Unrelated to AI:
See my latest Around the Web newsletter, 2023: my year in books, Watch my latest video essay.
[AI/ML + Bio] 🤖 🦠 🧬
[I]: 🧬 Genetic Discovery Enabled by A Large Language Model
Summary: This work demonstrates how Med-PaLM 2, a medically aligned LLM, can effectively analyze gene sets and identify causative genetic factors for various biomedical traits in mice, thereby facilitating novel genetic discoveries.
Technical details: the research outlines the utilization of Med-PaLM 2, an LLM specifically fine-tuned with biomedical text and clinician feedback, for genetic discovery. The model takes candidate gene sets, derived from mouse genetic models or genomic sequence comparisons, and processes these inputs to identify causative genetic factors for various biomedical traits in mice. Med-PaLM 2's architecture, based on transformer neural networks and trained on extensive text corpora, allows it to handle free-text queries effectively. The model's output includes hypotheses about gene-phenotype associations, which are subsequently validated through experiments. This study illustrates Med-PaLM 2's potential in bridging gaps in AI-assisted genetic discovery, showcasing a significant advance in the use of LLMs in biomedical research.
[II]: 🏥 Use of GPT-4 to Diagnose Complex Clinical Cases
Summary: The author of this paper assesses GPT-4's capability in diagnosing complex medical cases and compares its performance with that of medical journal readers, finding that GPT-4 correctly diagnosed 57% of cases, outperforming 99.98% of simulated human readers.
[III]: 🧬 CodonBERT: Large Language Models for mRNA design and optimization.
Summary: This study introduces CodonBERT, a LLM designed for mRNA sequence optimization, which demonstrates superior performance in predicting various mRNA properties and outperforms existing methods, including applications in flu vaccine development
Technical details: CodonBERT's architecture is particularly designed for mRNA sequence analysis. Key aspects of building the system:
Training Data: It was pre-trained on 10 million mRNA sequences derived from mammals, bacteria, and human viruses. These sequences were categorized into 13 hierarchical categories.
Input and Tokenization: CodonBERT takes the coding region of mRNA sequences as input, using codons (sequences of three nucleotides) as tokens. This approach allows for a more biologically relevant and precise representation of mRNA.
Output: The model outputs embeddings that provide contextual representations of codons. These embeddings can be further utilized with additional trainable layers for various downstream tasks.
Pre-training Tasks: CodonBERT was pre-trained with two main tasks:
Masked Language Model (MLM) Learning: This task involves learning codon representations, including the interactions between codons and the relationships between codons and sequences.
Homologous Sequences Prediction (HSP): This task is focused on modeling the sequence representation and understanding the evolutionary relationships between mRNA sequences.
Through these elements, CodonBERT is uniquely structured to capture the complexities and nuances of mRNA sequences, offering advanced capabilities for predicting protein expression, mRNA degradation, and other properties related to mRNA sequences.
[AI X Industry + Products] 🤖🖥👨🏿💻
[I]: 🤖Midjourney 6.0
The latest version of Midjourney was released a few weeks ago. And it’s pretty darn good.
[II]: 🪆AI Toy
Now, this is a clever use case for LLM. Bring Toys to life. A new startup named Curio are building AI Toys with Grimes.
[III]: 📷 Nikon, Sony and Canon fight AI fakes with new camera tech
Talking about mid-journey, “Nikon, Sony Group and Canon are developing camera technology that embeds digital signatures in images so that they can be distinguished from increasingly sophisticated fakes.”
[IV]: 💻Microsoft Edge is now an ‘AI browser.’
Microsoft has been adding more and more AI features to its Edge browser over the past year, and now the company is branding it the “AI browser.”
[V]: 👩 AI-created “virtual influencers” are stealing business from humans.
“Pink-haired Aitana Lopez is followed by more than 200,000 people on social media. She posts selfies from concerts and her bedroom, while tagging brands such as hair care line Olaplex and lingerie giant Victoria’s Secret.
Brands have paid about $1,000 a post for her to promote their products on social media—despite the fact that she is entirely fictional…” Full article link.
[AI + Commentary] 📝🤖📰
[I]: 💻 The Industries AI is Disrupting Are Not Lucrative.
This is a compellable critique of the current AI hype, at least as far as the author is concerned. The author questions whether the significant investment in AI development is economically justifiable. The author argues that the industries AI is poised to disrupt are not lucrative enough to warrant such extensive financial commitments. While part of me concurs with this viewpoint, another part disagrees, though I struggle to formulate a robust counterargument to the critique. My intuition diverges from the author’s opinion regarding the potential impact of Large Language Models (LLMs) on fields like science and medicine, which I believe could be profoundly disrupted. However, the author does raise another valid point...
“…Again, even assuming that’s all technologically possible, these examples, (such as) law and mental health, are extremely difficult to disrupt for structural reasons—not just because the vast majority of people want a final human overseer, but because a whole host of regulations, traditions, and legalities stand in the way.”
[II]: 🏥The Next WebMD
Talking about disruptions, this essay discusses the potential of LLMs to revolutionize healthcare by serving as an accessible and intelligent front door to medical services. It envisions a future where LLMs, integrated into healthcare service marketplaces, assist patients in diagnosing symptoms, selecting appropriate specialists, and navigating insurance complexities, thereby streamlining, and personalizing the healthcare experience.
[III]: 💻 The Where, When, and How of AI
Theory Ventures chats with folks from OpenAI, Lamini, and Mother Duck about the where, when, and how of AI. They chatted about genAI tools, sales, enterprise use cases for genAI tools, app development, amongst others.
[IV] 💎 The AI Goldrush - What Might Happen
Now that you can ‘write’ AI apps with just natural language (e.g. GPTs), what might happen next (essay). I for one think that, if we are talking about writing apps with natural language, there will be too much competition (since ‘everyone’ can do it), and hence not as valuable. Looking forward to how these things play out.
[V] 🛒 Marketplaces in the Age of AI
This essay from Andreessen Horowitz discusses the transformative impact of generative AI on marketplace dynamics. It suggests that AI will revolutionize both the demand and supply sides of marketplaces, enhancing user experiences through improved search functionalities and creating new opportunities for custom product creation. The essay also predicts significant changes in marketplace structures, with some businesses needing to pivot or potentially face obsolescence, while others will be empowered by AI to offer more dynamic and efficient services.