📈 How are companies using AI?

➜ Shopify announced a bunch of AI updates for their users:

–Semantic Search, a new AI-powered storefront search feature that can understand the customer’s intent.

–AI-powered Shop assistant that helps shoppers discover products they’re likely to buy.

–AI-enabled image editing features for product photos (generate images, match, or remove background).

–Guided search in their docs, enabling users to quickly find the apps and features they need. 

Shopify

➜ Boeing wants to patent an AI system that would inspect components for anomalies, serving as an assistant for inspectors and manufacturers in identifying what parts of an aircraft may go haywire. It would employ a computational model, like a neural network, trained to find faulty components in an aircraft.

Patent Drop

➜ The AI Grant startup accelerator site is a great resource to find interesting companies in the market. It helped me discover Hume, a startup that’s trained models which understand facial and vocal expressions, and can be used to build solutions such as a depression detector or toxic speech moderator.

➜ A better way to browse the web on mobile? The Browser Company has released Arc Search, a tool that browses the internet for you and builds a custom web page to present the results of its search – along with a handful of other interesting AI-powered features.

Arc Search

➜ Dopt (tool for building onboarding and education experiences) moved beyond text-driven, turn-based chatbots with AI assistants that provide contextual help directly within an app’s user interface. A great future direction for seamless, in-app AI assistants.

Dopt

➜ Rufus is Amazon’s genAI-powered shopping assistant that can guide customers through the whole buying process. It can answer questions, provide product comparisons, and make recommendations based on Amazon’s catalog and web information. Beta access to Rufus is launching in Amazon's mobile app for a limited group of US customers, with broader roll-out plans in the coming weeks.

Amazon

➜ Machine learning-based diagnostic test for ovarian cancer achieves 93% accuracy by analyzing blood metabolites – it’s better than existing tests, especially in identifying early-stage ovarian disease. The test employs a personalized probabilistic approach rather than a binary one, offering a more nuanced assessment of disease presence based on individual metabolic profiles. The team plans to commercialize the technology through a startup, aiming for FDA approval.

News-medical

➜ Gift Mode is an interactive gifting hub where AI and human curators help shoppers find the perfect gift. 

Etsy

➜ HYCU (data protection and backup services) has a new product that incorporates genAI to automate and simplify the integration process of their R-Cloud data protection product with over 60 different apps. The feature is based on Athropic’s Claude 2 model, and it reduces the integration process from hours to minutes.

Acceleration Economy

➜ Community Health Systems (CHS), one of the leading healthcare providers in the US, has enabled the use of generative AI for managing complex data and automating administrative tasks across its 71 hospitals by moving its data systems to Google Cloud.

Fierce Healthcare

➜ Liminal provides complete control over data in every interaction with generative AI through direct use, off-the-shelf software, or in-house applications – the model-agnostic security platform aims to enable regulated enterprises to safely use genAI without sacrificing data security, privacy, and sovereignty.

Datanami

➜ Really cool to see, because we've actually discussed an idea like this during our AI Primer workshops recently – A&M management consulting company offers the AI Tax Consulting Service, which combines real-time monitoring and analysis, tailored AI models developed together with tax professionals, and expert tax knowledge to help companies adapt to changes in tax regulations.

Alvarez&Marsal

➜ AskAva is the first generative AI assistant for contractor compliance and safety (still in beta, Q2 release plan). It uses OpenAI models to automate risk assessments within the Avetta platform. The goal is to minimize workplace incidents and address complex supply chain risk challenges by combining Avetta's proprietary data with genAI.

OHSOnline

💡 What are the nerds up to?

➜ Using genAI to manipulate live audio conversations for fraud – criminals can “audio-jack” conversations by:

--intercepting them,

--using speech-to-text to transcribe them,

--using LLMs to understand and modify the context (like changing bank account numbers mentioned in the conversation),
and then voice-cloning with text-to-speech to add the audio seamlessly.
Much more complex and effective method compared to traditional call scams, calling for completely new defensive mechanisms across IT infrastructures.

IBM

➜ Likely the first truly open model, including the dataset used for pretraining, full model weights for 4 variants of the model, 500+ checkpoints per model for evaluation, and full insights into training data and methodologies. The Open Language Model (OLMo) initiative by AI2 is a comprehensive framework aimed at fostering open research in AI.

The Allen Institute for AI

➜ The industry is abuzz with gossip about the new Mistral model, which is supposed to be almost as good as GPT-4. The model was leaked by an “over-enthusiastic” early adopter without attribution, and Mistral’s CEO confirmed that it’s their new project. 

X (Twitter)

➜ Shyam Pather, engineering leader who’s worked for Microsoft, Facebook, and Oculus among others, reports his findings from building a transformer model based on Andrej Karpathy’s video instructions. 

–He used it to answer the question “how exactly do the results of the attention computation turn into accurate predictions for the next token?” 

–His answer: “Each transformer block (containing a multi-head self-attention layer and feed-forward network) learns weights that associate a given prompt with a class of strings found in the training corpus. The distribution of tokens that follow those strings in the training corpus is, approximately, what the block outputs as its predictions for the next token.”

Shyam Pather

➜ Some people might still not know how much AI boosts productivity, so Ethan Mollick did an experiment where he completed five different marketing-related tasks in less than one minute. 

One Useful Thing

➜ Companies must immediately transition to asking custom interview questions to avoid candidates that use ChatGPT to blaze through technical interviews. 

–Professional interviewers asked devs original LeetCode questions, modified LeetCode questions, and custom questions. 

–Candidates using ChatGPT to cheat passed original questions 73% of the time, modified questions 67% of the time, but only 25% for custom questions. 

–Interestingly, none of the interviewers suspected that the devs were using ChatGPT (and they weren’t informed about it).

Interviewing.io

➜ Assistants are basically GPTs, but from Hugging Face and based on open models.

Hugging Face

➜ Machine learning models can effectively predict physiological skin conditions from images of stratum corneum cells (the outermost layer of our skin), providing a method to optimize dermatological treatments.

PubMed

➜ Machine learning method predicts growth of carbon nanostructures on metal surfaces to solve a big challenge in carbon nanotech. Combining theoretical models with experimental data, machine learning simulations figure out how to control the growth process. The method boasts great accuracy in experiments, providing a practical tool for precisely designing carbon nanostructures, and opening up new research opportunities in energy storage materials.

Phys

Thanks for reading!

I share these stories because I believe it's important for all of us to keep up with AI. To support my mission, share this newsletter on LinkedIn.