All 16 verified AI news entries from 2023, including 12 model releases, in date order with a link to each official source.
Showing 16 of 16 entries
MidjourneyCreative, Important
Midjourney V6 improved how the model handled long prompts, along with coherence, knowledge and image prompting. It became the default model on February 14, 2024.
Mistral AIModels, Important
Mixtral 8x7B used 12.9B of its 46.7B parameters per token, matched or beat GPT-3.5 on most standard benchmarks, and shipped under Apache 2.0 with a 32K context.
Image: Mistral AI
Google DeepMindModelsMajor
Gemini was trained to be multimodal from the start. Google said Ultra scored 90.0% on MMLU, and Pro went into Bard and reached developers through the Gemini API on December 13.
Image: Google DeepMind
OpenAIModelsMajor
At its first DevDay, OpenAI previewed GPT-4 Turbo at $10/$30 per 1M tokens with a 128K window, alongside the Assistants API, JSON mode, vision, the DALL·E 3 API and text-to-speech.
Image: OpenAI
xAIModels, Important
xAI's first assistant drew on real-time information from X and ran on Grok-1, which xAI said beat GPT-3.5 and LLaMA 2 70B on MMLU, HumanEval, GSM8k and MATH.
Image: xAI
Mistral AIModels, Important
The Paris startup's first model, a 7.3B-parameter LLM, beat Llama 2 13B on every benchmark Mistral reported and shipped with no usage restrictions.
Image: Mistral AI
OpenAICreative, Important
DALL·E 3 followed detailed prompts far more closely than DALL·E 2, used ChatGPT to write and refine prompts, and was set to reach ChatGPT Plus and Enterprise in October 2023.
Image: OpenAI
ElevenLabsVoice, Important
Eleven Multilingual v2 generated speech in 28 languages while keeping a speaker's voice and accent, and launched alongside the public release of Professional Voice Cloning.
Image: ElevenLabs
Stability AICreative, Important
SDXL 1.0 paired a 3.5B-parameter base model with a refiner in a 6.6B-parameter pipeline, generated 1024x1024 images natively and shipped open weights that ran on consumer GPUs.
Image: Stability AI
MetaModelsMajor
Llama 2 came in 7B, 13B and 70B sizes trained on 2 trillion tokens, with downloadable weights under a community license that allowed commercial use.
AnthropicModels, Important
Claude 2 improved coding and exam scores over Claude 1.3, accepted prompts of up to 100K tokens, and launched in the API and on claude.ai in the US and UK.
GoogleModels, Important
PaLM 2 came in four sizes from Gecko to Unicorn, was trained on more than 100 languages, and reached developers through the PaLM API and Vertex AI.
Image: Google
MidjourneyCreative, Important
Midjourney V5 arrived on March 15, 2023 as an opt-in model in Discord and quickly became the reference point for photorealistic AI images.
OpenAIModelsMajor
GPT-4 accepted text and image inputs, scored around the top 10% on a simulated bar exam, and launched in ChatGPT Plus and through an API waitlist.
Image: OpenAI
AnthropicModels, Important
After a closed alpha, Anthropic made Claude and a faster, cheaper Claude Instant available through a chat interface and an API, with Notion, Quora, DuckDuckGo and others as early partners.
OpenAIModelsMajor
gpt-3.5-turbo, the model behind ChatGPT, arrived in the API at $0.002 per 1K tokens, ten times cheaper than earlier GPT-3.5 models, alongside a hosted Whisper speech-to-text endpoint.
Image: OpenAI
