summary of open ai’s devday

keynote | nov 2023

michael raspuzzi
5 min readNov 7, 2023
source: screenshot from iphone demo + photo of new gpt release

there has not been a release as big as the iphone in 2007 until today.

sam altman, ceo of open ai, hosted their first developers’ day sharing big updates to features and plans for their ai models.

this marks one year since the chat gpt interface that broke the internet with 100M users in 2 months. they now have 100M active users each week. while it’s not technically a new unveiling, it shares major updates signalling what’s to come in the year ahead.

from a marketplace for custom gpts, like one specifically for canva workflows, to the new assistance api, this is the first major updates of many enabling ubiquitous ai tooling.

here’s a recap of the top takeaways for what was announced.

top 5 highlights, in case you missed it

listen from 12:40 onward for pricing update

1. larger context windows for a lower price.

two limiting factors for using gpt models have been size of input and the cost to run.

for the first one on limited context window for inputs: gpt-4 was previously 8k-32k. now gpt-4 turbo supports 128k tokens. that means it can take in about 300 pages of text… ie. full novels, textbooks, reports, legal documents, and more. larger inputs unlocks more robust applications.

the cost is also 2x–3x cheaper. chat gpt-4 was previously $0.03 per 1k token completion and now is $0.01 per 1k token. cheaper = more accessibility.

👉more on pricing updates here.

dalle-3 paints itself

2. chat gpt gets eyes and hands.

gpt-4 turbo will include access to the vision api. evolving from text input and text output only, gpt is now multimodal taking in audio, images, documents, and different data sources.

in terms of hands, it’s integrated with dalle-3 so image generation is seamless. one benefit of gpt v. midjourney is the textual understanding can help users design better prompts, getting one step closer to pure natural language input. rip gen 1 prompt engineers. text to speech will also be available.

👉more on the vision guide documentation here.

check out more at 8:48

3. chat gets a voice.

the new text to speech (tts) integration has two different model variants.

  • tts-1 is meant for real time transcription optimizing for speed.
  • tts-1-hd focuses on quality.

this enables a more natural way to interact with open ai’s models. the powerful application is combining vision based inputs (like a photo of an environment), with a reasoning based questions (chatting with it), to voice output. we’ll start to see spatial gpts everywhere.

👉 more on tts here.

check out from 26:24 for demo of sam building the startup mentor (aka sam gpt)

4. as apple has the app store ($80B annualy) open ai launches its gpt store (potential for $T+ annually).

different developers can move beyond plugin mode and into full application mode.

this brings together custom features released recently, like systems level prompting, as well as custom data integration.

some example gpts demo’d:

  • personal gpt to answer any questions based on personal knowledge graph
  • canva gpt for making a prototype of a poster and clicking out into editor
  • code gpt with’s curriculum data integrated into the chat interface

thinking in the oprah meme: you get a gpt. you get a gpt. and you get a gpt.

👉 more on gpt creator here.

more from 33:15 onwards

5. the assistants api will integrate agent like experiences onto any platform.

from natural language based data analysts to different smart assistants, this new api will enable platforms to have custom integrations with the back end power of open ai.

the more impressive thing is their tooling will make it easier to deploy solving challenges in state and prompt management or even limited capabilities.

four main features for assistants api include:

  • threading: persistent history in conversation
  • retrieval: built-in search
  • code interpreter: a working python interpreter in a sandbox environment
  • function calling: ability for multiple functions to take place

they demo’d what a travel booking assistant would look like on a travel website. it can:

  • remember the conversation from chat history
  • import and interpret documents like flight details or airbnb receipts
  • analyze those documents for a multi step function like splitting a tab per day including currency conversion

👉 more on assistants api in documentation.

other updates

  • updated knowledge for world events from september 2021 cutoff to april 2023
  • improved function calling to call multiple functions in one message to always return valid functions in json mode
  • model outputs can be deterministic with reproducible outputs beta ultimately reducing hallucinations

implications and applications

these updates increase the access of different ai models with less dev time to set up, support, and iterate. teams will be able to build more robust applications with this new tooling.

there will be an explosion of co-pilot this and smart assistant that ensuring truly personalized journeys on different platforms and unlocking new irl experiences.

  1. multi-modal (image and voice) and multi-model integrations will mean new physical interfaces and experiences in ways that we haven’t seen before with glass screens that fall flat in interactivity with their surrounding environment. let’s see what johnny ive can do here.
  2. whole new ecosystem to trade services faster and cheaper. the gpt store will enable a new generation of gpt developers who ensure that knowledge and content keeps getting distributed. in an education use case, rather than buying the latest text book or cram to study for an exam, every student will have a personalized bio or physics tutor. cramming will be one use case to get students started, but exploring curiosity will keep them learning.
  3. smaller teams can implement and iterate faster. in startups, it will be an advantage to keep the team small while building out agents for each use case to prove different concepts. the marketing gpt can generate its own social media campaigns, understand its own data, and iterate accordingly.

more resources

bonus: if you’re in the sf bay area, jump in to the ‘emergency’ hack to build and experiment.

or if you’re somewhere else ^ copy the idea and host your own local hackathon.

hey, i’m michael and i’m exploring the intersection of ai and alt education. the next generation of innovators deserve next generation tooling.

connect with me on linkedin to follow my build journey.



michael raspuzzi

building something new. previously @tks @harvard @culinaryai