The Incredible Powers of GPT-3.5
Michael Spencer
A.I. Writer, researcher and curator - full-time Newsletter publication manager.
Hey Everyone,
If you enjoy articles about A.I. at the intersection of breaking news join AiSupremacy?here. I cannot continue to write without community support. (follow the link below). For the price of a cup of coffee, Join 100 other paying subscribers.
https://aisupremacy.substack.com/subscribe
ChatGPT is causing both havoc and innovation. Banned on coding Q&A site Stack Overflow, you don't say? 1 Million free testers for OpenAI, that not so Open AI lab that is going full on commercial. It's a wacky world in A.I for sure as we head into 2023 in a few weeks time.
Wynter is an incredible B2B message testing platform. It's the fastest way to get feedback from your target customers and learn?how your messaging is resonating with them.?See their LinkedIn?here.
Get paid for your feedback, join Wynter's B2B Research Panel
Wynter is looking for people to join its research panel. Participate in B2B market research studies, get paid for your feedback and comments ($15-$100 per 5-15min). Super low time commitment.
~ Back to our topic:
While large language models are fine tuned, we’re still?waiting for GPT-4 to be announced. It’s been a long time since GPT-3 arrived, yet there are still new versions of it being released that show potential.
Released two years ago, OpenAI’s remarkably capable, if flawed, GPT-3 was perhaps the first to demonstrate that AI can write convincingly — if not perfectly — like a human.
Reinforcement Learning with Human Feedback
Now in December, 2022 we have?Davinci-003?and?ChatGPT. GPT-3.5 hit the world on November, 30th, with ChatGPT, a fine-tuned version of GPT-3.5 that’s essentially a general-purpose chatbot. What?Davinci-003 can do is also pretty stunning.
ChatGPT is more about a dialogue upgrade.?According?to OpenAI, GPT-3.5 was trained on a blend of text and code published prior to Q4 2021. Davinci-003 is their latest model builds on?InstructGPT, using reinforcement learning with human feedback to better align language models with human instructions.
Like GPT-3 and other text-generating AI, GPT-3.5 learned the relationships between sentences, words and parts of words by ingesting huge amounts of content from the web, including hundreds of thousands of Wikipedia entries, social media posts and news articles.
GPT-3.5 is Optimized to Play with Humans
What it can actually do at a decent level is also improving.
There’s the usual bragging of A.I. now such as it can help you?write code,?compose essays,?dream up stories, and?decorate your living room. Unlike?davinci-002, which uses supervised fine-tuning on human-written demonstrations and highly scored model samples to improve generation quality,?davinci-003?is a true reinforcement learning with human feedback (RLHF) model.
Various?Substack writers are also?trying to explain the jump from GPT-3 to GPT-3.5. Data scientists at Pepper Content, a content marketing platform,?report?that text-davinci-003 “performs better in understanding the ‘context’ behind a request and then using that to produce better content” while “hallucinating” less than GPT-3-based models. (Where it concerns text-generating AI,?hallucination?refers to an AI writing inconsistent, factually incorrect statements.)
Poetry, Chat and the Antics of Hype Pre GPT-4
Ars Technica?reports?that commenters on Y Combinator’s Hacker News forum used text-davinci-003 to write a poem explaining Albert Einstein’s theory of relativity and then re-write the poem in the style of John Keats. See:
If you want to understand Einstein’s thought
It’s not that hard if you give it a shot
General Relativity is the name of the game
Where space and time cannot remain the same
Mass affects the curvature of space
Which affects the flow of time’s race
An object’s motion will be affected
By the distortion that is detected
The closer you are to a large mass
The slower time will seem to pass
The farther away you may be
Time will speed up for you to see
While there’s a lot of hype about it on Twitter the demo of ChatGPT seems to be going well.
GPT-3.5 is therefore good publicity for OpenAI that is negotiating with Microsoft for more money.
OpenAI’s announcement email mentions the following improvements for?davinci-003:
Generative A.I. is being hyped as a significant step forward for the utility of Transformers land LLMs. It’s hard to tell how much this is hype and how much innovation will actually result from all of this.
Scale A.I. wrote?an SEO optimized blog?about Davinci-003 without any high level insights about what the trend actually means. Everyone wants in on the momentum, without really saying exactly what GPT-3.5 and GPT-4 will mean for society moving forwards.
I wouldn’t call it a breakthrough, it’s more like a playground at this point. OpenAI’s GPT-3 is a leader in large language model applications — but its conversational interface makes workshopping speeches and blog posts much easier.
Analyzing the poetic understanding of GPT-3.5 is not doing us any favors:
The Scale AI team even found that text-davinci-003/GPT-3.5 has a notion of meters like?iambic pentameter. See:
O gentle steeds, that bear me swift and
sure
Through fields of green and pathways so
obscure,
My heart doth swell with pride to be with
you
As on we ride the world a-fresh to view
The wind doth whistle through our hair so
free
And stirs a passion deep inside of me.
My soul doth lift, my spirits soar on high,
To ride with you, my truest friend, am I
Your strength and grace, your courage and
your fire,
Inspire us both to go beyond our sire.
No earthly bonds can hold us, only fate,
To gallop on, our wond’rous course create
However even a decade ago, many didn’t believe that A.I. would even touch or intersect with our more “creative” pursuits unique to human civilization. How wrong they turned out to be even in 2022.
AI chatbot?ChatGPT?has been trained to provide conversational answers to users’ queries and it’s also getting even more hype than the more impressive capabilities of Davinci-003. OpenAI has the distinction of creating foundational models that others build upon, but won’t be one of the more interesting startups in how it’s applied to a specific field.
Even as OpenAI partners with Microsoft it now appears like?Stability.AI is partnering with Amazon. This suggests that BigTech, and not startups will leverage Generative AI’s potential the most in the 2020s itself. It’s not clear how Generative A.I. startups scale or grow revenue without being essentially bought out in the Cloud. Thus all that feels new, is not always really new.
LLMs and GPT is a tweaking process of doing things better now with human feedback built in. It’s becoming clear just how powerful RLHF is to improve real-world performance. Like?InstructGPT,?GPT-3.5 was trained with the help of human trainers who ranked and rated the way early versions of the model responded to prompts. This information was then fed back into the system, which tuned its answers to match the trainers’ preferences.
Thus as quickly as GPT-3.5 appears, GPT-4 will soon be announced. And the speed of the iterations and the tweaking and the potential applications in various industries accelerates noticeably around the mid 2020s or like some believe, even today in mid 2022 or 2023.
According to industry sources, OpenAI quietly improved GPT-3 over time, making text-davinci-003 a notable public upgrade. ChatGPT certainly adds some novelty as we like to experiment with Chatbots to see what they can or cannot do.
The Future of A.I. Prompting
We don’t exactly know what the full potential and limitations of prompting are, but in a no-code scenario it’s a logical beginning. A.I. prompting other AIs to accomplish more complex tasks does make sense to some extent. At what point will we still need a human behind the wheel to create and accomplish things if this really will have an impact on the automation of tasks though?
Still, GPT-3.5 and its derivative models demonstrate that GPT-4 — whenever it arrives — won’t necessarily need a huge number of parameters to best the most capable text-generating systems today.
We still don’t know if this will even be overly useful, but we can hope it will. Venture Capital is making big bets on “Generative A.I.” as being transformative, yet we know how hype cycles usually turn out.
9-ChatGPT Tools
This list is from Ben, of Ben’s Bites in his awesome Newsletter. It gives you just a first glimpse at how ChatGPT is creating new tools around the web.
While it’s fun to think and write about Generative A.I., it will take many years to mature and may not be as impressive as forecasted.
GPTChat has captured the internet's imagination and it's incredible to see one million testers try out the limited research demo OpenAI provided. I'm averaging a lot of daily to weekly content across A.I. news, business, technology and the future impacts on society.
Thanks for reading!
If you enjoy articles about A.I. at the intersection of breaking news join AiSupremacy?here. I cannot continue to write without community support. (follow the link below). For the price of a cup of coffee, Join 100 other paying subscribers.
https://aisupremacy.substack.com/subscribe
https://aisupremacy.substack.com/subscribe
I'm consciously trying to support independent writers on Substack, that values their work.
Further reading:
A.I. Writer, researcher and curator - full-time Newsletter publication manager.
1 年Stack Overflow banned chatGPT here's why: https://www.theverge.com/2022/12/5/23493932/chatgpt-ai-generated-answers-temporarily-banned-stack-overflow-llms-dangers
A.I. Writer, researcher and curator - full-time Newsletter publication manager.
1 年To try ChatGPT go here: https://chat.openai.com/chat
C# - Dot Net Software Engineer
1 年Great piece! I love