A.I. Executive Briefing #7

The A.I. Executive Briefing is an expert weekly curation of A.I. news by our research team, shared externally now because we feel there’s too much hype & noise in the market. The same content will be distributed through?this substack.

NEWS ROUND-UP

1. A.I. to help Type 1 diabetics

2. x.ai?+ Tesla collab or conflict of interest?

3. New generative AI releases from Google, Meta & Stability AI

4. The arms race to collect data: OpenAI … and Hollywood?

5. Meta releases Llama2 and nabs several partnerships

6. Apple “Ajax” & iOS 15 AI Health Coach?

7. New AI features in LangChain, Wix, Loom

VENTURE NEWS

8. Hugging Face, open source AI platform raises additional $200m

9. Greenfly, a digital media company, Acquires Miro AI, an AI solution for analyzing sports photos and videos

10. Tractable, Insurance valuation AI, has been revived by softbank with a $65m Series E

11. Neura Robotics, AI robotics start-up, raises ￡48M in down round

12. Runway raises $27m & Cognaize raises $18m, both in Series A rounds to apply AI systems to financial operations.

13. AI startup Nomic raises $17M to build its open-source alternative to GPT-4

News Round-up

1. AI to help Type 1 diabetics

Type 1 diabetics have little to no insulin in their body. Currently, Continuous Glucose Monitors (CGM) are the most efficient way to manage insulin administration. The downside to this system is its reliance on smartphones & the internet to maintain the integrity of the system. The University of Bristol & University Hospital of Southampton?trained their new AI?on data from seven months in the lives of 30 simulated patients, and it learned how much insulin to deliver in a variety of real-life scenarios. It was able to figure out a dosing strategy on par with commercial controllers, yet it needed only?two months of training data to do so—less than a tenth required by previously tested algorithms. Even with inaccurate or missing data fed into the system during scenario testing, the insulin administration maintained performance, all while offline. Some community members are even taking experimentation into their own hands,?publishing models and systems?based on their own self-testing of insulin management. All in order to push innovation forward while the?current research?is under review by the FDA.

The healthcare AI space is seeing fervent enthusiasm in tandem with the overall desire to make sure people are being safe with their data and their bodies. According to a?recent poll?from Common Sense Media and?Impact Research, there is an?overwhelming desire among parents (82%) for a rating system that will help them evaluate the quality of AI programs. Disseminating code/research is equally as important as publishing independent reviews and the context that surrounds the research. Especially when concerned with systems that regulate bodily chemicals such as insulin.

2. x.AI + Tesla collab or conflict of interest?

Musk stated during his twitter space last week that?x.AI?will need to develop technology that “understands the physical world and not just the Internet,” and he thinks that?Tesla’s driving data will help it on that front. He also?announced?a super computer for Tesla titled “Dojo,” a supercomputer Tesla is developing for AI machine learning and computer vision training purposes. Additionally,?x.AI?will collaborate with Tesla both on the “silicon front” and on the “AI software front.” Many view this as a conflict of interest though, as Sen. Elizabeth Warren sent a letter urging the U.S. SEC to investigate Tesla and its board of directors over possible “conflicts of interest, misappropriation of corporate assets, and other negative impacts to Tesla shareholders” related to CEO Elon Musk’s Twitter takeover.

3. Generative AI: Bard, CM3leon & Stability Doodle

Google’s Bard?has received a?suite of multi-modal upgrades. On top of bolstered translation and language capabilities (via a text to speech feature), there is also in-prompt image analysis & UI improvements. The major win here is the?context analysis?that Bard is able to generate from images it is fed by users. This feature is also linked with a?google lens?integration further bolstering opportunity for Bard as a live in-headset assistant that can see what you see and speak to you in a way you can understand.

Meta?also announced a new multi-modal model called?CM3leon, which moves away from diffusion?to rely on transformers. This change has led to boosting speed, requiring 5x less compute, and being easily parallelizable. To train CM3Leon, Meta used a dataset of millions of licensed images from Shutterstock. The?most capable version?of CM3Leon has 7 billion parameters, over twice as many as DALL-E 2. The fact that Shutterstock was used to train CM3leon is especially interesting given their new partnership with Open AI. See below an example of outputs from CM3leon side by side against DALL-E 2 when fed the same prompts. It is clear that Meta’s new image training recipe can output meaningfully different results.

Stability AI?also released a new feature called?Stable Doodle?allowing for image generation context to be built from a simple doodle. Generating original high quality images from low quality sketches is a strong way to sidestep the idea that art will become less original as generative AI grows. Additionally, it is good to see that even though Stability AI is?shrouded in controversy, they are still pushing out new products.

The theme amongst these three releases is a new focus on context building from images or using text to enhance a given context of an image.

4. The arms race to collect data: OpenAI … and Hollywood?

We have a view that proprietary data will be the most valuable asset for businesses building in the A.I. ecosystem, and we’ve predicted from over a year ago that we may see an arms race to collect these proprietary datasets. We are seeing OpenAI take this exact strategy this week:

• The Associated Press?on Thursday said it reached a 2-year deal with OpenAI to share access to select news content + technology, marking one of the first officials news-sharing agreements between a major U.S. news company and an A.I. firm. As part of the deal, OpenAI will license some of the AP’s text archive dating back to 1985 to help train its artificial intelligence algorithms.
• Shutterstock?signed a 6-year training data agreement with OpenAI. As part of the agreement, OpenAI has obtained a license to access additional training data from Shutterstock, including its image, video and music libraries, along with associated metadata. In return, Shutterstock will enjoy priority access to the latest OpenAI technology and will continue directly integrating DALL·E’s groundbreaking generative text-to-image capabilities into the Shutterstock platform.

We’re actually seeing this same strategy being played out in Hollywood currently. This week,?Hollywood actors confirmed that they were going on strike?after the Alliance of Motion Picture and Television Producers (AMPTP) revealed a proposal from Hollywood studios that sounds ripped right out of a Black Mirror episode. As part of the proposal, Hollywood studios proposed that background performers should be able to be scanned, get one day’s pay, and their companies should own that scan, their image, their likeness and should be able to use it for the rest of eternity on any project they want, with no consent and no compensation.

5. Meta releases Llama2 and nabs several partnerships

Meta has?released Llama 2!?This new model is fully open-sourced and has been thoroughly embraced with a variety of partnerships including Microsoft,?Qualcomm,?Together AI?&?Scale AI. Meta’s partnership with Qualcomm is a clear step in the direction of edge computing. The two are working to optimize Llama 2 directly on-device – without relying on the sole use of cloud services. This goes back to the developments we saw in edge computing earlier this month. The theme of running newer models on smaller devices with less compute will not go away and it will become more pronounced.

The partnership with Microsoft seems to be driven by the open-source part of the pitch. Llama 2 is now available in the Azure AI model catalog, enabling developers using Microsoft Azure to build with it and leverage their cloud-native tools for content filtering and safety features.

Scale AI and Together AI seem to be using the popularity of the Llama model to their benefit. Together specifically launched the Together API and Together Compute, cloud services to train, fine-tune and run the open-source AI models. The CEO of Together, Prakash, claims typically “It’s $4 an hour for an A100 GPU on AWS — we have created a technology where we can host instances of a model for a user — for example, hosting a RedPajama 7 billion-parameter model on an A100 on our platform is 12 cents an hour.”

6. Apple “Ajax” LLM & rumors of an Apple AI Health Coach

Where is Apple in all of this AI and LLM frenzy? So Apple actually does have it’s own LLM called Ajax. Apple’s Ajax system is built on top of?Google Jax, the search giant’s machine learning framework. Apple’s system runs on Google Cloud, which the company uses to power cloud services alongside its own infrastructure and?Amazon.com Inc.’s AWS. It has been reported that Ajax doesn’t include any novel features or technology. The system is accessible as a web application for company-only use and Apple has no public GTM to release it to consumers.

That being said….with the Apple watch and all of the healthcare data Apple receives, it seems that?iOS 15 is rumored?to have some AI Health coaching updates, specifically related to bodily metrics acquired via the Apple Watch & language metrics it pulls from messages. Going beyond physical health tracking, there is opportunity for an Apple AI Health Coach to have mental health functionalities as well. Is Apple releasing a ChatGPT-like A.I. health coach in the near future?

7. New AI features in LangChain, Wix, Loom

Amidst the flurry of AI releases this past week, a few more are deserving of callouts.?Wix?has released a?text based website generator?that severely reduces the amount of technical skills needed to generate and build a website. It is important to note that Wix is passing the copyright check for the content generated and used to the end user generating the website.?Loom?released an AI tool?enabling for smarter transcription, task creation and intelligent chapters (it’s about time…)

These features are not earth shattering, but merely a requirement for companies to keep pace with the innovations that are taking place — especially with how models interact with image and video content. Democratizing the ease of building with AI will take distribution of these features and education for the next cohort of users.

Langchain?has also?released?Langsmith, a platform to help developers close the gap between prototype and production when iterating on LLM based application builds. The platform is built specifically to solve for the roadblocks that current LLM app builders face including API token tracking, detailed prompt tracking, understanding user behavior on their app & more. Generally, the debugging, testing & evaluations of LLM applications are made simple by Langsmith enabling more consice and functional applications to be built and tested.

Venture News

8. Hugging Face, open source AI platform raises additional $200m

Open source AI company Hugging Face has?reportedly?raised an additional $200m just 1 year after its $100m Series C. The round has yet to close, but the valuation is coming in between $3-4 Billion. With the support for Llama 2, Nomics & reports of this new funding round, it is fair to say that being open source is definitely a viable model to get funded in this space.

9. Greenfly, a digital media company, Acquires Miro AI, an AI solution for analyzing sports photos and videos

In an undisclosed round?Greenfly has acquired Miro AI. Currently, Greenfly empowers sports and entertainment organizations to provide live access to short-form digital media to their staff and partners. Miro AI has developed technology to analyze photos, videos & other media to then produce generative content off of. Greenfly will also take ownership of RunnerTag, an AI-powered athlete identification service used by sports photography firms, marathons, and other endurance events. The power of Sports and AI cannot be understated. Computer Vision, predictive LLMs & human expertise can help sports team leverage new tech to better optimize teams & help promote players while they are in season. Style and branding around performance will be the key differentiator in a world where all content is seamlessly generated.

10. Tractable, Insurance valuation AI, has been revived by softbank with a $65m Series E

Tractable currently processes some $7 billion in annually insurance claims through partners like Aviva, Geico and Admiral. It will be using its?fresh Series E capital?both to expand into Japan and to incorporate more of the latest advances in AI to expand its services from insurance assessments into repairs, maintenance and sales of the items it scans. 2 years ago, this company also raised a similar amount, but at a $1b valuation ans the valuation of this round is undisclosed. There have been a variety of new entrants’ into this digital valuation market —?Uveye, ProovStation, Ravin, Claims Genius, Innovation Group and many others build technology to assess vehicles and more. The only valid reasoning for Softbank to deploy into this company seems to be related to Tractable’s renewed focus on the Japanese market.

11. Neura Robotics, AI robotics start-up, raises ￡48M in down round

Raising a little over half as much as they raised for the 2021 Series C, Neura Robotics boasts a five year backlog of $450m in revenues. Neura is building intelligent robotic arms that can “interpret” data from its environment. The company is not vertically integrated to the extent that it will not work with third parties: Its AI can be used on any robot via an API it provides. Intelligent robotics companies are an already implemented aspect of manufacturing today, but enabling a “neuraverse” of smart robots connected by IoT is an interesting enough goal to keep the company around.

12. Runway raises $27m & Cognaize raises $18m, in Series A rounds to apply AI systems to financial operations.

In a new attempt to kill excel, Runway is?building a notion style financial dashboard?that applies AI to your financial transactions via api integrations. The company has over 50 customers across various industries, such as Stability AI, Superhuman, Lobe AI, Origin Protocol and Joby Aviation. Disintermediating financial operations by rewrapping the UI employees interact with to glean insights from financial data will help drive efficiency in financial operations. However, this is not the first, nor the last company, to attempt to get the financial industry out of spreadsheets.

Cognaize has built a proprietary model trained on 1.3m+ financial documents. Its use case is to ingest unstructured data like documents, bank statements, tax returns, paystubs, agreements, emails, etc. and to construct the data into a dashboard that makes it easier to understand and execute on. Their platform is constantly infromed and reviewed by the end user in order to esure accuracy & also to improve the model over time. Leveraging specialized use cases + continuous training by experts in that specific use case is a sure fire way to correctly leverage & scale an AI model. A document management system such as this one could see interesting competition for when platforms like Google or DropBox launch an LLM pre-trained on all your cloud stored documents instead of your?notes.

13.?AI startup Nomic raises $17M to build its open-source alternative to GPT4

Nomics currently has two main products, an open-source large language model similar to ChatGPT that’s called?GPT4ALL, which can run on small devices such as a laptop. The company’s second product is?Atlas, which is a tool that makes it possible to visualize unstructured datasets that can be used to build other LLMs. In an opensource show of support, Nomic announced a new partnership with?Hugging Face Inc. to create and distribute rich and interactive data visualizations. Branding as an open source project has a lot of benefits mainly being a decentralized team [read: users] reporting & often building improvements or features. We have seen great success from Meta and the Llama 2 release in this regard. Democratizing AI by allowing non technical people to read the code seems to be the best pitch right now.

Reader Questions

None for this week

Send us a message with any questions/comments/thoughts on anything A.I. related and we’ll try to answer them in our next release.