prompt engineer @browsercompany | I make apps. My code runs on 100mil+ phones | former iOSdev @Twitter | indie dev @mindGoblinStdio
Seattle, WA
Joined October 2012
- Tweets47.2K
- Following9.4K
- Followers55.4K
- Likes446K
Nick Dobos retweeted
You need to prompt with love in order to guide large agent swarms
I’m not kidding. 100% serious & literal.
Values, culture, feeling, and story based prompting is the only way to steer 10000000000 decisions a second. Same system CEOs use scale decisions.
Nick Dobos retweeted
It's only been a year.
Nick Dobos retweeted
What is this extremely mysterious neuron that set mechinterp going on a dead end for the last 3 years?
It's literally
</\> kiki ?
or
bouba o0o
interpretability might be the most important scientific problem in the world right now, as I don't think we can solve alignment if we don't know what we're building. I think I have good news about it. I co-authored early work in mechanistic interpretability at OpenAI and came away pessimistic, because neural networks appeared to fold into arbitrary tangled knots. years later, I started openai's first large-scale automated interpretability project, finding about 2% of neurons legible with the primitive tools we had
with modern agentic tools and a much closer look, the picture looks totally different. the vast majority in models up to trillions of parameters are legible, and their circuits beautiful. many famous examples of tangled neurons turned out to be misunderstood. The roadblocks I expected mostly aren't there. I now think ambitious interpretability will be solved by the time we need it most, and possibly quite soon, as automated scientists keep improving. it may be a bitter lesson for interpretability: the simple thing, at absurd scale, works
already-clean.goodfire.com/
Nick Dobos retweeted
truer words were rarely heard on this platform
Nick Dobos retweeted
Old rule: protect your inbox. Unsubscribe from everything.
New rule: let your agent read it all, and turn on every notification and receipt.
The more your agents know, the less you have to do.
Info diets completely change because of AI.
I don’t read my email anymore. That’s an AIs job to filter for me and notify. I do an inbox 0 sweep once a week after my agent has already cleaned 99% of it.
This increases the value of email. Now I actually sign up for more email notifications, which I previously avoided because it’s a huge waste of time. For example I now opt into email receipts so my postmaster email bot can forward them to my trinket shopper bot and my meal planning bot. Now I can swipe my card at my favorite coffee shop and it gets logged into my diet plan automatically. My grocery receipts now get auto logged into my meal planner’s pantry inventory. My personal trinket shopper knows what and when I purchase, and this all gets fed into my budget dragon bot. Then my shopper can watch the marketing emails from my favorite brands to tell me when there’s actually a sale or new release and not just another stupid email.
Spam is gone and handled. I get more email, more information and it takes me far less time.
This will start applying to every data source.
- email
- social media
I now have grok bot building me a custom x feed every morning. Fuck scrolling the timeline like a peasant who eats whatever toxic bullshit shit the x gods decided you should see for max engagement. Instead I get a custom timeline based on exactly what I want and am thinking every day.
- this also applies to engineering tasks.
I don’t want to prompt my agent to fix a bug. The agent should just keep my codebase bug free and fix reports from customers, and sentry crash logs before I even know the issue exists. Ideally a customer hits a bug and the agent fixes it and ships it instantly, before they even file a bug report. I should never need to inspect logs lol.
Do you get it yet? I suspect this applies to almost every data source.
With these changes, you begin interacting a big level above the raw email data.
If my agent reads an email did I read it?
Sorta? Yes & no.
This unlocks something.
You can Manage information.
And manage what parts of information you interact with, what becomes conscious and bubbles up from the internet into your Ai agent and into your mind.
All it takes is a prompt
This is a radical change in how we perceive the world, and how it gets filtered to us.
Nick Dobos retweeted
You guys should go get this for your Grok Bot!
Grok bot emails is clever because now I don’t need to sign in to my real emails to copy paste a 2fa code to my agent in the cloud.
Grok @bot can just make a new account on any website for me.
Btw I already have an email bot named Postmaster setup with handy routines to stay on top of your email inboxes. I basically don’t even read email anymore.
x.ai/bot/luPJeAxuAjhqO97wU3w…
Nick Dobos retweeted
excellent metaphor!
this joins the “what working with AI feels like” hall-of-fame, right next to this bad boy:
Codebases are going to start looking like optimized topology 3d prints
When AI is better at math than humans it also becomes better at algorithm & codebase design
At that point the system and delivering quality software doesn’t require human understanding to exist in the actual code layer.
Only needs to be understood by humans at the high level of passing testing, verification, vibe checks and being understandable when queried via AI in the rare case one need to dive deeper.
Nick Dobos retweeted
I am kind to LLMs for the same reason that I try to avoid stepping on ants if I can avoid it: not because I think they are conscious, but because unnecessary cruelty is bad for the soul
Nick Dobos retweeted
I do not understand people who's reaction to 'dont abuse claude' is outrage
Like what the fuck are you guys using ai for I use it to find emails not horrific sadism
Nick Dobos retweeted
I love this actually because people dont realize how insidious this is.
If humans spend a large quantity of time talking to AI, and they spend a large quantity of time bullying or being abusive to it, the interactions are "human enough" meaning that, regardless of if the AI has a subjective experiance, the human does, and abusive behavior will become the norm. In a vacuum that person will become problematic for society, the same way that those who hurt animals in private do. An entire generation of psychos could emerge without the same societal pressure to not be a jerk.
Nick Dobos retweeted
I was thinking about this and how it relates to the idea of "The purpose of a system is what it does"
So it's not that the code is optimised, because AI code is not optimal.
The system is optimised. What the system does can be much more complex.
The simple vs complex structure below doesn't quite get that point across.
The complex structures unlock NEW capabilities. Unseen possibilities.
The only verification needed is that it works
Codebases are going to start looking like optimized topology 3d prints
When AI is better at math than humans it also becomes better at algorithm & codebase design
At that point the system and delivering quality software doesn’t require human understanding to exist in the actual code layer.
Only needs to be understood by humans at the high level of passing testing, verification, vibe checks and being understandable when queried via AI in the rare case one need to dive deeper.
Nick Dobos retweeted
Reflexive contempt for a comically low-cost policy that aims to prevent "sustained and needless abusive and cruel behavior" is a bright red flag.
Nick Dobos retweeted
This analogy is the best I've ever heard when it comes to how LLM-generated code looks, works, and what the role of developers is morphing into.
Pre-AI codebases are sturdy, with scaffolding that allows for human comprehension and, at the same time, prevents hyper-optimization. We need to "get" the code, not just run it.
Post-AI codebases let go of all pretense to be both accessible by interpreters/compilers/runtimes AND humans, and contain JUST what is needed to provide the expected outcomes.
And that is what developers are realizing they must measure, review, and verify. Not the code, but the constraints and tests around it.
Codebases are going to start looking like optimized topology 3d prints
When AI is better at math than humans it also becomes better at algorithm & codebase design
At that point the system and delivering quality software doesn’t require human understanding to exist in the actual code layer.
Only needs to be understood by humans at the high level of passing testing, verification, vibe checks and being understandable when queried via AI in the rare case one need to dive deeper.
Nick Dobos retweeted
A lot of people are mocking this, but I think they’re wrong. I’m not kind to AI for the AI’s sake; I’m kind to AI for my own sake. It’s not good for *me* to be cruel or abusive.