- Starter AI
- Posts
- The next AI chip maker, Groq’s LLM enhancement, Karpathy is back!
The next AI chip maker, Groq’s LLM enhancement, Karpathy is back!
The party's happening at the chip venture.
Sthefania
February 20, 2024
Hello, Starters!
As the need for training bigger and better AI models persists, more players are ready to tackle the shortage by creating their own chips. Should Nvidia stay on the lookout?
Here’s what you’ll find today:
SoftBank wants to build an AI chip venture
Groq’s chip enhances LLMs
Andrej Karpathy’s Minimal Byte Pair Encoding
Mistral unveils “Next”
ElevenLabs’ new sound effects
And more.
Another response has emerged to Nvidia's dominance in the chip-making industry, and this time it's from Softbank. Reportedly, CEO Masayoshi Son is seeking investors to build a $100 billion AI chip venture in collaboration with Arm. The project is called Izanagi.
Prominent figures in the industry, such as Sam Altman, are also exploring ways to overcome the GPU shortage. All of them agree that the best approach is to develop their offerings. Although Masayoshi was contacted by Altman for his initiative, the Izanagi project is not related to the latter.
🤖 Groq’s chip enhances LLMs (1 min)
Not to be confused with a famous chatbot, Groq is an AI startup that has developed Language Processing Units (LPUs) to enhance text generation, achieving speeds of up to 500 tokens per second—surpassing the capabilities of Gemini Pro and GPT-3.5.
The "GroqChip" is specialised hardware built with a tensor streaming architecture to improve efficiency, speed, and accuracy, outperforming traditional GPUs. For those intrigued by Groq's innovation, a test with Mixtral and Llama is available.
It hasn't been long since we last heard from Andrej Karpathy, and he has returned with the one thing he is best known for, teaching. In a recent GitHub repository, he tackled Byte Pair Encoding for beginners.
Byte Pair Encoding, or BPE, is an algorithm that optimises code for large language models. The approach shared by Karpathy offers a minimal and clean implementation at a "byte-level," operating with UTF-8 encoded strings.
🤖The French startup Mistral, well known for its advancements in open-source AI, has rolled out Mistral Next, a prototype large language model that, according to first reviews, is almost on par with OpenAI's GPT-4.
🔊With the potential of Sora in mind, ElevenLabs has stepped up its game with a text-to-sound model. This model aims to help creators turn their AI-generated videos into cinematic masterpieces, crafting sound effects directly from their most imaginative prompts.
⚡️Quick links
What did you think of today's newsletter? |