AI News: Language Models, Hardware, Biocomputing and More (AI News, 14 December 2023, 1st Edition)

In this edition: 23 hot news, from new language models that outperform competitors, to AI hardware that can generate immersive experiences, to biocomputing that uses mini-brains to recognize speech, there is a lot to explore and learn in this fast-moving field. Stay tunned for today's AI Tools & Services edition, presenting 14 new ones!

?? Balls of human brain cells hooked up to a computer perform basic voice recognition

• Brainoware: A system that uses cerebral organoids -- mini-brains made up of nerve cells -- for artificial intelligence (AI) tasks;
• Adaptive learning: Brain organoids can recognize the voice of an individual among eight people who pronounce Japanese vowels, with an accuracy of 70 to 80 percent, after unsupervised training sessions;
• Biocomputing Challenges: Brain organoids have high power consumption, silicon limitations, and a short lifespan – about one to two months;
• Future prospects: Researchers hope that biocomputing can overcome the challenges of conventional AI and contribute to a sustainable, clean energy future.

?? https://www.newscientist.com/article/2407768-ai-made-from-living-human-brain-cells-performs-speech-recognition/

?? Role-playing with AI will be a powerful tool [Editor's Note: it already is]

Role-playing with AI is becoming a powerful tool for creating stories and games. AI can generate complex characters and scenarios, allowing users to explore imaginary worlds in a way never before possible. However, there are still challenges to overcome, including the need to improve AI's ability to understand and respond appropriately to context and user actions.

?? https://resobscura.substack.com/p/roleplaying-with-ai-will-be-powerful-tool

?? Safety Models: Promoting the Responsible Use of AI

This article discusses the importance of safety in AI:

• Addresses the risks associated with the irresponsible use of AI models;
• It emphasizes the need for independent and standardized assessments.

?? https://www.together.ai/blog/safety-models

?? Gavin Uberti - Real-Time AI and the Future of AI Hardware

• AI Hardware: describes the hardware building blocks of AI models, how model-specific chips differ from traditional GPUs, and what the challenges and opportunities are for AI hardware development.
• Etched.ai: Features a service that uses real-time AI to create interactive and immersive experiences, such as games, virtual and augmented reality, and simulations.
• Transformer revolution: discusses the need to build dedicated physical infrastructure to support the transformer revolution, which are generative AI models that can generate text, image, sound, and video from data;
• Overview: provides an overview of the field of AI, its key actors, trends and challenges, and how the Etched.ai positions itself in this scenario.

?? https://www.joincolossus.com/episodes/73767562/uberti-the-future-of-ai-hardware

???? The skills needed to build a generative AI team

• Generative AI: uses pre-trained large-scale language models (LLMs) to generate content from natural language instructions;
• Prompt Engineering: is the ability to write clear and effective prompts to customize LLMs for different applications;
• AI Team: should include domain experts, full-stack engineers, and appropriate tools to iterate and evaluate prompts;

?? https://humanloop.com/blog/how-to-build-the-right-team-for-generative-ai

?? The Funding Difficulty for AI Startups

An article that reports on the challenges that some artificial intelligence startups face in raising capital, due to competition from giants such as Meta Platforms and OpenAI. Some highlights are:

• Liquid AI, a startup developing a new type of AI model that can learn on the fly, not just during training, raised just $37.6 million, instead of the $100 million it had intended;
• OpenView Ventures, a venture capital firm that focused on AI startups, laid off most of its employees and stopped investing in new companies;
• Some investors are more cautious about AI startups, due to regulatory, ethical, and security risks, as well as uncertainty about the profitability and scalability of their products.

?? https://www.theinformation.com/articles/some-ai-startups-find-the-moneys-no-longer-so-easy

???? Snapchat+ launches AI-generated image creation and upload functionality.

Snapchat+ is a paid subscription that offers users access to exclusive features such as custom filters, stickers, special effects, and now, AI-generated images. The new feature allows users to create and send realistic images of people, animals, objects, and scenery, using only text or voice as input. The images are produced by an AI model trained on millions of photos, and can be edited and shared within Snapchat+. Some examples of AI-generated images are:

• A dog with butterfly wings;
• A beach with snow and penguins;
• A chocolate-covered pizza;
• A selfie with Cristiano Ronaldo.

The AI-generated images feature is one way Snapchat+ attracts and retains more users, who are looking for creative and fun ways to express themselves and communicate. However, it also raises ethical and legal questions about the use and ownership of the images, as well as the potential risks of abuse and manipulation. ?? https://techcrunch.com/2023/12/12/snapchat-subscribers-can-now-create-and-send-ai-generated-images/

?? Meta launches AI-powered Ray-Ban glasses that can see what you're looking at

• Glasses: Uses cameras and voice to analyze images with generative AI; can recognize objects, read labels, translate texts, make subtitles, and more;
• Launch: Available in early access mode starting today; scheduled for 2024; uses anonymized query data to improve AI services;
• Potential: May be useful for assistive, educational, research, and entertainment purposes; it is a precursor to future AI that mixes various forms of sensory data;
• Limits: Needs a voice command to activate and "see"; pause for a few seconds before responding; may vary in detail and accuracy; it can hallucinate.

?? https://www.cnet.com/tech/computing/metas-ray-ban-glasses-added-ai-that-can-see-what-youre-seeing

?? ChatGPT Vision: transcribing handwritten journal entries (or any other piece of handwriting) into digital text

Testing OpenAI's GPT-4 Vision model to recognize and transcribe handwriting into digital text:

• Accuracy: it can almost perfectly transcribe an example of a handwritten journal entry, surpassing other tools such as Apple iOS' built-in text recognition or Readwise;
• Context: uses knowledge of context to fill in gaps or correct handwriting errors while maintaining the meaning of sentences;
• Limitations: it may make some mistakes, such as swapping numbers or words, if there aren't enough contextual clues to verify them.
• [Editor's Note] If you don't have a ChatGPT Plus subscription, in theory, you still can try this out via Bing Chat or MicroSoft CoPilot. Update: as of today, both are refusing to to this; yet, https://bard.google.com/chat did do it ?? [perhaps not so great as GPT-4V, because it is still powered by PaLM 2, not Gemini Pro]

?? https://twitter.com/fortelabs/status/1734284384537333813

?? AI can help decide the legal fate of those who end up in a British court ("The UK now permits judges to use the “jolly useful” AI chatbot in court")

The UK Judicial Office issued guidance this Tuesday that allows judges to use ChatGPT and other AI tools to write court decisions and perform various other tasks.

• Orientation: Guidance is the first step in a set of future work to support the judiciary in its interactions with AI;
• Risks: The guidance recognises that AI responses may be inaccurate, incomplete, misleading or biased, and suggests that judges check the accuracy of AI responses before making decisions that alter the course of people's lives;
• Privacy: The guidance also warns of privacy concerns, pointing out that AI companies collect the results of user interactions. The guidance said judges should assume that typing something into a chatbot interface is the same as posting it for all the world to see;
• Example: In September, Lord Justice Birss of the Court of Appeal for England and Wales used ChatGPT to summarize legal theories he was unfamiliar with and copied and pasted the results into an official decision. Birss called AI a "very useful" tool.

?? https://gizmodo.com/uk-judges-now-permitted-use-chatgpt-in-legal-rulings-1851093046 Break

?? [still in the realm of justice] Lightspeed Venture Partners: LegalTech vs AI

• Use of artificial intelligence to automate legal processes and reduce costs;
• It offers solutions for contract management, risk analysis, legal research and compliance;
• Explores the opportunities and challenges of the $16 billion LegalTech market;
• It discusses trends and best practices for investing and innovating in this growing industry.

?? https://lsvp.com/legaltech-x-ai-the-lightspeed-view/

?? Better, cheaper, faster alignment of LLM with KTO

A service that uses a technique called Kahneman-Tversky Optimization (KTO) to align large-scale language models (LLM) with user data without compromising performance. Some key points are:

• KTO is inspired by economists Kahneman and Tversky's work on human decision-making;
• Does not require human preference data, but only feedback on whether an output is desirable or undesirable;
• Is equivalent in mathematical terms to the standard method of alignment, but it is much simpler and cheaper;
• Is made available as open source and also as a collection of 56 KTO-aligned models in different sizes and datasets. ?? https://contextual.ai/better-cheaper-faster-llm-alignment-with-kto/

?? notdiamond-0001: A model that determines whether a query should be sent to GPT-3.5 or GPT-4

• A service that uses a text classification model to choose between GPT-3.5 or GPT-4, depending on the task;
• The model was trained on hundreds of thousands of robust, cross-domain assessment data;
• Free under the Apache 2.0 license and can be accessed through a free API;
• It requires query-specific formatting and returns a label to GPT-3.5 or GPT-4;
• It provides documentation on how to integrate the model into your system and how to improve quality, reduce latency, and cost.

?? https://huggingface.co/notdiamond/notdiamond-0001

?? StripedHyena: A New Sequence Model for Language and Other Domains

• Uses a hybrid architecture of attention and convolutions with gates to process long sequences efficiently;
• It is competitive with the best open-source Transformer models in both short and long context tasks;
• It's faster and more memory-efficient than the Transformers for training, fine-tuning, and generation;
• It is optimized using new model grafting techniques, allowing you to change the architecture during training;
• It is based on research on scale laws and mechanistic design of alternative architectures.

?? https://www.together.ai/blog/stripedhyena-7b

?? Microsoft Releases Phi-2, a Small AI Language Model That Outperforms Llama 2 and Mistral-7B

• Phi-2: a text-to-text language model with 2.7 billion parameters, capable of running on a laptop or mobile device;
• Performance: comparable to or superior to other larger models, such as Llama 2-7B, Mistral-7B, and Gemini Nano 2, with less toxicity and bias in responses;
• Limitation: Licensed for research purposes only, not for commercial use;
• Competition: An indirect note to Google, showing that Phi-2 can also solve physics problems and correct student errors, like Gemini Ultra.

?? https://venturebeat.com/ai/microsoft-releases-phi-2-a-small-language-model-ai-that-outperforms-llama-2-mistral-7b/

?? Phi-2: The Surprising Power of Small Language Models [Phi-2's official MicroSoft blog post]

• Phi-2: a 2.7-billion-parameter language model that demonstrates exceptional reasoning and language comprehension capabilities;
• Quality data: a mix of synthetic and web data, selected based on educational value and content quality;
• Knowledge transfer: an innovative technique to scale the 1.5-billion-parameter Phi-1.3 model and incorporate its knowledge into Phi-2;
• Performance: cutting-edge results in various academic and internal benchmarks, comparable to or superior to models up to 25 times larger;
• Availability: The Phi-2 model is available in the Azure AI Studio model catalog to foster research and development in language models.

?? https://www.microsoft.com/en-us/research/blog/phi-2-the-surprising-power-of-small-language-models/

???? Answer.AI: a new kind of AI research and development lab

• Uses existing AI models to create products and services that are useful in practice;
• Founded by Jeremy Howard (former co-founder of Kaggle and fast.ai) and Eric Ries (creator of Lean Startup and Long-Term Stock Exchange);
• Received a $10 million investment from Decibel VC;
• It's a fully remote team of deep-tech generalists;
• It does not focus on creating Artificial General Intelligence (AGI), but on harnessing the potential of Applied Artificial Intelligence (HAI).

?? https://www.answer.ai/posts/2023-12-12-launch.html

?? Mamba-3B-SlimPJ: A Mamba architecture model with 3 billion parameters trained on the SlimPajama dataset

• Mamba: an architecture based on state space models (SSMs) that scales linearly with sequence length and has fast inference;
• SlimPajama: a clean, deduplicated version of the RedPajama dataset, with 600 billion tokens and the GPT-NeoX tokenizer;
• Mamba-3B-SlimPJ: a Mamba model with 3 billion parameters trained on SlimPajama, which matches the quality of some of the best 3 billion Transformers, such as the BTLM-3B-8K, with 17% fewer FLOPs;
• Evaluation: The model was evaluated on 10 natural language understanding tasks, using zero-shot or 5 shots, and obtained results comparable to or superior to Transformers.

?? https://www.together.ai/blog/mamba-3b-slimpj

?? GPT-3.5-turbo beats MistralAI mixtral-8x7b in reasoning and math tasks

A short test on Vercel's AI Playground showed that GPT-3.5-turbo is clearly superior to the MistralAI mixtral-8x7b in solving reasoning and math problems. Some of the differences observed were:

• Speed: GPT-3.5-turbo took about 2 seconds to generate each response, while MistralAI mixtral-8x7b took about 10 seconds;
• Precision: GPT-3.5-turbo got all the questions right, while the MistralAI mixtral-8x7b got two of them wrong;
• Clarity: GPT-3.5-turbo provided simple and straightforward answers, while MistralAI mixtral-8x7b used long and complex sentences;
• Creativity: GPT-3.5-turbo showed some ability to generate original examples and explanations, while MistralAI mixtral-8x7b repeated the statement of the questions.

?? https://twitter.com/mayowaoshin/status/1734557229779472564

?? PhotoMaker: Personalizing Realistic Human Photos through ID Stacking

PhotoMaker is an efficient text-to-custom image generation method that encodes an arbitrary number of input ID images into an ID stackup to preserve ID information.

• Efficiency: PhotoMaker takes about 10 seconds to generate each response;
• Precision: PhotoMaker can get all the questions right, while other methods can get two of them wrong;
• Clarity: PhotoMaker provides simple and straightforward answers, while other methods use long and complex sentences;
• Creativity: PhotoMaker shows some ability to generate original examples and explanations.

?? https://photo-maker.github.io/

?? Investing.com Uses AI to Copy Wholesale Competitors

An article by Semafor Media reveals that the financial news site Investing.com is creating stories with the help of AI, which often appear to be thinly disguised copies of human-written stories on other sites.

• Motivation: Investing.com, a Tel Aviv-based site, wants to leverage AI to produce content quickly and undermine the competitive advantage of other sites that pay humans to write articles;
• Method: Investing.com it appears to be using AI to rewrite content unique to competitors, without citing or crediting the original sources, but only disclosing that the stories were written with the help of AI;
• Impact: The competitors of Investing.com, such as FXStreet and The Motley Fool, feel threatened and undermined by the Investing.com, which they consider a form of plagiarism and a threat to journalism and the creation of original content;
• Challenge: Investing.com raises legal and ethical questions about the use of AI to copy stories, and how legal and ethical guidelines will adapt to AI tools.

?? https://www.semafor.com/article/12/10/2023/a-financial-news-site-uses-ai-to-copy-competitors-wholesale

???? OpenAI: A Billion-Dollar AI Company With a Derisory Revenue in 2022

• OpenAI has a hybrid non-profit and for-profit structure, which has raised questions about its mission and management;
• It faced a leadership crisis in November, when CEO Sam Altman was fired and then reinstated by the board of directors after protests from employees and investors;
• It has a valuation of $86 billion by private investors, but it reported only $44,485 in revenue in 2022, coming mainly from investment income.

?? https://www.cnbc.com/2023/12/12/openai-nonprofit-arm-45000-in-2022-revenue-company-worth-billions.html

What was the news that surprised you the most? Please share your thoughts in the comments section below! ?????

See all my articles / AI News, here: https://www.dhirubhai.net/in/filipebento/recent-activity/articles/