As Elad Gil posted yesterday, LLMs are really opening the door for startups: adoption speed, infrastructure (open source), and clear use cases with value.
At the same time, GPT-3 has room to improve as a model, and that's why we focus on "curation" as the fine-tuning dataset.
As OpenAI said before, 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 their models safer, more helpful, and more aligned, they 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.
Some of their previous research in this direction found that they can reduce harmful outputs by fine-tuning on a small "curated dataset" of human demonstrations.
Here, I found an opportunity in Glasp where people (50%+ are writers) are curating important information for all.
Since our community has more than 700k unique highlights so far, leveraging this high-quality curated dataset for fine-tuning could bring a better outcome for users, especially in AI-powered writing.
Plus, one of the upcoming issues that I see is that if everyone gets access to the same technologies, how could you differentiate yourself? Where and how will your identity be reflected in the output? Only in the prompt? Isn't it sad?
Since "what you collect says so much about who you are" (thanks Ben Silbermann), by using the user's own curated information for the dataset, the output could reflect one's identity and taste well! This must be exciting!
Also, curated information can cover one of the weaknesses of the current models such as the old dataset issues because people are curating information constantly (+ info matters to them). It could fill the dataset freshness gap!
This is why I'm so excited to build an AI-powered writing system on top of curation. We focus on "curation to creation"
Here's our demo on AI-powered writing with curation dataset:
This is a solid newsletter, but you have a massive blind spot on the COVID mRNA jabs. The only things they have been successful at are generating massive profits for Pfizer/Moderna and ruining trust in public health. <5% of eligible kids under 5 have gotten the jab and <5% of eligible adults have gotten the bivalent booster, which was only tested in 8 mice. European countries like Denmark are recommending against them for healthy young people due to side effects like myocarditis. Pfizer/CDC lied that these jabs would stop transmission and did not affect menstrual cycles or breast milk. This is regulatory capture and criminality that deserves investigation and punishment. Read Dr. Vinay Prasad, Alex Berenson, and el gato malo for more details. You should also know better than quoting Prof G - his company Section4 is imploding because he is too busy chasing clout on CNN+ ;)
Imagine if even 10% of the items you covered today did pan out, it would mean some good steps forward.
Yes, AI is super hot again now!
And thankfully, our AI-powered writing (from curation to creation) got featured in his newsletter today ;)
https://www.youtube.com/watch?v=blZQw9CM_D8
As Elad Gil posted yesterday, LLMs are really opening the door for startups: adoption speed, infrastructure (open source), and clear use cases with value.
http://blog.eladgil.com/2022/10/ai-startup-vs-incumbent-value.html
At the same time, GPT-3 has room to improve as a model, and that's why we focus on "curation" as the fine-tuning dataset.
As OpenAI said before, 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.
https://openai.com/blog/instruction-following/
To make their models safer, more helpful, and more aligned, they 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.
Some of their previous research in this direction found that they can reduce harmful outputs by fine-tuning on a small "curated dataset" of human demonstrations.
Here, I found an opportunity in Glasp where people (50%+ are writers) are curating important information for all.
Since our community has more than 700k unique highlights so far, leveraging this high-quality curated dataset for fine-tuning could bring a better outcome for users, especially in AI-powered writing.
https://twitter.com/_Glasp/status/1578118436453916672
Plus, one of the upcoming issues that I see is that if everyone gets access to the same technologies, how could you differentiate yourself? Where and how will your identity be reflected in the output? Only in the prompt? Isn't it sad?
Since "what you collect says so much about who you are" (thanks Ben Silbermann), by using the user's own curated information for the dataset, the output could reflect one's identity and taste well! This must be exciting!
Also, curated information can cover one of the weaknesses of the current models such as the old dataset issues because people are curating information constantly (+ info matters to them). It could fill the dataset freshness gap!
This is why I'm so excited to build an AI-powered writing system on top of curation. We focus on "curation to creation"
Here's our demo on AI-powered writing with curation dataset:
https://www.youtube.com/watch?v=blZQw9CM_D8
Happy learning,
Kazuki
This is a solid newsletter, but you have a massive blind spot on the COVID mRNA jabs. The only things they have been successful at are generating massive profits for Pfizer/Moderna and ruining trust in public health. <5% of eligible kids under 5 have gotten the jab and <5% of eligible adults have gotten the bivalent booster, which was only tested in 8 mice. European countries like Denmark are recommending against them for healthy young people due to side effects like myocarditis. Pfizer/CDC lied that these jabs would stop transmission and did not affect menstrual cycles or breast milk. This is regulatory capture and criminality that deserves investigation and punishment. Read Dr. Vinay Prasad, Alex Berenson, and el gato malo for more details. You should also know better than quoting Prof G - his company Section4 is imploding because he is too busy chasing clout on CNN+ ;)
fusion!
thanks for being another optimistic voice out here. I continue to think its a choice and a positive life one at that.