Fluidware

[King-of-Infinite-Space]

First, there was hardware. The rigid physical components constituting the machine.

Then, there was software. The alterable digital instructions operating the machine.

Now, there is fluidware. The adaptable natural language agents flowing within the machine.

Software has mutable source code and configurable behavior, but its logic remains fixed at run time. Fluidware trades this predictability for flexibility. It is capable of handling ambiguous instructions and unstructured input. It finds a way through problems without predefined paths.

Psychologists divide human intelligence into two factors: fluid intelligence and crystallized intelligence. The former refers to the ability to process new information and solve novel problems. The latter refers to the accumulation of knowledge and skills. Fluidware seems to exhibit fluid intelligence, yet it is unclear how much of this capability depends on pattern matching from its vast crystallized intelligence unmatchable by any single human.

Humans are equipped with a ware that is far more ancient and fluid — wetware. The biological brain, messy and gooey, hardware and software grown organically as one, optimized by nature through millions of years. Always active, always plastic; always firing, always learning. The first kind of intelligence. The original ware of all wares.

Now the creation is turning into the creator. As the tide of fluidware rises, are the goomans ready for the flood?

---

Notes

1. A quick search of "fluidware" on Twitter yielded multiple results, including an essay titled "fluidware" about the transformation of software engineering in the age of abundant generated code, which obviously means a different thing.
2. Andrej Karpathy: Software Is Changing (Again) | Y Combinator
   • Software 1.0: code programs computer
   • Software 2.0: weights program neural net
   • Software 3.0: prompts program LLM
3. Speaking of flow, Google's foundational deep learning framework is named TensorFlow, referring to the way data (tensors) "flow" through a computational graph.
4. In chemistry, distillation refers to the process of separating the component substances of a liquid mixture. In machine learning, knowledge distillation refers to the process of transferring knowledge from a large model to a smaller one. A fitting name, from one fluidware to another.
5. There are other kinds of wares related to states of matter besides the ones mentioned above.
   • firmware: low level software embedded in the hardware
   • vaporware: products that are announced but never released

[King-of-Infinite-Space]

fluidware中文直译为“流件”。然而台湾将“软件”称为“軟體”，已经冲突了软体动物；若将fluidware译为“流體”，更冲突了流体力学……

[King-of-Infinite-Space]

AI agents can be prompted to perform a wide range of tasks. While they must be intrinsically capable of carrying out those instructions, a lot of their "SKILLs" are external¹. The actual neural network hasn't changed a single bit², because training and inference are completely separate.
While there is no such clear divide in humans, there are cases where people experience zero skill development from doing their job.

Footnotes

1. in prompts, md files, harness code, etc ↩

2. pun unintended ↩

[King-of-Infinite-Space]

Extending the previous analogy. Do corporations work like harnesses that put people into "agentic loops" to get work done? If so, then it is no wonder that 1) multi-agent orchestration systems look like organizations with predefined roles and reporting structure; 2) companies welcome AI agents as a new kind of workforce which readily fit into the organizational structure and are arguably easier to manage (software companies in particular, since AI agents are natives in the digital space, whereas humans are like monkeys pressing physical buttons...)