- The 79
- Posts
- New AI model can take an input of 3 million words!
New AI model can take an input of 3 million words!
Hello! Here’s what you need to know about AI today:
👉 New Chinese AI model beats its American rivals
👉 ChatGPT sometimes thinks in Chinese and Persian
👉 Meta’s new AI model can translate speech across 100+ languages
and many more!
📧 Did someone forward you this email? Subscribe here for free to get the latest AI news everyday!
Read time: 5 minutes
CHINESE AI
MiniMax just shipped 3 competitive AI models
Source: MiniMax
What’s going on: MiniMax, a Chinese AI startup backed by Alibaba and Tencent worth $2.5 billion, launched 3 new models: MiniMax-Text-01, MiniMax-VL-01, and T2A-01-HD. The first one is a text-only model, while MiniMax-VL-01 can understand both images and text. The third model can generate audio and speech. MiniMax claims that MiniMax-Text-01, which has 456 billion parameters, performs better than Google’s Gemini 2.0 Flash on a couple of benchmarks.
What does it mean: China is trying hard to be at the top of the AI game. If MiniMax's models are really as good as they claim, it could mean more competition for big American AI companies. Minimax is really challenging rivals like Gemini and ChatGPT with providing a context window as big as 38 standard novels! (You can ask MiniMax-01 a question and paste a passage with a length as long as 3 million words).
More details:
MiniMax-Text-01 has an extremely large context window (input size). It has a context window of 4 million tokens.
MiniMax-Text-01’s context window is 31 times the size of GPT-4o’s!
T2A-01-HD can generate a synthetic voice with adjustable cadence, tone, and tenor in around 17 different languages, including English and Chinese, and clone a voice from just 10 seconds of an audio recording.
MiniMax’s new models are open-source and can be downloaded from GitHub and Hugging Face. However, T2A-01-HD is only available through MiniMax’s API and Hailuo AI platform.
CHATGPT
OpenAI's o1 reasoning model thinks in Chinese and no one has a clue why
Source: X
What’s going on: OpenAI's o1 has demonstrated an unusual behavior where it sometimes "thinks" in languages like Chinese or Persian, even when questions are posed in English. This phenomenon was observed shortly after o1's release, with the model switching languages during its reasoning process before reverting to English for the final answer. The exact reason for this behavior remains unclear as OpenAI has not provided an explanation.
What does it mean: We don’t actually know! But several theories have been proposed by AI experts. Some suggest that o1's training data, which likely includes a significant amount of Chinese characters could be the reason. Another perspective is that the model might find certain languages more efficient for solving specific types of problems. The last theory is that AI models like o1 operate on probabilities, and their reasoning might not follow human-like linguistic patterns but rather use whatever patterns were most effective in their training data.
More details:
Luca Soldaini, a research scientist from the “Allen Institute for AI”, emphasized the difficulty in verifying these observations due to the blackbox nature of AI models.
Performing some tasks like basic mathematic operations is quicker in Chinese because each digit is just one syllable, which makes calculations efficient.
AI models don’t process words directly. They use and process tokens. Tokens can be words, such as “barber” or they can be syllables, like “bar” and “ber” or they can even be individual characters in words like “b,” “a,” “r,” “b,” “e,” “r,”.
Matthew Guzdial, an AI researcher at the University of Alberta, claims that AI models don’t know what language is, or that languages are different, and they look at all of the languages in the same way.
💎📰 Google has partnered with AP (The Associated Press) to bring a real-time news feeds into Gemini, to enhance its information delivery.
🤖 A recent survey indicates that more teenagers are using ChatGPT for schoolwork, despite its known limitations (errors and hallucinations), with 26% of US teens reporting usage, double from two years ago.
💬 Meta has introduced a new AI model, SeamlessM4T, capable of translating speech across more than 100 languages, aiming towards real-time, simultaneous interpretation.
📈 As the number of Gen Z job applicants increases, companies are increasingly relying on AI agent recruiters (like Maki which recently raised $28.6 million) to manage the hiring process.
⚡ Meta has expanded its renewable energy portfolio by adding 200 megawatts of solar power from Engie, bringing its total to over 12 gigawatts to further pursue AI ambitions.
📞 OpenAI is testing a new feature allowing users in the US and India to sign up for ChatGPT using only a phone number, without an email.
🚀 Microsoft is launching “Microsoft 365 Copilot Chat” with agentic capabilities for businesses to enhance productivity by automating routine tasks.
🎓 Keele University researchers have developed a new AI tool that can detect fake news with an impressive 99% level of accuracy. (Read the paper)
🔗 LinkedIn's new AI tool, Job Match, helps job seekers by showing how well their profiles match job openings, making the job search more efficient.
🐍 Scientists are taking advantage of advancements in AI-driven protein design to create effective snake antivenoms.
Email marketing with AI
I'm looking for a [type of email] that will showcase the unique features and benefits of my [product/service] to [ideal customer persona] and persuade them to make a purchase. Please write the email for me.
Gemini’s answer
Together AI - Director, Machine Learning
NewsBreak - Machine Learning Engineer, Ad platform
Thank you for staying with us like always! If you are not subscribed, subscribe here for free to get more of these emails in your inbox! Cheers!