24 of the top 50 AI start-ups are bootstrapped, Apple and Microsoft double down on hardware, and EY’s launches an LLM

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

Product

1. Google Gemini is Closer than ever

2. IRS Leverages AI to Tax Corporate Partnerships

Reports

3. A16Z Generative AI report

4. BCG and HBS: AI Augments Consulting

Research

5. Three New Models from Stability, Adept and Deci

Hardware

6. Apple Extends Semiconductor Agreement with Qualcomm

7. Microsoft and Oracle Deepen their Cloud Integration

Healthcare

8. Protein Centric AI is Folding Open

VENTURE NEWS

9. EY Unveils Fruits of $1.4 Billion Artificial-Intelligence Investment

10. Databricks locks in $500M at a $43B Valuation

11. Defense AI startup Helsing Raises a $223M Series B

12. Enfabrica, which builds AI networking hardware, raises $125M Series B

13. AI Startup Writer Raises $100 Million to Pen Corporate Content

14. HiddenLayer Raises $50 Million to Protect AI Systems

15. Essential AI Raises $40m to build “LLM Related software”

16. Druid, a Conversational AI Platform for Enterprises, raises $30M Series B

17. Ello Raises $15M in Series A Funding to Help Children Learn how to Read

18. Rephrase AI raises $10.6M Fresh Investment to Grow its Synthetic Media Platform

19. LastMile AI Closes $10M Seed Round to ‘Operationalize’ AI Models

20. Deduce Raises $9M in Series A funding to Detect AI-Generated Identities

NEWS ROUND-UP

Product

1. Google Gemini Beta

Google has been teasing a new model called Gemini all year. In the last week, the company opened up testing of Gemini to select companies, indicating its public launch could be imminent. Gemini is a multi-modal model capable of outputting and ingesting Audio Visual data as well as text-based data.

Why it matters: Google has already released two other major AI products (not even including all the announcements from the Next Conference) Bard and their open source PaLM 2 model. The Gemini model being tested with selected enterprise companies is because this new model may NOT be open source and is likely for profit entirely. Bard was inferior to chatGPT on release, but as the PaLM 2 model came online to support the Bard product, it is now on equal footing with chatGPT. The Gemini model release is an extension of the ecosystem play that Google is taking with its cloud products. With Duet AI, Bard, Gemini & PaLM 2 I think a big question is: How, if at all, will Google tie all these products together?

2. IRS leverages AI to Identify Sophisticated Tax Fraud

Tax returns related to large corporate partnerships are being analyzed for audit & investigation with the use of AI & ML right now! The IRS is leveraging AI & a data science team to identify compliance risk on multiple dimensions including partnership taxes, income accounting & international tax. By the end of the month, the IRS will open examinations of at least 75 of the largest partnerships in the U.S. including hedge funds, real estate investment partnerships, publicly traded partnerships, large law firms, and other industries. On average, these partnerships each have more than $10 billion in assets.

Why it matters: The US government in general has been struggling between implementing AI technology and attempting to pass/discuss regulations regarding the same technology. The IRS was initially supposed to get a big bump in funding and staff, which was later denied by Congress, luckily AI is a good alternative to being denied extra staff. What is largely surprising is actually that they had a team of data scientists managing the implementation of the AI technology in order to select returns that need further investigation. The IRS utilizing experts, not letting the AI systems make any decisions & using it to support a lack of staff increases are all generally correct. I am sure that if this initiative goes well then there will be a steady increase in A.I. selecting returns for audits/investigations.

Reports

3. A16Z Generative AI report

A report focused on how users are interacting with generative AI tools and products has been released by the team at A16Z. They identified the top 50 most popular GenAI web products. Of the list almost half are bootstrapped start-ups that have not yet raised any funding, while 15% of the list have raised over $50M each.

“Excluding ChatGPT (which skews the data given OpenAI’s $11.3 billion raised), companies with a proprietary model have raised an average of $98 million. This compares to $20 million for companies that have fine-tuned an open-source model, and $9 million for ‘wrapper’ companies.”

It is also fairly clear that currently most of the generative AI applications are all web-based and any mobile application has not hit scale. Only 15 companies on the list have a mobile app, and almost all of them see less than 10% of total monthly traffic from their app versus their webapp. This is important to note as 55% of web traffic comes from mobile and over 90% of internet users will access the internet via a mobile phone.

Why it matters: AI on Mobile is the biggest hole in the market right now. This fact goes hand in hand with the small bootstrapped teams currently building in the space. They lack resources to dedicate to a mobile team & with the lack of edge computing technology. Edge computing will enable end-point devices like phones to run sophisticated AI applications without nuking the battery or putting the device under too much stress.

4. BCG and HBS: AI augments consulting

HBS conducted an experiment that involved 758 consultants comprising about 7% of the individual contributor-level consultants at the company. There were three main groups: no AI access, GPT-4 AI access, or GPT-4 AI access with a prompt engineering overview. There were 18 tasks given to each group which generally found that some tasks were made easier by AI and others were not.

• Consultants with AI completed 12% more tasks and completed them 25% faster.
• Consultants with those below the average performance were able to increase performance by 43% and those already above average saw an increase by 17% compared to their own scores.
• Consultants working on tasks outside the capabilities of AI saw a 19% error rate.

Why it matters: This study clearly found that for tasks that are appropriate for supervised AI use, there is a significant boost in productivity & low performers see some of the greatest benefits. This study is actually extremely significant because since its release BCG ended up finalizing a partnership with Anthropic in order to use Claude 2 within its teams to perform research, analyze data, and overall augment general productivity.

Research

5. Three new Models from Stability, Adept and Deci

Stability AI released Stable Audio, a model focused on generating audio based on text prompts. Currently, it is the most flexible and produces the highest quality output when compared to current Opensource audio models, including Metas Audiocraft. Stable Audio is trained on a dataset of over 800,000 audio files and has partnered with AudioSparx for copyrighted material.

Adept, which focuses on AI automating digital tasks, has released a new fully open sourced model. They went so far as to even publish the weights and biases of their model, which is a dimension that most “open source” models fail to do. Some interesting notes about Persimmon-8B are that It was trained on just 37% of data compared to similar models yet exceeds their performance, it has a 16k context window (which is larger than Llama 2 & chatGPT) & it enables custom C++/Python code for faster inferencing.

Deci AI’s new foundation model is about?4.8x faster than Llama 2 and 15x faster when using their own inference SDK. It has 5.7B parameters and is?open to download under the same license as Llama 2. It leverages original technology such as the Neural Architecture Search engine & DeciLM which varies the number of attention groups, keys, and values across transformer layers. This may be the first language model where the transformer layers are not structural duplicates of one another.

Why it matters: There are a lot of innovations coming out of these three models. Stability AI released a myriad of models this past summer and is not slowing down. The Stable Audio model can output custom lengths, is high quality & is protected from copyright infringement (thanks to their partnership) Adept made a huge splash upon its first release earlier this year. Since then they have been relatively quiet. The Persimmon model is truly open source and seems to be far more lightweight than its current competitors. Deci’s model seems to come straight for Meta as they are boasting a huge speed differential that seems to lie in a few key choices Deci made regarding its transformer infrastructure. Learning from existing open-source models that based on research and R&D done over the past 5 years, Deci has implemented more dynamic weighting in their neural net and a new internal search engine for its model.

Hardware

6. Apple Extends Semiconductor Agreement with Qualcomm

Apple Inc. is extending an agreement to get modem semiconductors from Qualcomm Inc. for three more years. This is a little surprising given Apple's massive push towards its new Apple Silicon Chips. However, the company is over 1/5 of Qualcomm's sales, a very significant chunk. The company also has been?working to replace?other semiconductors within the iPhone, including a key?Broadcom Inc.?part. Like Qualcomm, Broadcom counts Apple as its biggest customer. As part of the push, Apple has?staffed up?in Southern California, where Qualcomm and Broadcom both have offices, aiming to recruit chip talent.

Why it Matters: Although there is a trend to pioneer new chip designs, it's really hard to compete effectively when other companies actually have the manufacturing capabilities that are needed for a company the size of Apple. It is more than likely that as the semiconductor manufacturing market continues to evolve out of east Asia, ideally there will be new semiconductor forges established in North America. Big tech and emerging start-ups will have easier access to manufacture new chip designs.

7. Microsoft and Oracle Deepen their Cloud Integration

Microsoft and Oracle have been working together for the past four years, but have recently decided to deepen their partnership. Oracle will physically locate its Exadata (Oracle Exadata is an enterprise database platform that runs Oracle Database workloads of any scale) hardware in Microsoft’s data centers, speeding up applications that customers use. customers will have direct access to Oracle database services running on Oracle Cloud Infrastructure and deployed in Microsoft Azure data centers. The new?interoperability?will make it easier?for Oracle users to run Microsoft Azure’s AI?on top of the Oracle database.

Why it Matters: Microsoft is expanding its tentacles throughout the AI space. They have invested in several different LLMs, partnered with AMD to produce a new chip & have now integrated further with Oracle. This particular development is to compete with other LLM & database partnerships (think Databricks acquiring MosaicML), which makes building on-prem custom enterprise LLMs significantly easier. Currently, Microsoft has fulfilled its goal to secure a foothold in a variety of bleeding edge technologies associated with the current AI boom.

Healthcare

8. Protein Centric AI is Folding Open

Google’s Alphafold protein database was recently mined by researchers in order to uncover new protein shapes and relationships that were previously unknown. They even went so far as to structure a network graph of proteins including known & novel ones.

Additionally, Microsoft has open sourced its novel protein generating AI, EvoDiff. EvoDiff doesn’t require any structural information about the target protein and can be used to create enzymes for new therapeutics and drug delivery methods as well as new enzymes for industrial chemical reactions. Core to the EvoDiff framework is a 640-million parameter model trained on data from all different species and functional classes of proteins EvoDiff can also fill “gaps” in an existing protein design. If provided a part of a protein that binds to another protein, the model can generate a protein amino acid sequence around that bind, which meets a set of criteria.

Why it Matters: Relationships between protein structures vs protein sequences have not been visualized in the way that the researchers displayed the Alphafold Database. The insights here are generated from the unknown similarities between known proteins. Additionally, several brand new protein shapes were also uncovered given their work. Microsoft releasing an open source protein generating model is a huge boon for this sector of biochemical AI. This area was previously roadblocked by a lack of public databases and tools available to the community. Together with the new Alphafold developments and with the EvoDiff release, there will be a slow but heavy trickle of new products and drugs that will be produced by these AI systems and validated by qualified teams.

VENTURE NEWS

9. EY Unveils EY.ai and a $1.4 Billion Artificial-Intelligence Investment

EY is not the first nor will it be the last to announce its own 10 figure AI capital allocation. However, what's fun about EY is that it created its own large language model, EY.ai EYQ, and that it would train its 400,000 employee workforce on AI. As for EYs peer companies, KPMG said in July that it?planned to spend $2 billion in AI and cloud services?globally over the next five years. Also in July,?Accenture?announced a?$3 billion investment to expand its data and AI practice. PricewaterhouseCoopers in April said it planned to?invest $1 billion in generative AI and Deloitte last December said it was spending $1.4 billion on employee training on technologies including AI. That's a total of almost $9B!

10. Databricks locks in $500M at a $43B Valuation

This new round makes Databricks one of the biggest funding events for private tech companies this year amid AI-fueled optimism. The company has been quite active this year, acquiring Mosaic ML, reporting a 50% year-over-year jump in revenue as of July, and amassing 10,000 customers globally. The company has raised $4 billion since inception.

11. Defense AI startup Helsing Raises a $223M Series B

This investment could potentially make Helsing the largest European AI company and also the largest European defense tech unicorn. The round was led by General Catalyst. Swedish heavy industry and defense group Saab is also a strategic investor.

Helsing has expanded in Europe by winning government contracts. In June this year, the German government selected Helsing and Saab to provide AI-enabled electronic warfare capabilities for a jet fighter. There is a lot of development taking place around AI and defense with a large geographic diaspora of companies and developments. This fact makes sense as AI warfare and defense are inextricably tied to a government's ability to defend its people and its land.

12. Enfabrica, which builds AI networking hardware, raises $125M Series B

Enfabrica’s hardware, the Accelerated Compute Fabric Switch (ACF-S), can deliver up to “multi-terabit-per-second” data movement between GPUs, CPUs, and AI accelerator chips. The hardware can scale to tens of thousands of nodes and cut GPU computing for a LLMs by around 50%. This sort of hardware development is going to become increasingly important for LLMs that run across or that retrieve data from multiple chips.

13. AI Startup Writer Raises $100 Million to Pen Corporate Content

Writer Inc. helps businesses write and summarize content has raised $100M at a $500M valuation. Iconiq Growth led the funding round, with participation from investors including Insight Partners, Accenture and Vanguard Group Inc. (both Writer customers). Although the summarization and writing aspect of AI is a commodity, even at this point, if a company can capture and hold on to market share, it will be able to build increasingly specialized tools for its clientele.

14. HiddenLayer Raises $50 Million to Protect AI Systems

HiddenLayer, a cybersecurity startup focused on protecting AI systems from adversarial attacks, has raised $50 million in a funding round co-led by M12 and Moore Strategic Ventures. The company has the ability to detect and respond to attacks on AI models. HiddenLayer uses techniques to observe the vectors or mathematical representations of inputs and outputs, without accessing customers’ proprietary models directly. Protection from LLM attacks is extremely important. Data safety and privacy have consistently been noted as the public's primary fear of AI adoption.

15. Essential AI Raises $40m to build “LLM Related software”

The $40M funding for Essential AI is led by Thrive Capital. Before founding Essential, Parmar and Vaswani founded Adept, a popular machine learning research and product service (they just released their Persimmon model). Additionally, their only stated goal is to build “LLM related software.” The team just emerged from stealth and has no known product or known revenue.

16. Druid, a Conversational AI Platform for Enterprises, raises $30M Series B

Druid build conversation AI tools by integrating with enterprise databases. Druid can be hosted on-premises or in the cloud, which is key for companies seeking full control of all their data. And through APIs, Druid can connect to existing systems such as customer relationship management (CRM) and human resources information systems (HRIS). Druid currently has 150+ employees globally and its Series B round was led by New York-based investment firm TQ Ventures, with participation from an array of European investors, including Smedvig Capital, GapMinder, Hoxton Ventures and Karma Ventures.

17. Ello Raises $15M in Series A Funding to Help Children Learn how to Read

Ello is a subscription-based app aimed K-3 students that delivers five books every month for $24.99. Ello listens to the child read out loud and analyzes their speech to correct mispronunciations and missed words. The AI reading uses phonics-based strategies to teach them critical reading skills based on their reading comprehension & pronunciation. This application of AI technology is really awesome to see. Typically speech therapy is a positive application of AI coaching and the application of voice analysis is likely a field of AI language Ops in we will see a lot of innovation. The round was led by Goodwater Capital, with participation from Homebrew, Reed Hastings, Common Sense Growth, and Ravensburger.

18. Rephrase.ai Raises $10.6M Fresh Investment to Grow its Synthetic Media Platform

Rephrase.ai, a generative AI media company announced $10.6M Series A round led by Red Ventures with participation from Silver Lake and 8VC. The original goal of the platform was to build a text-to-movie platform, something that RunwayML has also stated. Although it is clear that this concept is not novel, the technology is and hopefully we will see a compounding effect of innovation as a result of competition and technology advancements.

19. LastMile AI Closes $10M Seed Round to ‘Operationalize’ AI Models

LastMile AI raised a $10M seed round from Gradient Ventures, AME Cloud Ventures, Vercel’s Guillermo Rauch, 10x Founders and Exceptional Capital also participated in the round. LastMile enables is a text-to-content platform that allows for the user to fine-tune and to select between multiple LLMs.

“Our goal with LastMile is to provide a single developer platform that encompasses the entire lifecycle of AI app development” - Sarmad Qadri, Co-Founder & CEO LastMile

20. Deduce Raises $9M in Series A funding to Detect AI-Generated Identities

Now that we have passed the event horizon of digital AI-generated information, we are going to need technology that is capable of parsing purely generative information and information that is related to humanity. Deduce uses patented technology and an activity-backed identity graph of 840M US-based email identities. The company uses real-time data to recognize patterns and flag AI-generated fraud.

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.