The Decoder

09/11/2024

1/ According to Similarweb, ChatGPT set a new record of 3.7 billion global visits in October, up 17.2 percent month-over-month and 115.9 percent year-over-year.

2/ Google's NotebookLM grew more than 200 percent to 31.5 million visits in October, driven by a viral AI podcast feature in which two synthetic voices discuss document content.

3/ Other offerings also grew, but only moderately: Microsoft's Copilot by 87.6 percent to 69.4 million visits, Perplexity and Claude by 25.5 percent each. Google's Gemini site reached 291.6 million visits, an increase of 6.2 percent.ebookLM

ChatGPT continues its strong growth, while a new podcast feature sends Google's NotebookLM usage soaring. AI services across the industry are seeing steady user growth, albeit at a measured pace.

09/11/2024

1/ Google has launched Vids, a new AI-powered video creation service, which is now generally available to Workspace customers.

2/ The built-in AI assists users with technical and creative aspects of video creation, such as generating scripts, designing transitions, and eliminating filler words.

3/ Vids provides various recording options, including selfie videos, screen recordings with voice commentary, and audio-only recordings, to help companies create professional multimedia videos for different purposes.

Google is adding a new AI-powered video creation service called Vids to its Workspace platform. The tool helps businesses create multimedia presentations using their existing content and data.

09/11/2024

1/ OpenAI redirected two million ChatGPT queries about the U.S. election results to verified news sources such as Associated Press and Reuters instead of providing direct answers, because ChatGPT could provide misinformation if it did.

2/ In the month leading up to the election, OpenAI also redirected approximately one million ChatGPT queries about the voting process to the official CanIVote.org website and blocked more than 250,000 attempts to generate DALL-E images of politicians such as President Biden, Donald Trump, and Vice President Harris.

3/ While OpenAI's caution with election information shows an awareness of potential misinformation risks, the company has been less transparent about ChatGPT's accuracy in general searches. But spreading misinformation is not just an election issue.

OpenAI sent two million ChatGPT requests about US election results to verified news sources rather than letting its AI system answer directly. This decision raises broader questions about AI systems' ability to provide reliable information across all topics.

09/11/2024

Suno has released the first audio samples of its upcoming v4 AI music generation model. The demos show improvements in natural sound, variety, and consistency compared to previous versions.

08/11/2024

1/ Amazon is negotiating a second billion-dollar investment with AI company Anthropic. The condition: Anthropic must increasingly use Amazon's Trainium chips instead of Nvidia hardware for AI training.
2/ The planned investment is similar to the first $4 billion deal from last year. The size of the new investment will depend directly on the number of Amazon chips used.
3/ There are technical hurdles to switching to Amazon chips: Trainium's software is less mature than Nvidia's CUDA platform. The deal would also tie Anthropic more closely to the Amazon ecosystem, as the hardware is not available to other vendors.

Amazon is negotiating a second billion-dollar investment in AI company Anthropic. This time, however, Amazon is setting conditions: Anthropic should increase its use of Amazon's AI chips instead of Nvidia's.

08/11/2024

1/ AMD has released its first open source language model with one billion parameters. It is based on the OLMo architecture, but was trained with less than half the number of training tokens and still achieves comparable performance.
2/ The three stages of training included a baseline model, two stages of supervised fine-tuning for specific skills, and adaptation to human preferences. In benchmarks, AMD's OLMo outperformed other open source chat models by an average of 2.6 percent.
3/ The release of OLMo is part of AMD's broader AI strategy, which includes investments in AI companies, acquisitions, and the development of specialized AI hardware such as the Instinct MI355X accelerator planned for 2025, with which AMD intends to compete directly with Nvidia.

AMD has released its first open-source language model with one billion parameters. The model builds on a previous version but uses significantly less training data.

08/11/2024

1/ A federal judge in New York has dismissed a lawsuit filed by news sites Raw Story and AlterNet against OpenAI, ruling that the plaintiffs failed to demonstrate sufficient harm from the use of their articles as training data for AI.

2/ The judge agreed with OpenAI's fair use argument, stating that ChatGPT synthesizes AI responses based on the training content, and the likelihood of an exact copy of an article is low. If it does occur, it is considered an error rather than a feature, and facts are not protected by copyright.

3/ This decision may influence other ongoing lawsuits against AI companies, such as the legal dispute between the New York Times and OpenAI, as well as lawsuits filed by music companies against AI music generators.

A New York federal judge has dismissed a lawsuit brought by news sites Raw Story and AlterNet against OpenAI for using their articles to train AI systems. The ruling could affect similar ongoing cases.

07/11/2024

1/ Microsoft is integrating its Copilot Pro AI assistant into Microsoft 365 Personal and Family subscriptions, starting with select countries in Asia and Oceania.

2/ Users will receive a monthly allocation of AI credits, and those who need more will need to upgrade to a separate Copilot Pro subscription. Along with the integration of Copilot, Microsoft is also raising prices for Microsoft 365.

3/ The inclusion of Copilot Pro in the standard bundles may indicate that separate marketing is not meeting expectations. Microsoft has not provided specific sales figures for Copilot.

Microsoft integrates its Copilot Pro AI assistant into Microsoft 365 subscriptions, suggesting a potential shift in strategy for the standalone AI service.

07/11/2024

1/ Anthropic, AWS, and Palantir have partnered to provide US intelligence agencies with access to Anthropic's Claude AI models through Palantir's AI Platform. The integration takes place in a highly secure IL6 environment for sensitive government data.
2/ Anthropic's terms of service explicitly allow the use of its products for "legally authorized foreign intelligence analysis," "identifying covert influence or sabotage campaigns," and "providing warning in advance of potential military activities."
3/ The collaboration is part of a wider trend, with other AI companies such as Meta and OpenAI also seeking to work with the defense sector. Anthropic is currently negotiating a new round of funding with a potential valuation of $40 billion.

A new partnership between Anthropic, Amazon Web Services (AWS), and Palantir gives US intelligence agencies access to Anthropic's AI models - another step in the growing connection between AI companies and the defense sector.

07/11/2024

1/ A study by German researchers found that 0.052 percent of profiles on X use AI-generated profile pictures created with StyleGAN 2, with a likely much higher number of unreported cases.

2/ An analysis of 1,000 accounts with AI-generated profile pictures revealed thematic focuses on politics (often related to Trump), Covid-19 and vaccinations, the war in Ukraine, lotteries, and financial topics such as cryptocurrencies.

3/ The researchers say the mass creation of these accounts and their identical metadata point to organized networks aimed at spreading disinformation and political propaganda.

A new study shows that accounts on X (formerly Twitter) that use AI-generated profile pictures make up about 0.052 percent of profiles and often share content about politics, COVID-19, and the war in Ukraine.

07/11/2024

OpenAI has acquired the domain Chat.com and is redirecting it to . Chat.com is one of the oldest domains on the web and was registered in September 1996.

OpenAI has acquired the domain Chat.com and is redirecting it to ChatGPT. Chat.com is one of the oldest domains on the web and was registered in September 1996.

06/11/2024

Black Forest Labs has added two new modes to its FLUX1.1 [pro] AI model: Ultra for high-resolution images up to 4K and RAW for a more natural snapshot-style look.

German AI startup Black Forest Labs has released its latest image generation model, Flux 1.1 Pro. The company claims the new model is six times faster than its predecessor while also enhancing image quality.

06/11/2024

1/ The KLING AI video generator introduces 'Custom Models', a feature that allows users to train their own video characters. Users upload between 10 and 30 videos of at least 10 seconds in length.
2/ The system uses video footage instead of images for training and can consistently display the trained characters in different scenes - from 360-degree rotations to complex shots with different camera angles.
3/ Developed by Chinese company Kuaishou, KLING version 1.5 offers 1080p HD video, multiple image formats and a 'Motion Brush' for precise motion control.

KLING, an AI video generator, is testing a new feature called "Custom Models" that significantly improves character consistency in generated videos.

06/11/2024

1/ OpenAI introduces Predicted Outputs, a feature for the GPT-4o and GPT-4o-mini that accelerates AI-powered text processing. Response times are two to four times faster than existing models, with large file processing now taking around 20 seconds instead of 70.
2/ The system allows developers to pre-populate an expected portion of the output. Saving 50 percent of output tokens reduces latency by around 50 percent. Tests in Python, JavaScript, Go and C++ have been successful.
3/ The feature is particularly suitable for repetitive tasks and minor document changes, but less so for completely new content. It does not support advanced API parameters such as multiple outputs or function calls, and is only available for GPT-4o and GPT-4o-mini.

OpenAI has introduced a new feature called "Predicted Outputs" for its GPT-4o and GPT-4o-mini language models that makes AI-supported text processing much faster than before.

06/11/2024

1/ Seattle-based start-up Moondream has released moondream2, a compact vision language model (VLM) with just 1.6 billion parameters that, despite its small size, can compete with much larger models in benchmarks.
2/ moondream2 accepts not only text but also images as input and can answer questions, extract text (OCR), count or classify things based on it. It is available as open source on GitHub and has caused quite a stir in the developer community.
3/ Moondream recently raised $4.5 million in a pre-seed funding round. The model joins the ranks of specialised and optimised open source models that deliver similar performance to larger models while using fewer resources, which is particularly relevant for use on smartphones.

Moondream, a startup based in Seattle, has released moondream2, a compact vision language model that performs well in benchmarks despite its small size. The open-source model could pave the way for local image recognition on smartphones.

05/11/2024

1/ Meta is making its open-source Llama models available to US government agencies and national security contractors, including Amazon Web Services, Microsoft, IBM, Lockheed Martin, Palantir, and Oracle.

2/ The partner companies are integrating Llama into existing services, such as analyzing aircraft maintenance documents, mission planning, and identifying enemy vulnerabilities.

3/ Meta is positioning itself as a patriotic tech company and argues that open-source AI models are crucial to maintaining the USA's technological leadership over China.

Meta announced it will provide its open-source Llama AI models to US government agencies and defense contractors, partnering with major tech and defense companies to integrate the technology into government services.

05/11/2024

1/ In a new white paper, Microsoft warns of the growing misuse of AI systems for fraud, child abuse and disinformation. The company proposes a six-pillar approach that combines technology, legislation, and stakeholder collaboration.
2/ The company is calling on UK politicians to take concrete action: Providers should be required to label AI-generated content, new laws should protect against electoral manipulation, and public-private partnerships should support victims.
3/ Microsoft is focusing on technical solutions such as Content Credentials to label AI content and is working with organizations such as StopNCII.org. Victims can use a central reporting portal, and young people can use the "Take It Down" service.

Microsoft has released a new white paper analyzing how generative AI is increasingly being misused for fraud, child sexual abuse material, election manipulation, and non-consensual imagery. The tech company outlines a comprehensive approach combining technology, collaboration, and legislation to add...

05/11/2024

1/ A Chinese research team has compiled the multimodal data set Infinity-MM with 40 million image-text pairs. This includes image descriptions, visual instruction data and synthetically generated data from AI models such as GPT-4o.
2/ The researchers used this data set to train the Aquila-VL-2B model, which, despite its comparatively small size of 2 billion parameters, achieved top performance in various benchmarks, for example in mathematical tasks or general image comprehension.
3/ The four-stage training with increasing complexity and the integration of synthetically generated data proved to be decisive for performance. Open source models thus also excel in vision tasks.

A team of researchers from various Chinese institutions has created Infinity-MM, one of the largest publicly available datasets for multimodal AI models, and trained a new top-performing model on it.