AI startup Mistral has confirmed the leak of a language model comparable in power to GPT-4
French AI startup Mistral has admitted that one of its customers accidentally published a powerful open source language model called "miqu-1-70b" online. By some estimates, this model outperforms the free ones and comes close to OpenAI's commercial GPT-4.
Here's What We Know
On 28 January, the model appeared on the HuggingFace platform and then on the 4chan imageboard. Mistral co-founder Arthur Mensch confirmed its authenticity.
According to him, the model is an old development of the company, which was retrained on the Llama 2 and distributed to some customers. However, one of them, "overly enthusiastic" about the technology, published the files on HuggingFace.
Mensch also made it clear that Mistral has made significant strides in language model development since the release of miqu-1-70b and will soon announce something more powerful. Experts speculate that we could be talking about a full-fledged open source analogue of GPT-4.
An over-enthusiastic employee of one of our early access customers leaked a quantised (and watermarked) version of an old model we trained and distributed quite openly.
- Arthur Mensch (@arthurmensch) January 31, 2024
To quickly start working with a few selected customers, we retrained this model from Llama 2 the minute we got...
The emergence of a free model of this calibre has the potential to fundamentally change the landscape of generative AI. It will put a lot of pressure on OpenAI's business model, which is based on paid subscriptions for access to GPT-3 and GPT-4.
Source: VentureBeat
