Original source: Silicon Rabbit Race
Author | Eric Gu、Zuri Wang
Image source: Generated by Unbounded AI
** Let’s take a look at 2023, we missed a few $Bs. **
A year ago today, ChatGPT was released, and it took only 5 days for the number of registered users of ChatGPT to reach 1 million, setting a new record on the Internet. Two months later, in January 2023, ChatGPT reached 100 million monthly active users.
The meteoric rise in the number of users heralds the rapid growth of generative AI this year, with a16z founder Marc Andreessen proposing “Software is eating the world” in 2011, and 12 years later, the phrase will probably be replaced by “AI is eating the world”.
In just one year, many companies have taken advantage of generative AI, raising two rounds of funds a year and valuations of more than one billion US dollars, while others have been hit by ChatGPT and their stock prices/valuations have fallen… Lay off employees to survive.
ChatGPT’s one-year anniversary, let’s take a look at 2023, we missed a few $B.
2022.11.30
**OpenAI releases ChatGPT powered by the GPT-3.5 model. **
Prior to this, Jasper, an AI-generated marketing content company, and Stability AI, an AI-generated image company, closed a new round of funding at a valuation of $1.5 billion and $1 billion, respectively, which was unique in the sluggish primary market in the second half of 2022 and ignited the enthusiasm of American investors for generative AI.
2022.12
OpenAI’s investment attractiveness has skyrocketed. **Eric Newcomer, author of well-known tech media, wrote: "Some high-profile investors in Silicon Valley have been considering buying shares from OpenAI shareholders. At the same time, according to people familiar with the situation, Microsoft has been in talks with OpenAI about the possibility of a direct investment. ”
2023.01
Microsoft continues to invest in OpenAI News, adding $10 billion, accounting for 49% of the shares, and OpenAI’s valuation reaching $29 billion. More details of the deal will be revealed – OpenAI will have the right to get its shares back if it can bring in $105 billion in returns for Microsoft and $150 billion in returns for other investors.
2023.02
**The battle between Google and Microsoft is emerging. **
Microsoft launched Bing Chat, a chatbot version of the search engine supported by GPT-4, and Google was caught off guard and announced that the Bard chat search engine was about to be launched in a hurry, and as a result, Bard made a mistake in the demo process, and the stock price plummeted by 9%, evaporating the market value of hundreds of billions of dollars. For a long time then, there was a debate about whether Google Search could hold itself in the era of generative AI.
In addition, Google shakes hands with Anthropic, a large language model company founded by former Google and OpenAI researchers. Google not only invested $400 million exclusively, but also partnered with it to implicitly challenge the combination of Microsoft + OpenAI.
2023.03
**Technology giants and start-ups run to enter. **
Bloomberg unveiled the 50 billion BloombergGPT model, at the Nvidia press conference, Lao Huang claimed that “the iPhone moment of AI has arrived”, and Bill Gates called for “the advent of the AI era” through a blog post, calling it “a revolutionary and disruptive product, just like the emergence of smartphones and the Internet”.
Adept AI, an AI agent company founded by the author of the foundational paper “Attention is All You Need”, and Character AI, an AI social networking company, have each completed a new round of financing and become unicorns. Musk quietly established xAI, a large model company.
Artificial intelligence is showing what it is capable of causing panic. **
Led by Elon Musk, more than 1,000 Silicon Valley entrepreneurs and scientists jointly called for “Danger! Stop all large-scale AI research immediately”.
2023.04
**The battle for large models has entered a white-hot state. **
In response to the menacing Microsoft, Google merged Google Brain and Deepmind into Google Deepmind, which was unanimous with the outside world. Amazon has jumped on the big model bandwagon with the launch of Bedrock, a hosting service for third-party models and its own large language model, Titan FM.
**The problem of lack of “core” is gradually emerging. **
Musk has a reserve of about 10,000 GPUs, and Corweave, which started as a blockchain mining business, began to provide computing resources for AI, successfully completing a new round of financing with a valuation of $2 billion.
**The cornerstone model lowers the threshold for application development, and generative AI applications begin to explode. **
Among the 268 projects in the Y Combinator Winter Incubator, a well-known incubator in Silicon Valley, 50% of the projects are related to AI. AutoGPT, the new top stream of AI, was born, with nearly 80,000 stars on GitHub, successfully surpassing PyTorch (65,000), and the development direction of AI Agent represented by AutoGPT has attracted attention.
In addition, because it is too popular with investors, the founder of Rewind AI, a personal historical search engine/memory assistant, recorded a video and raised it publicly, and its direction and linear growth of ARR (ARR reached 700,000 US dollars in a few months) made investors crazy, and 170 TS got soft, and finally NEA got the admission ticket and led a new round of Rewind investment according to 350 million US dollars.
The undercurrent under the heat of AI is not only, privacy, data security… With a variety of issues, Italy’s privacy regulator has temporarily banned ChatGPT’s use, saying it collects and stores information in an inappropriate way.
2023.05
Google and Microsoft Developer Conference were unveiled this month, full of gunpowder.
At the Google I/O conference, AI became the absolute core. The language model PaLM2, the AI platform Vertx AI, the one-click P-graph Magic Editor, the office assistant Duet AI, and the upgraded Bard were all unveiled. Microsoft announced Windows Copilot at Build, bringing AI as infrastructure to the core systems of Windows.
The lack of “core” problem made Nvidia’s financial report exceed expectations, and its stock price rose to 380 US dollars after hours, exceeding the all-time high, and also making Nvidia’s market value exceed one trillion, becoming the fifth company with a market value of more than one trillion after Apple, Google, Amazon, and Microsoft. Cloud vendors such as Microsoft, Google, and Amazon are scrambling to deploy AI chips and plan to increase supply in the second half of the year.
Companies that provide large models and AI services for enterprises have attracted attention, the CSO & CTO (author of the Transformer paper) of AI agent start-up Adept AI left to found Essential AI to help enterprises better use large models, and announced the completion of financing at a valuation of $50 million. Cohere, a large-scale enterprise company founded by Aidan Gomez, another author of the “Attention is All You Need” paper, has become a unicorn and has received support from Nvidia, Oracle, Salesforce and others.
AI concerns are further fermenting, and AI-generated images of “explosions near the Pentagon” have gone viral, causing the S&P 500 index to fall for a time. Sam Altman attended the hearing to discuss AI risks and regulatory issues with lawmakers.
2023.06
**The large model continues to show super gold absorption ability. **
Chips become the oil of the AI era. **
Former GitHub CEO Nat Friedman and well-known angel investor Daniel Gross purchased 2,512 H100 chips to build servers worth nearly $100 million, giving priority to providing computing services for startups backed by their fund AI Grant.
It has been revealed that the U.S. Department of Commerce is preparing to update the export controls on artificial intelligence chips implemented in October last year, and NVIDIA and AMD will be affected. According to WSJ, the Biden administration is planning to restrict Chinese companies from using U.S. cloud computing services, including AWS, Microsoft and other cloud service providers may be affected.
The AI controversy has spread to cloud data platforms. Databricks and Snowflake chose the same week to host their annual conference, transforming their relationship from friend and foe to major competitors on the AI battlefield. Against this backdrop, Databricks acquired MosaicML (ML Model Deployment), which was only one and a half years old, and gave a sky-high price of $1.3 billion.
2023.07
**Large model companies show their respective focus. **
OpenAI is assembling a new team led by its chief scientist, Ilya Sutskever, to develop ways to guide and control “super-intelligent” AI systems. The team has access to 20% of OpenAI’s computing resources.
Anthropic launched the latest version of its AI chatbot, Claude 2, and began to make it available for free to a wider user base. At present, Claude 2 supports attachment uploading, which can complete tasks such as file analysis and data organization.
** Elon Musk led 12 people to found X.AI. **
**Hugging Face raises $200 million at a valuation of $4 billion. **
Meta continues to open source the large model ecosystem and releases Llama 2, which is free and open for commercial and research use, containing three parameter variants: 7 billion, 13 billion and 70 billion. 2 trillion tokens were used for model training, and 1 million human annotation data was used for fine-tuning.
Apple, which has been silent in the era of generative AI, has finally heard the news that Apple has allegedly built an LLM framework called “Ajax” and created an “AppleGPT”. According to people familiar with the matter, AppleGPT is a replica of Bard, ChatGPT and Bing AI.
Google DeepMind made efforts in the direction of embodied intelligence, which lasted 7 months, and released the robot model RT-2, which has the ability of symbol understanding, reasoning and human recognition. The model is a collection of PaLI-X, PaLI, and PaLM-E.
2023.08
Although the number of visits to ChatGPT has been declining for 3 consecutive months, new players and new directions continue to emerge from the cornerstone model to the application.
Llion Jones, author of Attention is All You Need, left Google to co-founded Sakana AI, the former head of AI research at Stability, with plans to build the next generation of artificial intelligence.
Midjourney is a strong contender, with text-to-image company Ideogram AI closing a $16.5 million funding round led by a16z and Index Ventures. The founding team is composed of famous paper presenters such as Diffusion and Imagen.
OpenAI also appears to be actively looking for a solution to the traffic drop — announcing its first acquisition since its founding, Global Illumination, which has developed popular online role-playing games, in a variety of ways, such as strengthening its visual capabilities, moving into gaming, or training AI agents.
According to The Information, OpenAI’s ARR exceeded $1 billion, about half of which came from ChatGPT subscriptions and the other half from B-end customers. OpenAI announced the launch of the enterprise version of ChatGPT, which focuses on serving large enterprises, providing generative AI services without compromising data privacy and security. ChatGPT Business for small and medium-sized companies will also be available soon.
2023.09
**The keywords of the month are open source and computing power. **
LLaMA and Falcon open source models are firmly established. Meta announced that it will develop an open-source large model that benchmarks GPT-4, with several times larger parameters than LLaMA 2, and is expected to start training in early 2024; Falcon-180B, an open-source large model launched by Tll, the world’s leading technology center in Abu Dhabi, is free to commercialize, beating Llama 2 in benchmarks, and its performance is close to GPT-4, and it directly topped the Hugging Face ranking once it was released.
The shortage of computing power continues to be a headache. According to an exclusive report by The Information, Meta’s internal AI team is divided over computing resources. More than half of the 14 authors of the original LLaMA model have left. According to a report by Reuters, OpenAI is exploring the manufacture of its own AI chips and has even evaluated potential acquisition targets.
OpenAI is still leading the development of large models, with a valuation of $90 billion. ChatGPT multimodal upgrade, launch of voice and image functions, DALL· E 3 generates images with realistic details, which greatly reduces the threshold for users.
The Google model is still playing catch-up. Google has opened up an early version of Gemini to a small group of companies, designed to compete with GPT-4. The training power used by the model is 5 times larger than that of GPT-4, and the multimodal capability will be greatly improved. Bard updates access to Google**FamilyTr, which can access information from Gmail, Google Docs, Google Maps, Youtube, and more. At the same time, Bard also launched the model self-check function.
2023.10
**The end of the year is approaching, and the parties are approaching the moment to hand in their answers. **
In September, ChatGPT received 15.6 million downloads of iOS and Android apps worldwide, with total revenue approaching $4.6 million, a 20% growth rate and a decline from the previous month.
**GitHub exceeded 1 million paid users. **
Microsoft’s cloud computing business grew more than expected. As of September, Azure’s cloud business grew at 29%. Conversely, Google’s cloud business is not growing as much as expected.
AI search engine unicorn Perplexity ARR 3 million US dollars, with about 15,000 paying users, plans to raise according to a valuation multiple of 50 times and a valuation of 500 million US dollars.
**Significant progress has been made in embodied intelligence and AI agents. **
Google DeepMind open-sourced the Open X-Embodiment dataset, which contains more than 500 skills and 150,000 tasks demonstrated in more than 1 million scenarios. It is expected to be the cornerstone of embodied intelligence.
Adept.ai On HuggingFace, a small version of the model that powers its product, Fuyu-8B, has been open-sourced. Compared to other multimodal models, Fuyu-8B is easier to understand, scale, and deploy, and is specifically designed for AI Agent.
Imbue is one of the few companies with sufficient capital to develop a foundational model to create a dedicated model for the AI Agent, raising $12 million with participation from the Amazon Alexa Fund.
**AI regulation has ushered in an important update. **
For the first time, the U.S. has issued a new executive order on AI, requiring new safety assessments, fairness and civil rights guidance, and research on the impact of AI on the labor market.
2023.11
**2023 is the year of “AI”. **“AI” was named the most talked-about word of 2023 by dictionary publisher Collins. In the past year, the use of this word has increased fourfold. The word of the year for 2022 is “NFT”.
**OpenAI staged a palace fighting drama. **The first developer conference was held in the first half of the month, ChatGPT App Store, GPT-4.5, Assistant API… In the second half of the month, Sam Altman suffered a “backstabbing”, 9 percent of employees resigned, OpenAI was almost bought by Microsoft for “0 yuan”, and finally Sam Altman staged a “true return” and returned to serve as CEO to resolve a crisis.
**Musk’s xAI large model blossoms. **xAI releases its first large-scale model product, Grok. With a scale of 33 billion parameters, 8K context, and 2 months to train, Grok-1 surpassed ChatGPT-3.5, Inflection-1 and other models in benchmarks.
Nvidia has launched its latest H200 chip, which includes 141GB of next-generation “HBM3” memory, which is twice the performance of the H100. The B100 chip, based on the Blackwell architecture, is expected to be released in 2024. Microsoft has launched two chips, including the AI computing chip Maia 100 and the general-purpose computing chip Cobalt 100 Arm. OpenAI has improved and tested Maia with GPT-3.5 Turbo.
Looking back on the first anniversary of the release of ChatGPT, what are your biggest feelings?
