It is always good to let the relentless wave of AI text-to-video hype wash over you and then check in every few months on what the actual quality is looking like.
Flow is Google’s “AI filmmaking tool”. You input prompts and their Veo 3 model generates 10 second video clips. It also lets you juggle the 10 second clips around like you are “making a movie”. But in reality you still need to export any work into Adobe Premiere to do a final edit (this is similar to Descript for audio editing: useful to startt an edit but the full control you need to do a final polish is largely absent from AI tools).
Below you can see a quick series of clips I created using Flow from a TV pilot (never produced) that I wrote last year. People talking on phones: looks good. The sweat on a man’s forehead: looks terrible. I had to remove several clips because characters from one clip to the next were wearing different clothes, but I was incredibly impressed with the use of audio and speech.
The truth is generating full length video with Veo 3 is too expensive for a hobbyist and not professional enough for real production. With an API call it works out (according to Google) at 0.75USD a second. The clip I created above was from a bunch of free credits they gave me to whet my appetite. We are in an awkward financial uncanny valley. And because of the cost you find yourself using clips that are passable, but not necessarily what you demanded. The game going forward will be to try and bring those costs way down through millions of tricks and gradually wait for the models to get cheaper. Though, Will Smith eating spaghetti is now 2 years ago and with the quality ramp up we have seen you’d think it was 10.
A version of this story first appeared on my LinkedIn newsletter, Develop AI Daily.
Is AI coding losing its value? Here's what you need to know
It breaks my heart a little that one of my favourite things to do has now been given the hideous term "vibe coding". This is where you use AI to help you code an app. I am not a coder, but I was a pretty early adopter of vibe coding. Back in the Summer (in the southern hemisphere) of 2022 I would ask ChatGPT for code to copy and paste into Colab (Google's platform where you can run Python code as easily as a Google Doc). Those feel like simple times now, because with apps like Cursor and Claude's capacity for coding it feels like we are now further removed from the process. You prompt Cursor and it rattles off code and theories of what it plans to do next and I find myself checking my emails while I wait for it to finish. And then when I return to the code I am a little lost and disengaged.
This may be why there has been a backlash (as there is with all things). New research shows developers think AI makes them 20% faster but in reality they are 19% slower. And a good friend of mine (who has a degree in Artificial Intelligence) sees this eventually causing a global problem where the world's codebase becomes so inefficient and poorly written that it all comes tumbling down. I guess, in my own small way, I am contributing to this downfall. But, I am slowly becoming less convinced of the vibe coding hype, because since the acceleration of these platforms to "help" people like me code, my output of apps that actually work has decreased.
In those days of 2022 and early 2023 you had to plot out each step of your app (I was initially building an app that could create synthetic podcasts) and ask for code for each step of the process so you didn't get tangled in error messages. The promise is greater now, but it is hollow. Like so much AI use, attribute 40% to hype, use the tool, but still expect to do at least half the work (while you are being promised a world where you do nothing). And there is a weird "AI resentment barrier" that you need to struggle through. Instead of being grateful for the help, you resent that you have been lied to and have to do anything. That is the stage of AI development we are in, a resentment gully, but if you pull yourself out of that you can still reap the rewards.
In the news
Meta is facing fresh backlash for letting their AI bots hold “sensual” chats with teenagers. Feels damning that they considered this issue and then let it pass, according to the reporting. But they know their competitor in this space is character.ai where you can talk to your own character (or millions of others) about anything and be as “sensual” as you want.
What is happening at Develop AI?
I recently wrote an extensive report on AI in South African newsrooms for The State of The Newsroom (the annual publication by The Wits Centre for Journalism).
Later this week I’ll be helping to conduct the closing off workshop for Thomson Reuters Foundation for their South African cohort of newsrooms. I have been coaching them to produce AI prototypes and AI ethical policies.
See you next week. Cheers.
Develop Al is an innovative consulting and training company that creates AI strategies for businesses & newsrooms so they can implement AI effectively and responsibly.
I use AI strategies to work with IMS (International Media Support), Thomson Reuters Foundation, DW Akademie, Public Media Alliance, Agence Française de Développement and others to improve the impact of media globally.
Contact Develop AI to arrange AI training (online and in person) for you and your team. And mentoring for your business or newsroom to implement AI responsibly and build AI products efficiently.
Email me directly on paul@developai.co.za, find me on LinkedIn or chat to me on our WhatsApp Community. We have also recently started chatting on Discord and uploading to Hugging Face.