AI Weekly: 04/24/23

Good morning and welcome to this week’s edition of AI Weekly! In this week’s news, Adobe has announced tools to revolutionize video editing, Stability AI released their new open-source large language model, and Humane publicly demoed their first-of-its-kind wearable AI assistant.

We have also started to see businesses take advantage of their large first-party data sets, as two massive Q&A sites in Reddit and StackOverflow have announced that they will start charging companies that are building models for access to their APIs.

In notable funding news, Langchain just raised another round led by Sequoia valuing the company at over $200 million. Enjoy reading about the rest of this past week’s AI news below!

- ZG

Here are the most important stories of the week:

TEXT

Stability AI released an open-source language model, StableLM, with an Alpha version available in 3 billion and 7 billion parameters, and larger models to follow. Link.

• The StableLM models can generate text and code and are trained on a new experimental dataset three times larger than The Pile, with surprisingly high performance in conversational and coding tasks despite their small size.

• StableLM builds on Stability AI's experience in open-sourcing earlier language models with EleutherAI, including GPT-J, GPT-NeoX, and Pythia, which were trained on The Pile open-source dataset.

• The release includes a set of research models that are instruction fine-tuned for conversational agents, initially using five recent open-source datasets, which are intended for research use only.

• The models are designed to be transparent, accessible, and supportive, promoting trust, enabling adaptation without sharing sensitive data, and allowing for efficient, specialized, and practical AI performance.

• The StableLM models are available in the GitHub repository, with a technical report to follow, and the company is launching a crowd-sourced RLHF program and collaborating with Open Assistant to create an open-source dataset for AI assistants.

Elon Musk is working on "TruthGPT," an AI rival to ChatGPT that aims to be a maximum truth-seeking AI. Link.

• Musk wants to avoid AI destruction of humanity by creating an AI that cares about understanding the universe.

• Musk compared an AI's supposed lack of desire to destroy humanity to humans' protection of chimpanzees.

• Musk sees TruthGPT as a course correction to OpenAI, which he helped found but has since begun operating a for-profit subsidiary.

• Musk signed an open letter in March that urges companies to pause "giant AI experiments" that their creators can't "understand, predict, or reliably control."

• It is not clear how far along TruthGPT is, but Musk established a new AI company, called X.AI, in March.

Together, a Menlo Park-based company, announced RedPajama, a collaborative project between six organizations, will create leading fully open-source LLMs. Link.

• The project began with the release of a 1.2 trillion token dataset that follows the LLaMA recipe, which is a foundation LLM model.

• Together plans to release a full suite of LLMs and instruction-tuned versions based on the RedPajama dataset in the coming weeks that will be commercially viable and fully open source.

• RedPajama is part of a wave of open source AI, and the developers of RedPajama believe that it is essential that systems are transparent.

• There have been debates about the ethics of the datasets themselves, but Together's CEO, Vipul Ved Prakash, believes that these models capture the output of human society and should be open and usable by everyone.

• The RedPajama project brings together a larger community to these models, and broader access could allow a lot of brilliant people around the world to explore different directions of neural architectures, training algorithms, and safety research.

Snapchat is rolling out its AI chatbot My AI globally to all users, having previously been exclusive to Snapchat+ subscribers. Link.

• My AI runs on OpenAI's GPT technology and is designed to respond like a person messaging a friend, offering recommendations for places to visit or lenses to use.

• Users can name their AI and customise a Bitmoji avatar for it, creating a virtual companion.

• In group conversations, My AI will have a sparkle next to its name to distinguish it from real humans.

• Paying Snapchat+ subscribers will eventually receive AI-generated image responses to Snaps sent to My AI, in addition to text responses.

• The global rollout of My AI was announced at the Snap Partner Summit, which introduced several new features for the tool.

IMAGE/VIDEO

Adobe has developed new video tools with Firefly that enable users to enhance, manipulate and edit videos with a few typed commands without requiring a TV studio, millions of dollars of equipment or a team of editors. Link.

• Adobe Firefly, a tool developed by Adobe in March, uses AI to simplify the process of image editing, and has been trained using images from Adobe's own stock image site, public domain content and openly licensed work.

• Adobe Firefly will also be available to those working with videos with the promise that results will be "safe for commercial use."

• With Adobe Firefly, users can make specific adjustments or improvements to a clip such as brightening a subject's face, describe how they want a piece of on-screen text to look, find appropriate b-roll footage, generate music and sound effects that are cleared for broadcast and auto-generate storyboards and previsualizations.

• Adobe Firefly's capabilities are currently in beta and are not yet available to everyone.

• Adobe Firefly's video tools will be available later this year.

CODE/DEVTOOLS

LangChain, a startup that helps developers build complex applications on top of LLMs, has raised between $20 million and $25 million in funding from Sequoia at a valuation of at least $200 million. Link.

• The deal was pre-emptive and led by growth investor Sonya Huang, who has been a prominent evangelist for the generative AI movement.

• LangChain was originally launched as an open-source project in October 2021 and aims to help developers build more complex AI applications by combining the power of LLMs with other sources of computation and knowledge.

• The startup achieves this by providing users with the ability to split up complex tasks into easy-to-digest components called "chains" that can then be connected together.

• LangChain's dedication to openness and collaboration is a major differentiator for the startup.

• LangChain still has something to prove as it supports few large-scale production applications, and other players have emerged to layer on enterprise-level needs like scalability and security onto LangChain.

HARDWARE

Microsoft has reportedly been developing its own AI chips in secret since 2019. Link.

• The chips are designed to train large language models (LLMs) like GPT-4 and avoid reliance on Nvidia, the current key supplier of AI server chips.

• Nvidia's latest H100 GPUs are selling for over $40,000 on eBay, highlighting the demand for high-end chips.

• Microsoft's in-house efforts may significantly cut costs as it pushes to roll out AI-powered features in Bing, Office apps, GitHub, and elsewhere.

• Microsoft is reportedly planning to make its AI chips available more broadly inside Microsoft and OpenAI as early as next year.

• Other tech giants like Amazon, Google, and Meta also have their own in-house chips for AI, but many companies still rely on Nvidia chips to power the latest LLMs.

MULTIMODAL

Humane, a startup founded by ex-Apple employees, has given a live demo of a new wearable device with a projected display and AI-powered features intended to act as a personal assistant. Link.

• The device is built from the ground up for artificial intelligence and is completely standalone, meaning that it doesn't require a smartphone or any other device to pair with it.

• During the presentation, the device responded to voice commands like an Amazon Echo speaker and also supports gesture commands.

• The device is able to project a screen onto nearby surfaces and includes a camera that can identify objects in the world around it.

• There's also a translation feature demonstrated where the device reads out a sentence in French after the user says it.

• The compact, screenless device raises questions about its usability in public or during hurried situations.

HEALTHCARE

Moderna and IBM are partnering to incorporate AI and quantum computing into mRNA development, used to create Moderna's COVID-19 vaccine. Link.

• Moderna and IBM will invest in generative AI to improve product design and advance and accelerate mRNA research and science.

• IBM will allow Moderna to access its systems for quantum computing, an emerging technology that utilizes the principles of quantum mechanics to solve problems too complex for classical computers.

• Moderna and IBM scientists will use the AI model MoLFormer to help optimize mRNA and lipid nanoparticles, which surround and protect mRNA within the body.

• The partnership is meant to prepare Moderna for the era of quantum computing and for breakthrough advances with quantum computing.

• Moderna's CEO Stéphane Bancel said the partnership with IBM is aiming for game-changing technologies.

POLICY/LAW/ETHICS

Chinese researchers conducted an AI experiment giving it control of a satellite camera for 24 hours, marking the first time AI has been given unbridled control of an observational satellite without being fed any human interaction through prompts or tasks. Link.

• The AI only had control of the satellite’s camera, not its systems, and was tasked with seeing how effective AI could be at using China’s various remote-sensing satellites.

• During the experiment, the AI chose unusual targets to observe, including the Japanese port of Osaka and the Indian city of Patna, which was home to a border dispute between China and India in 2020.

• The experiment is considered "rulebreaking," but the researchers say they were ready to face whatever consequences it brought.

• While the experiment can provide useful information, it also adds to fears growing around AI, as the targeting of military-connected targets raises questions.

• It's unclear what type of AI was used in the experiment, and the AI was not designed to explain itself.

OTHER

Google has merged its two main artificial intelligence research units, Brain and DeepMind, to form a new unit called Google DeepMind. Link.

• The new unit will be led by Demis Hassabis, the co-founder and CEO of DeepMind.

• Jeff Dean, the previous head of Google Research overseeing the Brain team, will move to a new role as chief scientist.

• Google aims to accelerate its AI efforts while trying to be more efficient, as it announced its biggest job cuts ever in January.

• Building and running large AI algorithms requires significant computational resources, which is why most of the work is concentrated in big tech companies like Google and Microsoft.

• Google plans to build more conversational AI features into its core search engine, similar to what Microsoft did with Bing.

Reddit plans to charge companies for access to its API, which allows outside entities to download and process the social network's conversations for the development of AI systems. Link.

• LLMs, such as OpenAI's ChatGPT and Google's Bard, rely on Reddit conversations and images as training data to improve their algorithms.

• The move is one of the first significant examples of a social network charging for access to its conversations for AI development purposes.

• Reddit believes its data is particularly valuable because it is continuously updated and has promised to improve software tools for moderators and incorporate more machine learning into the site's operations.

• The new fees could go into the tens or hundreds of thousands of dollars, which could be a problem for companies that rely on outside data to improve their models.

• The move is part of Reddit's preparations for a possible initial public offering on Wall Street this year.

Stack Overflow, a popular internet forum for computer programming help, also plans to begin charging large AI developers for access to its 50 million questions and answers on its service. Link.

• The News/Media Alliance, a US trade group of publishers, including Condé Nast, which owns WIRED, has released principles calling on generative AI developers to negotiate any use of their data for training and other purposes and respect their right to fair compensation.

• AI developers use LLMs to generate strings of text based on word patterns learned from online sources such as Stack Overflow and Reddit.

• Data sets used in AI development are often built through unofficial means such as scraping content from websites.

• Stack Overflow plans to license its data to companies developing LLMs for big, commercial purposes, but will continue to license data for free to some people and companies.

Samsung is considering replacing Google with Microsoft's Bing as the default search engine on its devices, causing panic at Google, as an estimated $3 billion in annual revenue is at stake with the Samsung contract. Link.

• Bing’s recent addition of AI is making it a serious threat to Google's search business.

• Google is racing to build an all-new search engine powered by A.I. technology, called Magi, and is upgrading its existing search engine with A.I. features.

• The new search engine will offer a more personalized experience to users, attempting to anticipate their needs.

• Plans for the new search engine are still in the early stages, but the Magi project will add features to the existing search engine in the meantime.

• The search additions will keep ads in the mix of search results, which is important for Google since search ads are its primary source of revenue.

Microsoft has integrated its generative AI-powered 365 Copilot technology into its Viva employee engagement and experience platform. Link.

• Copilot in Viva will focus on streamlining workforce communications and creating better organizational alignment.

• Copilot in Viva Goals can draft objectives and key results (OKR) recommendations, summarize the status of OKRs, identify blockers, and suggest next steps.

• Viva Engage with Copilot can suggest post ideas for company intranet pages, and Copilot in Viva Glint can help summarize employee survey comments.

• Microsoft 365 Copilot combines GPT-4 with Microsoft Graph data and Microsoft 365 apps including Teams, Word, Outlook, and Excel.

• About 20 million people actively use Viva on a monthly basis.

OpenAI's ChatGPT requires expensive servers to calculate responses based on user prompts, potentially costing up to $700,000 a day for infrastructure. Link.

• Operational expenses for running the AI model exceed training costs when deployed at scale.

• Companies using OpenAI's language models pay steep prices, with one startup paying $200,000 a month for the AI to answer user queries.

• To reduce costs, some companies have switched to other language software providers.

• Microsoft is developing its own AI chip called Athena to reduce the cost of running generative AI models.

• The chip could be released for internal use by Microsoft and OpenAI as early as next year.

Insider has laid off 10 percent of its staff, just a week after urging writers to incorporate AI tools like ChatGPT into their workflow. Link.

• The layoffs speak to a new pattern of media layoffs and AI announcements going hand-in-hand, perhaps as a way of softening bad news to investors.

• Axel Springer, the site's owner, has laid off employees and noted that AI has the potential to make independent journalism better than it ever was or simply replace it.

• While AI may not be mature enough to start replacing human writing jobs, investors have made it clear that they love AI.

• BuzzFeed CEO Jonah Peretti announced that BuzzFeed News is shutting down after more than a decade of work but made sure to mention AI in his layoff memo to staffers.

• There seems to be a pattern of publications shifting towards AI in the midst of industry turmoil and layoffs to keep funding bodies happy.

OpenAI's ChatGPT has shown an ability to decipher the meaning behind Federal Reserve statements and exhibited predictive stock market capabilities in response to headlines, according to separate studies. Link.

• Richmond Fed researchers found that ChatGPT can reason similarly to humans when decoding Fed speak and outperformed BERT, another language model used in finance, without prior training.

• University of Florida researchers prompted ChatGPT to make investing recommendations based on financial headlines from after October 2021, showing a strong correlation between ChatGPT's responses and stock market moves.

• ChatGPT was able to provide a reasonable justification for its choices when sorting Fed policy statements into five categories: dovish, mostly dovish, neutral, mostly hawkish, and hawkish.

• GPT-4 was more likely to agree with humans than GPT-3 in categorizing Fed policy statements.

Google plans to incorporate the same AI tech powering the Bard chatbot into its advertising tools, according to an internal Google presentation for ad buyers cited in the Financial Times report. Link.

• The AI system will put together text, images, and video fed in by the company buying ads into something resembling a real ad campaign made by marketing professionals.

• There are concerns about ads spreading misinformation this way, as Bard famously got basic information wrong in its own announcement and some Google employees have had ethical concerns about the chatbot since before it launched.

• Google apparently will have defenses in place to make sure these AI-generated ad campaigns don't do anything weird.

• None of this has gone public yet, so it's unclear exactly how effective the AI-generated ad campaigns will be.

Sam Altman predicts the end of ever-larger AI models due to cost constraints and diminishing returns on performance gains. Link.

• One major driver of this shift is the exorbitant and unsustainable expense of training and running large language models (LLMs), which require powerful graphics processes and hundreds of millions of dollars’ worth of computing.

• Nvidia dominates the GPU market with 88% market share, and their latest H100 GPUs designed for AI and high-performance computing can cost over $30,000 per unit.

• Progress in AI will come from improving model architectures, enhancing data efficiency, and advancing algorithmic techniques beyond copy-paste scaling, rather than from giant models.

• Despite this, access to GPUs remains critical for AI-focused companies, and it can take months even for hyperscaler cloud providers to secure them.

• Altman suggests that model size is a “false measurement of model quality” and that focus should shift away from parameter count.

Salesforce is developing a generative AI-driven workflow system that uses Data Cloud, Einstein GPT and Flow tools to allow users to enter plain text requests that will be carried out by the software. Link.

• The system is designed to harmonise disparate business data and create complex, multi-system workflows.

• Salesforce is working with early design customers on new tools, including Einstein GPT for Flow, which creates workflows by interpreting plain text requests, and Data Cloud for Flow, which uses data-driven event triggers to initiate actions.

• The system uses a blend of available AI models from Google, OpenAI, Meta and Salesforce's own Einstein intelligence layer.

• The tooling is expected to go into early beta release in June and is planned for full release next year.

• The aim of publicising the development before beta release is to help customers understand the rapidly evolving tech world and prepare for the next generation of tools.

GPT Tip of the Week

Prompt: What are the key questions to ask when evaluating whether a new AI whitepaper is on a true technological innovation?

Response: When evaluating whether a new AI whitepaper is on a true technological innovation, here are some key questions to ask:

1. What is the underlying technology behind the innovation, and how does it differ from existing technologies?

2. How has the innovation been tested, and what evidence is there to support its claims?

3. Are there any potential ethical or legal implications associated with the innovation, and how have these been addressed?

4. Has the innovation been independently reviewed and validated by other experts in the field?

5. How does the innovation compare to existing state-of-the-art solutions, and what are its potential advantages and limitations?

6. What is the potential impact of the innovation on the broader field of AI, and how does it advance the state of the art?

7. What are the potential applications of the innovation, and what are the key challenges or barriers to adoption?

8. Are there any limitations or drawbacks to the innovation, and how can these be addressed?

9. What is the track record of the authors or research team behind the whitepaper, and have they produced other notable contributions to the field of AI?

By asking these key questions, you can gain a better understanding of whether a new AI whitepaper is truly innovative and has the potential to make a significant impact in the field of AI.