A.I. Executive Briefing #6
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. Elon Musk launches x.ai
2. Dukaan, Indian Enterprise e-commerce co, replaces 90% of support staff with chatbot
3. Anthropic releases Claude 2
4. New NFAP Policy Brief: AI and Immigrants
5. Scientist have mapped a fruit fly’s brain on a path to Strong AI
6. GPT details leaked
7. All of the internet now belongs to Google’s AI
8. China, US & the predictive power of AI
9. AI Babies, Brain Cancer, PTSD support & more
Venture News
10. Digital Ocean Acquires Paperspace for $111m Cash
11. Causally & Recursion, both drug discovery platform, raises $60m in Series B and $50m from NVIDIA respectively
12. Collective, a finance platform for freelancers, raises $50M in series B
13. Prolific, an AI testing company, raises ￡25M in series A
14. Raft.ai, an intelligent logistics start-up, raises $30m in series B
15. Pano AI, a wildfire detection start-up, raises $20m in series A
News Round-up
1. Elon Musk launches x.ai
In a Twitter Spaces event Wednesday evening, Musk explained his plan for building a “safer AI” with the launch of his new startup, x.ai. Musk has repeatedly voiced concerns about OpenAI ’s models and AI's potential for?"civilizational destruction." Rather than explicitly programming morality into its AI, x.AI will seek to create a "maximally curious" AI, he said. "If it tried to understand the true nature of the universe, that's actually the best thing that I can come up with from an AI safety standpoint," Musk said. "I think it is going to be pro-humanity from the standpoint that humanity is just much more interesting than not-humanity.”
2. Dukaan, Indian Enterprise e-commerce co, replaces 90% of support staff with chatbot
Suumit Shah , the CEO of Dukaan? , an Indian e-commerce company, posted on Twitter earlier this week announcing that the company had laid of 90% of their support team and replaced them with an AI chatbot. In the thread he shares some stats such as dropping resolution time from 2 hours to 3 minutes and dropping support costs by 85% (which happens when you lay of 90% of support staff..). Shah received a lot of backlash following this public announcement, with general negative sentiment seeming to be around the insensitivity of publicizing the laying off of nearly his entire support headcount.
3. Anthropic releases Claude 2
Anthropic , an LLM chatbot company rivaling OpenAI, has released its chatGPT4 competitor, Claude 2. The model is trained on data up until earlier this year, putting it 2 years more recent than GPT4 & supports a context window of 100k tokens, which breaks down to around 75k words or PDFs over 100 pages. Additionally, the API pricing hasn’t changed (~$0.0465 to generate 1,000 words) for Claude 2.0, where GPT4 can be significantly more expensive (ranging from $.03-$.12). With the process that these models are trained on becoming more public (see GPT4 leaks below), the commoditization of highly capable models will continue. The main constraint will continue to be acquiring enough compute to train and run the models.
4. New NFAP Policy Brief: AI and Immigrants
The National Foundation for American Policy published a new report on AI & Immigrants. The paper is a pretty interesting study done around the top 50 AI companies (43 are US based) covered by Forbes specifically focusing on Immigrants and secondarily 1st Generation Americans. We took the paper and fed it to Anthropic’s new model Claude 2 to summarize the stats and here are the top 5 most interesting highlights.
5. Scientist have mapped a fruit fly’s brain on a path to “Strong” AI
An interdisciplinary team of neuroscientists and computer scientists from schools including Cambridge and Princeton recently made a breakthrough in brain-mapping using A.I. by mapping a fruit fly’s brain — the first whole-brain connectome (a map of all neural pathways) created for any adult animal. For context, a fruit fly brain has about 150,000 neurons (compared to 86 billion for a human), and its entire brain is smaller than a single large neuron in the human brain. In a recent blog post from Bill Gates, he notes that this is a step toward superintelligent AIs, or “strong” AIs:
”Compared to a computer, our brains operate at a snail’s pace: An electrical signal in the brain moves at 1/100,000th the speed of the signal in a silicon chip! Once developers can generalize a learning algorithm and run it at the speed of a computer—an accomplishment that could be a decade away or a century away—we’ll have an incredibly powerful AGI. It will be able to do everything that a human brain can, but without any practical limits on the size of its memory or the speed at which it operates.”
6. GPT4 details leaked
Semianalysis broke the information regarding GPT4s training information (for $50), but it was later leaked online and is now available here (for free and in technical detail). GPT-4 is more than 10x the size of GPT-3 with a total of ~1.8T parameters across 120 layers. It was theorized earlier that GPT4 could potentially be a “Mixture of Experts” with 8 smaller GPT models that were trained on specific topics that can interact via switch transformer technology. However, we now know that it actually has about 16 experts making up the whole model. OpenAI's GPT-4 training was expensive from a hardware and a notional perspective. It was done on ~25,000 A100s (Estimated at $10k/ea) for 90 to 100 days. If their cost in the cloud was about $1 per A100 hour, the training costs for this run alone would be about $63 million. Today, the pre-training could be done with ~8,192 H100 in ~55 days for $21.5 million at $2 per H100 hour.
As we see OpenAI continue to develop their GPT models, we are getting a clearer sense of the kind of incremental change needed to approach AGI and GPT5. With Plug-ins, Functions, & now Code Interpreter; the path of the autonomous agent continuously clears.
领英推荐
7. All of the internet, including you, now belongs to Google’s AI
In light of all the lawsuits going on regarding AI and copy right infringement, 谷歌 ’s latest update to its privacy policy will make it so that the company has free range to scrape the web for any content that can benefit building and improving its AI tools. On the back of this change, Google has also announced the release of the NotebookLM, an LLM “grounded” in your own notes and on sources you choose. The use of NotebookLM is three fold: summarizing your notes, engaging in Q+A & Idea generation/content creation. Both their privacy change and this new release signify two major ideas: Everything you have posted online will be fed into an AI system and Personal AIs will be lead by note taking apps specialized to the topics you keep digital records of.
8. China, US & the Predictive Power of AI
The United States Government has identified and begun testing on five different LLM models, only one of which is confirmed by the company itself (Scale AI). Testing includes feeding the model with 60,000 pages of open-source data, including US and Chinese military documents. Based on this data questions like “Could the US deter a Taiwan conflict?”, and “Who would win if war broke out?” result in a series of bullet points with explanations within seconds. This predictive use of AI is outside of the current realm that we see touted on social media (simplifying tasks & customer service automation).
Speaking of China, there have been a few AI developments that are taking the A.I. world by storm. Huawei Cloud has released PanguWeather, used by the China Meteorological Administration, to successfully track the path of Typhoon Mawar off the east of Taiwan in May. The model was found to be ‘10,000 times faster’ than the most powerful forecasting tool currently in use. It returns estimates of temperature, wind speed , air pressure and other weather-related data, but no perception information. Luckily, researchers from Tsinghua University, have relased NowcastNet designed to predict precipitation levels for the upcoming six hours. Additionally, Alibaba Cloud presented at the World Artificial Intelligence Conference in Shanghai an image generator named Tongyi Wanxiang that will initially be available to enterprise customers in beta form.
Governments seem to be using A.I. for its prediction features— weather its for tracking storms or military planning, utilizing AI as a scenario analyzer seems to be the preferred use case for national governments.
9. AI Babies, Brain Cancer, PTSD support & more
Patients are already using chatGPT to self diagnose & Google is piloting PaLM2 with Mayo Clinic, testing its use in hospitals since the start of Q2. It seems that IBM Watson had to crawl so Google and 微软 could run. However, the doctors found Med-PaLM 2 included more inaccurate or irrelevant content in its responses than those of their peers, suggesting the program shares similar issues with other chatbots that have a tendency to confidently generate off-topic or false statements. As concluded before almost all AI systems related to healthcare should exclude patient data (given security concerns & HIPAA) and must be used in conjunction with a doctor or medical professional.
New advances in AI’s use case in mental health predictions, a new AI model analyzes data from Active Military personnel and Veterans pre- and post-deployment health assessments, including physical symptoms, mental health factors, and alcohol use to predict and detect PTSD. When tested on over 200,000 Army service members, the model correctly identified 82% of those who developed PTSD after deployment. Early identification could lead to better prevention and treatment. The current methods only identify 50% of PTSD cases.
CHARM, a new AI tool tool studies images to quickly pick out the genetic profile of a kind of brain tumor called?glioma, a process that currently takes days or weeks. While the tool is not as accurate as current genetic tests, the computer system can predict a tumor’s profile almost instantly. Early detection continues to be the number one use case of AI in diagnosis and is the best way to increase odds of surviving any chronic or highly deadly medical condition/disease.
AIVF, a company that has increased the success rate of In vitro fertilization (IVF) by ~30% using its AI-powered embryo evaluation software, called EMA. EMA is designed to process vast amounts of data — beyond what the human eye can detect — to simplify the embryo selection process. The IVF process is very daunting and requires a lot of human judgement during selection of fertilized eggs. AI as an assistant in assessing the fertilized eggs seems to vastly increase the already low rates of success for IVF.
Venture News
10. Digital Ocean Acquires Paperspace for $111m Cash
Paperspace offers cloud access to GPUs and CPUs in oder to train and deploy A.I. models. As a cloud services provider, DigitalOcean 's acquisition of Paperspace makes it an excellent addition to its portfolio of product offerings. As the physical access to high performance chips continues to be limited, the digital access opening up for those who do have the physical chips will become an extremely profitable line of business.
11. Causally & Recursion, both drug discovery platform, raises $60m in Series B and $50m from NVIDIA respectively
Other than earlier diagnosis, AI systems are being utilized for novel molecule modeling resulting in drug discovery as a popular use case.
Clients of Causally can use its cloud-based platform to identify targets for R&D, determine the specific biomarkers in the targets, and better understand the target compound or molecule, in order to determine what might be fixed with the right pharmaceuticals and other therapeutics.
Recursion received a cash injection from NVIDIA to utilize its proprietary 3 trillion gene and compound relationships data set to train foundation models on?NVIDIA DGX? Cloud?for possible commercial license/release on BioNeMo, NVIDIA’s cloud service for generative AI in drug discovery. Proprietary data in the chemical compound space is still the primary differentiator in the quality of model outputs as the open source datasets related to biochemical structures is not prolific enough yet.
12. Collective, a finance platform for freelancers, raises $50M in series B
Collective provides a variety of back office services for solo entrepreneurs and freelancers. They have implemented?large language models?to develop AI copilots. These copilots collaborate with a company’s tax experts, accountants, bookkeepers and relationship managers, to reduce the time related to operations like bank reconciliation and expense categorization. For operations that have straightforward and predictable expenses, automation is a key concept for reducing expenses and staying competitive. Building out Collective’s Copilot LLMs is a good play as things are still early and owning models that automate operations for a variety of businesses will be increasingly valuable over time.
13. Prolific, an AI testing company, raises ￡25M in series A
One of the top ideas that has emerged from government regulations is directly related to using 3rd parties to run stress tests and validate how an AI system works or what kinds of results a chatbot can produce. Prolific has been doing this kind of work for over almost 10 years and has recently raised ￡25M to expand their operations. They provide an online platform for their current network of over 120k AI systems testers. They already have an established process for selecting new ones & over 300 filters to help their testers tune what they are looking for. Similar services are provided by companies like Amazons Mechanical Turks and we should expect more of the existing LLM companies, especially ones that specialize in or do their own training, to build out networks similar to Prolific & Amazon.
14. Raft.ai, an intelligent logistics start-up, raises $30m in series B
Raft AI has rebranded from Vector AI and raised $30m. The company is focusing in on applying AI to accounts payable reconciliation and customs entry preparation for logistics companies. It seems as though reconciliation & smart PDF reading is a popular topic to be automated by AI systems. Applying these tools to industries that have not seen too much operational software come in, such as logistics and SMBs, is a good business play (maybe not venture backable in the long term), but will likely see margins compress as these product tools become commoditized features of larger incumbent systems such as intuit, adobe & other commonly used legacy systems.
15. Pano AI, a wildfire detection start-up, raises $20m in series A
Computer vision start-up, Pano AI, has raised $20m on the back of Canada’s worst wildfire season in recent history. Operating in the U.S. & Australia, Pano was able to detect fire and issue a warning 14 minutes before the first 911 call. They are in the business mainly of setting up camera systems and then applying AI systems to detect smoke and other signs of wildfires. Pano AI owns the cameras and infrastructure and sells licenses to its software to companies and organizations starting at $50,000. Field cameras getting AI layered into them will be a trend that increases as computer vision technology continues to advance. Pano is leading in detecting Natural disasters and Lilz, a Japan based company of a similar nature is focused on monitoring physical gauges that measure water/gas pressure & electrical currents.
Reader Questions
Q: “Could you explain what Edge Computing is in simple terms?”
A: Some devices like self-driving cars or smart speakers need to make decisions very fast using data they are collecting locally. They can't wait to send all that data to the cloud and receive back interpretations. So small computers called "edge servers" are placed very close to those devices. They allow the devices to process and understand their data quickly without waiting on the cloud. Alternatively, the computations that run on the cloud or on edge servers could potentially be run on the end device itself as well.
Edge computing complements cloud computing. It does not replace the cloud but rather provides another layer of computing infrastructure. The cloud is still needed for storage, processing, analytics, etc.
Key technologies enabling edge computing include 5G networks, mini data centers, AI chips and new software architectures like Kubernetes.
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.
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.