Is Artificial Intelligence safe? It depends who you ask...

 Over the last few days there's been some significant calls for a slowdown or pause in artificial intelligence research and development, or at least for a pause on public implementations of AI.

There's some significant rationale behind this - with the arrival of GPT4 (which I've been extensively experimenting with) we have seen another huge leap in AI capability.

We've also seen many large companies announce they are working to integrate this level of AI into their services in ways that allow the public to use (or misuse) this capability.

Some of this is extremely valuable - such as integrating a writing, formula and presentation assistant into Microsoft's Suite of tools. Some appears risky - such as Snapchat's release of an AI 'friend' into its paid service in February which, as the video, The AI Dilemma (linked), can incidentally be used to help sexual predators groom children (watch from 47 min - 49 min for this specific segment).

Also we've seen over a thousand AI luminaries and researchers call for a pause on AIs more sophisticated than GPT4 (letter here, article about it here) - which has received particular attention because Elon Musk signed, but is actually notable because of calibre and breadth of other industry experts and AI company CEOs that signed it.

Now whether government is extensively using AI or not, AI is now having significant impacts on society, These will only increase - and extremely rapidly.

Examples like using the Snapchat AI for grooming are the tip of the iceberg. It is now possible - with 3 seconds of audio (a Microsoft system) - to create a filter mimicking any voice, making all voice recognition security systems useless. 

In fact there's already been several cases where criminals are calling individuals to get their phone message (in their voice) or their initial greeting, then using that to voice authenticate the individual's accounts and steal funds.

Now this specific example isn't new - the first high profile case occurred in 2019. 

However the threshold for accessing and using this type of technology has dramatically come down, making it accessible to almost anyone.

And this is only one scenario. Deep fakes can also mimic appearances, including in video, and AIs can also be used to simulate official documents or conversations with organisations to phish people.

That's alongside CV fakery, using AI to cheat on tests (in schools, universities and the workplace) and the notion of outsourcing your job to AI secretly, which may expose commercially sensitive information to external entities.

And we haven't even gotten to the risks of AI that, in pursuit of its goal or reward, uses means such as replicating itself, breaking laws or coercing humans to support it.

For governments this is an accelerating potential disaster and needs to have the full attention of key teams to ensure they are designing systems that cannot be exploited by humans asking an AI to read the code for a system and identify any potential vulnerabilities.

Equally the need to inform and protect citizens is becoming critical - as the Snapchat example demonstrates.

With all this, I remain an AI optimist. AI offers enormous benefits for humanity when used effectively. However with the proliferation of AI, to the extent that it is now possible to run a GPT3 level AI on a laptop (using the Alpaca research model), there needs to be proactivity in the government's approach to artificial intelligence.

Read full post...

View summary...

Artificial Intelligence isn't the silver bullet for bias. We have to keep working on ourselves.

There's been a lot of attention paid to AI ethics over the last few years due to concerns that use of artificial intelligence may further entrench and amplify the impact of subconscious and conscious biases.

This is very warranted. Much of the data humans have collected over the last few hundred years is heavily impacted by bias. 

For example, air-conditioning temperatures are largely set based on research conducted in the 1950s-70s in the US, on offices predominantly inhabited by men and folks wearing heavier materials than worn today. It's common for many folks today to feel cold in offices where air-conditioning is still set for men wearing three-piece suits.

Similarly, many datasets used to teach machine learning AI suffer from biases - whether based on gender, race, age or even cultural norms at the time of collection. We only have the data we have from the last century and it is virtually impossible for most of it to be 'retrofitted' to remove bias.

This affects everything from medical to management research, and when used to train AI the biases in the data can easily affect the AI's capabilities. For example, the incredibly awkward period just a few years ago when Google's image AI incorrectly identified black people as 'gorillas'. 

How did Google solve this? By preventing Google Photos from labelling any image as a gorilla, chimpanzee, or monkey – even pictures of the primates themselves - an expedient but poor solution, as it didn't fix the bias.

So clearly there's need for us to carefully screen the data we use to train AI to minimise the introduction or exacerbation of bias. And there's also need to add 'protective measures' on AI outputs, to catch instances of bias, both to exclude them from outputs and to use them to identify remaining bias to address.

However, none of this work will be effective if we don't continue to work on ourselves.

The root of all AI bias is human bias. 

Even when we catch the obvious data biases and take care when training an AI to minimise potential biases, it's likely to be extremely difficult, if not impossible, to eliminate all bias altogether. In fact, some systemic unconscious biases in society may not even be visible until we see an AI emulating and amplifying them.

As such no organisation should ever rely on AI to reduce or eliminate the bias exhibited by its human staff, contractors and partners. We need to continue to work on ourselves to eliminate the biases we introduce into data (via biases in the queries, process and participants) and that we exhibit in our own language, behaviours and intent.

Otherwise, even if we do miraculously train AIs to be entirely bias free, bias will get reintroduced through how humans selectively employ and apply the outputs and decisions of these AIs - sometimes in the belief that they, as humans, are acting without bias.

So if your organisation is considering introducing AI to reduce bias in a given process or decision, make sure you continue working on all the humans that remain involved at any step. Because AI will never be a silver bullet for ending bias while we, as humans, continue to harbour them.

Read full post...

View summary...

Does Australia have the national compute capability for widespread local AI use?

There's been a lot of attention on the potential benefits and risks of artificial intelligence for Australia - with the Department of Industry developing the Artificial Intelligence (AI) Ethics Framework and the DTA working on formal AI guidelines.

However comparative less attention is often placed on building our domestic AI compute capability - the local hardware required to operate AIs at scale.

The OECD also sees this as a significant issue - and has released the report 'A blueprint for building national compute capacity for artificial intelligence' specifically to help countries that lack sufficient national compute capability.

As an AI startup in Australia, we're been heavily reliant on leveraging commercial AI capabilities out of the US and Europe. This is because there's no local providers of generative AI at commercial prices. We did explore building our own local capability and found that the cost for physical hardware and infrastructure was approximately 10x the cost of the same configurations overseas.

There's been global shortages of the hardware required for large AI models for a number of years. Unfortunately, Australia tends to be at the far end of these supply chains, with even large commercial cloud vendors unable to provision the necessary equipment locally.

As such, while we've trained several large AI models ourselves, we've not been able to locate the hardware or make a commercial case for paying the costs of hosting them in Australia.

For some AI uses an offshore AI is perfectly acceptable, whether provided through a finetuned commercial service or custom-trained using an open-source model. However, there's also many other use cases, particularly in government with jurisdictional security requirements, where a locally trained and hosted AI is mandated.

A smaller AI model, such as Stable Diffusion, can run on a laptop, however larger AI models require significant dedicated resources, even in a cloud environment. Presently few organisations have the capability to pay the costs for sourcing the hardware and capability to run this within Australian jurisdictions.

And that's without considering the challenge in locating sufficient trained human resources to design, implement and manage such a service.

This is an engineering and production challenge, and will likely be resolved over time. However with the high speed of AI evolution, it is a significant structural disadvantage for Australian organisations that require locally hosted AI solutions.

If Australia's key services have to rely on AI technologies that are several generations behind world-standards, this will materially impact our global competitiveness and capability to respond.

As such, alongside worrying about the ethics and approaches for using AI, Australian governments should also reflect on what is needed to ensure that Australia has an evolving 'right size' national compute capability to support our AI requirements into the future.

Because this need is only going to grow.

Read full post...

View summary...

To observe future use of AI, watch Ukraine

I'm not given to regularly commenting on global events, such as major wars, through this blog. However, given the active use of advanced technology by both Ukraine and Russia in the current war, I am making a brief exception to comment on the use of advanced technologies, including AI.

The Russian invasion of Ukraine has seen some of the most advanced uses of technology on the battlefield in history. Some of this involves AI, some other technologies, but all of it demonstrates the practical potential of these technologies and should be considered within any nation's future defense planning.

Now drones have been used in previous conflicts in Afghanistan and elsewhere, with militia forces modifying off-the-shelf products into aerial scouts and IEDs. Meanwhile US forces have used custom-built human-controlled Predators for over ten years to pinpoint target enemy commanders and units.

However, the war in Ukraine represents the largest deployment and the most versatile use of drones to-date by regular armies opposing each other in the field.

For example, both Ukraine and Russia have successfully used autonomous watercraft in assaults. These unmanned above and below water drones have been used to inflict damage on opposing manned vessels and plant and clear mines at a fraction of the cost of building, crewing and maintaining manned vessels.

The Ukrainian attack on Russia's Black Sea Fleet in port exhibits signs of being one of the first uses of a drone swarm in combat, with a group of kamikaze above and below water drones, potentially with aerial drone support, used in concert to damage several Russian ships.

While perhaps not as effective as was hoped, this is a prototype for future naval drone use and heralds that we are entering an era where swarms of relatively cheap and disposable drones - partially or completely autonomous to prevent signal blocking or hijacking - are used alongside or instead of manned naval vessels to inflict damage and deter an attacker, acting as a force magnifier.

The use of autonomous mine layers offers the potential for 'lay and forget' minefields astride enemy naval routes and ports, to again limit and deter enemy naval movements. Again the lower cost and disposability of these drones, compared to training explosive handling human divers and building, maintaining, defending and crewing manned naval minelaying and removal vessels, makes them a desirable alternative.

We've also seen extensive use of aerial drones by both sides in the conflict to spot and target enemy combatants, allowing more targeted use of artillery and manpower and helping reduce casualties by lifting aspects of the fog of war. Knowing the numbers and strength of your opponent on the battlefield greatly enhances a unit's capability to successfully defend or assault a position.

While many of these drones are human controlled, there's been use of AI for situations where direct control is blocked by enemy jamming or hacking attempts. AI has also been used for 'patrol routes' and to 'return' home once a drone has completed a mission - complete with ensuring that drones fly an elusive course to conceal the location of their human controllers.

The Ukrainian war has even seen the first public video recorded aerial drone on drone conflict, with a Ukrainian drone ramming a Russian drone to knock it out of the sky and remove the enemy's best source of intelligence.

These examples of drone use aren't totally new to warfare. 

Hundreds of years ago unmanned fire ships were used in naval combat - loaded with explosives and left to drift into enemy warships and explode. 

Similarly, early aerial combat by humans, in World War One, saw pilots take pistols and bricks aloft to fire at enemy planes or drop on enemy troops below. Just as drones today are being modified to carry grenades to drop on infantry positions or used to ram opposing drones.

The war will help advance the technology and improve the tactics. If nothing else Ukraine is a test ground for learning how to effectively use drones within combined forces to improve overall military effectiveness and reduce casualties.

And artificial intelligence is becoming increasingly important as a control alternative when an enemy blocks signal or attempts to take control of an army's drone assets.

We need to put these learnings to use in our own military planning and acquisitions, so that Australia's military becomes capable of fighting the next war, rather than the last.

Read full post...

View summary...

Do AIs dream of electric explainability?

One of the primary arguments against artificial intelligence in many processes related to decision-making is lack of explainability.

Explainability (also known as interpretability) refers to being able to explain how a machine learning AI model functions to produce a given output in a way that “makes sense” to a human being at an acceptable level.

Think of how in maths classes at school where you may have been asked to 'show your working' - to write out the steps you took to get from the initial problem to your solution.

For learning AIs, those that are trained on massive datasets to reach a level of capability, explainability can be highly challenging.

Even when the initial AI algorithms used for the machine learning and the data set used to train it are made fully explainable, the internal method by which the AI goes about deriving a solution may not be fully explained and hence the AI doesn't meet the explainability test.

When presented with identical inputs for a decision process, different AIs, using the same initial algorithms and trained on the same training dataset, might form very different conclusions.

Now this may appear similar to how humans make decisions. 

Give two humans the same information and decision process and, at times, they may arrive at completely different decisions. This might be due to influences from their past experiences (training), emotions, interpretations, or other factors.

When humans make decisions, it is possible to ask them how they arrived at their decision, and how they weighed various factors. And they may be able to honestly tell you.

With AIs it is also possible to ask them how they arrived at a decision.

However, the process by which they use to respond to this request is the same as they used to arrive at their decision in the first place. The AI model is not self-conscious and as such there's no capability for self-reflection or objective consideration.

In fact, most machine learning models only have an attention span of only a few thousand words, So, they may not even recall making a decision a few minutes or days before. AI doesn't have the consciousness to be aware that 'they' as an entity made the decision. 

This is unlike a human, who might forget a decision they made, but be conscious they made it and able to 'think back' to when they did to provide reasons for their decision-making.

Asking an AI to explain a decision is not necessarily providing explainability for that decision. What you are getting is the machine learning model's probabilistic choices of letters and words. These may form what may seem to be a plausible reason, but isn't a reason at all.

You can even simply tell a machine learning AI that it made a given decision and ask it why it did, and it will write something plausible that justifies that decision.

At a basic level I can easily explain how a machine learning AI, such as ChatGPT or Jurassic, arrives at a given output. It takes the input, parses it through a huge probability engine then write an output by selecting probabilistically likely words. 

For variability it doesn't always select the highest probability every time, which is why the same input doesn't always result in the same output.

However this doesn't explain how an AI makes a 'decision' - AKA prefers one specific option over other options. It does explain wht the same AI, asked the same 'question' (input), may produce diametrically opposed decisions when asked to regenerate its response.

The AI isn't interested in whether a decision is 'better' or 'worse' - simply that it provides an output that satisfies the end user.

There's a Chinese proverb that describes this perfectly:

“A bird does not sing because it has an answer. It sings because it has a song.”

This is why no current machine learning models can be explainable in their decision-making. And why we should not use them in situations where they are making decisions.

Now if you wish to use them as a way to provide information to assist decision-making, or to help write up the decision once it has been made, they have enormous utility.

But if you want explainability in decision making, don't use a machine learning AI.

Read full post...

View summary...