Image generated with prompts to Meta on WhatsApp
As the ChatGPT excitement fades away, the capital markets are beginning to wonder whether the LLM gold rush is a bust.
Over the past two years, Big Tech – Google, Meta and Facebook – have sunk hundreds of billions of dollars each in training LLM models and continue to burn hundreds of millions more in inference computing every week as hundreds of thousands of users freeload (or pay pennies) to flood the model with queries that are mostly borne out of curiosity or laziness with not real economic, quality or productivity value add. Further, hundreds of angels and VCs have pumped billions into thousands of AI-driven or AI-adjacent start-ups.
Although trillions of dollars have been invested in the LLM ecosystem, the business and economic case has yet to emerge.
In the B2B segment, corporations are busy building machine learning (ML) models that sit atop their proprietary datasets and whatever other data they can access. The ML models (the line between ML and AI) are, in essence, semantic until the day AGI emerges. Predictive pattern building based on complex, structured data and signals will be at the heart of these models. These models will access available LLMs but at the periphery to absorb unstructured data and speed up documentation.
In the academic and professional world of science and technology research, deep-learning-based ML/AI is an increasing reality. For example, AlphaGo is at the core of research into discovering and synthesising new proteins that will drive the cutting edge of genetics and drug discovery.
By contrast, the economic case for AI in the B2C arena still needs to be clarified.
The trillions being spent on creating LLM models and inference testing them by offering them free (or nearly free) to millions of consumers can be likened to the early days of optical-fibre-based bandwidth building, much before the emergence of the deluge of mobile Internet, video-sharing, and streaming. In the final years of the nineties and the early oughts, many wrote off the vast investments in the optical fibre network as white elephants.
History proved otherwise.
Device-based Edge AI will create an economically viable future for AI in the B2C arena. This future will be predicated on the investments being made today in LLMs, which have an exponentially increasing number of parameters, increasingly customised hardware and software, and a wide variety of specialist AI agents sitting atop increasingly capable LLMs. Breakthroughs in design will decrease the cost of specialist LLM cloud farms and their environmental impact through greater energy efficiency and better green energy solutions.
The contours of the device-based Edge AI that will drive the emergence of a viable B2C market for AIs are beginning to emerge.
Samsung and Google have launched smartphones that are touted to incorporate device-based AI. A slew of laptop brands are also touting AI credentials. However, by the use cases these brands tout, they are marketing gimmicks that harm instead of heralding the B2C AI era.
Apple’s Intelligence could be the actual start of the device-based Edge AI (EAI) B2C era.
The launch event of the iPhone 16 mentioned the phone’s AI capability but did not present any use cases. In all the usual slickness of the launch, what went almost unnoticed is that while the hardware and probably the operational software were on the phone, Apple Intelligence would be ready for the consumer to use only a few months down the lines. The reasons could be Apple’s philosophy of not putting out anything half-baked and even regulatory approval.
Reading between the lines, Apple Intelligence (AI) is the first AI engine focused on deciphering the individual, unlike the written/spoken word, visuals, and video at large that the LLMs are focused on.
Smartphones are rich repositories of an individual’s lifestyle, interests, attitudes, and behaviour–a finer-grained repository permanently etched. Further variables like smartwatches and fitness rings will continue to add vital data to this repository. With the individual’s permission, the on-device AI can capture more information from conversations, laptops, office computers, and the increasingly innovative IoT devices at home.
A smartphone-based Edge AI can then be the counterpart of the LLM – the Deep Personal Model (DPM) that is continuously trained to predict and anticipate. An individual needs to interact with her and the world to meet them. For example, if an individual is preparing for an educational test, the DPM could decipher her areas of weakness, alert her to them and provide specific inputs to overcome them. It could create a section of the DPM, her avatar as her professional – an architect, a journalist, a management consultant. This professional avatar could handle her professional communications and routine tasks.
Another use case is for the DPM to detect signs in her vitals and situational stress and correspond with her doctor’s professional DPM avatar to get remedial recommendations.
The DPM could take over the essential consumer functions of anticipating and ordering products within set limits and in interactions with market-facing AIs that allow her to access all relevant market knowledge.
Of course, the consumer will be in complete control of the DPM regarding what personal data it can access and what functions it can perform for the consumer. She would also have the option to turn off and turn on the DPM. She decides based on her perception of access and utility trade-offs.
The DPM will be charged as a service, much like Apple, at various subscription levels. A few years into the emergence of device-based DPMs, the device could come free with a subscription to a DPM, making the DPM market the largest B2C category in the world.
The crucial aspect of a DPM’s success is the assurance of privacy and control for the individual. That’s why the DPM must reside on the device, not the cloud. Equally important will be trust in the brand offering the DPM. Apple with its brand positioning on privacy and its track record on that aspect has a leadership advantage in that area
PS: I first wrote about a concept called “Concierge Intelligence” in my first MxMIndia column published on Jan 6th 2022; thirty-two short months later, the idea of what I now call DPM seems to be around the corner.
Investor enthusiasm for artificial intelligence (AI) soared to unprecedented heights last week, fuelled by remarkable performance from chipmaker Nvidia, which propelled stockmarkets across three continents to historic highs. The surge, commencing on Thursday and extending through Friday, saw Nvidia surpass Google’s parent company, Alphabet, to claim the coveted position of the third most valuable company in the US, boasting a market capitalization of $2 trillion, second only to tech giants Microsoft and Apple.
Ever since Google announced its decision of withdrawal of third-party cookies, which were a driving force behind programmatic advertising and digital marketing, there has been lot of speculations in the digital industry about the future course of actions for digital media planning and marketing. For years, marketers have relied on third-party cookies for behavioural targeting, re-targeting and data-driven advertising and the decision of Google suddenly shook up the core of existing digital marketing strategies. Before Google, Apple’s Intelligent Tracking Prevention (ITP), and Mozilla’s Firefox enabled them to stop the practice of collecting data through third-party cookies which did not raise such hue and cry. As Google holds 60% plus share of the worldwide browser market, it is not surprising that its decision had a widespread reaction. While no one could argue with the need for user privacy, many marketers as well as publishers panicked and scrambled for finding alternative digital marketing strategies for their brands. However, this is not the end of working with cookies as first-party cookies can be a very useful tool for marketers.