From zero to hero: tension at the heart of AI, a better way to measure AI progress, how AI saved me $12,000, why thinking AI is the new generative AI

Written by Fola Yahaya

Thought of the week: the tension at the heart of AI

Brought up on a diet of film classics, I used to fantasise about an alternate reality in which I was walking down the red carpet en route to receiving the Palme d'Or for my magnum opus. However, lacking any discernible creative talent and acutely aware of the terrible economics of film-making, I chose to set up a translation agency – ironic, given the devastating impact AI is now having on that particular industry.

So, I was genuinely excited by the creative possibilities that text-to-video generators, such as OpenAI’s Sora , offer to the creatively challenged. While Sora has only been available to a select band of ‘creators’ due to ethical considerations, competitors such as Runway and, more recently, a host of Chinese companies such as Kling and Hailuo have brushed aside such concerns and opened their tools to the masses.

With some unexpected time to kill over the weekend, I tried out Hailuo (formerly Minimax), a Chinese text-to-video generator. On a scale of zero to 100, my film-making skills hover around zero. However, with a few mouse clicks, Hailuo seemingly managed to bump me up to a solid 20.

Struck down with blank page syndrome, I asked ChatGPT to recreate some fighting scenes in the style of Martin Scorsese’s black and white masterpiece, Raging Bull. As a topical twist, I asked it to use Vladimir Putin and Volodymir Zelenskyy as the main protagonists. Hailuo then promptly ;-) created around 15 six-second scenes of cinematic weirdness.

The video below was then put together with Elevate, a web-based video editor, and Udio, an AI music generator.

Okay, my creation is clearly not going to get me to Cannes. The lighting is weird, Putin does some odd things with his hands, and the characters are all over the place. And yet, for an hour’s work, it isn’t awful. Moreover, I’m confident that if I had a good week and an ounce of talent, I could learn how to create the right prompt to ensure a decent story, maintain continuity, and use interesting camera angles, ultimately levelling up my film-making skills to a solid 25. (See the film below, for example, which was clearly made by someone who knows what they’re doing. Yes, it’s still evidently AI-generated, but hey, it’s not bad for five days of work and an outlay of apparently less than $50).

So what were my key takeaways from my weekend of playing with Hailuo?

1. Budgets will tumble. Instant and good-quality AI imagery and code is already impacting graphic designers, photographers and coders. AI also automates and reduces the cost of a whole range of intermediate steps in the creative process. So expect a race to the bottom, with agencies competing to secure work and while combatting clients’ perceptions about cost, time and quality.
2. Competition in AI is crucial. The Chinese and the open-source movement provide a vital counterpoint to heavily censored, US-centric tools. While guardrails around what users can create are necessary, diversity in AI development helps push the boundaries of innovation.
3. Bias and lack of training data are significant issues. These systems are trained on a fairly narrow cannon of video data. This means that well known figures like Vladimir Putin can be rendered accurately, whereas minorities are often rendered as an amalgamated pastiche.
4. AI video is (for now) incredibly challenging to execute well. Consistent output is nearly impossible to achieve at the moment. Even more concerning, a six-second scene that Minimax generated included a random R-rated element that was not part of my prompt. The value will clearly come from automating intermediate steps such as colour grading.

My experiments with Hailuo reinforced my conviction that we all need to lean in, play with AI toys and learn how retool our processes for the near future.

Moreover, even though film-making, especially in developing countries, is a significant part of my business, I’m not rethinking my business model.

This is because even as AI transitions from toy to tool, there will always be a significant market and premium for authentic stories well told.

How an AI lawyer saved me ￡10K

When we pivoted a few years ago from being a translation agency to a full-service communications agency, we made the classic mistake of signing up for a bunch of tools that claimed to help us manage media relations for our clients. One of them, Cision , offered access to a global database of journalists who, on the surface, we could spam with client press releases. Like many things in business, tools that claim to automate very human processes are rarely effective, and we cancelled about 10 months into an annual subscription for a service that had been, frankly, useless.

Sadly, we (or rather, I) hadn’t read the small print about a 90-day notice period, and Cision gleefully sent me a bill for a $15,000 annual subscription that they had not so kindly auto-renewed. Very long story short, both parties settled on a $2,500 payment. Throughout the process, AI was critical. It:

• wrote the ‘legal’ letters
• provided case law on unreasonable terms
• gave me compelling advice that I didn’t have a leg to stand on, as the terms and conditions were clear in the contract
• advised me on my negotiation tactics
• saved me the ￡1,000 or so in legal fees I would have had to pay to get the same outcome

The lesson I learned from this wasn’t just to always read the fine print – I'm not a lawyer, so it would all be Greek to me – but rather to use AI. That's exactly what I did when I had to sign a dense and complex cloud services contract this week. I scoured the OpenAI GPT Store for a contract reviewer, uploaded the contract, and sat back while the robot did its beautiful thing. See the outcome in the image below.

I duly sent these back to the provider, who modified the terms and ultimately reduced my risk. The specific GPT I used was called Contract Reviewer, but the OpenAI GPT Store has 100,000 mini ChatGPTs trained in specific domains and functioning as emergent agents.

Better still, you can create your own GPTs by uploading your documents, favourite books or curricula – with the obvious caveat that, irrespective of what OpenAI promises, your data may enter the public domain. Configure your ChatGPT settings with your particular parameters, such as jurisdiction, company or personal details, and preferences.

By creating your own mini agents, you’ll experience the shift from AI being a mere toy to a truly useful tool that can add real value.

From tool to tyrant: where AI is now and where it’s heading

I find Sam Altman’s five-level typology of AI progress vague and hard to relate to. Instead, I propose a much more relatable, four-level scale of progression:

1. Toy – Most companies and people are still ‘wasting’ time, money and effort playing with ‘vanilla’ AI toys. They know that AI can be useful but are unsure how to quantify and create processes that add value.
2. Tool – AI becomes a useful co-pilot that, once instructed, does your bidding and saves time, money and effort. Think of an AI lawyer or a co-pilot that sits within your desktop application, able to format documents, analyse your data and interact with your applications. Also known as AI agents, this is the value most AI tool vendors, such as Microsoft and Salesforce , are currently pushing.
3. Taskmaster – AI systems act as managers, assigning tasks to workers or other AIs and monitoring their completion. Similar to Uber drivers being given new work, AI taskmasters are ruthlessly efficient managers that coordinate our activities.
4. Tyrant – AI decides what, how and why things should be done, self-orchestrating resources to achieve a desired outcome. Humans remain in the loop depending on the area and degree of regulation. Examples include an autonomous factory producing goods, an unmanned nuclear power station, an AI surgeon performing operations, or an AI monetary committee setting interest rates based on real-time data.

Though the term 'tyrant' is clearly negative, it effectively conveys the inherent risks of allowing AI to operate processes autonomously.

AI video of the week

What we’re reading this week

• How AI Could Transform the World for the Better – a long (15k words!) and rambling essay by Dario Amodei, the head of Anthropic (maker of Claude). Following in the footsteps of Sam Altman, another lofty AI vision from the tech ivory tower, that imagines a future in which machines care, understand and prioritise human well-being. Nice in theory, but totally falls apart in the face of the profit imperative.
• Adobe adds AI tools that build 3D scenes, animate text, and make ‘distractions’ in images and video disappear.
• More silicon snake oil – after an underwhelming robotaxi reveal, Elon Musk gets caught out using humans to control his Optimus robots at a key Tesla event.
• Big Tech continues to go nuclear with another tech giant, Google, signing a nuclear power deal to ensure its data centres can keep firing.

Tools we’re playing with

• Hailuo – the closest thing we have to Sora, and for free.
• ChatGPTs GPTs – a wealth of (largely) free vertical AI systems that can save you time, money and effort.
• TANGO, an AI system created by researchers at the University of Tokyo that generates realistic human speakers, movements and gestures to match audio input.

That's all for this week. Subscribe for the latest innovations and developments with AI.

So you don't miss a thing, follow our Instagram and X pages for more creative content and insights into our work and what we do.