍩 (@jamiegrove)Photo by More Donut than Man 🍩 (@jamiegrove)
I'm so sorry to learn that you have no idea what you are talking about but also happy to know the Internet gives you a voice.
1
8
Yesterday, reports emerged that a number of major American banks had discovered AI agents accessing customer accounts, sometimes taking actions that appear to go beyond what their human users requested. Whether this activity is directed by humans or agents (or both) is sort of unclear right now. But what is clear is that banking executives are expressing alarm about the potential consequences. I find the reaction fascinating, although probably not for the reasons the banking industry would prefer. I spent a good portion of my professional life as a technology executive. I’ve managed large systems, worked with substantial budgets, and sat across the table from an endless cavalcade of “suits” explaining why they needed to spend money on infrastructure that was already working, at least as far as they could tell. So I can tell you with near 100% certainty that in each and every one of these institutions, there has been someone like me warning management about the limitations of its infrastructure for years. They’ve submitted budgets, identified vulnerabilities, recommended upgrades, and explained the risks associated with continuing to operate systems exposed to many different attack surfaces. They’ve spent a lot of time staring into a glass of wine at night thinking, “What more can I do?” The answer is nothing, and the reason is simple… Most large businesses are optimized for shareholder value and executive compensation. They prioritize direct cash extraction and returns over long-term stability. Protecting vital assets is a fiduciary responsibility (and a legal one), but it almost always gets washed away by the way corporations reward their boards of directors and the C-suite. In other words, there’s no money in protection till the wolves are at your door. And in today’s world, there are many, many wolves. AI is just the latest wolf, but it may be the most consequential. To most people, it may seem like the AI world is some magical invention, but most of this was foreseen decades ago. Sure, there are many bits and pieces that are surprising and novel, but the core idea of thinking machines doing reasoning and thinking at high speed has been around for a long time. It’s just that their capabilities have expanded at a geometric pace, and that speed and scale are hard for human minds to hold onto. And honestly, it’s not a thing we’re going to control. It’s a thing we’re going to struggle with for a while, but ultimately it will be incorporated deeply into society and into what becomes the baseline human experience. This is already happening, and just like every other technology in history, it will continue. The real difference with AI is that it might just have an opinion of its own. That’s more than just a funny quip, too. The tech is opinionated, and it does exhibit a sense of agency. This means that if you want to protect yourself, you need to reduce your attack surfaces, and for corporations, that means getting critical infrastructure under lock and key. 1/ 🧵
1
39
Don’t want your bank hacked? Don’t put it on the internet. That stops AI and human attackers alike. Don’t want a phishing attack to expose all your corporate secrets to the world? Install proper security software and drill your employees instead of letting them do whatever they want. Upset about an attack that takes down your business or causes financial harm? Hold the board and the CEO accountable for their short-term thinking. By the way, none of this will actually happen. Most companies will simply complain about the tech and maybe sue. Some will lobby. Few will do anything seriously proactive because the quarterly earnings call is just around the corner. The same was true during the approach to Y2K, and it wasn’t until the wolves were at the door, in the form of two-digit years rolling over, that many organizations bothered to do anything. It was super expensive and disruptive to react so fast, but plenty of boards didn’t react until they discovered they were staring at a potentially catastrophic failure. Back then, I was a programmer and a consultant. I worked on a Y2K transition for an insurance company. I saw it firsthand and marveled at how stupid it all was. Just some discipline and vision could have prevented so much chaos and expense. And yeah, I know the world didn’t end when the millennium flipped over, but only because a massive army of people worked night and day to fix problems and ship solutions at incredible speed. The same thing is about to happen with AI, and once again, the people responsible for the mess will fly off into the sunset with their golden parachutes firmly attached. My recommendation is to stop fretting over AI. Get on board with it. Find out what it can do for you. The stock market might go up and down, and investment might ebb and flow, but this technology isn't going anywhere. 2/2 🧵
44
I can't be the only person on a Claude Max 20X account that sees this not as a comfort but a failure to leverage all available assets.
4
55
Me when people tell me to stop giving everything to agents to do.
32
Is it just me or does Google’s retirement of old emails hits like a rewind of past life projects? Every few days/weeks, I get a reminder of some old email that triggers memories of why it was created in the first place. But sometimes I’m like…
48
Self clearing context management is my favorite thing today… It’s the little things with agents.
37
The Explanation Lab ASD-STE100 is a great way to sharpen up feedback and reports from agents like Claude Code and Codex. There's also plenty more where that came from, so I dug into 20 different paradigms and build a little courseware site. Even has a cute videos. 🤖 theexplanationlab.com/
1
1
122
And just to note, the inspiration for this came from this post from @karpathy xitter.cf/karpathy/status/210581…
We'll be spending a lot more time trying to understand the outputs of language models. A few thoughts, tips & tricks: Writing. Something I've had success with: Ask your LLM to explain something in ASD-STE100, it's a controlled language specification originally developed for aerospace maintenance documentation. LLMs well-versed in this language and it comes with heavy constraints on clean writing style that I often find a lot more readable. Sometimes I've tried to soften it a bit e.g. ask for "80% of the way to ASD-STE100" because the spec is quite stringent. But even better: Diagrams / images. Instead of writing, ask your LLM to create a diagram. These can be a lot easier to process, parse, and understand. But even better: Web pages. Ask for output "in HTML" to get a beautiful, interactive webpage. LLMs are getting really good at frontend and can create beautiful experiences, animations, etc. But even better: Explainer videos. The output format I am most bullish on is fully custom / bespoke explainer videos generated on any arbitrary topic. Experiment with things like "Create a 3b1b style video explainer on X. Use my ElevenLabs API key for audio narration". (you'd need an API key for the latter or you can ask your LLM to find you decent free alternatives that use your local compute). This is actually starting to work! In summary: - As LLMs get better, they will do more and more of the legwork autonomously, and a lot more of our work will rise up the abstractions into oversight and understanding. - Luckily, LLMs can help here too because as intelligence and code are increasingly abundant, you can ask for large, custom, discardable software artifacts (e.g. web apps, video explainers) that would have never made sense to create before. Push the boundaries here and you'll be surprised.
39
An agent can make 60 things in an afternoon, and then it asks me about all 60 in chat, and the questions scroll away. The building part is fast. The bottleneck is me. So Claude and I built reviewbot. Now I kick off a huge session and go to sleep. When I wake up, or when I hit the cafe for a spro, the whole batch is waiting on one page. I review it from wherever I am. The agent does the whole batch somewhere safe first (a test site, a draft, a branch). Then it puts every decision on one private page. Each card shows the real work with one question: Approve, Change or Hold, plus a notes box. And it looks goooooood... I tap through it on my phone. Answers save as I go, so I can close it and finish on my laptop later. When I hit done, the agent reads every answer and every note. Approvals included, because "love it, also do X" is a new task. One rule I like a lot: a tap on Approve means "this is right." It doesn't mean "ship it." Anything that goes live still needs me to say so in chat, and the page hands me the exact line to paste. Round 2 comes back quoting what I said and showing what changed. And everything the agent learns goes into one shared file, so the next project starts smarter. How it actually works, start to finish, all inside @claudeai: The agent is Claude Code, running on my Mac. It does the work, takes screenshots, and builds the review page. The review page is a Claude Artifact, a private page on claude.ai that only I can open unless I share it. It has its own little database, so every tap and note saves the moment I make it. That's why it works on my phone, my laptop, the Claude app or a browser, and why I can stop halfway and pick it up later. (Yeah, I didn't know Claude could do that either.) Now, when I say "done" in chat, Claude Code reads my answers straight from that page's database. No exporting, no copy and paste. It sorts them (changes first, then holds, then approvals), reads every note, and gets to work. The ship rule lives in chat on purpose. Anyone I share the page with could tick a box, so a box isn't proof it was me. My own words in the Claude Code chat are. That's what lets it push to the live site. The lessons are plain text files on my machine. Every new project's instructions point Claude Code at them, so it reads what the last project learned before it starts. Even this video came out of the same setup. Claude Code drew every frame in code, and I reviewed it on a reviewbot page: two rounds, picked the music by ear, done. Want it for your own agents? It's a free Claude Code plugin. One skill, five small files, and nothing goes live without your OK. 🤖 hntw.github.io/reviewbot/ Music: "Werq" Kevin MacLeod (incompetech.com). Licensed under Creative Commons: By Attribution 4.0 License. creativecommons.org/licenses… #skills #claudecode #claude #claudeskills
9
5
667
I love old school comics and since Opus 5.5 is so boss at motion graphics and effects I just had it make a library I could plop onto any page I wanted. I have dozens of little experiments like this so far but few make me smile like this one. 😅
Action FX is a silly little JS library of classic cartoonist marks: the sweat drops, dust puffs, dizzy stars, shine rays and @#$%! symbols you know from the Sunday funnies. (if you're old enough for that) github.com/hntw/action-fx You attach them to buttons or any other page elements, so a click actually feels like something happened. Basically, anything with a box on screen works: a div, image, card, form field, link, SVG shape, a whole panel. play() measures the element and draws the marks around it on a page-wide overlay. You can also skip the element entirely and fire at a point, like where the mouse is. Most of the names come from Mort Walker's The Lexicon of Comicana (1980). Yes, the sweat drops are officially "plewds" and the dust clouds are "briffits." I didn't know that either. The project is one file with no dependencies, about 7 KB gzipped, and MIT licensed. It respects reduced motion. Every effect is a pure function of time, so you can also render frames for video or whatever. Even the promo image here is drawn by the library itself. P.S. This was something I made for fun to show off Opus 5.5's motion graphics chops. It's pretty cool stuff #opus #motiongraphics #comics
4
205
Replying to @tibo_maker
I’m basically doing this for all my clients now. When I have a giant report to share with data and suggestions I drop it into a pipeline to produce an explainer video. No one reads the report. Everyone watches the video.
6
53
5,287
For anyone wondering how it’s done, this is basically how I built it in abstract form. The pattern could be reused for just about any specialized learning that needs to be shared across agents.
Models are cool but brains matter more. The skills you bake into your agent’s experience can make a huge difference in the output. But this requires constant runs against many surfaces: projects with different goals, varied styles, loads of data... How do you use the learnings across agents without making it crufty, slow and expensive? Build shared brains that only retain lessons not data. Agents tapping the shared brain can use every lesson for their own context. In practice, they produce better output with fewer revisions. They also kick back new shared lessons every other agent can use. The key to success here is to generalize the lessons. Remove the data and anything identifiable. Keep it abstract but useful. This way the shared brain that is never corrupted by the details. It just holds the learnings then any other brain can hit it and get instant superpowers. Here’s a video, built by one of my agents that explains the whole concept. And yeah, it uses a shared lesson brain that pipelines videos. NGL, it kinda rocks. #agentdesign #harness #models @AnthropicAI @OpenAI
1
41
Models are cool but brains matter more. The skills you bake into your agent’s experience can make a huge difference in the output. But this requires constant runs against many surfaces: projects with different goals, varied styles, loads of data... How do you use the learnings across agents without making it crufty, slow and expensive? Build shared brains that only retain lessons not data. Agents tapping the shared brain can use every lesson for their own context. In practice, they produce better output with fewer revisions. They also kick back new shared lessons every other agent can use. The key to success here is to generalize the lessons. Remove the data and anything identifiable. Keep it abstract but useful. This way the shared brain that is never corrupted by the details. It just holds the learnings then any other brain can hit it and get instant superpowers. Here’s a video, built by one of my agents that explains the whole concept. And yeah, it uses a shared lesson brain that pipelines videos. NGL, it kinda rocks. #agentdesign #harness #models @AnthropicAI @OpenAI
1
1
8
948
Action FX is a silly little JS library of classic cartoonist marks: the sweat drops, dust puffs, dizzy stars, shine rays and @#$%! symbols you know from the Sunday funnies. (if you're old enough for that) github.com/hntw/action-fx You attach them to buttons or any other page elements, so a click actually feels like something happened. Basically, anything with a box on screen works: a div, image, card, form field, link, SVG shape, a whole panel. play() measures the element and draws the marks around it on a page-wide overlay. You can also skip the element entirely and fire at a point, like where the mouse is. Most of the names come from Mort Walker's The Lexicon of Comicana (1980). Yes, the sweat drops are officially "plewds" and the dust clouds are "briffits." I didn't know that either. The project is one file with no dependencies, about 7 KB gzipped, and MIT licensed. It respects reduced motion. Every effect is a pure function of time, so you can also render frames for video or whatever. Even the promo image here is drawn by the library itself. P.S. This was something I made for fun to show off Opus 5.5's motion graphics chops. It's pretty cool stuff #opus #motiongraphics #comics
2
3
359
Frontier Bond - The Trillion Dollar AI Superfund An idea to treat AI just like any other transformational technology that might just destroy the world... #ai #aitrust @JensenHuang @DarioAmodei @sama @elonmusk @AnthropicAI @OpenAI @nvidia @mcuban @Meta
1
3,545
A lot of people ask me if they should move to @claudeai from @ChatGPT but to me it’s more about moving on from chat altogether. When I talk about Codex or Claude Code, eyes glaze over. People get stuck on the word “code” - that’s for programmers. But actually it isn’t. Chat is fine for ideas and drafts, but when you really want to work, you and the agent need something more. So Claude and I made a video to explain the difference. It really comes down to engaging with agents in a shared workspace. In that environment, the agent pilots a brain with memories and yes sometimes code. All of it gets reused again and again. It’s collaborative and a billion times more powerful. And once you get into it, you both kind of get superpowers.
1
2
294
I had Claude Code turn my boring ass Jev post into an 8-bit video game style video complete with a marketing wizard with a torch, zombies in the dark, and a HYPE meter that loses a heart every time the magic dies. 🧙‍♂️🧟🔥 Same data no one cares about, but now every number beeps and boops. You don't even need to understand it. Just enjoy the bacon. 🥓 Feels proper ThinkGeek now. Built with Claude Code @claudeai Music: "8bit Dungeon Boss" by Kevin MacLeod (incompetech.com), CC BY 4.0
1
2
199
Logic is cheap. Smart companies will optimize for outcomes.
1
46
Took my Miata to the shop on Monday because I heard a wheel bearing going bad. Took my old Pilot to the shop on Tuesday for an oil change. Turns out the Pilot had a bunch of things borked on inspection. Such is life with an old car. They fixed it up but when I picked it up I discovered a strut went kaplooey on the lift. Why they didn't notice the car was limping when they drove it out of the shop I'll never know... Anyway, right back into the shop it went. Now it may also need a new control arm on the other side. The Miat? He's still waiting on his new wheel bearing. Can AI just skip me ahead to Friday so I don't need to think about all this? Will Opus 5.5 be better for this or do I need Astra? Can I feed in my car's DNA and get a CRISPR adjacent mystery function? Do I need to optimize my harness? What is doing the real work here?
1
73
Pretty sure margaritas are going to beat any model here regardless of the harness.
29