New releases, benchmarks, pricing and API changes from GPT, Claude, Gemini, DeepSeek, Qwen and other large language models, updated daily.
Stability AI just released an experimental version of Stable LM 3B, a compact 3-billion-parameter language model tailored to run locally on laptops and handhelds. Despite its small footprint, it outperforms prior 3B models and even beats some 7B open-source models on common reasoning benchmarks. The base model is open-sourced under CC-By-SA 4.0 on Hugging Face, with an instruction-tuned variant coming soon.
OpenAI just added voice conversations and image analysis to ChatGPT. Users can now snap photos to troubleshoot broken grills or plan dinner recipes, and chat back and forth using five realistic synthetic voices. The multimodal upgrade rolls out to Plus and Enterprise users over the next two weeks.
OpenAI is reportedly rushing to roll out multimodal image capabilities to GPT-4, aiming to preempt Google's upcoming Gemini model. The update, dubbed GPT-Vision alongside a next-gen model codenamed Gobi, will finally make visual inputs like sketches and charts widely available. Why it matters: Google was primed to steal the AI spotlight this fall, and OpenAI is moving aggressively to spoil those plans.
Meta is developing a powerful new AI model designed to match OpenAI’s most capable system, according to a Wall Street Journal report. Expected to vastly outperform Llama 2, training is slated to start in early 2024 to help businesses build complex text generation and analysis tools. Meta has not commented.
Stability AI just rolled out Japanese InstructBLIP Alpha, its first Japanese vision-language model. Built on top of its 7B Japanese LLM and the InstructBLIP architecture, it handles image captioning and visual QA while recognizing Japan-specific landmarks like Tokyo Skytree. It is currently available on Hugging Face for research purposes.
Baidu just officially launched its Ernie Bot chatbot to the public following regulatory clearance in China, racking up over 1 million users within its first 19 hours. The ChatGPT rival handles text analysis, summaries, and creative brainstorming, alongside an ecosystem of new plug-ins. While available globally in app stores, users still need a mainland Chinese phone number to register.
OpenAI just launched fine-tuning for GPT-3.5 Turbo, allowing developers to customize models with their own proprietary data. Early testers managed to shrink prompt sizes by up to 90% by baking instructions directly into the model, speeding up API calls while slashing operational costs. Why it matters: Developers can now achieve consistent, tailor-made outputs without blowing their token budget.
OpenAI just detailed a new system using GPT-4 to handle content moderation and lighten human workloads. Policy experts write rules into prompts, compare the model's flags against human labels, and refine definitions on the fly. OpenAI claims it cuts policy rollout time down to hours, though human oversight is still strictly required to catch model bias.
OpenAI just rolled out its custom instructions feature to all users, including the free tier on web, iOS, and Android. It lets you set response preferences, tone, and role context once so you don't repeat yourself in every prompt. Steerability is officially becoming a default expectation rather than a paid perk.
Anthropic just launched Claude Instant 1.2, filtering the capabilities of its flagship Claude 2 down into its faster, cheaper text model. It gains a 100K token context window, handles formatting and safety better with reduced hallucinations, and bumps coding benchmark scores from 52.8% to 58.7% via the API.
Stability AI just launched Stable Code Alpha 3B, its first generative AI model tailored for coding. Trained on 560B tokens of code, the suite includes base, instruction, and long-context variants with a 16,000-token window capable of reviewing multiple files at once. Why it matters: It gives developers a lightweight, accessible open-source assistant for everyday programming.
Amazon just launched Agents for Bedrock in preview, allowing enterprises to build conversational AI that connects to internal systems and executes tasks via API calls. Alongside the agent rollout, AWS added new third-party foundation models to the service, including Anthropic's Claude 2, Stability AI's Stable Diffusion XL 1.0, and Cohere's enterprise models. Why it matters: AWS is turning Bedrock into an enterprise orchestration platform to counter OpenAI's plugin push.
Apple is quietly testing a ChatGPT-like internal chatbot that employees call "Apple GPT." Powered by its home-grown "Ajax" large language model framework, the tool is currently limited to internal prototyping and summarization, though Bloomberg reports the company is eyeing a major AI announcement next year.
Microsoft just expanded its AI partnership with Meta, becoming the preferred cloud partner as Meta makes Llama 2 commercially available. Azure developers can now deploy and fine-tune 7B, 13B, and 70B models with built-in safety tools, while Windows devs get native optimization via ONNX Runtime and DirectML to run models locally.
Stability AI just released Stable Beluga 1 and 2, two instruction-tuned language models built on Meta's LLaMA and LLaMA 2. Inspired by Microsoft's Orca paper, the team used just 600,000 synthetic data points—roughly 10% of Orca's dataset size—to climb to the number one spot on Hugging Face's Open LLM Leaderboard. Both models are available for download under a non-commercial research license.
Google just dropped its biggest Bard update yet, expanding the AI chatbot across Europe and Brazil with support for over 40 languages, including Chinese and Spanish. Users can now upload images alongside prompts using Google Lens, listen to spoken responses, adjust reply styles, and export code directly to Replit.
Anthropic just released Claude 2, featuring a 100K token context window that can digest hundreds of pages in a single prompt. The new model doubles safety performance while significantly improving coding and math benchmark scores. It is now live for users in the US and UK through a public web beta at claude.ai, with API pricing unchanged from Claude 1.3.
Google is now testing its specialized Med-PaLM 2 chatbot at research hospitals like the Mayo Clinic, racing Microsoft to dominate healthcare AI. Trained on medical licensing exams, the model can answer clinical questions and summarize records, though Google researchers admit it still hallucinates and isn't ready for real-world clinical decisions.
Alibaba and Huawei just launched major AI updates to challenge global rivals. Alibaba Cloud debuted Tongyi Wanxiang, an image generator taking on Midjourney and DALL-E, while Huawei introduced its industrial-focused Pangu 3.0 model targeting freight safety and weather forecasting instead of content creation.
OpenAI just made the GPT-4 API generally available to all paying developers, rolling out the 8K context model right away. It is also sunsetting older completion models by January 4, 2024, pushing everyone toward the structured Chat Completions API that now handles 97% of its GPT traffic.
