- Inclined.ai
- Posts
- š¤ OpenAI Wasnāt Kidding
š¤ OpenAI Wasnāt Kidding
PLUS: Bard Is Catching Up
What's up? You're reading Inclined AI. TIME released an AI focused cover for their next issue. Thatās cool but does anyone even read the articles from them anymore?
Here's the front cover news:
OpenAI publishes new research
Bard gets a fresh update
Runway is rolling in fresh cash
Another AI Godfather is wary of AIās future
CHATGPT PICKS SOME LOW-HANGING FRUIT
Sam Altman told us all that GPT-4 is the companyās focus right now, because thereās a lot of easy improvements they can make to this model. Yesterday, they announced the first development in that effort.
The goal is to limit hallucinations in Large Language Models through a new form of supervision.
The result is a GPT-4 model thatās better at solving math problems without the Wolfram plugin attached.
What exactly did they do? Letās break it down.
To train a neural network on data you need to offer it a reward. When DeepMind trained models on games like Brick the reward was obvious: the points you get from scoring in the game.
Training a Large Language Model to do what it does for ChatGPT is a more complex reward system, but one thing never changed and that was where in the process they gave the feedback/reward.
What weāre talking about is Outcome Supervision versus Process Supervision.
It turns out Process Supervision shows a lot of promise for getting a model to reflect on itās chain of thought and work towards eliminating hallucinations in its output.
How? Itās the same thing your math teacher made you do to learn complex formulas. OpenAI is making LLMs show their work and giving it feedback at every stage of the process.
The trade off is obvious. More feedback means more human labeling. But the potential for Process Supervision is significant.
You may notice that ChatGPT can fall flat sometimes when given certain prompts. Some times this drop in performance is from something called an āalignment tax.ā
In order to get the model to behave and align to our values, it requires a drop in output.
OpenAIās research in Process Supervision shows a negative alignment tax for math problem solving. Thatās incredible and means we can better align models in the future without losing performance.
The company has not fine-tuned ChatGPT with this method, but that will come.
Regardless, itās refreshing to see this company publish some new stuff and remind us that the best is yet to come.
Have you ever been neck-deep in a project, only to grind to a halt because you can't find that one crucial bit of information you need?
That's where Albus steps in.
Imagine a tool thatās a bit like a smart assistant, but instead of reminding you about appointments or dictating text messages, it has the answers to all your HR-related questions.
That's Albus for you. It's your sidekick for those "I know I saw that somewhere" moments.
Albus is like having your own personal librarian who knows everything about your project, the kind of team member who's always there, ready to dish out the answers you need from the pile of information you've accumulated across Google, Slack, Notion, Jira and more.
Just a quick chat with Albus, and you've got the golden nuggets you need.
So, are you ready to banish the frustration of endless searching and focus on what you do best?
Give Albus a spin and get 20% off today.
GOOGLE PUTS BARD ON THE SAME TRACK AS CHATGPT
Bard is running behind in this race.
ChatGPT is in the lead even though itās giving Bing Chat a piggyback ride and Claude managed to find some more energy with its new 100K context window. But Bard is running behind, itās no longer crawling.
Google announced a new set of updates that should get everyone excited. And I mean everyone. No more waitlist across 180+ countries.
But the real meat of the new update is in images and apps.
Remember when OpenAI promised that weād be able to upload images as part of our prompts into GPT-4. Yeah, that hasnāt happened.
Google is making that a reality with the help of Google Lens. They didnāt mention how well it performs, but even pushing it to product before OpenAI would be a win for them.
Bard is also getting āappsā not āpluginsā soon that will help it match the performance and market opportunity of Bing. One of them is Adobe Firefly and thatās meant to help match the image generation thatās possible in Bing Chat.
See, Bard is running.
Before we overpromise and build too much hype, a quick reminder: we have not tested any of these new updates. The jury is out on whether Bard delivers a performance to match ChatGPT.
What we can say is that weāve tested the new Generative Search Engine feature and it holds up. Bing Chat is collecting dust on desktop and weāre back to searching on Google like we never left.
Weāre one lap into a marathon race. Donāt underestimate Google.
Bard is improving and the distribution channels for their AI features is unmatched. Expect these tech titans to exchange more blows over AI that show a real fight is taking place.
Donāt stress, weāthe usersābenefit from all this competition and OpenAI/Microsoft is bound to respond soon.
Letās see what haymaker they toss out.
Quick Nuggets
š«” Yoshua Bengio is the 2nd AI Godfather to admit worries over AIās current direction
āØ The young guns building startups are pushing for a space in the AI conversation
š¤Æ 1.7M installs: Character AI had a great first week in the App Store
š£ļø ChatGPTās non-English skills need a lot of work if weāre going to get AGI
š Damaged goods are precisely what Amazon wants to find with itās new AI system
šø Runway AI gets a massive investment from Google that comes at a price
š¾ Baidu unveils venture fund to help China push AI development closer to American counterparts
š Donāt take your chatbots shopping, the end results can leave you underwhelmed
š„ Fresh Products
Speechki - ChatGPT plugin to turn text into audio (link)
Rep AI - uses two AI tools to convert Shopify sales (link)
FireTexts - AI helps you craft the best text for any occasion (link)
Hoku - unveils new tool to roast your meal photos (link)
Top VC Funds - describe your startup & AI finds the funds (link)
Roll - mimic quality video production on your phone w/ AI (link)
Planit Earth - new way to plan trips (link)
Interviews by AI - prepare for your next job interview (link)
Good Content, Long Time Ago
The opening title card did say a long time ago, so itās fun to picture what that would look like from our context. Wild West Chewbacca is definitely Big Foot.
Thatās it for today. I hope you enjoyed the latest edition of inclined.ai - Davis.
Check Out Our Sunday Edition
You can dive into more AI news and topics with us every week by subscribing to our premium edition.
Weāve written about the following:
If youāre not already subscribed, thatās okay. Weāre offering a free 7-day trial so that you can read this one. Thatās how excited I am to post it.