AI Newsletter #89 - Reimagining browser interface with natural language

Welcome to Nural's newsletter focusing on how AI is being used to tackle global grand challenges.

Packed inside we have

  • AI startup Adept wants Transformers to replace the Mouse and Keyboard
  • Tiny, caterpillar-like soft robot for minimally invasive surgery
  • The Meta Oversight Board says Facebook’s automated image takedowns are broken
  • Prompt injection attacks against GPT-3

If you would like to support our continued work from £1 then click here!

Marcel Hedman


This weeks newsletter is brought to you by Wynter -
Get paid up to $800 for your contribution

Wynter, a customer feedback tool, is looking for data science professionals (CEOs, senior leaders, MLEs, data scientists) to join its paid research panels.

Wynter is looking for machine learning engineers to join its research panel: - Participate in research activities (1-15 mins) - Get paid up to $800 for your contribution - NO ongoing commitment.
Join Wynter’s Paid Data Science Panels | Learn More


Key Recent Developments


AI startup Adept wants Transformers to replace the Mouse and Keyboard

ACT-1: Transformer for Actions
At Adept, we are building the next frontier of models that can take actions in the digital world—that’s why we’re excited to introduce our first large model, Action Transformer (ACT-1).

What: Adept is a new AI start-up formed of a few researchers who left Google Brain. In Adept's vision, future PC use will be without keyboard and mouse and we will entirely interact with speech based commands interpreted by language models. Earlier this year, they raised $65m to make this happen.

They have recently released ACT-1 as the first iteration of their "natural-language interfaces" and have trained it to use a web browser.

Key Takeaway: As technology develops, it becomes even more important to reimagine the fundamental ways we interact with our daily tools. Adept have begun on this journey to reimaging how we interact with our browsers using natural language. See below for a list actions it has already been trained to do. What more complex actions can you imagine for the future? Does this promise a future where anyone who can articulate their ideas in language can implement them, regardless of expertise?

  • What might ordinarily take 10+ clicks in Salesforce can be now done with just a sentence - See video on blog
  • Search a site for non-structured data based on a search criteria
  • Providing commands on Google sheets interface (even if you don't know the formuale) e.g. "Make a profit and then a profit margin column and then highlight the lowest two values"

The Meta Oversight Board says Facebook’s automated image takedowns are broken

Link

What: The Meta oversight board was established in 2019 as an independent group of experts to whom Meta and its users can refer appeals over platform and content decisions, providing another avenue for more complex concerns.

Having taken on a new set of cases, the board have found that Facebook's automated image takedown process is broken. Following the removal of an image, it is added to Meta’s Media Matching Service database, which means Meta’s systems flag it automatically for takedown when users post it. Images will stay in this database even if it is later found that an image was incorrectly taken down.

The board have stated "By using automated systems to remove content, Media Matching Service banks can amplify the impact of incorrect decisions by individual human reviewers."

Key Takeaway: The board semi-independent board was established by Meta for this exact reason; to ensure that Meta's policy's act in the best interest of its users. Other examples of high profile decisions by this board include the banning of Trump from its platforms.

Are these more democratic styles of governance the future of tech? Are they effective?


Prompt injection attacks against GPT-3

Prompt injection attacks against GPT-3
Riley Goodside, yesterday: Exploiting GPT-3 prompts with malicious inputs that order the model to ignore its previous directions. pic.twitter.com/I0NVr9LOJq- Riley Goodside (@goodside) September 12, 2022 Riley provided several examples. Here’s …

What: With new technology, comes a new wave of bad actors attempting to exploit weaknesses for their own gain. Following the advent of large language models and GPT-3, the same has occurred. Users have found certain prompts can get the model to either completely ignore the request, or leak what the initial request was.

Key Takeaway: New technology, new attacks, increased security. This is the constant cycle. OpenAI and other developers of LLMs bear this burden as the drivers of this new emergent technology and it is more critical than ever before following the increasing number of players who are commercialising the models.


Tiny, caterpillar-like soft robot for minimally invasive surgery

Tiny, caterpillar-like soft robot folds, rolls, grabs and degrades
When you hear the term ‘robot,’ you might think of complicated machinery working in factories or roving on other planets. But ‘millirobots’ might change that. They’re robots about as wide as a finger that someday could deliver drugs or perform minimally invasive surgery. Now, researchers have develo…

What: "When you hear the term "robot," you might think of complicated machinery working in factories or roving on other planets. But "millirobots" might change that. They're robots about as wide as a finger that someday could deliver drugs or perform minimally invasive surgery."


AI Ethics & 4 Good

🚀 Meta Faces Mounting Questions from Congress on Health Data Privacy As Hospitals Remove Facebook Tracker

🚀 Meta moved PyTorch to a new, independent PyTorch Foundation, under the Linux Foundation

🚀 The US regulator, the FTC Is Closing in on Runaway AI

🚀 Benchmarking AlphaFold-enabled molecular docking predictions for antibiotic discovery [Paper]

🚀 YouTube Shorts is Making Ads Vertical, Uses Machine Learning to Adopt it from Landscape Videos

Other interesting reads

🚀 Dogs are more attuned to actions rather than to who or what is doing the action - ML analysis of a dog's brain

🚀 NIH Injects $130M to Provide Quality Tools, Data for AI in Medicine

🚀 Computer scientists and chemical engineers at the Tokyo Institute of Technology in Japan have developed a machine-learning algorithm that can reverse-engineer a smell based on its chemical makeup

🚀 Free AI-based data annotation suite

🚀 AI Art: World’s First Bot-Generated Graphic Novel Hits the Market


Cool companies found this week

Health

White Lab Genomics - Data science for DNA and RNA therapies. They leverage graph knowledge technology and machine learning to help customers discover and design new genomic therapies. Recently raised $10m in funding

High-res satellites

Stability AI - Building openAI tools and the company behind the viral text-to-image generator Stable Diffusion.


...and Finally

Minirobots to perform minimally invasive surgery

AI/ML must knows

Foundation Models - any model trained on broad data at scale that can be fine-tuned to a wide range of downstream tasks. Examples include BERT and GPT-3. (See also Transfer Learning)
Few shot learning - Supervised learning using only a small dataset to master the task.
Transfer Learning - Reusing parts or all of a model designed for one task on a new task with the aim of reducing training time and improving performance.
Generative adversarial network - Generative models that create new data instances that resemble your training data. They can be used to generate fake images.
Deep Learning - Deep learning is a form of machine learning based on artificial neural networks.

Best,

Marcel Hedman
Nural Research Founder
www.nural.cc

If this has been interesting, share it with a friend who will find it equally valuable. If you are not already a subscriber, then subscribe here.

If you are enjoying this content and would like to support the work financially then you can amend your plan here from £1/month!