AI Top-of-Mind for 9.24.24 — AI Adoption Framework
Today: McKinsey on AI adoption, Apple Intelligence and UI-JEPA, avoiding hallucinations, the uncanny valley, and are we heading to 1984?
Top-of-mind, some great guidance from ‘McKinsey’ on how to best adopt AI within an organization. From their report ‘A data leader’s operating guide to scaling gen AI’ showing the phases of IT evolution:
And this is tied to observations by Enrique Dans on AI assistants. He writes:
It is also worth pointing out that not too long ago, managers asked administrative staff to carry out tasks such as typing texts or preparing presentations, then took on these tasks themselves, and now are going back to entrusting it to a third party, except this time, it’s an AI assistant.
Next, if you’ve been following Apple’s Intelligence announcement, you may be wondering about the delays in its actual release. Ignacio de Gregorio offers some thoughts on the ‘why?’ He dives into what is termed ‘UI-JEPA,’ and from his post:
· The key difference between JEPAs and Generative models is that they make predictions in the latent space, avoiding the need to predict unnecessary things.
· In other words, it doesn’t predict Steph Curry by generating every single detail about him (something LLMs would do), but it predicts the vector that represents Steph Curry, ensuring it only needs to predict the essential components of what makes ‘Steph Curry’ different from the rest.
The above only touches on his full analysis, and as a deeper dive, the Apple paper is a start. One example is the latency advantage of Apple’s approach, key when expecting a timely response:
Still on models, Will Lockett penning in ‘Predict’ has some very critical views as to the essence of o1-preview. He notes:
It just automates a relatively successful pre-existing prompt technique, known as “chain-of-thought” prompting, on top of an AI training method that has existed for even longer.
His post includes background on ‘alignment,’ how o1 and GPT-4o are trained differently, and why OpenAI make the following statement:
“Strawberry” (o1) does better at answering questions involving science, programming, and mathematics than ChatGPT-4o, but regarding many other queries, ChatGPT-4o does better.
Continuing on the same thread, the ‘NY Times’ looks at the evolution of LLMs and their math abilities that include checking for correctness. The article dives into the company ‘Harmonic’ and their ‘Aristotle’ model that is said to offer better verification and protection from hallucination than OpenAi’s o1.
Onto the graphics front, and you’ve probably heard of the ‘uncanny valley’ as it relates to CGI. Nir Zicherman asks the question — is the uncanny valley even real — and offers some thoughts. He includes a few potential outcomes, including where technology may reduce or even eliminate our responses:
With that in mind, check out Jim Clyde Monge’s tutorial on how to create a great brand video. He dives into three techniques and includes links:
- Generate AI videos offline using your own machine:
- Use AI video generator platforms:
- Hire AI experts to create videos for you:
He adds:
According to Deloitte’s 2023 Creator Economy in 3D survey, 94% of brands working with content creators are either currently using or have plans to use generative AI. 55% of those brands are already using generative AI in some capacity, and the other 39% have plans to use it within the next year.
And to close, are we heading to a ‘1984’ scenario? Will Lockett in ‘Predict’ pens:
· Larry Ellison, the billionaire founder of the IT giant Oracle, thinks this sounds like a great idea. In fact, he wants to use AI to make “1984” a reality and turn the US into a surveillance state, with him, unsurprisingly, reaping the profits.
· He stated that AI is on the verge of ushering in a new era of mass surveillance and that Oracle is ready to serve as the technological backbone for such an application.
I’d add that there are also elements of ‘Minority Report’ in play here.
