Genevieve Conaty

I ♥ all forms of experience design, especially restaurants, stand-up comedy, and voting

[Genevieve Conaty @genbyday]

Happiness is a cold vaccine

[Anaridis Rodriguez @Anaridis]

Boston health care workers dancing on the sidewalk to celebrate the arrival of the first covid-19 vaccine.

[Reza Jamei @RezaJamei46330]

How do we teach LLMs to generate original scientific thought? One missing ingredient may be simulating the internal state of a scientist who doesn't yet know the answer. A toy version of the idea is studied in detail in the paper linked in the first comment. Here is a high-level summary:
When humans learn something — a skill, a fact, a concept — they generate an internal thinking corpus that leads them to accept it, along with preference data for which thought patterns were more or less helpful. Benjamin Franklin famously practiced this deliberately: he'd read an essay, set it aside until he forgot the exact wording, reconstruct it from the ideas alone, then compare the gap.
In contrast, we typically don't train LLMs this way. The vast majority of LLMs' training data is language meant for communication, not internal thinking. Research shows (o1, STaR) massive benefits from limited versions of adding preference data over reasoning to the training pipeline of models after training (doable for mechanically verifiable answers).
Still, LLMs have a huge repository of knowledge and even problem-solving skills that they're trained on without training on the process that discovered those methods and facts. It's of course not easy to generate genuine reasoning data that could lead to discovery using a model that already knows the answer. Prompting a model to pretend it doesn't know something leaves its world model intact — it's role-play, not genuine uncertainty. I've experimented with creating lesions — deliberately inducing a temporary misconception — in an LLM's weights so that it answers questions about one issue poorly, then letting it generate many thinking paths around the issue and grading those traces based on whether they make progress toward recovering the un-lesioned knowledge. For example: fine-tune a model to believe organisms can deliberately rewrite their own DNA to adapt (Lamarckian evolution), then prompt it with contradictory evidence like the Luria-Delbrück experiment, and grade whether it notices the tension. The goal is to simulate the kinds of situations we care about: cases where a model does not know a method or holds inconsistent beliefs.
Early results in a small-scale study are promising:

1. training a clean model on thinking preference data generated from 55 training lesions increased preference toward recovery on held-out pairs from 56 different lesions in 76-80% of the test cases.
2. when we lesion both the base model and the preference-trained model on held-out misconceptions and compare their greedy-decoded responses, the trained model scores about 1.2-1.4 points higher on a 1-10 GPT-4 judge scale.

The primary metric here is transfer on held-out lesion identities, while the generation result is supplementary evidence under lesion stress. The cleaner test — whether the clean trained model produces better reasoning without being lesioned — remains future work. Neither metric is ideal and may not translate directly to realistic generative tasks, but I think this is about as much as an experiment at this scale can show. Given the scalability of the experiment, I think this method is a good candidate to produce a meaningful improvement in the quality of thinking in LLMs, especially in the area of generating original scientific thought.

[Genevieve Conaty @genbyday]

@JeremyKonyndyk @USAID You have been a voice of reason and guiding light from the earliest days of the pandemic! I have been hoping you would opt to serve in this administration, so I am tickled pink to hear this wish came true. Break a leg, and I’ll stay for the bread 💕

[Genevieve Conaty @genbyday]

How I wish that this had only been "Stupid Watergate" https://t.co/jlJNk8pUd3

[Eric Topol @EricTopol]

Corrected @realDonaldTrump's last tweet.
@Twitter @Jack It should be his last tweet.
Please make that happen.

[Ben Rhodes @brhodes]

America has always had competing stories. I believe in the one we saw last night in Georgia, and it's imperative for that story to persevere and ultimately prevail.

[Jessie Brill 💚🌊☀️ @BioJessieB]

Thanks for the reminder, @jonlovett. And just so you know the @votesaveamerica community has been an amazing support system today. So thanks for that too. Cc: @jonfavs @TVietor08 💙

[SF QuakeBot @earthquakesSF]

A 3.6 magnitude earthquake occurred 4.35mi S of Muir Beach, CA. Details: https://t.co/bTSPtuObPC Map: https://t.co/0WqR5wgl63

[Genevieve Conaty @genbyday]

I too am practicing law on my own

[Genevieve Conaty @genbyday]

“Cancel Trump’s 2nd inaugural” ✅ thank you @benwikler and @WisDems 💙

[Ben Wikler @benwikler]

Our state party entered 2020 with the WISCO plan:

Win the spring election; use it as a rehearsal for fall
Inspire/recruit/train at the Dem National Convention
Save the veto—stop GOP supermajorities
Cancel Trump's 2nd inaugural
Organize for the long term

[Genevieve Conaty @genbyday]

P.S. @staceyabrams was born in Madison, Wisconsin 💙

[Ben Wikler @benwikler]

There's a lot of totally correct talk about how @StaceyAbrams was pivotal to winning Georgia. Folks: Stacey & her team were pivotal to flipping Wisconsin too. And every other battleground. They worked with us to build massive, supercharged voter protection teams—starting *early.*

[Genevieve Conaty @genbyday]

That’s how we do! So proud to have played even the tiniest role in the WISCO plan 💙 thank you again to @JillianSinger_ @ctrlcori @haley_arata and @tylercgreen for running the best volunteer program I’ve ever been part of 💙 and thank you to all my fellow volunteers!

[Ben Wikler @benwikler]

THREAD: The Biden win over Trump in Wisconsin is the result of thousands of heroes, working in the trenches for years.

To everyone who made this possible—THANK YOU. Here's one story of the path to this victory, and some of the people & groups involved. Everything mattered.

[Genevieve Conaty @genbyday]

The Clone Wars! I am ready 💙💙💙

[Genevieve Conaty @genbyday]

Yeah it worked in Wisconsin! Personally I would gladly donate to join 50 Princess Bride reunions 💙

[Genevieve Conaty @genbyday]

Something to do while you wait ...

[Genevieve Conaty @genbyday]

I always ask @staceyabrams this question 💙💙💙

[Wajahat Ali @WajahatAli]

If @staceyabrams helps Dems get Georgia and tie the Senate, which is still doable, Biden should call her and ask, "What do you want? And thank you."

[Genevieve Conaty @genbyday]

Team, I made a playlist for those of you (like me) who HATE WAITING https://t.co/mLYTf4IiH9

[Haley Arata @haley_arata]

And none of this happens without VOLUNTEERS! These volunteer leaders brought their heart - and they brought it home for WI. @pammie823 @carlyalys
@dinaalhadeff @achang @deblamb67 @jleeolson
@Gainington @bunnyhugger75
@adamskopp @jesscatlin @randieru
@genbyday & those w/o Twitter!

[Genevieve Conaty @genbyday]

Shout-out to Steph, one of our last-minute WisDems texters last night - blasting out texts to encourage voters to stay in line to vote in Milwaukee 😍💙

[Genevieve Conaty @genbyday]

This was our bad-ass WisDems Staff team for us texters 💙💙💙 I would happily vote for any of them in 2024 or 2028 💙💙💙

[Haley Arata @haley_arata]

A flipped #Wisconsin. A finite margin. A kickass #DigitalOrganizing team I’m SO glad I was a part of. S/o my WI Text Team family @JillianSinger_ @tylercgreen @ctrlcori

[Genevieve Conaty @genbyday]

Meet some of my very favorite people 💙

[Haley Arata @haley_arata]

And none of this happens without VOLUNTEERS! These volunteer leaders brought their heart - and they brought it home for WI. @pammie823 @carlyalys
@dinaalhadeff @achang @deblamb67 @jleeolson
@Gainington @bunnyhugger75
@adamskopp @jesscatlin @randieru
@genbyday & those w/o Twitter!