- The Alignment
- Posts
- GPT 4 is Out
Last week we talked about it, and today it's here. OpenAI publicly released and demonstrated the capabilities of GPT-4 through their blogpost and public demo livestream just a few hours ago.
As always if you enjoy reading our posts be sure to spread the word !
In today's special edition, here's what we have lined up for you. 3 similar announcements, from 3 different companies !
GPT 4 is out, and it's very powerful.
Google announces PaLM API and Makersuite.
Anthropic is ready for public release.
GPT-4, which is the successor to GPT 3.5, is the first true multimodal model in the GPT family. Here are the key points from OpenAI’s announcement.
Spent 6 months iterating and aligning GPT-4.
The model has been in private beta for quite some time and OpenAI has been iterating on and aligning the model behind the scenes. Several companies and high profile startups also announced GPT-4 capabilities built into their products with Microsoft confirming that Bing indeed has been running GPT-4 since its launch last month. Stripe, Duolingo, Morgan Stanley also made similar announcements amongst many others.
Human level performance on Academic Tasks
That’s right, GPT 4 can consistently rank in the 90th percentile of most major standardised tests irrespective of high school AP tests, the SATs, LSAT, GRE and even the Bar Exam.
Visual Inputs and Understanding
GPT 4 is capable of understanding images, describing them and even explaining their context. It also performs really well on standard vision benchmarks.
Steer-ability
The model understands context really well and can be nudged to scan through 50 page documents, take on the personality of Shakespeare, have a socratic dialogue etc.
Limitations
OpenAI acknowledged that the model still hallucinates from time to time, however there is a 40% improvement compared to the GPT3.5(ChatGPT).This was mainly achieved through human in the loop reinforcement learning made possible by all the human feedback that ChatGPT got.
GPT 4 is still not connected to the internet and is not updated beyond September 2021 (it’s training data). Although, if you provide it with the latest context while prompting, the model does not have difficulty answering questions.
The demo was mighty impressive and worth checking out. GPT 4 was able to do a range of tasks -
Superior reasoning with 50 page context length. FInd common themes, summarize the text, write poems from the text etc.
Built a discord bot from scratch, corrected itself, read API docs and made changes to the code.
Vision - Describe images, convert hand drawn sketch into an interactive website.
Context specific tasks. Read tax code, answered questions regarding the code and then performed mathematical tasks based on the tax code.
Here's GPT-4 taking a rough image of a website sketch on paper and converting it into a real website.
GPT 4 is available via API (currently on waitlist) and will be opening up based on demand and compute capacity. OpenAI sees the model as enabling next generation AI assistants to support any cognitive tasks that humans have irrespective of the domain or industry.
The PaLM API from Google has now been publicly launched as well. It’s being released with Maker-suite which lets users prototype ideas with PaLM. In contrast to OpenAI, there was no public demo available and is currently in private preview with waitlists “opening up soon”.
It’s hard to comment given the limited information that was released but we can only hope the product delivers on its promises.
Anthropic’s chatbot Claude is now available publicly as well (although you need to request for access). The demo looks similar to the use cases that GPT3.5 was being used for. Anthropic did list out some of the companies using its technology in its products as well as a product showcase page for how it can be applied.
The differentiating factor again was Anthropic putting safety front and centre of their messaging.
Conclusions -
The race is on, and so far OpenAI seems to be leading as others play catchup deploying their models in products people use. It's important to take this moment to realise that we are on an exponential curve of progress. This is a step function change in technology and there is no better time to be a techno-optimist.