Why you shouldn’t get ChatGPT to do maths, how to get AI to write in your voice, Adobe's AI video editing, and the 'Mechanical Turks' driving AI hype

Thought of the week: why you shouldn’t get ChatGPT to do your maths

My son is in the midst of preparing for his GCSE maths. This means treading the fine line between a) worrying about whether he’s working hard enough (whilst finding novel places to hide his PlayStation) and b) reassuring him that at the end of the day it gets dark and it will all be fine. During a rare revision session, he shouted that he just didn’t ‘get’ something called ‘position vectors’, and in desperation he asked for my help, then skulked off to hunt down his PlayStation.

Given that it’s been almost 40 years since I resolutely struggled with GCSE maths, and ever the AI advocate, I suggested we “ask ChatGPT”. Never one to miss a learning opportunity with my kids, I asked him to co-construct this P.O.W.E.R prompt (see here for more on this):

Long story short, ChatGPT provided an almost poetic explanation that inspired my sceptical son to rework his original approach and answer so that it matched the bot’s. So far so transformative. However, according to the Cambridge exam board, not only was ChatGPT wrong, but the explanation was just plain idiotic.

Yet, it’s common currency that the models underlying LLMs are pretty thick when it comes to solving maths problems because they are trained on words (i.e. large language model). Expecting them to be good at maths is like assuming librarians are brilliant at calculus just because they are heavy readers.

Despite this, a growing number of people are outsourcing their thinking to a black box digital oracle that is unaccountable, notoriously bad at basic maths and loves to make stuff up. No doubt AI will get better at this stuff, but we need to be mindful that current consumer AI tools are still very limited, which means your kids still need to learn to do their homework, and we still need to write our reports and do our taxes.

Adobe offers a glimpse of the (near) future of AI video editing

Desperate to stay relevant, the global leader in design tools, Adobe, is hyping up its AI creds. This week, it was showing off the integration of Adobe Firefly into its video editing software, Premiere Pro. New features include the ability to replace things in videos with simple text prompts and a generative extend feature which lets editors add filler frames to extend shots.

What's really interesting is the as yet unreleased option to plug in Sora and other AI generation models, such as Pika and Runway, to ‘extend’ a scene. This means that filmmakers may need to shoot less B-roll (extra footage that adds context or meaning to a scene), and that editors will eventually be able to ask an AI to fill in the blanks.

The trend is definitely moving towards getting AI to do more of the heavy lifting in video editing, which will further impact the livelihoods of videographers and stock footage creators who rely on their unique content to stand out.

How to get ChatGPT to write in your voice

From the overly formalised language to the rookie mistake of not telling your AI to use US or UK English, you can sniff out ChatGPT content a mile off. Yet for near instant production of inconsequential guff, AI reigns supreme. However, sometimes needs must, so to get your content to pass the sniff test, use these tips to ‘fine-tune’ your favourite AI generator to write in your voice.

1. Feed it your writing samples. Upload a substantial corpus of your own writing. Include a diverse range of your work, such as blog posts, emails, social media updates and more. The more examples you give, the better the LLM can analyse and learn your writing style, including your tone, word choice, sentence structure and pacing.

2. Use few-shot learning with contrasting examples. Few-shot learning means training the model with a small number of representative examples. To make this technique even more effective for style training, provide contrastive examples that highlight the key differences between your writing style and others. For instance, you could include a paragraph written in your voice and a similar one written in a generic style, clearly labelling which is yours. This helps the LLM identify and prioritise the elements that define your unique voice.

3. Engage in interactive fine-tuning. Instead of just feeding the LLM static examples, try engaging in a conversation with it to provide real-time feedback and guidance. Start by giving the model a writing prompt and letting it generate a response. Then, offer specific feedback on what it got right and wrong in terms of capturing your style. Repeat this process iteratively, allowing the LLM to learn and adapt based on your input. This interactive fine-tuning approach can be particularly effective for teaching nuanced aspects of your voice.

4. Experiment with different prompt formats. The way you structure your prompts can significantly influence the LLM's ability to understand and emulate your style. Experiment with various prompt formats, such as providing explicit instructions ("Write this in the style of [Your Name]"), using a question-and-answer format, or even roleplaying as yourself. Find the prompt style that yields the best results in capturing your unique voice and stick with it for consistency.

5. Train on different types of content. To create a well-rounded model of your writing style, train the LLM on a variety of content types. In addition to the examples mentioned earlier (blog posts, emails, etc.), consider including scripts, fiction, poetry or any other forms of writing you engage in. By exposing the LLM to a diverse range of your content, it can develop a more comprehensive understanding of your style and apply it flexibly across different contexts.

Why it’s so hard to separate hype from reality

AI start-ups are having to fight tooth and claw for an increasingly dwindling pool of venture capital money. As the hype cycle moves on, it’s unsurprising to see the lengths that companies offering little more than a slick interface slapped on top of one of the major LLMs will go to. A great example of this is Devin, ostensibly an AI coder that could autonomously code and, a month ago, was the hottest AI start-up since, well the last one. However Devin, like many before it, was revealed as being pretty useless without human intervention.

Similarly, Amazon recently abandoned its 'Just Walk Out' stores which promised to save customers the massive inconvenience of queuing for a few minutes by using AI and sensors to charge them invisibly. In reality, the ‘AI’ was a team of humans in India manually labelling and charging people for the stuff they put in their carts. These examples, and a raft of other failed or underwhelming AI launches, suggest that many supposed intelligent apps are no more than the modern equivalent of the infamous Mechanical Turk, an ingenious chess-playing machine constructed in 1770, which appeared to be able to play a strong game of chess against a human opponent, but that in fact concealed a human chess master hiding inside.

With global AI investments and valuations forecast to reach the high 100s of billions, the stakes are high, but be wary of the hot new thing.

Are we on the cusp of a new era of humanoid robots?

A day after announcing it was retiring Atlas, its hydraulic robot, Boston Dynamics has introduced a new, all-electric version of its humanoid machine. The next-generation Atlas robot is designed to offer a far greater range of movement than its predecessor.

“We designed the electric version of Atlas to be stronger, more dexterous, and more agile,” the company said in its press release. “Atlas may resemble a human form factor, but we are equipping the robot to move in the most efficient way possible to complete a task, rather than being constrained by a human range of motion. Atlas will move in ways that exceed human capabilities.”

The new version has been redesigned with swivelling joints that the company claims make it “uniquely capable of tackling dull, dirty, and dangerous tasks.” Though Boston Dynamics has been releasing scary robot videos for the last 10 years without much to show for it, I sense that we are getting very close to wider deployment of useful robots for three reasons:

1. Robots are getting cheaper

You can now get a cheapish domestic robot like Hello Robot's Stretch for $25k. Open-source projects like Stanford's Mobile ALOHA now allow anyone to download a kit and build a usable robot with off-the-shelf components.

2. Plug and play brains

AI models like Google's RT-2 and Covariant's RFM-1, and imitation learning techniques are enabling robots to quickly learn new skills, serving as 'general-purpose robot brains' that adapt to their environments.

3. More data allows robots to learn more skills

Initiatives like the Open X-Embodiment Collaboration are pooling robot demonstration data, enabling models like RT-X to learn skills 50% more successfully across different robots compared to individual lab efforts.

Given the above, and hype cycles aside, expect humanoid robots doing things that humans shouldn’t really being doing (like mine clearing or waste collection) within a few years.

Other interesting AI news this week

• Stanford University releases its annual AI Index Report. Nothing earth-shattering, but useful if you need to wave some stats/charts at people (like this one)!

The Humane AI Pin, a $700 over-hyped AI wearable slated by practically everyone, but critically savaged by YouTuber Marques Brownlee (20 million followers), as the 'Worst Product Ever Reviewed'.

• The World Economic Forum outlines nine ways AI is being used to fight climate change.

• More bad (and painfully obvious) news for translators, as the Guardian reports that more than four in 10 translators said that their income has decreased because of generative AI.

• Wall Street banks are also likely to reduce salaries and hiring for junior roles by up to 66% due to AI advancements, as reported by The New York Times.

• The end of influencers? Despite earning a record $16 billion in revenue in 2023, TikTok is looking to further cut costs by helping customers promote their products via AI-generated influencers.

Tools that we're playing with this week

• Udio: Oh, how fickle am I? One week I’m crowing about Suno, next minute it’s another new kid on the block. Udio is a music generator that is currently in beta, and generously allows you to create up to 1,200 songs per month. It has cool features such as lyric generation.

• OpenRouter: A useful tool for comparing the output of a single prompt across different models.

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.