Orca, a 13B Model, is Raising the Bar...
Redesign your Room w/ Gen-AI, Google's Catching Up, Optimising Algorithms, Digital IPs & More!
Greetings everyone,
š Deiniol here.
"SuperAGI" is now available and accepting requests for a user-friendly interface that enables fast and reliable development and deployment of autonomous agents. Meanwhile, Apple has released an improved autocorrect feature that is personalised to your preferences. So, if you often use the word "ducking," your keyboard will learn to suggest it correctly...
SuperAGI from nowhere and the choice of swear word from a trillion-dollar company. Happy days.
Here is your weekly dose of AI broken down into 3 highlights, 2 tools and 1 demo/tutorial.
Please feel free to share with your closest. Hope you enjoy š¦
"The most important thing artificial intelligence can do for us is to help us understand ourselves better."
- Yoshua Bengio | 2018 Turing Award Winner & AI Pioneer
1 - Google is catching up.
Bard updates: More accurate responses, export to Google Sheets
š¶ āThen I saw its (inter)face, now Iām a Bing-liever. I couldnāt Bard it if I triedā¦ā š¶
Well, that was positively horrible. However, I do use Bing Chat every day. And, with GPT4 slowly seemingly succumbing to a detrimental bout of RLHF, Bard is back on my radar.
Bard (Googleās answer to ChatGPT & Bing Chat), is on a mission to conquer the realms of logic and reasoning, proving that even machines can excel in math and coding. Google has introduced a fascinating technique called "Implicit Code Execution" (inspired by System 1 and System 2 thinking) empowering Bard to not only write code but also execute it. This allows Bard to generate more accurate responses to computation-based problems. According to Google, Bard has witnessed a 30% enhancement to computation-based word and math problems in their internal challenge datasets compared to its previous version. Impressive, right?
So, Bard will get better at answering prompts like:
What are the prime factors of 15683615?
Calculate the growth rate of my savings
Reverse the word āLollipopā for me
Still, Bard may occasionally stumble and generate incorrect code or omit executed code from its responses. Google acknowledges these imperfections and is working to fine-tune Bard's performance. After all, Rome wasnāt built in a dayā¦
Holy sheet! I almost forgot - you can now directly export Bard outputs to Google Sheets!
2 - This aināt a gameā¦
Google DeepMind repurposes game-playing AIs to optimise algorithms.
Improvements in the efficiency of some fundamental algorithms have proved challenging for both human scientists and computational approaches. But not for AI!
Google DeepMind recently launched AlphaDev. A specialised version of AlphaZero, used to improve sorting algorithms by making a better hashing function for small byte ranges (9-16), reducing the load by 30% (in laymanās terms - this takes less pressure off microchips, computation and energy use).
š Important Links
Also, DeepMind's original Alpha series of AI models, originally trained to play games, have been repurposed for other tasks with remarkable accuracy.
AlphaZero, for example, was exposed to data from Google's Borg system (which manages task assignment at data centres) and was able to identify patterns in data centre usage and incoming tasks, producing new ways to predict and manage that load. Reducing the amount of underused hardware by up to 19%!
Similarly, MuZero was able to reduce the bitrate of YouTube videos by 4% and improve sorting algorithms compared to the standard ones in the library Google was using. While not world- breaking, these results suggest that AI models originally trained for one task can generalise their approach to other, unrelated fields.
3 - The (open & small) LLM that learns from GPT-4, is guided by ChatGPT and then outperforms ChatGPT
Orca 13B LLM (No, not the whale.)
Orca LLM, developed by Microsoft Research, is a 13B model that challenges the notion that only large-scale models like GPT-3.5 and GPT-4 can deliver high-quality AI performance.
Orca focuses on learning the reasoning process of larger models. It learns from the outputs of GPT-4, which include detailed explanations, step-by-step thought processes and complex instructions. It receives guidance from ChatGPT to facilitate its learning.
Orca uses a diverse range of imitation data, carefully selecting and sampling from large-scale sources, to promote progressive learning. In benchmark tests that evaluate complex reasoning abilities, such as Big-Bench Hard (BBH), Orca outperforms other advanced models like Vicuna-13B by over 100% and achieves a 42% improvement on AGIEval.
Furthermore, Orca performs on par with ChatGPT in the BBH benchmark and demonstrates competitive performance in professional and academic exams like the SAT, LSAT, GRE, and GMAT in zero-shot settings, without the need for additional guidance. However, Orca still falls behind GPT-4 in terms of overall performance.
The research indicates that learning from detailed explanations and step-by-step processes, whether generated by humans or advanced AI models, is a promising approach to enhance the capabilities and skills of the models.
TLDR;
Orca significantly outperforms other smaller, open-source models. In some settings, it can match or even surpass the quality of ChatGPT, although a substantial gap with GPT-4 still remains.
Open-source is making progress.
š Tool 1
Room GPT | Redesign your room
Take a picture of a room and instantly redesign it in 9 different themes. In seconds.
This product is bananas.
I now want new furniture for my entire house.
Addiction has taken hold and I have already bought more credits to redesign practically everything I see. Until youāre also addicted, give it a shot for free.
š Tool 2
Rephrase.ai | Text-to-video powered by generative AI and youRephrase.ai is a text-to-video generation platform that allows you to create professional-looking videos with a digital avatar in minutes. While there is competition, Rephrase is really quite good.
š¢ The tool offers:
multiple stock avatars,
easy text-to-voice conversion,
multi-lingual speech,
pre-uploaded templates,
the option to monetise your avatar by building a digital IP š¤Æ
š§āš« Demo/Tutorial
Google Cloud | Seven Free Generative AI Courses
Google Cloud has released a new set of generative AI training courses that are available at no cost.
Available courses
Introduction to Generative AI (1 day)
Attention Mechanism (1 day)
Introduction to Image Generation (1 day)
Create Image Captioning Models (1 day)
Encoder-Decoder Architecture (1 day)
Have a day up your sleeve and want to learn more?
If you found value in this weekās newsletter, please support us by sharing with your closest and liking our publication at the bottom. It helps a lot š¤
Until next week, earthlings, bots and cyborgs, code out!
Deiniol š¦šŗ
Something exclusively for you guys:
Use the code AISUITE for 10% off all plans.
šĀ We got codesnippets.ai 28 new subscribers in just a few hours:
Get your business, podcast or AI Art in front of thousands of email subscribers, Twitter followers and AI enthusiasts, startup founders, aspiring entrepreneurs and artificial intelligence students who visit our website frequently.