Black Swans have started to be more and more common in these times. January witnessed the introduction of DeepSeek, a Chinese company that published the DeepSeek R1 model as a web app, mobile app, and open source code.
The impact of this event was truly earthshaking, with western tech companies losing their stock worth and users testing and comparing the models.
But it doesn’t mean that DeepSeek was the only interesting event of January.
January Special: the DeepSeek Drama
January saw the sudden premiere of DeepSeek’s R1 model, delivered by a Chinese company of the same name. The model is claimed to have been trained for $6 million dollars (compared to the $100 million required to train GPT-4) and requires only one-tenth of the electricity to run.
Contrary to OpenAI and Anthropic models, DeepSeek was not only released as an online service and on mobile apps, but also in an open-source version, distributed on the MIT license. As such, anyone can use the model for scientific or commercial purposes. And any company can run it internally for any purpose, or enrich their products with DeepSeeks capabilities, with no need to pay for API tokens.
The arrival of the model caused disruption on the markets, especially in the high-tech segment. Nvidia alone, arguably the biggest beneficiary of the AI revolution, lost 17% of its stock value – roughly $600 billion. And this is only one of the victims.
DeepSeek has also challenged the paradigm of super-costly AI development, raising questions if the western approach to AI training is, indeed, a good path to follow.
The model has also raised doubts regarding security. The data of the people using the web app or mobile app are stored outside of EU and US jurisdictions, raising many questions. The model has also shown that it is censored, refusing to talk about matters considered sensitive by the Chinese government.
3 Jan Meta uses Generative AI to understand user intent
In two recently published papers, Meta, the company behind Facebook, has revealed how it uses generative models to better understand user intent. The papers show that the models are better at understanding context and more efficient than classic approaches.
More can be found in VentureBeat.
3 Jan Meta shuts down its own Facebook and Instagram bots
The company launched AI-powered personas on Facebook and Instagram in September 2023. These bots published some content and interacted with human users, all with their AI capabilities. The company decided to delete these bots because users began to rediscover them, and screenshots of their interactions quickly went viral.
More can be found in The Guardian.
6 Jan Sam Altman: “Virtual Employees” may join workforce this year
According to the CEO of OpenAI, more powerful AI agents will be able to join companies and the workforce later this year. The change is made possible due to new features of the models, including the ability to move a cursor, type, or interact with an internet browser. According to McKinsey consulting firm, up to 30% of all work activities done today will be automated by 2030.
More can be found in The Guardian.
9 Jan VLC Media Player adds automatic AI-generated subtitles
The system can now generate subtitles in more than 100 languages, all offline. The feature does not need to be connected with the internet, with all the AI work done locally on the machine.
More can be found in The Verge.
9 Jan Google can now turn Discover feed into podcast
Daily Listen, the new feature from Google, creates an up to 5 minute long audio summary of “Discover.” The feature is comparable to the NotebookML Audio Overview feature.
More can be found in The Verge.
14 Jan OpenAI adds reminders and tasks to ChatGPT
The new features of the platform let users add some tasks as well as schedule future actions and reminders. This may include an everyday morning weather summary or a reminder about bills to pay.
More can be found in the Verge.
15 Jan Tech companies pay creators for unused footage
Google, OpenAI, and Moonvalley, among others, are paying content creators from YouTube, Instagram and TikTok up to $4 per minute of unused video content to acquire data for training their video models.
More can be found in artificialintelligence-news.
21 Jan Tencent Unveils Hunyuan3D 2.0
The model unveiled by a Chinese tech company lets users create detailed 3D objects from text. For a human artist, this process tends to take days, if not weeks. The model delivers a 3D shape within seconds. The system contains two components – the Hunyuan3D-DiT model that turns the text into a 3D object, and HunYuan3D-Paint, which adds surface details and colors.
More can be found in VentureBeat.
22 Jan ByteDance shows UI-TARS outperforming GPT-4o
ByteDance, the company behind TikTok, releases an agent that can take control of a computer, use a mouse pointer and internet browser to interact with online services. The agent understands the Graphical User Interface and can take step-by-step actions as one would expect from an AI agent.
The agent was described in a Research Paper co-authored by researchers from Tsinghua University and ByteDance.
22 Jan The announcement of Project Stargate
The project is a joint venture between OpenAI, Oracle, and Softbank, and aims to deliver the computing and storage resources for new AI development initiatives. The project aims to build a network of powerful data centers throughout the US.
More can be found in The Guardian.
27 Jan DeepSeek targeted with malicious attacks
Due to malicious attacks on DeepSeek’s app, the company had to temporarily limit the registration of new accounts to owners of Chinese telephone numbers.
More can be found in The Verge.
28 Jan $11 trillion disappeared from US stocks
The arrival of the Chinese DeepSeek model heavily impacted the US stock market, with significant drops in the value of leading tech companies. For example, Nvidia shares dropped by 17%, resulting in a $600 billion loss of value for the company.
More can be found in Reuters.
28 Jan Chinese chatbot DeepSeek censors itself in real-time
According to user reports, the model shares its reasoning with the user, yet after realizing that the facts to be shown do not align with official Chinese messaging, the chatbot quickly deletes the answer.
More can be found in The Guardian.
28 Jan OpenAI launches ChatGPT for government
The ChatGPT Gov version is more secure and robust, so government agencies may access it and use it in their daily tasks. The platform will give access to the latest models. Government instances may launch, access, and manage ChatGPT Gov using their own Azure Cloud instances only, making both the data and security management easier.
More can be found in The Verge.
29 Jan DeepSeek blocked from app stores in Italy
Due to growing concerns regarding data privacy in DeepSeek’s ecosystem, app stores in Italy blocked the mobile app of the company. The app disappeared from the App Store and Google Play, with users seeing information that “the app download is ‘not supported’ or ‘not available’ in the country or area you are in.”
More can be found in The Guardian.
29 Jan Microsoft makes DeepSeek’s R1 model available on Azure AI
The Chinese open source model is now available in a model catalog in the Azure Cloud service. With that, Microsoft users may now integrate the model into their services and AI-powered workflows.
More can be found in The Verge.
29 Jan Microsoft and OpenAI accuse DeepSeek of data theft
Microsoft identified suspicious activity that suggests large-scale data extraction using the OpenAI API. Sources cited by the media believe that users may have exploited usage policy by extracting large amounts of data.
More can be found in artificialintelligence-news.
30 Jan Authors Guild will give its own certification for non AI-written books
The Authors Guild, one of the largest US associations of writers, is going to launch its own certification program for books that were written without using AI. According to the guild, this certification will give readers reliable information on who (or sometimes – what) created the book they are reading.
More can be found in The Verge.
