Oredev Conference - Day 3 2024

Learning and Improvement Conferences

If you haven’t already, be sure to check out my blog posts from my first day at the Öredev conference. and my second day at the Öredev conference

Learned a lot at my last day at the Öredev conference but this time the approach I took to this blog post is more of a brain dump. I was getting a bit tired so my notes from today were (unfortunately) not quite as good as yesterday’s.

Got some great ideas from today and was pinging my husband with ideas throughout the day. Hoping to start putting some of them into action in a week or two.

Creation as Curation

Laura Herman shared some of what her team is doing at Adobe including comparing AI-generated content with human-generated content and working on an industry-wide proof so people know something is not AI-generated.

How does AI-curated content (from Instagram) compare with human-generated content?

Let’s see how the visitors perceived this.

What differences did visitors notice? Instagram chose easily recognisable objects. Easy to parse. Prevalence of faces.

Example of a photo/art of a couch that looked like it was for sale (the style of the photo). At this time, Instagram was pushing Instagram shop.

Human chose the concept of loss. (Something that might be harder for an AI to choose images for?)

How does Instagram affect artists?

For example: An artist went viral with a specific post. But then he noticed that he felt he got locked into a specific type of art (similar to the viral one). It stifled his creativity

Generative AI

Generative AI has changed the creative process.

The power of human taste - this is hard to encode into Gen AI. Example of humans being able to tell the difference between a tune and random notes put together (that sound off).

Content Authenticity Initiative - aim to show which content is AI generated

Current limitations

Currently also looking into policy.

Focus switching from detecting fake to proving authentic.

Now working on the C2PA with other large tech companies niches Foogle, Microsoft IBM.

Debiasing

Eg. If someone types in “woman” then their tool takes into account how a woman looks like where the user is from (eg. Skin tone).

They are building in diversity and representation into their Gen AI tool at Adobe.

SecDevSecOpsSec

Jesper Olsen highlighted in his presentation the importance of security, having developers and security relate to each other and the effect AI will have on security going forward.

There are conflicting priorities between DevOps and security.

AI is turbocharging the speed and scale of cyber attacks, which means it is much more difficult for companies to protect themselves.

Cyber criminals are more likely to target easier targets. We also hace Cyber inequity - Small Medium Enterprises aren’t equipped to protect themselves from cyber threats.

“We want to have a developers mindset in. Security’s mind and we want to a security mindset in a developers mind.”

Ai brings new risks for business - Enterprise AI applications uses a specialised tech stack.

The EU is leading AI innovation

So you think a new tool will help? Here are some things to remember…

Isabel Evans shared her research into how companies/teams/people look into test tools. She also shared the guiding questions she has created to help teams make better decisions around tooling decisions.

Evans learned in her research that a good UI is not enough to guarantee happiness - too much emphasis is placed on the attractiveness of a UI, and not enough in the usefulness.

“Many people forget to ask themselves WHY they are buying a tool.”

-Isabel Evans

12 Heuristics (at present…)

Note: these questions may not always be applicable but are a good general indication of questions that could be useful.

Evans noticed that the more expertise someone had, the more likely they were able to use visualisations effectively.

Read more about Evans’ thoughts on her blog here. Read more about her research into this topic on her GitHub

Research questions

Untangle Your Spaghetti Test Code

Michael Kutz shared some mistakes that people make when writing test code and explained why they are bad. He then suggested different ways we can structure our test code with multiple examples and explaining the strengths of each approach

Why does the quality of test code matter? Lot of things we do in test code - we would never do in production. code. There is a tendency to be sloppy in test code

What do good tests look like?

“The intention (of a test) should be immediately obvious.”

Common approach to writing tests is Arrange Act Assert.

Here are some bad ways to go about writing tests:

Approval testing is a better fit if you have a lot of things you need to compare.

Behaviour over-specification definition

Consistent and concise test case name example

QA for ML: How we can trust AI with Food Sustainability

Serg Masis shares how QA for ML is different other types of software and then goes into how they are using AI for food sustainability at his workplace.

“Quality is about trust.”

“We shouldn’t forget that models can be right for the wrong reasons.”

Masis used the example of a crop growth model (see image below). In machine learning we have Model evaluation: error analysis (looks for strengths and weaknesses in the ML model).

Masis highlighted the benefits of Agentic AI as you can input data/information as well as guardrails. An example of using Agentic AI includes RAG

Main tasks for QA in AI diagram

Model interpretation: Feature importance diagram

Learning and Improvement Conferences