- Starter AI
- Posts
- Profluent’s gene editing, DeepMind’s “Many-shot” training, Meta’s road to profit
Profluent’s gene editing, DeepMind’s “Many-shot” training, Meta’s road to profit
Is the AI gold rush actually coming?
Sthefania
April 25, 2024
Hello, Starters!
Thursday has arrived again, and with it comes a whole new round of AI insights that are worth the time to read. Do you think we're witnessing history in the making with all of these developments?
Here’s what you’ll find today:
Profluent plans to edit the human genome with AI
DeepMind presents “Many-Shot In-Context Learning”
Zuckerberg’s approach to GenAI profit
Snowflake unveils Arctic LLM
Andrej Karpathy updates “llm.c”
And more.
Profluent, an AI company, has recently achieved groundbreaking results in gene editing using AI-designed molecules. The researchers trained large language models on massive CRISPR-based gene editing data, expanding the diversity of gene editing systems and creating editors with higher activity and specificity than current models.
Through the release of OpenCRISPR-1, Profluent aims to increase innovation in biological design, potentially leading to new treatments and advancements for patients in need.
The term "many-shot" may sound familiar to you, as Anthropic has also shared insights into it. However, a study from Google DeepMind provides a different perspective. In this study, researchers demonstrate the potential of extending the context window of large language models by incorporating hundreds or even thousands of training examples into the prompt.
This approach offers an enhanced method of training models, making it possible to fine-tune them without the need for major adjustments. It allows developers to prompt freely, without worrying about not having enough context window space.
💰 Zuckerberg’s approach to GenAI profit (2 min)
Mark Zuckerberg is realistic about Meta's path to profit. During the company's latest first-quarter earnings report, he discussed with investors Meta's current interest in generative AI development.
Despite the company's increased profitability, Zuckerberg warned investors of slow growth, noting that it may decrease in speed as Meta continues to redirect its efforts towards AI and the metaverse. He claimed it's likely that it may take them years before actually seeing significant returns.
❄️Following the rise of enterprise-oriented models, Snowflake has released"Arctic LLM," a generative AI model carefully crafted to deal with that kind of workload, available for free under an Apache 2.0 licence. Arctic LLM is a potential rival to Databricks' DBRX.
🧑💻Andrej Karpathy has provided an update on the progress of training large language models, sharing how the "llm.c" method can now train GPT-2 using only 2000 lines of C code on GPUs, with speeds matching PyTorch.
⚡️Quick links
What did you think of today's newsletter? |