add ai reservations, predictions

src/content/kb/tech/ai-reservations.md

@@ -0,0 +1,82 @@
+---
+title: Where are the category-defining AI products?
+description: Certainties, reservations, and predictions for AI products of the near future.
+date: Jan 26 2025, 02:19 -0400
+area: tech
+---
+
+Two things are certain in this decade for AI:
+
+1. AI will become more accurate, safer, and steerable in the coming years.
+2. The first wave of AI products will need a new crop of sharp _technical_ product visionaries, who do not carry a misguided sense of expertise from the SaaS era of previous decade. Inevitably, this crop of managers will get _commoditized_ very quickly.
+
+Unfortunately, we still do not have a category-defining AI product.
+And here's my (incomplete) list of reservations, sprinkled with predictions of how the markets will evolve.
+
+## Chat is a Local Minima
+
+Chat is almost certainly not the right interface for AI.
+Quite noticeably, chat just requires too many human touch points.
+Such an interface with friction from humans is a no-go for widespread deployment to reach [economies of scale](https://en.wikipedia.org/wiki/Economies_of_scale) with AI products.
+In the quest for [Artificial General Intelligence](https://en.wikipedia.org/wiki/Artificial_general_intelligence) (AGI), we may have forgotten the classic Unix philosophy behind great software,
+
+> Write programs that do one thing and do it well.
+
+We are stuck in a local minima for AI products with a chat interface which tries to do too much at once. But chat is a promising avenue for continued development.
+
+A chat interface is a frictionless buy-in for early adopters trying to test the capabilities of an AI system.
+With almost no vendor lock-in, chat interfaces are valuable but leaky user funnels.
+
+Nevertheless, frictionless buy-in allows reaching critical mass for data collection and user feedback, getting broad coverage of possible user inputs.
+I am convinced that the easiest way to make AI safe is by making all data in-distribution, i.e. all possible data distributions expected in-the-wild are amply available during model training.
+In fact, in the absence of effective techniques to search for strong neural network priors or the lack of convincing progress over the last decade, I would stretch to claim that data engineering may be the only way.
+Fortunately, [scaling laws](https://en.wikipedia.org/wiki/Neural_scaling_law) are alive and kicking.
+
+In the absence of a category-defining product but with a critical mass of early adopters, the next best thing is to build an API platform and a distribution channel.
+This approach is akin to the [App Store](https://www.apple.com/app-store/) model which allowed iPhones to gain a competitive advantage very quickly after the launch, which to this day remains significant despite regulatory pressure.
+
+However, the App Store model will _not_ apply to AI products precisely because of the lack of vendor lock-in. Any invention of a category-defining product that may arise from the democratized access to AI from major API providers is unlikely to benefit the provider.
+
+API access is a race to the bottom without unique capabilities that take advantage of strong [vertical integration](https://en.wikipedia.org/wiki/Vertical_integration).
+
+## AI Retrofitting is an Off-Ramp
+
+[Robotic Process Automation](https://en.wikipedia.org/wiki/Robotic_process_automation) (RPA), or the colloquial term "agents" is likely to be the first segment to fall to AI.[^a16z]
+
+[^a16z]: a16z also bets on RPA [in their thesis](https://a16z.com/rip-to-rpa-the-rise-of-intelligent-automation/).
+
+I really wish for AI to finally deliver on the promise of agents soon.
+God knows how much human capital has been wasted on boring corporate grunt work.
+Agents are also the perfect starting point to gain a strong moat from vertical integration - the cascading errors due to the inevitable chaining of agents can be better mitigated by keeping an end-to-end control of the input-output pipeline, without relying on opaque third-party providers.
+
+However, RPA is an example of sustained innovation rather than a [disruptive innovation](https://en.wikipedia.org/wiki/Disruptive_innovation).
+Enterprises with the widest customer base will win the first wave due to distribution advantage and the shortest cycle to bottomline impact.
+
+This retrofitting dynamic of making existing processes more efficient by replacing with AI, will hinder the invention of a category-defining product.
+A large market already exists for RPA, and unfortunately inertia will lead a large chunk of human capital to divert away from the grand goal for short-term results.
+
+There is a silver lining - I expect that short-term success in RPA will sustain continued investment in AI research.[^datadog]
+AI still needs a lot of work, and the market's appetite to keep the money flowing will be crucial.
+
+[^datadog]: Datadog, a cloud monitoring platform, is creating [a new AI research lab](https://www.linkedin.com/posts/atalwalkar_ai-researcher-datadog-ai-research-datadog-activity-7288619424205406209-0Ni-/) betting on agents, led by a veteran ML researcher.
+
+## The Missing Unit Economics of AI
+
+Unlike [SaaS](https://en.wikipedia.org/wiki/Software_as_a_service), the marginal cost of onboarding a new user for an AI product will remain high owing to the high CapEx of renting or owning GPUs.
+These costs are certain to fall in the future, maybe even drastically to the point of today's commercial-grade servers.
+
+Until then, we'll need alternate pricing models to sustain the AI momentum. The unit economics of AI agents can surely not use "per seat" pricing typically used by SaaS models.
+
+Borrowing a pricing dynamic from manufacturing, the cost reduction from automation often gets passed straight to consumers, leaving manufacturers with little advantage.
+
+The saving grace here will be a large enough demand that offsets the shrunken revenues, e.g. easy access to software creation will lead to an increased demand for software, a perfect flywheel. Moreover, entire industries like law and healthcare are ripe for software disruption that will keep the flywheel going.
+
+## The Future of Computing
+
+Steve Jobs' vision was crucial to elevate personal computers from a mere hacker's toy to a polished productivity tool for the non-technical masses.
+Despite drawing much awe, AI deployment is too raw in its current form.
+
+Perhaps, we are all looking in the wrong direction. The first breakthrough consumer AI product will not be in software but rather a paradigm shift in hardware. A good hardware product allows for vendor lock-in, strong vertical integration, and clear unit economics to kickstart a consumer AI revolution.
+What if the [Humane AI Pin](https://humane.com), even though universally hated, was looking in the right direction?
+
+In retrospect, [Google Glass](https://en.wikipedia.org/wiki/Google_Glass) walked so that [Meta glasses](https://www.meta.com/ai-glasses/) could run (more realistically just jog though).

src/content/thoughts/202501181556.md

@@ -0,0 +1,5 @@
+---
+date: Jan 18 2025, 15:56 -0400
+---
+
+What is life if not just a series of deadlines.

src/content/thoughts/202501201011.md

@@ -0,0 +1,5 @@
+---
+date: Jan 20 2025, 10:11 -0400
+---
+
+Play is the best form of rest. Or as they say, [rest in motion](https://mindingourway.com/rest-in-motion/).

src/content/thoughts/202501252144.md

@@ -0,0 +1,5 @@
+---
+date: Jan 25 2025, 21:44 -0400
+---
+
+> I suppose in the end, the whole of life becomes an act of letting go, but what always hurts the most is not taking a moment to say goodbye. - Pi Patel (Irrfan Khan), [Life of Pi](https://letterboxd.com/film/life-of-pi/)