Surfing the Singularity : "Please hold for the next available agent"

-


Our imaginations, having been so stimulated by the "innovation trigger" of early interactions with ChatGPT and its LLM kin, having experienced the illusion of the algorithm reading your mind, we have now firmly entered into the period of inflated expectations. Any day now we expect a knock on the door to be informed by some HAL Junior that not only are we now out of a job, we've also got 20 minutes to evacuate the premise before its bulldozed to make way for another solar farm and data center. AGI is only just one product announcement away, or maybe two, but certainly three at most... 

Nose Deep 

There is a strong desire on the part of companies trafficking in AI to generate not just chatbot hallucinations but also customers for real business use cases, meaning revenue, and now. To do that we're going to need hardware, fast, lots of it, and gigajoules to power it. So AWS buys a new data center in PA adjacent to a 2.5GW nuclear power plant.[1] Not to be outdone Microsoft re-revs up Three Mile Island (albeit with a catchy rebranding laughable by 1970's standards), with 100% of the power going to their regional AI data centers.[2] 

Three Mile Island nuclear power plant, aka the "Crane Clean Energy Center".
  
After vigorous expectations the trough of disillusionment will soon follow. Already Microsoft hints that demand for AI-oriented chips is waning.[3] Practical, as you'll have a hard time getting them anyway - the data-center grade GPU chips on which AI computation rely are in short supply - NVIDIA via their TSMC outsource manufacturing partner is fully booked for Blackwell GPU orders for the next 12 months.[4] AWS has recently announced to customers (like me) new limitations on availability of certain NVIDIA GPU instances. (Consider also that AI competes with crypto for these scarce GPUs.) Intel suggests it will ship mass quantities of chips for AI-ready PCs and other mobile devices in 2025, but the stock traders are not yet buying it, with the stock currently fallen over 50% year-over-year. In the end, and as evidenced by the long term investments, we of course expect the march of techno-progress to continue, but in the short run, aligning expectations with reality may remain a challenge.

The August 2024 Gartner Hype Cycle for Emerging Technologies. 
Generative AI - weee! [5]

What does OpenAI say about all this? First, the desire to be non-profit has bumped up against the realities of scaling up the models. Will they continue to scale up, yielding better and deeper performance on the road to artificial general intelligence simply by scaling up, or will they hit a theoretical wall?[6] Sam Altman says succinctly: "there is no wall".[7] The nuclear-powered race is on, be it sustainable or not.

"Your wait time is now less than..."

But as we argued in the last blog [8], we don't need dystopia-inducing super-human AGI in order to make productive and disruptive use of artificial intelligence technologies - a domain-tuned artificial capable intelligence (ACI) is enough.[9] Or a collaborating set of them. 

OpenAI's strategic product roadmap is more than a little vague [10], but in theory after chatbots capable of basic reasoning comes the age of agents - think: allowing Alexa to auto-restock your pantry via a hotline to Bezos when it overhears you say you're low on sugar. Such "AI" does such a good job doing basic thinks like, oh I dunno, controlling the lights in your home now, what could go wrong?! Truth is, today's LLMs perform only so-so on standardized benchmarks, and while they improve all the time [11], the current state of the art is not yet ready to be trusted and at times seems like snake oil.[12]

Today's agents tend to be domain-specific and tailored to narrow purpose - Salesforce.com agents for common customer interactions, ServiceNow agents helping the human agent perform repetitive or summary tasks in handling case loads, but not replacing the human.[13,14] Google Gemini can add events to your calendar, help you plan travel, but is not yet trusted to actually borrow your credit card and book it. Keeping the human-in-the-loop will remain for now, as a stepping stone to full automation.

If you visit agent marketplaces like Agent.ai or SwarmZero.ai, you'll see on the order of hundreds of agents available to handle what are largely small, mundane, and repetitive tasks. There are similar domain agent marketplaces on OpenAI's site, Anthropic's, GitHub, Hugging Face, and more. Let's go along with the current norm and define "assistants" as gaggles of agents loosely collaborating to accomplish more complex tasks, perhaps as part of a hybrid AI-human team or for some cases ultimately on behalf of the entire organization, and yet, still not requiring full-on AGI. (Consider what just one techno-savvy entrepreneur with a diverse collection of AI auto-orgs might do.)

The missing elements are reliable agent accuracy, which yields trust, and the hardware and power to run it all. Trust, unfortunately in the near term, may play second fiddle to profit, as the AI snake oil is sold to companies and governments and ultimately end users, most of whom barely understand it.

In fact, the scientists themselves barely understand it. The deep learning networks that power today's LLMs are generally black boxes, layers upon layers of neural networks, numeric weights and matrix computations, where its pretty difficult to tell where any given word, image fragment, or concept is held in the vast space of the model, and how with various feed-forward and back-propagation processes in the network it is used in computing responses.

A GPT model formed by combining successive attention and neural net layers. Input comes in at the left, and its black boxes all the way down.[15]
   

Black box or not, as Sam Altman says, deep learning just works.[16] Sort of - AGI is unlikely without strong ontological and reasoning abilities and a tactile understanding of the physical world.[17] And deep learning itself is not without its problems. If the training data is biased, so will be the results. Trainers have to be alert to overfitting the model to the training data in a way that makes the model ineffective on new data. And implementors need better tools which help introspect and observe the model to provide verification, to illuminate the black box. Until then, any technology which cannot be understood is indistinguishable from magic.

Hell-o Operator

AI is a broad term, encompassing many technologies, machine learning being just one of them, and deep learning based on neural networks being an even further niche. In many ways, given the black box nature of the solution, AI has become a substitute word for "automation", and/or "program", or "algorithm". And the ill-defined AI landscape is moving fast. Twelve months ago the buzz was about the emergence of the "prompt engineer" role in lieu of computer programmers, and today, not so much. Instead we now have thin but actionable (i.e. product-oriented) definitions like "agent" and "assistant" and a new suite of tools and cute icons to put on enterprise architecture diagrams. This is not to even mention the human and organizational impact of new agent-based workflows characterized by iterative, non-waterfall business processes - not something well understood or appreciated outside of software engineering circles.

In this turbulent time, with vendors leapfrogging each other's capabilities and performance, there is no and cannot be any real standardization, no agreed abstractions on which to base a unifying orchestration layer. Move fast and break things, fix them later if they live long enough. Let the prototype knowingly become the short-lived product, and iterate, maybe. Think: sqrt of web time. Think: ChatGPT + IFTTT.[18] That is not an enterprise IT solution, nor one manageable for most individuals. That is a fine mess.

Thankfully, we'll soon have AI assistants to fix it for us. 

Readers unfamiliar with the nuclear accident at Three Mile Island in 1979 can read the summary here: https://en.wikipedia.org/wiki/Three_Mile_Island_accident
[9] "The Coming Wave", M. Suleyman, Crown Pub., 2023
[10] https://www.theneurondaily.com/p/openais-leaked-agi-roadmap
[11] 12 Days of OpenAI, Day 12: https://www.youtube.com/watch?v=SKBG1sqdyIU 
[12] "AI Snake Oil", Narayanan & Kapoor, Princeton U. Press, 2024
[13] https://www.salesforce.com/news/stories/einstein-sales-agents-announcement
[14] https://www.servicenow.com/standard/resource-center/data-sheet/ds-virtual-agent.html
[15] https://miro.medium.com/v2/0*-8c-MXmNvcvTLdHH.png We recommend the following video for those not familiar with this architecture:  https://youtu.be/KJtZARuO3JY?si=Muq2xRdSTaa9LMXb
[16] https://ia.samaltman.com/
[17] Yann LeCun on Lex Fridman podcast, https://www.youtube.com/watch?v=5t1vTLU7s40
[18] https://ifttt.com/chatgpt