A.I. Executive Briefing #4

A.I. Executive Briefing #4

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. Google’s release of new AI products & its distribution advantage

2. Dark Chemical Discovery: New transformer model can predict novel compounds

3. Reconstructing 3D indoor spaces with NeRF

4. Stanford analysis on A.I. model compliance with draft E.U. A.I. Act

Venture News

5. Captions, an AI video Suite for Creators, Raises $25m Series B

6. Databricks acquires MosaicML, which allows companies to build LLMs on proprietary data

7. Finance automation company Ramp Acqui-hires Cohere.io Team

8. Celestial AI, Inter memory chip data transmission start-up, raises $100M in series B

9. Thomson Reuters Acquires Casetext, a legal language focused LLM, for $650M in cash

10. Meituan “buys” Light Years Beyond, the ‘OpenAI for China,’ for $234M

11. Stability AI, generative AI start-up Secures Convertible Note & Runway raises an additional $141m

12. Inflection AI, which launched the personable chatbot Pi, Raises $1.3 Billion


News Round-up

1. Google’s release of new AI products & its distribution advantage

At the end of last year when the GPT and LLM craze started to enter the mainstream, it felt as though many of the new entrants into the space were pushing out new products and partnerships at a rapid pace. However, incumbents like 谷歌 have been pioneering at dripping out advancements in the A.I. space for a while now. Google published the transformer model research paper in 2017, which is the foundation of the LLMs dominating the headlines today. Google started a stream of product announcements earlier this year and we are finally seeing the releases associated with them.

  • The company is?adding?OCR (Optical Character Recognition) technology to the Chrome browser that can convert PDFs to text that makes them more accessible, particularly if you want a screen reader to read them aloud. The tool will also provide image descriptions.
  • They have also launched the private beta of “help me organize” in Google sheets to help automate table creation via plain language input. This is one example of the Duet A.I. tools Google has promised to incorporate natively across Google workspace.
  • YouTube is bringing in the team from Aloud, which was part of Google’s Area 120 incubator to enable AI-powered language dubbing. The tool first transcribes your video, giving you a transcription that you can review and edit. Then, it translates and produces the dub.

I mentioned this on a podcast I was on earlier this year, but I still believe Google’s distribution power is completely unrivaled and underestimated in this space. As we see Google start to get their act together here, I think their distribution advantage will outcompete any tech advantages from competitors — i.e. end users are not going to care if a competitor’s feature is 30% better if Google can deliver the functionality through their wide range of products.


2. Dark Chemical Discovery: New transformer model can predict novel compounds

Earlier this year, this was seen as a the major blocker to the application of AI to the biochemical space. Enveda has released their own version of a transformer model trained on over 1 million mass spectrometry (MS) spectra of molecular structures. Currently, their dataset seems to be proprietary, which may be for the best at the moment given the contentiousness of developing novel chemicals and molecules that interact with our biology. Advancements specifically related to predicting novel molecular structures help to overcome what is seen as one of the primary barriers in the space. Additionally, Insilico Medicine has a suite of LLM and A.I. products that work in tandem to accelerate the entire drug discovery, testing and approval process. Their suite is powered by 英伟达 s BioNeMo cloud service for generative AI in drug discovery. Recently, they designed a drug to treat a rare respiratory disease, which is zipping into Phase 2 trials just 2.5 years after project inception. This trend of faster drug deployment we saw kick up with the mRNA COVID-19 shots will hopefully be replaced by more efficient and safe testing of novel drugs.


3. Reconstructing 3D indoor spaces with NeRF

Another A.I. research paper has been released from Google centered around using a series of images of an indoor space to generate a full 360 perspective of the space. Without full perspective of a space, their NeRF model seems to be able to generate new perspectives. This release seems to be in indirect competition with NVIDIAs Neuralangelo research that reconstructs surfaces generally, but similarly, from still frame images from several perspectives, but not all perspectives. The application of Googles indoor prediction technology has wide implications including reconstruction of spaces never before seen outside of select photos or even taking imagined still images into the 3-Dimensions by generating 360 perspectives of the rooms the spaces. Other applications could even include reconstructing currently inaccessible crime scenes based on images or (potentially even) descriptions from eye witnesses.

No alt text provided for this image


4. Stanford analysis on A.I. model compliance with draft E.U. A.I. Act

In the past 2 weeks, we have seen some more development on the regulatory front regarding AI systems. Most recently the Hugging Face CEO has told the US lawmakers that open-source A.I.s “are critical to incentivize and are extremely aligned with the American values and interests.” As much as this may sound like something any AI Open source company CEO is supposed to say, he’s not far off the truth. The EU AI act that we have been tracking in the Executive Briefings is the first and currently the most stringent laws regulating AI today. 美国斯坦福大学 did an analysis of the law and ranked the current AI models today in their compliance and guess who was on top… Hugging Face! Turns out that these models are such a black box to so many regulators, lobbies & law makers (shocker) that the most compliant companies are actually the most transparent (I.e. open source in execution and not in name only).

No alt text provided for this image

Funding News

5. Captions, an AI video Suite for Creators, Raises $25m Series B

Captions has a suite of seven unique AI tools making the video editing process more efficient for creators. A more similar target market to Youtubes creator suite, Captions does also currently have a translation/dubbing tool. The company has over 3M users and is backed by A16Z, Sequoia Capital & KKR .

6. Databricks acquires MosaicML, which allows companies to build LLMs on proprietary data

MosaicML was acquired by Databricks for $1.3B & the Databricks Mosaic Research team will continue to build out their AI. Investors like Lux Capital could have potentially returned $175m from ~$20M in total funding to the company. MosaicML provides software that lets startups use their own data to train and deploy LLMs and lets developers retain control over their models. This tool combined with the fact that DataBricks, which hosts datasets from companies makes their offering much more fearsome. Databricks has been valued at almost $40B in 2021 and this acquisition is more than likely necessary to maintain that valuation for their long awaited IPO.

7. Finance automation company Ramp Acqui-hires Cohere.io Team

Finance automation company? Ramp ?has acquired?Cohere.io,?a startup that built an AI-powered customer support tool for an undisclosed amount. Cohere was founded in 2020 and raised $3.1M in seed funding. Interestingly one of the founders was an employee of Ramp before leaving to start Cohere.

8. Celestial AI, Inter memory chip data transmission start-up, raises $100M in series B

First order, LLMs require loads of data to train on. Second order, LLMs require high end and custom data chips to train and run models. Third order, enter Celestial AI who just raised $100m in their series B. As the compute power needed to run and store data associated with LLMs, the transfer of bits between these chips is the next level of efficiency that can be unlocked. Using light to transfer data, Celestial’s tech can beam information within & between chips, which alleviates bandwidth put on compute and memory.

9. Thomson Reuters Acquires Casetext, a legal language focused LLM, for $650M in cash

Casetext, Part of Thomson Reuters originally focused on providing access to lawyer annotated legal documents. Using the documentation that they have collected, over the years, they slowly built up an AI legal offering called Counsel AI. CoCounsel does document review, legal research memos, deposition preparation, and contract analysis. 汤森路透 acquired this as a part of their ultimate strategy to integrate A.I. offerings into their legal, tax and accounting verticals. They have gone so far to say that they are allocating up to $10B for M&A in A.I. for the next two years. Corporate VCs are very eager to incorporate A.I. at all costs

10. Meituan “buys” Light Years Beyond, the ‘OpenAI for China,’ for $234M

The founder of 美团 , web based shopping platform, founded Guangnian Zhi Wai or “Light Years Beyond” less than 6 months ago. Meituan announced that it acquired this new start-up for $234M. To take this deeper, the investors of Guangnian Zhi Wai include Sequioa China (now split off into HongShan) & Qimai, which is also controlled by the CEO of Meituan. It is also important to note that the start-up had around over $250m of cash and no product telling us that the company has probably been hampered by the high-end chip shortage China is facing & also that this was more a restructuring than an acquisition.

11. Stability AI, generative AI start-up Secures Convertible Note & Runway raises an additional $141m

Stability AI has been under controversy since Tayab Waseem, a claimed founder of the company, filed a complaint against Stability AI and its co-founder and CEO Mohammad Emad Mostaque on May 18. Since then, the company has been attempting to raise at a $4B valuation as they seek to compete with start-ups like Midjourney , who just launched their most realistic version (5.2) & Runwayml who just raised $141m series C extension with participation from Nvidia & Google. Luckily for Stability AI, Ashton Kutcher has lead their sub $25m convertible note round.

12. Inflection AI, which launched the personable chatbot Pi, Raises $1.3 Billion

Inflection AI founded in 2022 by Reid Hoffman , Karen Simonyan &?Mustafa Suleyman pioneered Pi. Pi is a chatbot focused on being personable, conversational & “human” rather than solely transactional and responsive. Inflection AI raised $1.3b at a $4b valuation lead by Reid Hoffman, Eric Shmidt, Bill Gates & NVIDIA. This makes it one of the largest funding rounds for an A.I. company to date. The company itself has a hardware edge given its partnership and investors Bill Gates and NVIDIA and even tease that their publically available model is one of the smallest one they are planning on launching.

No alt text provided for this image

Reader Questions

Q: “Could you explain the difference between supervised, unsupervised, pre-trained, and fine-tuned models in a simple way?”

A: We’ll do a post on this next week. Might even do a short webinar with members of our data science team if that’s of interest


Q: “What are OpenAI’s Profit Participation Units (PPUs) that are mentioned in their comp packages? Are these stocks?”

A: This article on OpenAIs comp structure was referenced with this question. So from what i gathered here, it looks like this is how it works:

  • PPUs are not even equity - its an agreement to get a % of profits and you have no idea when profits will be generated (they would still need to repay all of their VC financing first)
  • Your PPUs are capped at 10X. So if OpenAI values your PPUs at $2M, your payout is capped at 10x (so $20M)
  • Investors can purchase PPUs from employees during certain liquidation events or financing rounds
  • PPUs have no value unless a "profit is generated" --> again very vague on numbers and timeline

This seems like a shit deal for employees IMO, but the language here is very vague and anyone from OpenAI is free to correct/clarify here


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.

Enjoy the relaxing read! ?? What's on your reading list? ??

回复

要查看或添加评论,请登录

Andrew Hong的更多文章

社区洞察

其他会员也浏览了