I've spent a two-week vacation to fine-tune a large language model on my writing from the last decade to produce what I lovingly call #IAN (intelligence artificielle neuronale). I wrote a Substack post about what it is and how I made it! Check it out :)
Jan Hendrik Kirchner
@janhkirchner
phd student in comp neuroscience @ mpi brain research frankfurt, universalprior.substack.com,
waluigi theorist
Science & TechnologySan Francisco, CAJoined March 2018
Jan Hendrik Kirchner’s Tweets
phrenology vindication arc
Quote Tweet
Brain shape might trump connectivity
psypost.org/2023/06/landma
Show this thread
1
1
𝗖𝗹𝘂𝘀𝘁𝗲𝗿𝗶𝗻𝗴 𝗶𝘀 𝗺𝗮𝘁𝗵𝗲𝗺𝗮𝘁𝗶𝗰𝗮𝗹𝗹𝘆 𝗶𝗺𝗽𝗼𝘀𝘀𝗶𝗯𝗹𝗲!
Kleinberg (2002) stated three axioms that any clustering procedure should satisfy and showed there is no clustering procedure that simultaneously satisfies all three.
Intuition for this striking result👇
9
79
426
the distinct cadence of text generated by a chatbot coerced into spouting nonsense
Quote Tweet
sampling only ever demonstrates the presence, not the absence of capability. would be surprised if sota models couldn’t play tic-tac-toe optimally after finetuning (which surfaces capabilities that are already there)
Quote Tweet
When @GaryMarcus and others point out that GPT-4 is bad at chess and therefore not close to AGI, it falls flat for me.
But when I can’t coax GPT-4 to defeat me at *tic-tac-toe*, I start to think there’s something even more deeply wrong than I realized.
poe.com/s/KxQMDTGMzBIT
Show this thread
6
8
minmaxing in poker is nontrivial
Quote Tweet
2
dang
Quote Tweet
Replying to @ESYudkowsky and @davidad
I wanted to understand this exchange better. I quoted the two tweets in #chatgpt4 and asked:
Can you please explain what their disagreement is, and which position is more rational.
This was the response:
The first tweet suggests that a type of machine learning model,… Show more
1
there’s a lesson about scaling here
Quote Tweet
Commentary at greater length:
- I'm encouraged that somebody ran right out and tried this.
- It's not clear (to me, yet) that it worked all that well, or better than expected; I have not yet signficantly updated my model of how technically hard interpretability is.
- It is… Show more
Quote Tweet
new research from OpenAI used gpt4 to label all 307,200 neurons in gpt2, labeling each with plain english descriptions of the role each neuron plays in the model.
this opens up a new direction in explainability and alignment in AI, helping make models more explainable and… Show more
Show this thread
63
80
1,018
We applied GPT-4 to interpretability — automatically proposing explanations for GPT-2's 300k neurons — and found neurons responding to concepts like similes, “things done correctly,” or expressions of certainty. We aim to use Al to help us understand Al: openai.com/research/langu
363
1,273
5,585
[IMPORTANT ❗] I feel a sense of duty to warn people about social media posts they'll be seeing in AI in the upcoming months.
Web3/crypto/salesy vibes came to AI big time. Huge monetary incentives are at play so many of them flocked to the space...
1/ 🧵👇
4
16
71
Show this thread
We’re developing a new tool to help distinguish between AI-written and human-written text. We’re releasing an initial version to collect feedback and hope to share improved methods in the future.
715
2,970
9,494
Super excited to share an initial version of a tool to help distinguish AI-written and human-written text! Would love to hear feedback for improving the tool in the future!
3
9
72
Everyone has a right to know whether they are interacting with a human or AI.
Language models like ChatGPT are good at posing as humans.
So we trained a classifier to distinguish between AI-written and human-written text.
But it's not fully reliable.
24
30
161
Show this thread
Awesome work by , , and many others!
It's been amazing to see how quickly it came together!
2
1
8
Show this thread
(I know the link doesn't look legit, I promise it's legit! All the better domain names were taken.)
Show this thread
The post has a bunch of pretty pictures and schematics, but really is just a thinly veiled attempt to get people to label data for me. Check out fashionator.xyz and tell me about your aesthetic preferences!
1
Show this thread
I wrote about fashion and ended up with a deep(-ish) meditation on all things pretty! Check it out!
1
1
7
Show this thread
would the past have been less dangerous if the atomic bomb had been released incrementally? (slowly ramping up in units of equivalent tnt)
1
3
has written a beautiful piece about the procrastination support group that he organized for me
2
4
This one was a ton of fun to write and work on in general!
Quote Tweet
@janhkirchner has written a beautiful piece about the procrastination support group that he organized for me
universalprior.substack.com/p/simulator-mu
3
A subset A ⊆ ℝ² is called a *cloud* around 𝑥 if every line through 𝑥 has a finite intersection with A.
[Komjáth 2001]: The following are equivalent:
(1) Three clouds cover ℝ².
(2) The continuum hypothesis holds.
26
198
1,328
The application of ChatGPT I'm most happy about is that finally my parents have heard of where I work
36
105
4,693
We’ve trained language models to be better at responding to adversarial questions, without becoming obtuse and saying very little. We do this by conditioning them with a simple set of behavioral principles via a technique called Constitutional AI: anthropic.com/constitutional
27
326
1,277
Show this thread
ChatGPT is incredibly limited, but good enough at some things to create a misleading impression of greatness.
it's a mistake to be relying on it for anything important right now. it’s a preview of progress; we have lots of work to do on robustness and truthfulness.
899
4,253
28.5K
Show this thread
🙏 to my colleagues for making it feel like i get to walk into the modern bell labs every day.
please consider joining us at openai! being in the room for research breakthroughs is awesome, the problems are interesting, the people are great, and we ship.
youtu.be/AyOnug-3OKM
130
200
4,094
While I still haven't gotten back into the rhythm of writing regularly (🥲) my good friend Sam is much better at it and is bringing some very exciting research discussions to paper. Highly recommended! snellessen.substack.com/p/introduction
3
New post in our Best of Science Blogging feed - “Drug Addicts and Deceptively Aligned Agents - a Comparative Analysis” by the excellent and Nadia Montazeri
4
9
If everything's a big manifold, why do neurons often code for human-interpretable factors? In arxiv.org/abs/2210.01768 we show the most efficient biological representation puts different factors in different neurons. This lets us build machines that disentangle too! 1/9
6
159
652
Show this thread
New Paper! "What are the Red Flags for Neural Network Suffering?" by and . Big thanks to gardeners , , , , , and many others for their reviews!
1
8
18
Show this thread
At EAG SF and interested in AI safety? Stop by OpenAI at the career fair or DM me. We're hiring across teams (incl. Trust and Safety, Security) - and we are always interested in hearing what you want to work on & what you think we should work on!
1
7
24
it’s been a blast reading Ava’s thoughts on this! She has a fantastic knack at getting art out of DALLE
Quote Tweet
Over the past few weeks, I've been wondering about the impact & utility of DALL-E & related algorithms.
My awesome friend @janhkirchner just joined @OpenAI & I got to play around with DALL-E.
With that, here's a piece I wrote for
Jan's substack. Enjoy!
universalprior.substack.com/p/hello-dall-e
1
2
Over the past few weeks, I've been wondering about the impact & utility of DALL-E & related algorithms.
My awesome friend just joined & I got to play around with DALL-E.
With that, here's a piece I wrote for
Jan's substack. Enjoy!
2
4
(I've quietly published a short post on that already some days ago - just another reason to subscribe to my Substack universalprior.substack.com/p/a-quick-one- )
2
Show this thread
When I was young(er) I started coding because I wanted to build AI. That's pretty difficult, so I pivoted to "being part of the team that builds AGI". Now I'm happy to announce that I'm approaching my goal - I've joined OpenAI (Alignment Team) 🥳 Looking forward to exciting times
8
98
Show this thread
(unless you're working in a field like physics or mathematics where all the low-hanging might *actually* have already been picked)
Show this thread
























