So many interesting articles in the AI space this week. Here are three ideas that caught my eye.
It’s currently hard to make money with LLMs
Less than a year in to this Generative AI revolution and we’re starting to see even the really large tech firms (with their own cloud compute on demand and at better than wholesale prices) struggle to find a price point for their LLM enabled products that breaks even. The Wall Street Journal (and hence a paywalled article) seem to have been the first to break the story that Microsoft, Google and Adobe are losing money on tools like GitHub Copilot (an AI powered coding assistant for software engineers) and Adobe Firefly (gorgeous and almost magical AI enhanced still and video editing).
AI often doesn’t have the economies of scale of standard software because it can require intense new calculations for each query. The more customers use the products, the more expensive it is to cover the infrastructure bills. These running costs expose companies charging flat fees for AI to potential losses.
Tom Dotan and Deepa Seetharaman, The Wall Street Journal
For instance, GitHub Copilot is a flat fee of $10 per user per month. Average costs are apparently running at $20 per user per month. What’s the fix? Given that we’re all anchored on $10 per month now they can hardly put the prices up. And ‘pay per call’ models or slowly down the service once some limit is hit (which apparently is how Adobe is controlling costs with Firefly) seem to defeat the purpose of trying to hook developers on a code writing assistant. The most common strategy appears to be using models that are computationally less expensive but still up to the task - effectively the “don’t use a Ferrari to drive to the corner shop” approach.
All of these reasonably public learnings from the Copilot experiment have likely fed into the $30 per user per month price tags for the Google and Microsoft personal assistant products.
What does this all mean for startups utilising Gen AI APIs in their SaaS products? Watch out, and do your compute cost calculations carefully.
And remember that once folk wake up to how much using Gen AI really costs, bad actors will go looking for insecure services to farm. So even if you think your costs will stay low because your users will be using short prompts, remember to put on your black hats and do some scenario planning in case you get hacked - not for credit card details but for bootlegged compute. And yes, something similar may have happened to me in a past life…
Deepfakes for political gain
According to Gizmodo and Freedom House, politicians the world over are already using deepfaked video and audio to introduce ‘alternative facts’.
Accounts affiliated with the campaigns of former president Donald Trump and Florida governor Ron DeSantis, both seeking the Republican Party’s nomination for the 2024 presidential election, shared videos with AI-generated content to undermine each other’s candidacy. One video included three fabricated images of Trump embracing Dr. Anthony Fauci, who led the federal government’s COVID-19 response and remains deeply unpopular among critics of pandemic mitigation measures. By placing the fabricated images alongside three genuine photos, the video muddied the distinction between fact and fiction for Republican primary voters.
The Repressive Power of AI, Freedom House 2023 Freedom on the Net report
For me, even more disturbingly than US politicians smearing each other within their own party (which is pretty disturbing given the outsized impact the US still has over the rest of the world), was this story on state media outlets in Venezuela using social media to distribute videos depicting faked news anchors from a nonexistent channel spreading pro government messages.
I can imagine how insidious this type of content would be. You don’t even realise you’re watching it, it’s just on in the background in the hairdresser or a pub. And bingo, you have a vague but convincing feeling that the government really is doing a great job these days.
To quote the Freedom House report “Even if deepfakes are obviously fabricated or quickly exposed as such, they still contribute to a decaying information space.” Sobering stuff.
The power of lobby groups in shaping AI policy
In Australia this week, as a nation we sadly chose not to give our indigenous people a voice in our parliament. But if you have (a lot of) money, you don’t need to change the constitution in order for your opinions to shape policy, you just form a lobby group.
A sprawling network spread across Congress, federal agencies and think tanks is pushing policymakers to put AI apocalypse at the top of the agenda — potentially boxing out other worries and benefiting top AI companies with ties to the network.
How a billionaire-backed network of AI advisers took over Washington, Politico, October 2023
The Politico article is a good read and, while you have to be sceptical and remain curious about everything you read these days, it does a convincing job of highlighting how money from companies and individuals who stand to benefit from AI licensing regimes and regulation that focus on the existential risks of AI over the present harms, seems to be flooding Washington and funding education session for influential lawmakers.
Here’s hoping for some sunshine in Victoria over the coming week. Wishing you all a peaceful and productive weekend.