AI Weekly: 01/23/23

Here are the most important stories of the week:

TEXT

ChatGPT has been banned in the following places due to falsehoods, socially biased info, and other undesirable outputs:

• ICMC in Honolulu

• NYC public schools

• WeChat (as a plugin)

• Stack Overflow

IMAGE

• Meta & Shutterstock partnering; Meta using Shutterstock’s collection to develop and train their own models. Link.

  • Shutterstock previously partnered with OpenAI and LG AI Research

  • Why this matters: tech companies will want to partner with large sources of data

    • Opportunities for strategic data

• Mark Hamilton and his colleagues developed STEGO, an architecture and training method for semantic segmentation that separates objects in a picture and labels them in an unsupervised way. Link.

• New AI image generator by Baidu called ERNIE-ViLG (essentially Chinese version of DALL-E 2). Link.

  • Chinese regulations: restrictions of deep fakes

• Three artists filed a class action lawsuit against Stability AI, DeviantArt, and MidJourney over Stable Diffusion. Link.

  • Main legal issues revolve around web scraping

  • “Stable Diffusion contains unauthorized copies of millions – and possibly billions – of copyrighted images. These copies were made without the knowledge or consent of the artists," their lawyer, Matthew Butterick, alleged.

• Getty Images is also suing Stability AI over alleged copyright violation. Link.

  • Getty believes that Stability AI “unlawfully copied and processed millions of images protected by copyright” to train its software

  • Getty Images has “commenced legal proceedings in the High Court of Justice in London” against Stability AI

SPEECH

• Joshua Browder’s DoNotPay autonomously navigates phone menus and converses with customer service representatives in a deepfaked version of one’s voice. Link.

  • How it works:

    • System interacts with corporate voice portals using an instance of GPT-3.5 that was fine-tuned on customer service prompts

    • Resemble.AI’s Clone service generates a synthetic version of someone’s voice

    • When with a human rep, it generates conversational responses and feeds them into Clone using GPT-J, an open source language model from HuggingFace

• Microsoft’s VALL-E can closely match the quality of any person’s voice if given a three-second audio sample. Link.

  • Builds on an audio compression technology known as Encodec to analyze and breakdown a person’s voice into its fundamental components, which it then uses to create other phrases with a similar sound and tone.

• DoNotPay has an AI defending someone in court to fight a traffic ticket. Link.

• ElevenLabs is deploying its own generative model which lets users design entirely new synthetic voices and apply different tones. Link.

  • Use cases:

    • Books —> transfers to audiobooks

    • News —> written pieces into audio pieces

    • Video games —> voicing NPCs

    • Advertising —> experimenting with different delivery styles and voices

OTHER

• DeepMind makes reinforcement learning (RL) model called DreamerV3 that learns a world model from experience. Link.

  • RL agents are tiny compared to language models

    • Trained in parameter sizes ranging from 8m to 200m

  • Demonstrated a reliable scaling law: increased model size leading to monotonic improvements in final performance and data efficiency

  • Efficient to train

  • Main problem with RL still exists: reality doesn’t have a great simulator

  • Verticals that will benefit from RL:

    • Video game testing

    • Predictive consumer behavior (retail)

    • Finance (trading algorithms)

• Absci Corporation has used generative AI to de novo design antibodies against three distinct targets in a zero-shot fashion. Link.

  • This is a big deal. A Rensselaer Polytechnic Institute paper in 2015 said, “the holy grail of antibody design is to accurately and reliably predict the sequences of antibodies that will bind with high affinity and specificity based solely on the sequence or composition of the antigen.”

• DeepMind’s AlphaFold 2 allows researchers to quickly and accurately predict protein structure from existing mRNA sequences. Link.

• La Trobe University researches have developed an AI algorithm that can instantly determine whether a person is above the legal limit by analyzing a 12-second clip of their voice. Link.

  • It’s called Audio-based Deep Learning Algorithm to Identify Alcohol Inebriation (ADLAIA)

• Cornell engineers have simplified and re-enforced models that accurately calculate fine matter contained in urban air pollution. Link.

  • UNRELATED/RELATED IDEA: B2B AI software for lead and asbestos testing

• KoBold Metals investing $150 million to develop a copper mine in Zambia. Link.

  • Problem: the metals needed to meet skyrocketing demand for EVs and renewable power plants are in short supply

  • KoBold uses ML to identify overlooked locations for finding copper, cobalt, nickel, lithium, and other rare earth elements.

  • Backed by Sam Altman, Jeff Bezos, Richard Branson, Bill Gates, a16z + more

DEALS

• Monogram Health, a value-based specialty provider of in-home evidence-based care and benefit management services for patients living with polychronic conditions, including chronic kidney and end-stage renal disease, today announced it has closed on $375 million of new funding. This round included investments from leading strategic investors CVS Health, Cigna Ventures, Humana, Memorial Hermann Health System, and SCAN. Link.

• CloseFactor has raised a $15.2m Series A round led by Vertex Ventures with participation from Sequoia. CloseFactor uses ML for automated sales research and contextual insights, reducing research time for account executives. Link.

• Butter Payments, developer of a payments management platform designed to mitigate accidental payment failure, announced that it raised $21.5 million in Series A funding from Norwest Venture Partners, with participation from Atomic. Link.

• ElevenLabs has raised a $2m pre-seed round led by Credo, a Czech VC. ElevenLabs wants to transfer text-to-speech and audio-to-audio in any language, any voice, with the full range of emotions. It could be used for everything from creating audiobooks to dubbing movies. Link.

• myInterview, the global leader in smart video interviewing software, today announced that the company has secured an $11m Series A funding round. This funding will allow myInterview to expand its product offering to meet the growing demand from leading brands for recruiting solutions that help better attract and engage talent on scale. Link.

• Los Angeles-based PLAI Labs, a new startup founded by former MySpace and Jam City founders Chris DeWolfe and Aber Whitcomb, said on Thursday that it has raised $32M in a seed funding round. The round was led by a16z (fromerly Andreessen Horowitz). The company said it is using a combination of artificial intelligence (AI) and crypto to create a new company focused on entertainment, social engagement, and "digital ownership". Link.

• Plai Labs, a company that builds social platforms for Web3, announced that it raised $32M in a seed funding round led by the investment firm a16z. Plai Labs, headed by former executives from gaming studio Jam City and social platform Myspace, has plans to use the money to grow its team and hire developers to create decentralized social platforms that combine Web3 and AI technology. Link.

• California-based AI creative tools platform, Vizcom, announces the closing of their 5 million USD seed round of fundraising. The purpose of the fundraising, according to the company, includes expansion of their product development projects, hiring more engineers for the company, and marketing and promotion of their products and services, among others. Link.

• SNOW has seen its AI avatar creation service gain popularity, with more than 200k users signing up in just 15 days. The service allows users to create up to 200 images in more than 12 different styles by registering 10-20 selfie images. The company’s AI camera app, SNOW, is available in South Korea. Link.

• Gigged.AI has secured its latest investment, a £1.6m seed round led by Par Equity alongside existing investor Techstart Ventures, and Nile HQ. Gigged.AI helps businesses source and onboard on-demand tech talent, while enabling SMEs and larger corporates to retain existing staff. Link.

KEY QUESTION:

Can/do large language models (LLMs) like GPT-3 endanger Google’s search engine business?

• GPT-3 has 175 million parameters —> requires ~350 GB to store (many reports say ~800 GB)

• Knowledge available on the internet —> requires an estimated 5 billion GB

Search points us to many answers LLMs with fixed memories can’t

• Potential technology: retrieval augmented generation

  • Finds relevant documents online and then uses an LLM to process the query and documents into an answer

    • Could provide a current alternative into current web search

• Examples include:

  • Meta’s Atlas

  • DeepMind’s RETRO

• Search engine giants well positioned to execute on this tech, but their business models rely on users clicking on ads placed next to their search results

  • If answers to search were just text outputs, where would ads fit into the picture?

  • Would Bing be more likely to embrace a business model upheaval than Google?

• Google still has many moats (i.e., control over Chrome & Android operating systems)

GPT Tip of the Week:

Prompt: What determines defensibility for a generative AI platform that builds on top of an existing foundation model?

Answer: A generative AI platform that builds on top of an existing foundation model can be considered defensible if it has the following characteristics:

• Unique data: The platform has access to unique data sets or has the ability to generate unique data sets that are not available to other platforms.

• Proprietary algorithms: The platform uses proprietary algorithms or has made significant modifications to the foundation model that cannot be replicated easily by others.

• Intellectual property: The platform has patents, trademarks, or other forms of legal protection for its technology.

• Network effects: The platform has a large user base or ecosystem that would be difficult for others to replicate.

• Branding: The platform has a strong brand that is associated with the AI technology it provides.

It's important to note that defensibility can change over time, as technology advances and new competitors enter the market, a company should continuously evaluate its defensibility and adapt its strategy accordingly.