AI-generated images surpass human photographs in history, ChatGPT Enterprise, competition amongst A.I. personality companies, and more
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
Regulation
Research
Hardware
VENTURE NEWS
NEWS ROUND-UP
Product
1. AI Google Search Features, Scrapbooks & Over 20 Generative Life Coaches
Google DeepMind has been working with Scale AI on 21 generative A.I. personalities to help with personal and professional tasks, including tools to give users life advice, ideas, planning instructions, and tutoring tips.
The company has also launched new AI features linked to the search experience . Auto-generated page summaries are available while you are visiting a page, new and improved in-search answers & enhancements to searches related to coding languages.
A new AI Google Photos feature, Memories , is to be viewed as a “scrapbook-like timeline.” You’ll find Memories that have been generated by AI and created to only show content you likely want to see.
Why this matters: The new personalities that Deepmind is building with Scale AI seem to be in direct competition with Character AI and are eerily similar to Metas “personality algorithms .” Character AI is currently one of the most dominant generative AI and chatbot products. With an average visit at over 30 min and fast approaching 100m monthly visits . The website allows you to chat with a variety of chatbot personalities and specialized assistants. The customizable chatbot marketplace is in high demand as end users do not want to spend time training a general model to be specialized, but rather select from a menu of already specialized models.
2. Open AI partners with Scale AI for GPT3.5 Fine Tuning
Scale AI customers will be able to manipulate their data using Scale’s platform and then fine-tune GPT-3.5 with their data and further customize the model with features like the ability to reference or cite their proprietary data in its responses. This partnership announcement came one day after OpenAI said that it would allow users to fine-tune?GPT-3.5 Turbo using its fine-tuning API.
Why this matters: Scale AI has announced partnerships with both Google and OpenAI to create specialized models. Although the use cases are slightly different, it's obvious that Scale AI is the winner either way. Unlike Google’s internally developed life coaches, this partnership is a little different. Allowing Scale AI’s customers access, not only to the GPT3.5 turbo model but also the ability to easily apply new finetuned weights to the model makes it infinitely more likely to be used for specific purposes. The partnership will not only continue to cement OpenAI’s GPT models as the standard LLM but will also increase access for Scale AI’s clients.
3. 15B Generative Pictures Posted, Longform Deepfakes & Adobe updates
Over 15 billion images ?have been created by text-to-image algorithms since 2022.?It took photographers 150 years, from the first photograph taken in 1826 until 1975,?to reach the 15 billion mark . These images were generated by DALLE-2, Adobe Firefly & Midjourney with over 80% as a product of stable diffusion model technology.
In fact, Adobe is making it even easier to use generative AI in its suite of products. Firefly is now?integrated into Adobe Express enabling text-to-image and text-to-effect prompting capabilities in over 100 languages. The company also beefed up its video suite offering boasting over 200M static and dynamic assets available for use by creators.
However, RunwayML is not about to let Adobe capture its market share of AI enabled video creators. With a community of over 10k creators, the company has launched “Watch” to allow users to share and consume longer-form videos created with Runway tools. The videos can be up to 10 minutes long and can be shared on social media or embedded on websites.
Why this matters: The threat to authentic content seems to be coming faster for image content than video and written content. 15B generative images in around a year is a figure that will only grow, likely exponentially, as tools become more democratized and easier to use. Adobe is a prime example of this trend as they continue to build out their Firefly feature set and availability. Adobe has over 30M paying users and sooner or later all of them will have produced an image that was either a product of or enhanced by AI tools. The natural extension of generative images is generative videos. The RunwayML CEO is targetting “long form” (10 minutes is long for AI videos, but not necessarily for human-produced content) creators to build out tooling for independent filmmakers. Democratizing video tooling to enable a new generation of filmmakers is a rosey vision, but we will be watching to see how well equipped independent video/film creators are with AI tools when standing against the establishment in Hollywood.
Regulation
4. No Copyrighting “AI”rt
Intellectual property law is historically only granted to works created by humans, and this most recent ruling by Federal Judge Stephen Taylor upholds that. Copyright law has “never stretched so far” to “protect works generated by new forms of technology operating absent any guiding human hand,” U.S. District Judge Beryl Howell found.
Why this matters: This topic has been debated since generative AI took the internet (and the world) by storm. The sheer amount of volume of content requires clear guidance from lawmakers and upholders regarding how society should treat and interact with this content. with this clear cut precedent case, we are one step closer to having certainty around navigating are increasingly AI generated world.
5. Beijing Instates AI Boundaries & Tencent “Joins” the AI Race
Beijing has spent years laying the groundwork for the rules that took effect earlier this month. The country’s cabinet put out an AI roadmap in 2017 that declared the development of the technology a priority and laid out a timetable for putting government regulations in place. Here is a table below detailing the differences between US and Chinese AI regulations.
Despite the seemingly strict rules from China, Tencent revealed this week that it will launch its proprietary foundation model later this year, claiming it will rank among the best in China. Tencent has been testing the AI, named Hunyuan, across gaming, advertising, cloud computing, and financial technology.
Why this matters: The difference between regulation in the US and China multiplied by supply-demand mismatch in GPUs needed for the AI boom (bubble?) results in an inefficient market where companies that get involved will either see benefits from being early or will get crushed by the actively shifting landscape. Tencent launching an AI model or AI product to its users would benefit greatly from its distribution advantage. Tencent's social media platforms have over one billion users. More than 129 million people pay for a Tencent Video subscription.
Research
6. New Opensource Coding LLMs from Meta & Stability AI
Stability AI has released an open source model, StableCode , which supports languages like Python, Java, C++, and more, leverages 560B code tokens & has a context window of 16k tokens. Around the same time, Meta released their open source coding model, Code Llama LLM , which is set to be released sometime this month or early next month.
Why this matters: Between OpenAI’s Codex, Github Copilot, StableCode, Code Llama, and more, the LLM for coding and software development seems to be self fulfilling its predicted commodity status. While the open source dimension to the AI movement remains, any potential “margins” from owning a propriety coding focused model. However, enterprise, start-up & solo teams leveraging these models to achieve a higher throughput of quality code pushed to development will see the most benefit from these
7. Metas Multi-Modal Multi-Lingual Model
SeamlessM4T is a foundational multilingual and multitask model that seamlessly translates and transcribes across speech and text. SeamlessM4T supports speech recognition, Speech-to-text translation, Speech-to-speech translation, Text-to-text translation, and Text-to-speech translation, all for nearly 100 input languages and output languages.
领英推荐
Why this matters: Meta is on a model release frenzy. We have learned that they specialize in Multi-Modal models; releasing Generative Music and Generative Art models & they specialize in supporting lots of languages , which they originally started teasing in May. Now we have a Speech translation model that seems to perform better than Google’s Audio PaLM and OpenAI’s Whisper model. These releases do not bode well for companies looking to profit off increasingly commoditized Models and are a win for any company attempting to design application building platforms that can access all available models.
8. AI Outsmarting Human Detection
Two new sets of research have been released that confirm AI is becoming harder to detect.
Researchers from London generated 50?deepfake?speech samples in each language and played them to 529 participants. About 27% of the time, listeners thought the deepfake speech was real. In a separate study, Researchers from Microsoft & UC Irvine recruited 1,400 participants to test websites that used CAPTCHA puzzles, which account for 120 of the world’s 200 most popular websites. The AI bots’ accuracy ranges from 85-100%, with the median above 96%. These results substantially exceed the range of human accuracy we observed (50-85%).
Why this matters: In Unwiring AI #3 we talked about cryptographically verified microphones as the pinnacle of security against deepfake voice technology. Unfortunately, most people will be relying on their ears, which means just under 1 in 3 people are viable to be fooled by AI generated voices. This fact is a massive red flag, as voice models convene on smoother tonality and more dynamic responses to conversational prompts, this statistic will only get worse. AI bots fooling captchas seem to be even more problematic as this sort of technology is easier to scale and deploy. Although this is a boon for folks building functional AI agents, whether this development is beneficial for society depends on what the agent is doing. Given that these bots can beat captchas more consistently than humans can, we already are outdating aspects of the modern website cyber security stack. These violations will (hopefully) spawn new advancements in bot detection.
Hardware
9. IBM’s Brain Chip & GPU Market Share
Just in case you were unsure of how human-like technology is becoming, IBM has unveiled a new "prototype" of an analog AI chip that works like a human brain and performs complex computations in various deep neural networks (DNN) tasks. IBM says the state-of-the-art chip can make artificial intelligence remarkably efficient and less battery-draining for computers and smartphones.
The current market dynamics of the global semiconductor market spend is changing. Currently, it is dropping in the US and rising in China, which should come as no surprise given the billions of dollars Chinese companies are spending on GPUs. It is surprising to see Japan on such a steep decline and the other nations falling flat overall.
Why this matters:
VENTURE NEWS
10. OpenAI acquires AI design studio Global Illumination
In an undisclosed deal, OpenAI acquires Global Illumination , which leverages AI to build digital experiences. This acquisition is likely fueled by the fact that UX in emerging tech products is often unintuitive for new users. Global Illumination is backed by VC firms Paradigm, Benchmark, and Slow, Global Illumination’s team designed and built products early on at Instagram and Facebook as well as YouTube, Google, Pixar, and Riot Games.
11. AMD Acquires Mipsology To Ramp Up AI Inference
AMD has acquired French startup Mipsology to strengthen its AI inference software capabilities as the chip designer mounts its biggest challenge yet to AI chip powerhouse Nvidia. This acquisition is in-line with AMD’s AI strategy to consolidate the industry overall and drive more developers to build on their hardware as opposed to using NVIDIA’s hardware.
Model inference is the process of using a trained model to make predictions on new data. For example, if you've trained a machine learning model to identify animals in pictures, model inference would allow the model to automatically identify animals in new pictures that it has never seen before. This technology is critical to the advancement of LLMs and other AI systems.
12. Hugging Face Raises $235M in Series D
Bringing its total funding to over $330M this year alone, Hugging Face has raised another round from a variety of strategic investors such as Google, Amazon, Nvidia, Intel, AMD, Qualcomm, IBM, and Salesforce. The company has 10,000 customers today, it claims, more than 50,000 organizations on the platform, and its model hub hosts over 1 million repositories. Huggingface has partnered and collaborated with companies like ServiceNow, NVIDIA, Amazon, and Microsoft to expand its offerings and services to its user base.
13. Databricks is looking for a $100m cash injection
After a mouthwatering $1.3B stock acquisition of MosaicML, Databricks is now looking for a fresh $100m in cash to capitalize on enterprise demands for a reliable data platform to fuel a new generation of artificial intelligence applications. The company is still a net negative cashflow business with losses at almost $1B after the last 2 years. That being said, the company is reporting a 70% YoY revenue growth rate. With large losses mounting and aggressive revenue growth, an injection of cash could help to maintain Databricks edge assuming they allocate the capital appropriately to maintain their momentum.
14. Anthropic raises $100M from Korean telco giant SK Telecom
In another 9-figure deal, Anthropic raises $100 million in funding from one of the largest mobile carriers in South Korea,?SK Telecom (SKT) . The funding news brings their total funding, this year alone to over $850M across a total of four rounds. Claude 2.0 has been quite popular as a free alternative to GPT-4. SKT has been developing large language models for the Korean language for years. The company started building a Korean version of BERT in 2018 and launched a chatbot “A.” in beta last year and upgraded its chatbot by adding a feature, “Chat T,” which is powered by Microsoft’s Azure Open AI cloud computing service, to “A.” in June.
Some stats on how Claude Ai does, its average site duration is just over 5 minutes with around 16m page visits. Something to note is that since the release of Claude 2.0, their website traffic has increased by several orders of magnitude, and fundraising on the back of intense growth and momentum bodes well for the company.
15. Modular, AI software developer platform, raises $100M in Series A
Modular has raised $100 million in a funding round led by General Catalyst with participation from Google Ventures, SV Angel, and Greylock. Modular’s engine, currently in closed preview, lets developers import trained models and run them up to 7.5 times faster versus on their native frameworks. Modular’s other flagship product, Mojo, is a programming language that aims to combine the usability of Python with features like caching, adaptive compilation techniques, and?metaprogramming .
16. Helm.ai, working on unsupervised learning for AI Raises $55M in Series C Funding
Helm is an autonomous driving startup focused on robotics automation. Self-driving has been a bit of a hot topic since SF debuted self-driving cars with Cruise & Waymo. Townships actively altered how big fleet sizes were given any accidents the cars were involved in. Several issues such as unexpected stopping due to people placing cones in front of the cars have not stopped them. With companies like Helm improving the computer vision technology used by autonomous vehicles, highway, urban, and last-mile delivery should be seeing compounding effects on how well these vehicles navigate the roads.
17. EvolutionaryScale, A drug discovery start-up, raises $40M
Earlier this month Meta announced that they are disbanding their protein-folding, which created the first database of more than 600mn protein structures, in favor of other AI projects. Afterward, this team has now started EvolutionaryScale and raised $40M at a $200M valuation led by Lux Capital. There are a couple of interesting things about this company to keep an eye on:
18. Ideogram AI, a generative AI art start-up, raises $16.5M in Seed Funding
Generative Art is a very popular use case for AI models. As we saw above, stable diffusion is the most popular given that a majority of AI-generated images are a result of stable diffusion technology. While Stability AI is focused on releasing general LLMs and Coding LLMs, a new start-up has decided to take up the competition with Midjorney and Stability AI’s image model. Ideogram AI raised $16.5M in a seed round off of a seemingly novel feature that other models do not have; reliable text generation?within?the image, such as lettering on signs and for company logos. Here are some examples below!
19. DynamoFL raises $15.1M in Series A to help enterprises adopt ‘compliant’ LLMs
Companies struggling with equipping their workforce with AI while also banning publically used tools have created a space in the market for enterprise private or closed LLMs. Companies are worried about their confidential data ending up with developers who trained the models on user data; in recent months, major corporations?including ?Apple, Walmart, and Verizon have banned employees from using tools like OpenAI’s ChatGPT. DynamoFL has raised $15.1M to bring an LLM developmental platform for enterprises to build models based on documents in their private cloud servers. The company specifically emphasizes its compliance focus and close work with legal teams.
20. Caden, a data monetization startup, Raises $15M Series A
Caden has raised 15M to offer controls that allow users to fine-tune which data they share with third parties. Caden pulls in analytics data on what movies users watch, where they travel, what they buy, and more, and lets those users opt-in to monetize their data in various ways. The company is integrating data availability to be used in model training while enabling users to take ownership of their data. Although this marketplace play will require buy-in from both sides, the ethos behind this product is an encouraging step in how retail can take control of their data and leverage it to assist in the creation of the next generation of AI systems.
21. Voiceflow, a platform for building conversational AI experiences, raises $15M
Voiceflow is a “Figma-quality” collaborative design platform for building AI agents. A company can design, test, and deploy an AI agent without being locked into a particular model or technology, mixing and matching different models — including speech recognition models — to create conversational AI experiences. This sort of LLM app-building extraction layer is similar to Amazon's Announcement of Bedrock . Moving up and down the AI stack away from building transformer models is a cost-effective way to earn a first-user advantage in those building multi-model-based applications such as AI agents.
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.
Senior Managing Director
1 年Andrew Hong Fascinating read.?Thank you for sharing.