The Decoder

30/01/2025

1/ OpenAI has rolled out an updated version of GPT-4o in ChatGPT, extending its knowledge base through June 2024. The refresh aims to provide more current and contextual responses across topics.
2/ The model's visual capabilities have also gotten an upgrade.
3/ It can now better understand spatial relationships in images and make stronger connections between diagrams, graphics, and text.

OpenAI has rolled out an updated version of GPT-4o in ChatGPT, extending its knowledge base through June 2024.

30/01/2025

1/ OpenAI and the US National Laboratories are entering into a partnership in which OpenAI, in collaboration with Microsoft, will install an o-series AI model on the Nvidia Venado supercomputer at Los Alamos National Laboratory (LANL).
2/ The approximately 15,000 scientists at the National Labs plan to use the AI models in a variety of research areas, from medical applications, cybersecurity and critical infrastructure protection to developing new approaches in energy research and deepening the understanding of fundamental mathematics and physics.
3/ A key area of collaboration is the labs' nuclear security program, which aims to reduce the risk of nuclear war and secure nuclear materials and weapons worldwide. OpenAI wants to examine every use of AI models in this area and sees the collaboration as the beginning of a new era in which AI advances science, strengthens national security and supports government initiatives.

OpenAI and Microsoft are bringing their latest AI technology to US national laboratories, with plans to install an o1 or similar o-series model on NVIDIA's Venado supercomputer at Los Alamos National Laboratory (LANL).

30/01/2025

1/ Meta CEO Mark Zuckerberg outlined ambitious plans for the company's AI development in the latest quarterly report. The new Llama 4 language model will have agentic capabilities and secure Meta's leadership in open source.
2/ The Meta AI assistant is expected to reach more than a billion people this year. Meta is focused on a personalized approach that adapts the AI to the context, interests, and culture of the user.
3/ Meta plans to make significant investments in its AI infrastructure, including nearly a gigawatt of computing capacity this year and a two-gigawatt AI data center. By 2025, Zuckerberg expects to have developed an AI engineering agent with the skills of a good mid-level engineer.

Meta CEO Mark Zuckerberg shared insights into the company's AI strategy for 2025 during his latest quarterly report, focusing on the new Llama 4 language model and expanding the Meta AI assistant.

30/01/2025

1/ The Trump administration is considering tightening restrictions on the sale of Nvidia chips to China, especially for the H20 model developed specifically for the Chinese market. The talks are still at an early stage.
2/ Before leaving office, the Biden administration had presented stricter export rules for AI accelerators and AI models, introducing a three-tier licensing system. A complete export ban applies to China. Nvidia fears a negative impact on its revenue in China.
3/ Anthropic CEO Dario Amodei defends the export controls on chips to China to prevent the country from acquiring the same AI capabilities as the US and becoming militarily dominant. He sees the need for such controls confirmed by Deepseek's recent progress.

The US government is discussing further restrictions on Nvidia's chip sales to China, potentially targeting the company's H20 chips - products specifically designed for the Chinese market to comply with existing US trade rules.

30/01/2025

1/ SoftBank is in talks to invest up to $25 billion in OpenAI, potentially matching Microsoft as the AI company's largest investor with a total investment of $40 billion. This would be one of SoftBank founder Masayoshi Son's largest deals yet.

2/ Microsoft has added the AI model R1 from Chinese startup Deepseek to its Azure AI Foundry platform and GitHub. Microsoft plans to offer optimized versions of R1 directly on Windows PCs equipped with Neural Processing Units (NPUs) from Qualcomm and Intel.

3/ Deepseek caused a sharp drop in US financial markets this week, as its R1 model can be trained and run much more cheaply than leading OpenAI models. Nvidia lost about $589 billion in market value in a single day - the biggest loss for any US company in history.

SoftBank is in talks to invest up to $25 billion in OpenAI, potentially matching Microsoft as the AI company's largest investor.

30/01/2025

1/ In a recent test by Newsguard, the Chinese chatbot Deepseek struggled to identify fake news, with an 83% failure rate in recognizing or even actively spreading misinformation when tested without reasoning capabilities and internet access.

2/ Deepseek repeated false claims 30% of the time, placing it in the lower middle range compared to other tested chatbots.

3/ The chatbot frequently echoed the Chinese government's stance without prompting and used "we" to align itself with Beijing's views.

A recent Newsguard test found that the Chinese chatbot Deepseek had trouble handling fake news, failing to recognize or actively spreading misinformation in 83 percent of cases.

29/01/2025

1/ Anthropic CEO Dario Amodei acknowledges that Deepseek's new competitor model performs similarly to seven to ten-month-old US models at a lower cost, but not to the extent some suggest.

2/ Amodei believes Deepseek's true technical innovation lies in their Deepseek-V3 model released in late December, rather than the currently discussed R1 model.

3/ Although Deepseek has reduced AI model development costs, Amodei estimates that the company's GPU reserves are within a factor of 2-3 of major US AI companies, indicating that despite efficiency improvements, overall investments remain substantial.

Anthropic CEO Dario Amodei wants to clear up some misconceptions about Claude 3.5 Sonnet. The AI model cost far less to develop than recent rumors suggest, and it wasn't built using more advanced, secret models as some have claimed.

29/01/2025

1/ Alibaba has released Qwen2.5-Max, a new language model trained on over 20 trillion tokens of data, which the company claims is a record-breaking amount. The model uses a mixture-of-experts (MoE) architecture.
2/ In some benchmark tests, Qwen2.5-Max outperforms leading AI models such as Deepseek-V3, GPT-4o, Claude 3.5 Sonnet, and Llama-3.1-405B.
3/ Users can access Qwen2.5-Max through Alibaba Cloud's API or test it in the Qwen Chat chatbot.

Alibaba has developed a new language model called Qwen2.5-Max that uses what the company says is a record-breaking amount of training data - over 20 trillion tokens.

29/01/2025

1/ Alibaba introduces Qwen2.5-VL, a multimodal visual language model that processes text, images, and videos, with improved handling of diagrams, icons, graphics, and layouts.
2/ Qwen2.5-VL serves as a visual assistant, analyzing screen content and providing instructions for tasks like booking flights and navigating complex interfaces.
3/ The largest version, Qwen2.5-VL-72B, performs on par with leading AI models across benchmarks.

Alibaba has added a multimodal visual language model to its Qwen2.5 series, marking another step in the Chinese tech company's effort to compete in the commercial AI space.

29/01/2025

1/ OpenAI claims Chinese AI startup Deepseek used its models without permission to train competing products, highlighting tensions around AI development practices.

2/ Microsoft and OpenAI discovered signs that Deepseek might have used OpenAI's proprietary models to improve its own alternatives through a technique called "distillation," which violates most providers' terms of service.

3/ Microsoft's security team also noticed unusual patterns last fall and blocked suspicious accounts possibly linked to Deepseek.

OpenAI claims it has found evidence that Chinese AI startup Deepseek used its models without permission to train competing products, highlighting ongoing tensions around AI development practices.

29/01/2025

OpenAI's ChatGPT Pro subscription service is growing faster than its enterprise business, bringing in at least $25 million a month - or $300 million a year - according to The Information, which says this is a conservative estimate.

28/01/2025

OpenAI just launched "ChatGPT Gov," a specialized version of its AI assistant designed specifically for US government agencies.

28/01/2025

1/ A new open source AI music tool called YuE lets anyone turn lyrics into songs, offering a free alternative to commercial services like Suno and Udio.
2/ According to developer Ruibin Yuan, the system can create songs up to five minutes long in multiple languages and musical styles.

A new open source AI music tool called YuE lets anyone turn lyrics into songs, offering a free alternative to commercial services like Suno and Udio.

28/01/2025

1/ Deepseek has launched Janus Pro, a major upgrade to its multimodal AI system, featuring improved training methods, expanded datasets, and larger model sizes.
2/ The new version incorporates 90 million additional examples for multimodal understanding from various sources, as well as 72 million synthetic training examples for image generation, bringing the ratio of real to synthetic data to 1:1.
3/ Janus Pro introduces a larger 7B model size, which outperforms its predecessor in both understanding and generating images.

Deepseek has completely overhauled its multimodal AI system Janus. The new version, Janus Pro, improves on its predecessor through refined training methods, expanded datasets, and larger model sizes.

28/01/2025

"We will obviously deliver much better models."

Meta has established multiple emergency response teams after Chinese AI company Deepseek demonstrated AI models that are both more efficient and significantly cheaper to operate than Western alternatives.

27/01/2025

1/ Alibaba has introduced two new open-source language models, Qwen2.5-7B-Instruct-1M and Qwen2.5-14B-Instruct-1M, which can handle context windows of up to one million tokens, a significant increase compared to previous models.

2/ The models showed promising results on various benchmarks for tasks with long contexts, especially when dealing with sequences longer than 64,000 tokens. However, the practical applications of such large context windows remain a subject of debate.

3/ Alibaba's release of these open-source models, along with other advances in AI from China, is putting competitive pressure on established AI providers in the United States.

Alibaba's Qwen team just added two new members to its Qwen2.5 family: Qwen2.5-7B-Instruct-1M and Qwen2.5-14B-Instruct-1M.

27/01/2025

1/ Researchers from the Chinese University of Hong Kong, Shenzhen, Alibaba's Qwen, and the Shenzhen Research Institute of Big Data have discovered that OpenAI's o1-mini model improves its performance through self-criticism, unlike most AI systems that deteriorate when attempting to correct their own errors.
2/ The researchers developed a new testing method called RealCritic, which ensures that AI models can not only identify their mistakes but also rectify them effectively.
3/ While most models struggled with self-criticism, o1-mini showed consistent improvement. When critiquing each other's work, all models showed improvement, with o1-mini leading the way.

Researchers at the Chinese University of Hong Kong, Shenzhen, along with teams from Alibaba's Qwen and the Shenzhen Research Institute of Big Data, have found something interesting about OpenAI's o1-mini model. While most AI systems get worse when trying to fix their own mistakes, o1-mini usually im...

27/01/2025

1/ Perplexity AI has proposed merging with TikTok's US operations, creating a new holding company called "NewCo" that could be up to 50% government-owned after an IPO. The deal values the company at a minimum of $300 billion.
2/ Under the plan, ByteDance would contribute TikTok's US business without its recommendation algorithm. Existing investors could maintain their stakes or cash out. Perplexity, recently valued at $9 billion, would gain access to TikTok's vast video content.
3/ The proposal comes as TikTok faces increasing scrutiny over its future in the US. Perplexity believes structuring the deal as a merger rather than a sale may be more palatable to ByteDance. Other tech companies like Microsoft have also expressed interest in TikTok's US presence.

Perplexity AI, the company behind an AI-powered search engine, has put forward a plan to merge with TikTok's US operations. The proposal comes as TikTok faces mounting pressure over its future in the United States.