Pie & AI: Pune

my prediction for 2024 (yes, i have only one) 💡 Local ML is going to be huge. It will be in part driven by the adoption of Apple Silicon and other innovative hardware, but also on raw CPU and mobile devices In many cases except for the largest of LLMs, local inference will become a viable alternative to hosted inference.

Today, we are adding Stable Video Diffusion, our foundation model for generative video to the Stability AI Developer Platform API. The model can generate 2 seconds of video, comprising of 25 generated frames and 24 frames of FILM interpolation, within an average time of 41 seconds. Developers interested in utilizing Stable Video Diffusion through an API can access it now on the Stability AI Developer Platform. Learn more here: bit.ly/3Rymnw9

Ok, this is bonkers, AI can now create unique music from texts🤯 This is Suno AI. Anyone can generate their own unique music with title, lyrics, and style in seconds just from texts! 10 wild examples (and how to use it):

🧑‍🎨 Original cartoon by Maritsa Patrinos.

The following LLM trends are undeniable. - small models are becoming more and more powerful - LLMs can work with non-Nvidia chips (AMD, Google, Amazon) - open-source explosion of new models and fine-tune This means working on LLMs becomes more accessible and mass-market, and our fundamental understanding of these models will vastly improve. This also translates to dissipating hype and fear (doomers). Hopefully, technological progress happens so fast that the regulators won't catch up, and regulatory capture by a handful of players will be prevented. TLDR: regulatory capture may soon be a non-issue as OSS AI models explode.

Following up on the Gemini API launch, here's a notebook to help you compare Gemini and ChatGPT API responses with a diff view! colab.research.google.com/github/googlec…

Some predictions for 2024 – keeping only the more controversial ones. You certainly saw the non-controversial ones (multimodality, etc) already 1. At least 10 new unicorn companies building SOTA open foundation models in 2024 Stars are so aligned: - a smart, small and dedicated team can reach close to OpenAI level in a few months as Mistral, Deci.AI or 01.AI showed - non-AI startups are struggling to raise money while VCs are eager to join the AI revolution - knowledge around training large models keeps spreading from teams to teams each time a new SOTA model is created leading many new venture to form - consequence: we now understand much better the push for regulatory capture that we witnessed from early AI-labs in 2023 (e.g. the idea of licences to train models, etc) 2. 2024 will be reality-check for older AI-unicorn pioneers Optimistic bet: we'll see a few older AI unicorn startups successfully transitioning to sustainable business models by leveraging widespread public adoption of AI 3. Model quality will be harder and harder to evaluate in 2024 With the surge of models and the saturation of open-benchmarks, users will tend to fall back on "brand quality perception". It will become essential for new teams to not be perceived as cheating on public evals and leaderboards 4. The return of academia Academia is back as we saw at NeurIPS 2023. With many private and open-source labs closing the doors on publishing their results and data, academia rise again in visibility and is shining with many impactful papers in 2023 and exciting new work coming 5. Dangerous times for annotation companies It's much easier to quickly spin and iterate on a pay-by-usage API than to hire and manage annotators. With model performance strongly improving and the privacy guarantee of open models, it will be harder and harder to justify making complex annotations contracts. 6. The rise of synthetic data We're running out of human data, there are many copyright questions on these and our largest models are already reaching human level annotation on many tasks. The next step coming is about large and good quality synthetic data.

Really enjoyed NeurIPS! After attending great sessions around LLMs, I documented a huge list of interesting LLM papers that were either presented or mentioned. Here is a list of some of my favorite papers in no particular order. I have included papers that won awards and are pushing ideas that we will continue hearing more about: -- Chain of Code: Reasoning with a Language Model-Augmented Code Emulator - arxiv.org/abs/2312.04474 -- Towards Revealing the Mystery behind Chain of Thought: A Theoretical Perspective - arxiv.org/abs/2305.15408 -- Scaling Data-Constrained Language Models - arxiv.org/abs/2305.16264 -- Language to Rewards for Robotic Skill Synthesis - arxiv.org/abs/2306.08647 -- Tree of Thoughts: Deliberate Problem Solving with Large Language Models - arxiv.org/abs/2305.10601 -- Why think step by step? Reasoning emerges from the locality of experience - arxiv.org/abs/2304.03843 -- Toolformer: Language Models Can Teach Themselves to Use Tools - arxiv.org/abs/2302.04761 -- Reasoning with Language Model is Planning with World Model - arxiv.org/abs/2305.14992 -- ToolkenGPT: Augmenting Frozen Language Models with Massive Tools via Tool Embeddings - arxiv.org/abs/2305.11554 -- DecodingTrust: A Comprehensive Assessment of Trustworthiness in GPT Models - arxiv.org/abs/2306.11698 -- QLoRA: Efficient Finetuning of Quantized LLMs - arxiv.org/abs/2305.14314 -- Direct Preference Optimization: Your Language Model is Secretly a Reward Model - arxiv.org/abs/2305.18290 -- Are Emergent Abilities of Large Language Models a Mirage? - arxiv.org/abs/2304.15004 -- Reverse Engineering Self-Supervised Learning - arxiv.org/abs/2305.15614 -- Learning Transformer Programs - arxiv.org/abs/2306.01128 -- OpenAssistant Conversations -- Democratizing Large Language Model Alignment - arxiv.org/abs/2304.07327 -- Privacy Auditing with One (1) Training Run - arxiv.org/abs/2305.08846 -- Language Models, Agent Models, and World Models: The LAW for Machine Reasoning and Planning - arxiv.org/abs/2312.05230 -- Large Language Models as Zero-Shot Conversational Recommenders - arxiv.org/abs/2308.10053 -- Zephyr: Direct Distillation of LM Alignment - arxiv.org/abs/2310.16944 -- I am also putting together a year in review later in the week or early next week so you will see other cool and important LLM papers that were published throughout the year. Stay tuned! Feel free to comment with your favorite papers as well.

@icreatelife Something worth pondering over 😅

What would you do if AI were never invented?

Phi-2 by @MicrosoftAI is now the #1 trending model on @huggingface (hf.co/models). 2024 will be the year of smoll AI models!

A lot of development in the #Indian #AI space, for Indian languages and use cases recently. • @SarvamAI OpenHathi-7B-Hi-v0.1-Base: A 7B modelparameter, based on Llama2, trained in Hindi, English, and Hinglish - sarvam.ai/blog/announcin… • KissanAI’s Dhenu 1.0: World’s first Agriculture #LLM, Dhenu 1.0, trained on extensive high-quality datasets, especially focused on Indian agriculture practices. The model is designed to be bilingual, trained on 300k instruction sets in English and Hindi, to support English, Hindi, and Hinglish queries from farmers  - youtube.com/watch?v=vEBR1e… • @Krutrim AI: Bhavish Aggarwal, CEO of Ola, unveiled #KrutrimAI, India’s pioneering full-stack AI solution. Key features include its meaningful proficiency in 20 Indian languages, and its ability to generate content in 10, including Marathi, Hindi, and Kannada, emphasizing cultural context and inclusivity. Krutrim, trained on over 2 trillion tokens, comes in two flavors— Krutrim and Krutrim Pro - youtube.com/watch?v=5BhN0Q…

Less than 144 hours since, DomoAI dropped video-to-video. Anyone can turn video into different styles. The most creative and fun ones I've seen are usually anime or 8-bit game styles. 8 favorite examples (And how to get started):

One thing that stood out to me at #NeurIPS2023 was there were a lot more application papers than before. This was not the case with my last neurips (2018). I'd like to encourage this wave by highlighting one such paper. LayoutGPT: Compositional Visual Planning and Generation with Large Language Models arxiv.org/abs/2305.15393 1. What it does With a text prompt, for example, "A living room with a sofa, a coffee table, ...", the app generates an image as well as the layout of the objects - their absolute position in pixel coordinates. e.g. {height: 81px; width: 93px; top: 119px; left: 15px;} 2. Why it's useful You might wonder how this is different from stable diffusion. Stable Diffusion: Text -> Image. LayoutGPT: Text -> Layout -> Image. Adding the layout step and before using generative models gives you control over where things are placed. The author suggests many practical applications such as text-guided dense layout generation, inpainting, scene synthesis/completion, etc. In my experience, marking human pose keypoints is quite difficult, yet this model does a pretty good job at it. They even showed how their method can be extended to 3d spaces. 3. How it's done Doing something new? This means you need a new dataset. What they did was take bounding boxes from MS COCO and use them to create the object-coordinate dataset. When the user inputs the prompt, it searches similar images and labels on the fly and adds them in-context.

@The_DailyAi Humans and AI coexisting together!

Share Your Best Ai Generated Image of the Day Here . Day 192/200.

The Top ML Papers of the Week (Dec 11 - Dec 17): - Gaussian-SLAM - LLMs in Medicine - Mathematical LLMs - Beyond Human Data for LLMs - Weak-to-strong Generalization - Towards Fully Transparent Open-Source LLM ...

@paulg Something worth pondering over!

This is an important point. A lot of the most successful people, not just in startups but in many fields, do what they do because there's something they want to exist, or some mystery they want solved, and no one else is working on it.

Magnific AI has revolutionized AI art forever. It takes AI generated images to a whole other level and users are discovering mind-blowing use cases. 10 crazy examples (Plus HUGE giveaway):

@minchoi Its exciting how @tldraw and @vercel are leveraging AI to create web apps quickly from scratch and making it so simple even for someone with limited coding experience.

⚡️Vercel v0 is now open to everyone. And It's incredible! You can generate UI with texts or images in matter of seconds. And people are already finding amazing ways to use it. 6 examples:

Huggingface has over 437K+ AI models hosted on it, including model demos hosted by OpenAI, Google, META, etc. If you're looking to explore the world of AI, here are 7 great apps that you can try out for free: 👇

@langchain Interesting article! Adding it to the reading list!

🦜🧮Building An LLM-Powered Analyst This great article deep dives on how to empower LLMs with external tools using OpenAI functions At a **19 minute** minute, it's pretty comprehensive It examines two use cases: extraction to get structured output and routing to use external information for questions s/o Mariya Mansurova Blog: towardsdatascience.com/can-llms-repla… Code: github.com/miptgirl/miptg…