What Did OpenAI Do This Week? - 23/07/2023 [+28 LINKS]

DEAR ‘OPENAI ROCKETSHIP’ - ARE THE FORCES WITH YOU?

OpenAI's head of trust and safety, Dave Willner, announced that ‘the rocketship of OpenAI’, will be um…rocketing without him (well, apart from an advisory role for now) to claw back time with his family as the company after just 18 months in the role. OpenAI then went on to blast out a host of product news and updates on safety, privacy and performance just to prove they are still, you know, interested in such things. You don’t have to have watched Scandal to know that something smells off with this scenario (breakdown, harassment, corporate disagreement), but the fact no replacement, only cover, has been named suggests serious issues vs. simply more family time needed. OpenAI is clearly going through a ‘high-intensity phase in its development’ (Willner), which will no doubt be music to +3.3 billion Android users’ ears when they finally get to start using ChatGPT this week, but does that usually take down the trust and safety personnel? No. No, it does not.

One of the newest features announced this week was ‘Custom Instructions’ for ChatGPT Plus subscribers on an opt-in basis, everywhere but the UK and EU (promised soon to all users). It’s designed to help you type a little less or as Sam Altman put it ‘one small step towards more personalized AI, but a surprisingly fun/useful one.’ The feature is in beta and works everywhere ChatGPT does (so likely to be super helpful on mobile devices). It gives you a place to tell your chatbot the things it should always know about you and how you’d like it to respond to your questions. You’ll need to answer two questions i) what would you like ChatGPT to know about you to provide better responses and ii) how would you like ChatGPT to respond?

Joanne Jang, who works on model behaviors and product at OpenAI, explained that the easiest way to think about it is as a sort of permanent preamble to your queries. Instead of crafting long questions with all the context and information required, you can just add that context and information to your custom instructions, and it’ll be there every time. Her practical example of its use for Teachers brought it to life …

“If you’re a teacher, you can put ‘I teach third grade’ into your custom instructions so that, every time you ask for interesting facts about the Moon, the bot can tailor its answer to the right age group”.

Whilst OpenAI tried to improve user experience this week, a study by Stanford and Berkeley researchers concluded that OpenAI's language models performed worse in some areas in June compared to March. For instance, it’s reported that GPT-4's accuracy in identifying prime numbers dropped from 97.6% to 2.4%. The non-peer-reviewed study also examined the performance of GPT-3.5 and GPT-4 in areas like solving math problems, answering dangerous/sensitive questions, generating code, and visual reasoning. GPT-4 showed less willingness to answer sensitive questions in June, and both models had more formatting mistakes in code generation. The paper highlights the issue of model drift, or a decline in the models' accuracy and performance over time. It wasn’t new news to some users who had already reported the model becoming less intelligent.

OpenAI's response has been decidedly subtle. Vice president of product, Peter Welinder denied intentional changes to make the models "dumber," claiming that users may notice more issues over time simply because they use ChatGPT more, and Logan Kilpatrick, head of developer relations promised to review reports of declining GPT-4 capabilities.

Want to know what this means for you and +28 news links this week? Subscribe now… Need more reasons? Forbes recommends WDODTW for one, and in a short time WDODTW is already amongst the top 100 AI Substacks:

SO WHAT?