AI Week In Review 24.11.09
Hunyuan-Large, new Claude Haiku 3.5, Mistral batch API, Meta's MobileLLM, SmolLM2, Runway Gen3 Alpha Turbo Advanced Camera Control, FLUX 1.1 [pro] Ultra and Raw, Hertz-dev. Trump's AI policy.
AI Tech and Product Releases
Tencent has released Hunyuan Large, an MoE (mixture-of-experts) model with 52B active and 389B total parameters. That makes Hunyuan-Large (Hunyuan-MoE-A52B) the largest open-source Transformer-based MoE model in the industry. This landmark SOTA LLM was trained on 7.5T tokens, including 2T tokens of synthetic data. Hunyuan-Large-Instruct features impressive performance in math (77 on MATH), coding (90 on HumanEval), and general knowledge (89.9 on MMLU).
Hunyuan-Large is available to run via an API as well as open source download, and you can try it out on HuggingFace spaces.
Anthropic Claude Haiku 3.5 is now available via API, but there was a severe backlash when users realized the Haiku 3.5 API price was raised by 4x, making it far more expensive than its predecessor version and its competitors Gemini 1.5 Flash and GPT-4o mini. Anthropic justifies it on the basis that this model is more capable, but we’ve been used to AI model prices coming down.
Mistral launched a new API for content moderation. The API uses a fine-tuned model (Ministral 8B) to classify text into nine categories of harmful content, including hate speech and personally identifiable information, across multiple languages.
Mistral also launched a Batch API to offer a lower cost method to use their LLMs. Mistral will process high-volume batch API requests to Mistral models at 50% lower cost than that of a synchronous API call.
Meta open sourced MobileLLM, their series of small LLMs sized at 125M, 350M, 600M, and 1B parameters. It’s available on HuggingFace.
HuggingFace introduced SmolLM2, version 2 of their SmolLM models, in 135M, 360M and 1.7B parameter sizes.
Infosys released Topaz BankingSLM and Infosys Topaz ITOpsSLM for businesses, banking, and IT operations.
Runway’s Gen-3 Alpha Turbo now has Advanced Camera Control, with dramatic zoom, panning, truck across scenes, and more:
“Choose both the direction and intensity of how you move through your scenes for even more intention in every shot.”
FLUX 1.1 [pro] now has Ultra and Raw modes, producing HD images. Ultra mode enables image generation at four times the resolution of standard FLUX1.1 [pro]. Raw mode captures the genuine feel of candid photography, with a less synthetic, more natural aesthetic.
Microsoft announced that Outlook users can now create personalized themes using generative AI through Copilot. The feature allows users to customize email aesthetics based on location and style preferences.
Apple has released iOS 18.2 to public beta. The update introduces AI-enabled features like Genmoji for AI emoji generation, Image Playground for image creation, ChatGPT integration with Siri, and visual search capabilities.
OpenAI made the o1 full model available last week briefly before pulling it down, but it got jailbroken in the interim. The o1 full model got positive feedback from users, “This model feels significantly better for coding than previous versions.”
Top Tools & Hacks
Standard Intelligence has introduced hertz-dev, an open-source conversational audio model. Hertz-dev is an 8.5B parameter model that takes in audio and can output audio with a real-world average latency of 120ms, 2 times lower than any other audio model. The SI startup team has shared demos and open-sourced checkpoints of their audio-only base model.
Grok 2 API is now in public beta. To encourage developers to use their new API, X.ai is offering $25/month free of Grok API credits. The grok-beta model available via API has a context length of 128,000 tokens and supports function calling and system prompts.
AI Research News
Tencent AI researchers published a technical report on their new Hunyuan-Large AI model: “Hunyuan-Large: An Open-Source MoE Model with 52 Billion Activated Parameters by Tencent.” How they trained this MoE model to get superior results:
High-Quality Synthetic Data: They used 1.5T tokens of carefully crafted synthetic data to train this model, as part of their 7T token training set.
KV Cache Compression: They utilized Grouped Query Attention (GQA) and Cross-Layer Attention (CLA) strategies to significantly reduce memory usage and computational overhead of KV caches, improving inference throughput.
Expert-Specific Learning Rate Scaling: They set different learning rates for different experts to ensure each sub-model effectively learns from the data and contributes to overall performance.
AI Business and Policy
OpenAI bought Chat.com, redirecting it to its chatbot, ChatGPT. The domain was previously acquired by HubSpot co-founder Dharmesh Shah for $15.5 million before being sold to OpenAI, for shares rather than cash.
Anthropic teamed up with Palantir and AWS to sell AI to defense customers. The collaboration will allow U.S. intelligence and defense agencies access to Anthropic’s Claude family of AI models, enhancing data analysis capabilities within secure environments.
Scale AI has announced Defense Llama, a version of Meta’s Llama models fine-tuned on military doctrine and intelligence operations, aimed at U.S. government customers.
OpenAI loses another lead safety researcher, Lilian Weng. Weng served as OpenAI’s VP of research and safety since August and was with OpenAI for seven years. Her exit, announced on X, continues a trend of AI safety experts leaving OpenAI, raising concerns over OpenAI’s commitment to AI safety.
Anysphere, the maker of AI-powered coding assistant Cursor, has attracted significant interest from VCs, with unsolicited offers valuing the company at up to $2.5 billion, a substantial increase from its $400 million valuation four months ago. Cursor has had surging popularity and rapid growth, with its revenue increasing from $4 million annualized recurring revenue (ARR) in April to $4 million a month.
The Beatles AI-Enhanced Song “Now and Then” has been nominated for Grammys.
Exiger Alum Launches Conflixis to Tackle Corruption Risks in Healthcare with AI. Conflixis aims to help hospitals manage legal compliance and reduce operational costs, securing a $4.2 million seed round in the process. The startup will use AI to identify corruption risks in healthcare by analyzing data from various sources including OpenPaymentsData.com.
UnifyApps Raises $20 million Series A for Its AI Chatbot Platform. UnifyApps has secured significant funding and notable clients by offering a solution that integrates company data with SaaS apps to build reliable AI chatbots.
Amazon is considering increasing its investment in OpenAI rival Anthropic. The potential new multi-billion-dollar investment would require Anthropic to use Amazon-developed silicon on AWS for AI training, a condition that differs from their current preference for Nvidia chips. This comes as Anthropic faces financial pressures and aims to secure additional funding at a high valuation of around $40 billion.
Coatue Management is raising $1 billion to invest in AI-focused companies, with funding primarily from institutional investors. Coatue, managing $50 billion in assets, has already invested in several AI startups like Glean and Skild AI.
Nvidia exceeds Apple’s market capitalization to become the world’s largest company, driven by a global AI push. Nvidia is now valued at $3.43 trillion, surpassing Apple’s $3.38 trillion; this marks the second time Nvidia has overtaken Apple in market value.
Google is constructing a new AI-focused data center in Saudi Arabia. The hub aims to support research into Arab language AI models and unspecified “Saudi-specific AI applications.”
The other election winners: Perplexity's AI chatbot provided accurate real-time election insights during the high-stakes presidential election, with integrated live data from reputable sources like Democracy Works and the Associated Press. During and after Election Day, ChatGPT directed around 2 million users to trusted news sources and sent about a million people to CanIVote.org for voting information.
AI Opinions and Articles – Election Special
The biggest news this week of course was the election of Donald Trump to be our next US President. A Trump second term shakes up the Federal regulation and significantly reshapes the landscape for AI governance. What to expect out of Trump’s AI policies:
Trump said he will undo President Biden's 2023 AI executive order that established the American Innovation and Security Initiative (AISI) and initiated some AI regulations.
Trump will move towards a light-touch regulatory regime on AI and tech in general. Trump’s speeches and the GOP Platform have stated: Republicans support AI Development rooted in Free Speech and Human Flourishing.
Bipartisan AI bills with limited scope around shared concerns (like copyright protection) are likely to pass, but comprehensive AI regulations will not pass Congress.
Vice President-elect JD Vance and Tech leaders supporting Trump, including Elon Musk, have been supportive of open-source AI.
The Trump administration will want to be America First in AI and will pursue a “Make America First in AI” policy with Manhattan-type projects to develop AI for national security.
Protectionist policies, tariffs and tighter export controls on China could impact AI technologies and their development in some countries, with potential geopolitical ramifications.
There is not likely to be global cooperation in regulating AI.
This policy is a green light for open-source AI and for American tech companies to keep moving forward on AI towards AGI. AI optimists will be happier than AI doomers over this ‘light-touch’ pro-AI-development policy.
There are undoubtedly risks related to AI. One of the biggest: A partisan group of crazy people use AI to infect every part of the information economy with left wing bias. Gemini can’t produce accurate history. ChatGPT promotes genocidal concepts. The solution is open source. – JD Vance, March 2024