GOOOOOOD MOOORNING VIET... DATAWAKE!

I hope you’re all having a great “little saturday” as we say here in Sweden. Half way through the week and we’re allowed to have a drink.

I will get coffee

Here's today's zesty agenda:

• ChatGPT getting holiday vibes 🎄

• Meta releases audio foundation paper 📜

• Somewhat fun memes

ChatGPT getting holiday vibes

(Source)

In recent weeks, users of ChatGPT-4 noticed a peculiar trend as the AI model appeared to become "lazy," refusing certain tasks or providing simplified responses. OpenAI acknowledged the issue, emphasizing that the unintentional change in model behavior occurred since November 11th and is under investigation. Some AI researchers have humorously proposed the "winter break hypothesis," speculating that large language models like GPT-4 might simulate seasonal behaviors, potentially influenced by data patterns indicating decreased activity in December.

While unproven, the notion is taken seriously due to the unpredictable nature of AI language models. Reports indicate instances of increased refusals and simplified responses, sparking discussions about potential causes and fixes, offering a glimpse into the evolving and sometimes quirky realm of large language models. The ongoing exploration of peculiar AI behaviors reflects the dynamic and rapidly evolving nature of this field.

So from now on, I will tell Mr Fancy “it’s sometimes friday”-GPT that it’s always middle of september..🤣

Meta releases audio foundation paper 📜

(Source)

Meta released Audiobox a while back (read more here). Now they have also released a research paper going in more detail.

Tackling the limitations of existing large-scale audio generative models, Audiobox stands out for its capability to generate various audio modalities with enhanced controllability. Overcoming challenges such as synthesizing novel speech styles based on text descriptions and broadening domain coverage, this model introduces description-based and example-based prompting, allowing independent control over transcript, vocal, and other audio styles during speech generation. Notably, Audiobox sets impressive benchmarks in speech and sound generation, achieving 0.745 similarity on Librispeech for zero-shot TTS and 0.77 FAD on AudioCaps for text-to-sound. What's more, the integration of Bespoke Solvers significantly accelerates generation speed without compromising performance, marking a significant stride in the evolution of AI-powered audio creation and paving the way for novel vocal and acoustic styles.

Dank Memes

That's all for today folks! If you want more news or just some dank memes, make sure to follow the newsletter and on X!