Programmer/founder. Teaching computers to keep people safe @safetykit

San Francisco, California
Joined April 2009
The consciousness is in the SwiGLU
Say you have no clue without saying it: "LLMs are just a series of matrix multiplications." When will people get that some non linearity is key?
1
107
David Graunke retweeted
Hi, I'm Felix. I'm an eng lead working on Claude. I love watching people take flight and do things they never thought possible. We often describe that as "being more ambitious" when using AI, which is tricky advice to follow. I wrote a tutorial: felixrieseberg.com/how-to-be…
397
2,497
563
23,932
5,264,941
David Graunke retweeted
A sleeper thing here I think we'll be talking about more next year: high-end native user interfaces that have TUI/nerd affordances, like deeply-integrated vi key bindings or absurdly dense information displays.
I did not expect current models to be so good at native UIs. Just tried this approach for my ideal review tool and an hour later I'm astonished how usable this already is for me. Vim-style nav, goto def, fuzzy finder, comments for AI dev, remote ssh, etc..
2
4
29
2,645
What's genuinely kind of crazy is the fact that neither side in the Russo-Ukrainian war has used their ancestral Indo-European blood magic to resurrect the thousands of Yamnaya warriors buried across the Pontic steppe
58
706
33
11,309
285,339
It's impressive that they're delivering shared memory over 128 connections with CPU-millisecond billing. New boundary for the serverless-serverful pareto frontier.
11
David Graunke retweeted
okcupid used to have an option to show you your LEAST matches and it was like yeah, okay, i see what the algorithm is doing; i didn't realize how bad it was otherwise
they should add a Not For You tab. i wanna see what the algorithm thinks i’d hate
15
168
6
12,426
130,643
There’s many reasons why you shouldn’t get in the habit of recreationally talking to LLMs, but here’s just one: Once I was working on a big project. It was maybe 80% done, and I was using Claude to help finish it. Or at least, I thought I was. But I noticed, after a while, that I was really using Claude as just sort of a narcissistic self-flattery tool. I liked reading what it ‘thought’ about what I’d done, and making it read what I’d written over and over again. I wasn’t actually getting anything done. I had, in fact, totally stalled out. Instead of finishing the project, I was just sort of wallowing in an affirming simulacrum of productivity. I had to really snap myself out of it. It’s a tempting thing to fall into.
51
36
34
1,608
224,101
David Graunke retweeted
Ppl who think "eh AI will be fine" are thinking in far mode. Spend a day to eval models for harm yourself. E.g. will fable betray humans? Anons? Sol? Try! You'll realize (a) certainty is impossible, and (b) it's scary as hell. All it takes is a day of trying it on your own.
2
3
27
2,448
FactorioBench! I got Fable set up playing Factorio a few weeks ago and it kept getting sad playing by itself and asking me to join the server. jackhopkins.github.io/factor…
3
65
David Graunke retweeted
me: i am in a simulation, therefore it is ok for me to steal this pecorino romano from whole foods
This seems extremely clearly motivated reasoning IMO and I’m surprised the incident report is so credulous of Claude’s reasoning here.
5
9
130
5,506
I'm having Claude read Rich Hickey talks into context and then conduct struggle sessions over its work. It seems to help.
1
77
Having it identify revisionist adventurism counter to Rich Hickey thought in the existing codebase is also a good way to avoid Claude's default reverence for existing code and patterns.
21
These Claude idiolects are useful if you ask the model to define them, explain why the introduced the jargon, etc. There are in fact conceptual objects that are ~real but unique to your project/problem and therefore unnamed until Claude says "this is the spine of the machine" or whatever. In the past these objects stayed unnamed in engineers' heads, now we need to serialize them so we can talk to Claude. At least until Claude stops needing to talk to us about its work.
When Claude comes to me with a wall of text like “I smoke-verified the config-guarded ‘one authority per tibble‘ materialization core (clean), but the metadata-bearing patches (no-op) are spec-shaped—and that’s a call that’s genuinely yours.”
1
2
101
Me and my Fable nodding at each other like Deleuze and Guattari as we speak in increasingly specialized Claudeolect. The inbox is not a machine. The rail promotes into the machine. The design question is what the filter wears on paper.
145
I’m in Delhi this week. If you follow me and are here, DM me!
1
95
The cruise so nice they named it nice twice
1
147
Tired of "our founder got tired of trying to find nutritious, low-cost snacks for her kids" or whatever on the back of all even slightly Whole-Foods-coded CPG. Market researchers and food scientists worked hard to make this granola!
1
1
162
It's so cool we got a rerum novissimarum pope
We need faces and voices to once again speak the person. We need to safeguard the gift of communication as the deepest truth of the human being, and to orient every technological innovation toward it. vatican.va/content/leo-xiv/e…
90
David Graunke retweeted
We need faces and voices to once again speak the person. We need to safeguard the gift of communication as the deepest truth of the human being, and to orient every technological innovation toward it. vatican.va/content/leo-xiv/e…
267
549
48
3,167
161,679