Google, OpenAI and Mistral released latest versions of their cutting-edge AI models in only 12 hours.
Meta may also be in the combo with its upcoming Llama-3 model, and OpenAI's highly anticipated GPT-5 is within the pipeline.
What began as a purely area of interest category dominated by ChatGPT is now flooded with alternatives that transcend Big Tech and the open- and closed-source divide.
Google Gemini Pro 1.5
Google's Gemini Pro 1.5 caused a stir and led to advances in long-context understanding that challenge Claude 3 Opus, which leads the best way on this category.
Our next-generation AI model, Gemini 1.5 Pro, is now available in public preview @GoogleCloud'S #VertexAI Platform.
Its Long Context Window is already helping corporations analyze large amounts of information, construct AI-powered customer support reps, and more. → https://t.co/CLMN3wNmeP pic.twitter.com/RpRVUul3eg
– Google DeepMind (@GoogleDeepMind) April 9, 2024
With the flexibility to process as much as 1 million tokens, Gemini Pro 1.5 can process large amounts of knowledge concurrently, including 700,000 words, an hour of video, or 11 hours of audio.
Its Mixture of Experts (MoE) architecture increases efficiency and performance through the use of specialized models for specific tasks.
Google's list of Gemini models is kind of complex, but essentially the most powerful model for typical tasks is ranked here.
New version of GPT-4 Turbo
OpenAI then released a new edition, GPT-4 Turbo, with superior math and image processing.
An X post states: “GPT-4 Turbo with Vision is now generally available within the API. Vision requests can now also use JSON mode and performance calls.”
GPT-4 Turbo with Vision is now generally available within the API. Vision requests can now also use JSON mode and performance calls.https://t.co/cbvJjij3uL
Below are some great ways developers construct with vision. Write us a solution 🧵
— OpenAI developer (@OpenAIDevs) April 9, 2024
OpenAI expects to release GPT -5 soon, in addition to its text-to-video model Sora, which currently has no significant competitors (although that may change).
Mixtral 8x22B
Perhaps the largest surprise, nonetheless, got here from mistralwho boldly released their Mixtral 8x22B model as a free downloadable 281GB file via torrent.
magnet:?xt=urn:btih:9238b09245d0d8cd915be09927769d5f7584c1c9&dn=mixtral-8x22b&tr=udp%3A%2F%https://t.co/2UepcMGLGd%3A1337%2Fannounce&tr=http%3A%2F%https://t.co/OdtBUsbeV5%3A1337%2Fanounce
— Mistral AI (@MistralAI) April 10, 2024
With a formidable 176 billion parameters and a context length of 65,000 tokens, this open source model based on the Apache 2.0 license is claimed to outperform Mistral's predecessor Mixtral 8x7B, which had already outperformed competitors resembling Llama 2 70B in various benchmarks.
Mixtral 8x22B's advanced MoE architecture enables efficient calculations and improved performance in comparison with previous iterations.
Meta Llama 3 is coming
Not to be left behind, reports suggest that Meta could release a small version of its highly anticipated Llama 3 model as early as next week, with the total open source model still scheduled for July.
Llama 3 is predicted to be available in a wide range of sizes, from very small models that compete with Claude Haiku or Gemini Nano to larger, fully responsive and reasoning models that compete with GPT-4 or Claude 3 Opus.
Model multiplication
A generative AI ecosystem once dominated by ChatGPT is now flooded with alternatives.
Virtually every major technology company is involved, either directly or through significant investments. And the more players join the fight, the more hope that one faction can dominate the market dwindles.
We are actually also seeing the gap closing between closed source models from OpenAI, Anthropic, Google, etc. and closed source alternatives from Mistral, Meta and others.
Open source models are still hardly accessible to the final population, but that may probably change too.
Do any of those models represent an actual advance in machine learning, or are they simply more of the identical but higher? It is determined by who you ask.
Some, like Elon Musk, predict that AI will surpass human intelligence inside a yr.
Others, like Meta's chief scientist Yann LeCun argue that AI is miles behind us when it comes to robust intelligence measurements.
LeCun declared in February about current LLMs: “So mainly they’ll’t invent latest things. They will reflect roughly every thing they were trained on from public data, meaning you possibly can get it from Google. People have said, ‘Oh my God, we want to manage LLMs because they’re going to be so dangerous.’ That’s just not true.”
Meta goals to create an “object-driven” AI that higher understands the world and attempts to plan and reason about it.
“We are working hard to work out find out how to make these models not only talk, but in addition think and plan.” . . have memory,” explained Joelle Pineauthe vice chairman of AI research at Meta.
OpenAI Chief Operating Officer Brad Lightcap also said his company is concentrated on improving AI's ability to know and handle more complex tasks.
“We will begin to see AI that may tackle more complex tasks in additional sophisticated ways,” he said at a recent event. “I feel over time we’ll see the models turn into longer, by some means more complex tasks, and that implicitly requires improving their considering skills.”
As 2024 approaches summer, the AI community and society at large can be watching closely to see what groundbreaking developments emerge from the laboratories of those tech giants.
By the top of the yr there can be quite a various selection.
