The generative AI revolution is here. Total value of generative AI startups hit $48bn, up 6x from 2020
The generative AI revolution has been built on a foundation of academic and corporate research. Large AI models like ChatGPT require an extraordinary amount of computing power to train on. Given the extraordinary costs of training large language models and other generative AI systems, industry has surpassed academia as the developer and deployer of the largest AI systems.
Big Tech companies like Google, Microsoft, Meta and Amazon are also among the leading publishers of academic research on AI.
The main semiconductor company benefiting from the AI boom is Nvidia.
Nvidia makes 80% of all AI chips
ChatGPT requires an estimated 10,000 Nvidia chips to run.
OpenAI is the leader of the generative AI pack.By some measures it may also be the fastest growing startup of all time. Generative AI systems like Copilot are already generating massive productivity gains for software engineers.
Investors are betting that AI will revolutionize drug discovery.
Funding for AI-led drug discovery has risen 3,800% in the past five years to $2bn. AI was leveraged by biotechs and big pharma to develop and optimize vaccines and therapeutics during the COVID crisis.
More data on the generative AI mega boom. There are now at least 539 generative-AI startups. Not counting OpenAI, they have so far collectively raised more than $11bn. OpenAI has raised over $11b.
In terms of the geography of AI innovation, it’s a two horse race between the US and China. Europe barely features. Academic publishing in AI is also dominated by China and the US.
The generative AI revolution has been built on a foundation of academic and corporate research. Large AI models like ChatGPT require an extraordinary amount of computing power to train on. Given the extraordinary costs of training large language models and other generative AI systems, industry has surpassed academia as the developer and deployer of the largest AI systems.
Big Tech companies like Google, Microsoft, Meta and Amazon are also among the leading publishers of academic research on AI.
The main semiconductor company benefiting from the AI boom is Nvidia.
Nvidia makes 80% of all AI chips
ChatGPT requires an estimated 10,000 Nvidia chips to run.
OpenAI is the leader of the generative AI pack.By some measures it may also be the fastest growing startup of all time. Generative AI systems like Copilot are already generating massive productivity gains for software engineers.
Investors are betting that AI will revolutionize drug discovery.
Funding for AI-led drug discovery has risen 3,800% in the past five years to $2bn. AI was leveraged by biotechs and big pharma to develop and optimize vaccines and therapeutics during the COVID crisis.
More data on the generative AI mega boom. There are now at least 539 generative-AI startups. Not counting OpenAI, they have so far collectively raised more than $11bn. OpenAI has raised over $11b.
In terms of the geography of AI innovation, it’s a two horse race between the US and China. Europe barely features. Academic publishing in AI is also dominated by China and the US.
👍4🔥1
The 100 most cited AI papers for 2022.
A detailed analysis of the most cited papers for the last three years allows good insights into the organisations and countries publishing the most impactful AI research right now.
In the recent discussion about the impact of AI R&D from various Big Tech labs, can now contrast volume and impact.
Volume top-5:
Google AI
Tsinghua University
Microsoft Research
Carnegie Mellon University
MIT
Impact top-5:
Google AI
Meta AI
Microsoft Research
Berkeley AI
DeepMind
A detailed analysis of the most cited papers for the last three years allows good insights into the organisations and countries publishing the most impactful AI research right now.
In the recent discussion about the impact of AI R&D from various Big Tech labs, can now contrast volume and impact.
Volume top-5:
Google AI
Tsinghua University
Microsoft Research
Carnegie Mellon University
MIT
Impact top-5:
Google AI
Meta AI
Microsoft Research
Berkeley AI
DeepMind
👍1
Apple will have its self-developed 5G modem chip, code-named Ibiza, made on TSMC's 3-nanometer process, with risk production to begin in the 2nd half of 2023, media report, citing supply chain sources, and adding the chips will debut in the #iPhone 16 in 2024.
工商時報
蘋果5G晶片傳台積電通吃 下半年試產
蘋果投入龐大資源研發5G數據機（modem）晶片，可望提前在明年導入iPhone 16系列手機。蘋果目前iPhone採用的5G數據機晶片都是向高通採購，但高通執行長艾蒙（Cristiano Amon）近日出席2023年世界行動通訊大會（MWC 2023）時表示，蘋果可能在iPhone 16系列搭載自...
If you’re waiting forever on Bing Chat’s waitlist … don’t. Try http://you.com/chat. It provides the same chat interface with internet access, and seems to be quite a bit more creative in my (limited) testing.
We’re getting really tired of Bing refusing to answer anything that it perceives to be negative, however harmless, even in "Creative Mode". IMHO the lobotomy guard rails went too far.
Screenshots:
1) YouChat. 2) Bing Chat, creative mode.
We’re getting really tired of Bing refusing to answer anything that it perceives to be negative, however harmless, even in "Creative Mode". IMHO the lobotomy guard rails went too far.
Screenshots:
1) YouChat. 2) Bing Chat, creative mode.
VC dollars shifted from subsidizing your taxi ride and burrito delivery to LLMs and generative AI compute.
Chinese team’s Ethereum Layer 2 network Scroll raised $50 million in new funding at $1.8 billion. Investors included Polychain Capital, Sequoia China, Bain Capital Crypto, Moore Capital Management, Variant Fund, Newman Capital, IOSG Ventures and Qiming Venture Partners.
The Block
Ethereum L2 Scroll reaches $1.8 billion valuation in new funding round: Source
Scroll, an Ethereum Layer 2 network that uses zk-rollups technology, raised $50 million in a new funding round.
Prediction: LLMs will play a big role in the 2024 US presidential election. The two major candidates will have ChatGPT-like bots you can talk to and it will seem like a Zoom call with the actual candidate. If just one of the two candidates has such a bot, he wins easily.
❤3👍2🔥1🤡1
Microsoft's latest AI assistant built on OpenAI is meant for its Dynamics products, taking on Oracle, SAP, Salesforce with a tool to help marketers, finance staff and customer service agents.
Dynamics 365 Copilot follows GitHub Copilot and the Bing bot.
Of course, the stakes are different if a business-focused copilot spits out incorrect info or engages in creepy behavior. One other thing Microsoft put into the announcement: its next AI-focused news will relate to productivity software (ie Office) and come March 16.
Dynamics 365 Copilot follows GitHub Copilot and the Bing bot.
Of course, the stakes are different if a business-focused copilot spits out incorrect info or engages in creepy behavior. One other thing Microsoft put into the announcement: its next AI-focused news will relate to productivity software (ie Office) and come March 16.
Bloomberg.com
Microsoft’s Latest AI Assistant Is Meant for Marketers, Customer Reps and Work Apps
AI ‘copilot’ will help answer customer calls, summarize sales meetings and generate marketing pitches
Google designed and developed their own optical switches for their network
This dramatically reduces their networking cost of ownership by 40% and capex by 30% and it improves performance 32%.
The switch is based on MEMS mirror arrays only 3db loss is nutty.
8 optical switches per rack, 32 of these racks across each datacenter.
This is their datacenter spine
Used for training PALM too because the data movement is rigid.
You know how much each accelerator needs to talk to others and how many bytes will be transfered before training.
They claim 30% less capex over 6 years and 40% less open from power savings, inclusive of reliability.
This dramatically reduces their networking cost of ownership by 40% and capex by 30% and it improves performance 32%.
The switch is based on MEMS mirror arrays only 3db loss is nutty.
8 optical switches per rack, 32 of these racks across each datacenter.
This is their datacenter spine
Used for training PALM too because the data movement is rigid.
You know how much each accelerator needs to talk to others and how many bytes will be transfered before training.
They claim 30% less capex over 6 years and 40% less open from power savings, inclusive of reliability.
Stability AI has acquired the industry leader in AI-powered imaging tools Clipdrop App.
Stability AI
Stability AI Acquires Init ML, Makers of Clipdrop Application — Stability AI
Stability AI, the world’s leading open source generative AI company, announces that it has acquired Init ML, makers of the globally popular imaging tool Clipdrop.
Google has launched a new AI model, and it's incredible.
Introducing PaLM-E, a multimodal language model across robotics, vision and text.
This is a pretty massive 562B model and beats all previous ones!
APPLICATIONS
1. Robotics
Instruct a robot to "bring me the rice chips from the drawer". Includes multiple planning steps as well as incorporating visual feedback from the robot's camera.
2. Visual Question Answering.
PaLM-E's a generalist across language, vision robotics.
Introducing PaLM-E, a multimodal language model across robotics, vision and text.
This is a pretty massive 562B model and beats all previous ones!
APPLICATIONS
1. Robotics
Instruct a robot to "bring me the rice chips from the drawer". Includes multiple planning steps as well as incorporating visual feedback from the robot's camera.
2. Visual Question Answering.
PaLM-E's a generalist across language, vision robotics.
palm-e.github.io
PaLM-E: An Embodied Multimodal Language Model
Project page for PaLM-E: An Embodied Multimodal Language Model.
❤2
Trends in AI March 2023 is out
1. LLaMA from MetaAI , an open-source set of GPT-like pretrained models to work upon. Licensed to be used only for research purposes and not academic though.
2. "Consistenty Models" from OpenAI, a generalization of diffusion to model an arbitrary jump in diffusion steps, achieving image generation in 1 step without adversarial training.
3. PaLM-E from Google AI and UMI LAB, an embodied 562B language model that takes actions in the physical world and shows substantial positive transfer learning across modalities.
4. "In-context Instruction Learning" from LG AI Research, showing that instruction learning can be done *in-context*.
5. Evaluating GPT models on machine translation by Microsoft Research showing how chatGPT and davinci models perform on par with many SOTA NMT and commercial translation services, and has complimentary strengths.
6. "Composer: Creative and Controllable Image Synthesis with Composable Conditions" by Alibaba: next level composable controllability for diffusion-based image generation.
7. Prismer from Nvidia AI an efficient modular mixture of experts multimodal language model that can answer and reason about images and text.
8. "Augmented Language Models: a Survey" by MetaAI , everything you need to know about LMs + tools, memories, search.
9. "Symbolic discovery of optimization algorithms" by GoogleAI and UCLA , symbolic program search to learn an optimizer, speeding up training 2x vs Adam.
10. "MarioGPT: Open-Ended Text2Level Generation through Large Language Models" by University of Copenhagen Research. Just that, generating Mario levels by prompting a GPT-2 model + embedding it on a novelty search evolutionary computing loop to generate even cooler worlds
1. LLaMA from MetaAI , an open-source set of GPT-like pretrained models to work upon. Licensed to be used only for research purposes and not academic though.
2. "Consistenty Models" from OpenAI, a generalization of diffusion to model an arbitrary jump in diffusion steps, achieving image generation in 1 step without adversarial training.
3. PaLM-E from Google AI and UMI LAB, an embodied 562B language model that takes actions in the physical world and shows substantial positive transfer learning across modalities.
4. "In-context Instruction Learning" from LG AI Research, showing that instruction learning can be done *in-context*.
5. Evaluating GPT models on machine translation by Microsoft Research showing how chatGPT and davinci models perform on par with many SOTA NMT and commercial translation services, and has complimentary strengths.
6. "Composer: Creative and Controllable Image Synthesis with Composable Conditions" by Alibaba: next level composable controllability for diffusion-based image generation.
7. Prismer from Nvidia AI an efficient modular mixture of experts multimodal language model that can answer and reason about images and text.
8. "Augmented Language Models: a Survey" by MetaAI , everything you need to know about LMs + tools, memories, search.
9. "Symbolic discovery of optimization algorithms" by GoogleAI and UCLA , symbolic program search to learn an optimizer, speeding up training 2x vs Adam.
10. "MarioGPT: Open-Ended Text2Level Generation through Large Language Models" by University of Copenhagen Research. Just that, generating Mario levels by prompting a GPT-2 model + embedding it on a novelty search evolutionary computing loop to generate even cooler worlds
👍6
Samsung has hired ex-TSMC exec Lin Jun-cheng as vice president to speed up development of its advanced packaging technology, media report, adding Lin worked at TSMC from 1999 to 2017, laying the groundwork for TSMC's 3D packaging technology.
www.businesskorea.co.kr
Samsung Electronics Brings in TSMC Executive as Vice President for Packaging Development - Businesskorea
Samsung Electronics hired Lin Jun-cheng as a vice president, a veteran engineer who worked for nearly 19 years at TSMC, Samsung’s archrival in the foundry business. The recruitment of Lin was aime...
Sam Altman, СЕО of OpenAI invested $180 million into a company trying to delay death
MIT Technology Review
Sam Altman invested $180 million into a company trying to delay death
Can anti-aging breakthroughs add 10 healthy years to the human life span? The CEO of OpenAI is paying to find out.