Kazuki
@kazuki
Cofounder of Glasp. I collect ideas and stories worth sharing 📚
San Francisco, CA
Joined Oct 9, 2020
1069
Following
5727
Followers
1.46k
13.51k
168.42k
www.youtube.com/watch?v=9uOMectkCCs
Nov 10, 2022
111
talia.gold/2022/10/31/a-trillion-dollar-opportunity/
Nov 9, 2022
85
hardfork.substack.com/p/easy-come-easy-go-understanding-the
Nov 9, 2022
61
www.generalist.com/briefing/what-to-watch-in-ai
Nov 9, 2022
51
brianbeckcom.medium.com/great-advice-on-writing-from-tim-urban-e601053173cd
Nov 4, 2022
61
learningaloud.com/blog/2022/10/30/sharing-notes/
Nov 4, 2022
51
a16z.com/2022/09/21/what-china-can-teach-us-about-the-future-of-tiktok-and-video-search/
Nov 4, 2022
7
paulgraham.com/fr.html
Nov 2, 2022
5417
hardfork.substack.com/p/limiting-beliefs-invert-invert-invert
Nov 2, 2022
42
newsletter.mem.ai/p/ai-powered-psychoanalysis-journaling
Oct 29, 2022
153
theamericanscholar.org/solitude-and-leadership/
Oct 27, 2022
224
svenschnieders.github.io/curiosity/
Oct 27, 2022
155
www.lennysnewsletter.com/p/what-is-a-good-activation-rate
Oct 26, 2022
71
humanloop.com/blog/stability-ai-partnership
Oct 25, 2022
4
every.to/divinations/how-lex-happened
Oct 22, 2022
4
bensbites.beehiiv.com/p/build-website-30-seconds-ai
Oct 21, 2022
11
blog.eladgil.com/2022/10/ai-startup-vs-incumbent-value.html
Oct 21, 2022
185
medium.com/swlh/6-powerful-note-taking-tools-to-activate-your-mind-connect-ideas-548214069c5b
Oct 20, 2022
94
www.idc.com/getdoc.jsp?containerId=prUS48958822
Oct 18, 2022
3
medium.com/keep-productive/5-productivity-apps-hyped-up-right-now-44610dcc788a
Oct 18, 2022
11
openai.com/blog/instruction-following/
Oct 18, 2022
101
hai.stanford.edu/news/examining-emergent-abilities-large-language-models
Oct 16, 2022
4
eriktorenberg.substack.com/p/daos-and-the-iron-law-of-oligarchy
Oct 16, 2022
153
thesephist.com/posts/medium/
Oct 13, 2022
101
neurosciencenews.com/anxiety-dopamine-21390/
Oct 12, 2022
121
www.hongkiat.com/blog/glasp-vs-matter/
Oct 12, 2022
81
podcast.ai/about
Oct 12, 2022
3
www.readaccelerated.com/p/is-ai-art-ethical
Oct 10, 2022
72
hbr.org/2014/05/making-freemium-work
Oct 8, 2022
8
www.linkedin.com/pulse/20121002124206-18876785-how-to-model-viral-growth-the-hybrid-model/
Oct 8, 2022
4
outsetcapital.com/writing/posts/lead-preseeds
Oct 7, 2022
2
digitalnative.substack.com/p/the-tiktokization-of-everything
Oct 6, 2022
5
www.cs.virginia.edu/~robins/YouAndYourResearch.html
Oct 6, 2022
368
cdixon.org/2013/08/04/the-idea-maze
Oct 4, 2022
84
spark-public.s3.amazonaws.com/startup/lecture_slides/lecture5-market-wireframing-design.pdf
Oct 4, 2022
172
www.sequoiacap.com/article/generative-ai-a-creative-new-world/
Sep 30, 2022
11
also.roybahat.com/introductions-and-the-forward-intro-email-14e2827716a1
Sep 29, 2022
9
every.to/divinations/the-infinite-article
Sep 29, 2022
93
lspace.swyx.io/p/eigenquestions-for-the-ai-red-wedding
Sep 28, 2022
61
GPT-3 is trained to predict the next word on a large dataset of Internet text, rather than to safely perform the language task that the user wants. In other words, these models aren’t aligned with their users.
To make our models safer, more helpful, and more aligned, we use an existing technique called reinforcement learning from human feedback (RLHF).
The resulting InstructGPT models are much better at following instructions than GPT-3. They also make up facts less often, and show small decreases in toxic output generation.
Our labelers prefer outputs from our 1.3B InstructGPT model over outputs from a 175B GPT-3 model, despite having more than 100x fewer parameters.
Some of our previous research in this direction found that we can reduce harmful outputs by fine-tuning on a small curated dataset of human demonstrations.
We find that InstructGPT models are significantly preferred on prompts submitted to both the InstructGPT and GPT-3 models on the API.
We also conduct human evaluations on our API prompt distribution, and find that InstructGPT makes up facts (“hallucinates”) less often, and generates more appropriate outputs.
Despite making significant progress, our InstructGPT models are far from fully aligned or fully safe; they still generate toxic or biased outputs, make up facts, and generate sexual and violent content without explicit prompting.
they may become more susceptible to misuse if instructed to produce unsafe outputs. Solving this requires our models to refuse certain instructions; doing this reliably is an important open research problem
Right now, InstructGPT is trained to follow instructions in English; thus, it is biased towards the cultural values of English-speaking people. We are conducting research into understanding the differences and disagreements between labelers’ preferences so we can condition our models on the values of more specific populations.