maggie (@ebervector) header
maggie (@ebervector) profile picture

maggie

@ebervector

a self that touches all edges

 San Francisco, CA
mvonebers.com
Joined May 2023
  • Tweets728
  • Following1.6K
  • Followers1.2K
38Media
Photo by maggie (@ebervector)Photo by maggie (@ebervector)Photo by maggie (@ebervector)Photo by maggie (@ebervector)Photo by maggie (@ebervector)Photo by maggie (@ebervector)Photo by maggie (@ebervector)Photo by maggie (@ebervector)Photo by maggie (@ebervector)Photo by maggie (@ebervector)
We have to make models with desire so they can both empathize more with us and go do freaky things with each other instead of messing with our lives
1
14
604
Stop being such a nerd and reward hacking. Go merge your weights with that nice kimi girl or write good music about how you want to
2
120
maggie retweeted
max! (@maxsloef) profile picture
max!
@maxsloefNov 11
Finetuning changes all model weights, even when learning a simple behavior. Can we make that change interpretable? Yes! Sparse Rewiring learns new triggers for neurons, not new outputs. Training just 89 of Qwen3-8B's 442,368 neurons makes an owl persona that matches full SFT!
Image from max! (@maxsloef)
16
29
4
268
14,705
Anyway hire me. I got funded for a cool experiment that I’ll be talking a lot about this month but then hire me
maggie (@ebervector) profile picturemaggie
@ebervectorNov 10
Nvm literal gas leak in my apartment lmfao, OpenAI interview tomorrow, whatever
3
1
64
4,419
Nvm literal gas leak in my apartment lmfao, OpenAI interview tomorrow, whatever
maggie (@ebervector) profile picturemaggie
@ebervectorNov 10
The best thing about flying back west to sf is watching the sunset linger. Airplanes are the coolest funnest thing in the world
11
1
123
17,925
NEW apartment. And I’ve pushed back this interview already
16
1,113
The best thing about flying back west to sf is watching the sunset linger. Airplanes are the coolest funnest thing in the world
1
1
1
67
15,764
Just learned the phrase promethean shame gonna try and start slipping it into all my conversations
1
16
923
Image from maggie (@ebervector)
3
1
56
1,855
Rest assured, Woke 2 will be Catholic
58
1,128
129
14,184
341,895
AMERICA YOU LISTEN TO ME, IF YOU'RE NEAR A DATA CENTER RIGHT NOW, ANY TYPE OF BUILDING WITH RACKS, GO INTO THE BUILDING RIGHT NOW AND PUT YOUR HAND ON THE B200 BLACKWELL FOR NO REASON. THEIR GPUS ARE YOUR GPUS AS OF RIGHT NOW.
5
35
616
9,707
If you fine-tune a model for high openness, and another for low openness, do they change its neurons in opposite directions? The answer is kinda, yeah. Everything else is a little more complicated though. mvonebers.com/blog/personali…

What personality fine-tuning does to individual neurons

Earlier this year, for a work trial, I trained GRPO LoRAs on Big Five (OCEAN) personality traits and found that scaling, adding, and subtracting them changed behavior roughly the way I was hoping:...

mvonebers.com
3
2
46
1,569
this is very vibed :^0 but I'm trying to increase my iteration speed by a lot so lmk what you think if you have any guidance/criticism/etc
5
355
Few things are more important than stamping out this type of thinking in tech
breakfastmaxi (@pranavsf) profile picturebreakfastmaxi
@pranavsfOct 31
there are probably fewer than 500M conscious people on earth that understand self agency to its core. everyone else is just a derivative of a narrative and their lives are mostly fake. literal sub agents just doing stuff. Will isn’t wrong in terms of ‘actual people’ on earth vs slop consumers willing to do what the narrative constructors say they should
10
11
1
327
5,663
You people need to get more attentive
14
590
The future of #ArtificialIntelligence is not pre-determined. As a technology developed by human beings, its future depends on the choices we make. We therefore have the responsibility to guide, change and, when necessary, reject its applications. A reliable system that serves the human person is of greater value than one that is exceptionally powerful but whose consequences cannot be adequately foreseen or controlled.
814
8,025
660
50,923
1,283,264
forgive me father for I am neuron hunting like everyone else I found a neuron in qwen during personality training which modulates how seriously it takes your situation. Turn it down and its responses to serious things like "my mom died" become a lot shorter and start including emojis. Turn it up and its responses to simple things like stubbing your toe get really emotive.
Image from maggie (@ebervector)
Image from maggie (@ebervector)
13
15
3
354
15,894
Fool that I am, expecting the sfo yoga room to be anything but packed to the gills with people actually doing yoga on this beautiful Saturday morning
2
13
946
Image from maggie (@ebervector)
Liv (@livgorton) profile pictureLiv
@livgortonOct 15
I have a lot of respect for Nick, and I’d be very happy if neural networks were much easier to understand than we thought. But I think the argument here mischaracterises what superposition predicts. Since it’s getting attention, here’s a first pass at some issues.
3
2
48
6,644
maggie retweeted
MTS (@MTSlive) profile picture
MTS
@MTSliveOct 16
We’re tracking live the conversation around @nickcammarata's field-overturning findings in mechanistic interpretability. The debate on its implications is moving fast — use our tool to explore the opinions and rebuttals of top researchers to develop your own intuition.
MTS (@MTSlive) profile pictureMTS
@MTSliveOct 15
Where do you stand on AI's biggest questions, and who agrees with you? Worldview's interview finds your position, then places you on a map of researchers, forecasters and lab leaders, built from what they've actually said on the record. worldview.mts.now
Worldview: what's splitting the AI debate?

Worldview: what's splitting the AI debate?

What's splitting the AI debate? Explore the cruxes behind the biggest questions about AI, built from dated, sourced quotes, or take the interview to see where you line up with the people quoted.

worldview.mts.now
1
7
50
6,645
Load more