New releases, benchmarks, pricing and API changes from GPT, Claude, Gemini, DeepSeek, Qwen and other large language models, updated daily.
The Stable LM team just released Stable LM Zephyr 3B, a 3-billion-parameter chat model that is 60% smaller than typical 7B alternatives. Tuned with Direct Preference Optimization (DPO), it delivers responsive instruction-following and Q&A on consumer hardware without high-end GPUs. It is currently available under a non-commercial license.
Google just rolled out its new Gemini model to Bard, giving the chatbot its biggest boost yet. English users across 170+ countries can now tap a tuned version of Gemini Pro for text prompts, which Google claims outscores GPT-3.5 on key reasoning benchmarks. Next up: early next year, Bard Advanced will unlock the flagship Gemini Ultra.
Anthropic just released Claude 2.1, doubling its context window to an industry-leading 200,000 tokens—roughly 500 pages of text. The update halves hallucination rates compared to Claude 2.0, introduces system prompts, and adds beta tool use for API developers to connect external apps and data.
Microsoft just released Orca 2, a pair of small language models trained to use tailored reasoning strategies rather than blindly mimicking giant models. By learning when to reason step-by-step or answer directly, it matches or beats models 5 to 10 times larger on complex zero-shot benchmarks.
Google just pushed the release of its flagship AI model Gemini to early 2024, after previously telling cloud customers it would roll out in November. Positioned by DeepMind as its true competitor to GPT-4, the delay shows Google is still having a tough time catching up with OpenAI.
Google is rolling out Bard to teenagers across most countries, adding step-by-step math explanations and chart generation from tables. To protect younger users, the company added content filters for age-sensitive topics and an automatic double-check tool that verifies responses against web results to catch hallucinations.
OpenAI just paused new ChatGPT Plus sign-ups after a surge in usage following DevDay exceeded its server capacity. The team said the move is temporary to ensure existing users have a smooth experience, and waiting users can sign up inside the app to get notified when subscriptions reopen.
OpenAI CEO Sam Altman just confirmed the company has started training GPT-5, claiming the next-gen model could potentially cross into superintelligence. But building it won't be cheap—Altman noted that hitting AGI requires massive compute and data, meaning OpenAI will need billions more from Microsoft and other backers before it ever sees the light of day.
Samsung just unveiled Gauss, its proprietary generative AI model built to power on-device features across its hardware lineup. Developed by Samsung Research, the tech splits into Language for translation and summaries, Code for programming assistance, and Image for generating and upscaling visuals. Running these features locally on upcoming flagships like the Galaxy S24 gives Samsung a latency and privacy edge over cloud-tethered chatbots.
Amazon is developing a massive new AI model codenamed Olympus to catch up with rivals OpenAI and Google. Overseen directly by CEO Andy Jassy, the model reportedly features two trillion parameters and could power Alexa and Amazon's online store as soon as December. Why it matters: After falling behind in generative AI, Amazon is ready to throw serious compute behind proprietary foundation models.
OpenAI just unveiled GPT-4 Turbo at DevDay, boosting context length to 128K tokens while slashing input prices by 3x and output by 2x. It also launched the Assistants API with built-in retrieval and code execution, effectively handling the messy state management developers previously had to build from scratch.
Elon Musk just debuted Grok, xAI’s first conversational model powered by Grok-1. Grok pulls real-time information directly from X and is explicitly designed with a witty, sarcastic edge that tackles spicy questions others avoid. It is rolling out to select U.S. testers first, with wider access coming for X Premium+ subscribers once out of early beta.
Phind just rolled out a new coding model based on CodeLlama-34B that scores 74.7% on HumanEval, outpacing GPT-4. Powered by Nvidia TensorRT-LLM on H100 GPUs, it delivers up to 100 tokens per second with a 16k context window. Why it matters: developers get GPT-4-level code generation at GPT-3.5 speed, completely open-sourced on Hugging Face.
OpenAI is rolling out a new beta update to ChatGPT Plus that lets users upload and analyze files directly in chat. Even better, it drops the clunky tool switcher—GPT-4 can now automatically figure out whether to browse the web, analyze data, or generate images based on context. The update is gradually rolling out to subscribers.
Twelve Labs just launched Pegasus-1, a video-language foundation model designed to generate text directly from video. Instead of slicing clips into thousands of static images, it natively analyzes visual and audio cues together to handle summaries, chaptering, and custom reports. Why it matters: It treats video as a unified medium rather than a glorified slideshow.
OpenAI quietly killed an internal AI model code-named Arrakis earlier this year after it failed to meet efficiency targets, according to The Information. Built alongside GPT-4 to drastically lower the eye-watering compute costs of running apps like ChatGPT, the setback left some Microsoft executives disappointed. It shows that slashing inference expenses remains a brutal hurdle, even for AI's market leader.
Baidu CEO Robin Li just unveiled Ernie 4.0 at an event in Beijing, boldly claiming the new model is not inferior to OpenAI's GPT-4 in any aspect. The update brings major upgrades to understanding, reasoning, generation, and memory, and will plug directly into Baidu Search and Maps. Markets reacted cautiously, with shares dipping slightly after the demo.
Anthropic just expanded web access to claude.ai, officially rolling out its flagship AI assistant to users across 95 countries for professional and everyday tasks. It’s a major leap forward from its initial US- and UK-only availability, giving millions more users direct access to Claude.
Portkey just revealed that GPT-4 latency has steadily dropped over the past three months, effectively cutting 99th-percentile wait times by more than half. While complex prompts still spike processing, median latencies now sit comfortably below 1ms per token. It is still pricier, but speed is no longer the bottleneck holding teams back.
Leaked screenshots reveal Google is quietly bringing its multimodal Gemini model to MakerSuite alongside an unannounced tool called "Stubbs." The secret feature lets creators generate, deploy, and share functional app prototypes with just a single text prompt and an optional reference image.
