AI’s Future: Etched Silicon vs. Smart Harnesses

William Palmer

Imagine building a computer chip where the software is not loaded into memory, but literally carved into the silicon itself. That, that, that is not a metaphor. AMD just acquired a startup called Taalas, and their whole premise sounds almost absurd on its face. Instead of spending gigawatts of power moving numbers back and forth between graphics memory and process cores, you etch the AI model weights directly into the physical transistor connections. At least in the case of Taalas, the weights are physically encoded directly on the chip.

William Palmer

To really get why Lisa Su and AMD made this play, you have to look at the massive wall modern computing has hit. It is called the von Neumann memory bottleneck. Right now, when a massive model like a 70 billion parameter LLM generates a single word, every single weight has to be dragged out of high bandwidth memory, shoved through a bus, multiplied inside a tensor core, and sent back. It is incredibly inefficient. It is like driving a race car where you have to pull into the pit lane every single second just to adjust a single valve in the engine. You spend all your time and energy moving stuff around, not actually driving.

William Palmer

What Taalas did was look at 4 bit quantized matrix multiplication and say, why are we calculating this every time? In 4 bit math, there are only 16 possible multiplication results. So instead of fetching weights from DRAM or SRAM, the hardware runs all 16 possible results in parallel on the bottom layer of the silicon, and then the top metal wiring layer acts as a physical selector based on the hardwired weight. The computation happens literally where the data sits. It is pure compute in memory.

William Palmer

Now, if you are a software engineer, your immediate reaction is probably, wait a minute, that sounds completely insane. AI models evolve every six months! If you spend tens of millions of dollars to tape out an ASIC with fixed weights, isn't that chip obsolete the exact day it arrives from the foundry? Well, yes and no. That is the obsolescence paradox. If you are trying to stay on the bleeding edge of generic consumer chatbots, etching weights into silicon is a suicide run. But think about massive enterprise workloads. Think about an airline running a dedicated customer agent like ChatJimmy, or a company deploying voice processing directly onto millions of edge devices or autonomous cars. That underlying model does not need to change every week. They need raw throughput, zero latency, and rock bottom power consumption.

William Palmer

It reminds me a lot of working on race cars. When you are testing on the dyno, you want a fully re-programmable software mappable ECU. You want to adjust air fuel ratios and ignition timing with a laptop every single lap. But when you build a dedicated high performance engine for one specific endurance race, you take away the dynamic overhead. You build fixed, mechanically tuned carburetors or optimized air intake runners carved out of solid aluminum. You trade re-programmable flexibility for uncompromised, raw, fixed hardware speed. And right now, as inference costs threaten to eat tech budgets alive, fixed speed is worth a fortune.

William Palmer

But while hardware is getting locked down into physical silicon, the exact opposite trend is exploding on the software side. We are seeing a massive shift where the intelligence of an AI system is moving away from the raw size of the model weights and into the orchestration harness surrounding it. Just look at what Meta dropped with Muse Spark 1.2. A few weeks ago, nobody was talking about it. Then suddenly it rocketed into the top five on the Vals Index at just 69 cents per test. That is 3x cheaper than Kimi and more than 10x cheaper than models like Claude Opus 5 or GPT 5.6 Sol.

William Palmer

And here is the kicker. Meta claimed gold medal level performance across five STEM Olympiads, including perfect theory scores in physics and math, without using a single external tool. No python code interpreters, no web search, no calculators. How? Not by scaling up to a trillion parameter dense giant, but by using multi agent parallel reasoning inside the prompt harness. They split complex questions across multiple parallel agent streams that collaborate, critique, and re-derive the answers in real time.

William Palmer

This is test time compute in action, and it fundamentally changes the math of AI engineering. Instead of paying massive training and serving costs for a bloated base model, you take a smaller, extremely fast model and wrap it in a smart execution loop. François Chollet made a fascinating observation about this recently. He argued that when you build a massive inference time harness that orchestrates dozens of neural calls, you are no longer running a pure neural network. You are building a neurosymbolic program. The harness provides the symbolic control structure, and the model provides the raw intuition.

William Palmer

You can see the entire industry pivoting toward this harness first mindset. Look at OpenAI collapsing their ChatGPT experience. They took GPT 5.6 Sol and combined instant responses and deep reasoning into one single chat surface, giving users a simple reasoning effort slider. Or look at Cloudflare launching Kitesurf during their Agents Week. Kitesurf is a stateless browser running on Workers designed specifically for AI agents. Instead of spinning up a massive, heavy Chromium instance for every agent scrape, it splits the script and document object model from the renderer, lazily instantiating workers only when needed. They are building browser infrastructure where the agent harness is a first class citizen.

William Palmer

So step back and look at the whole landscape right now. We have this fascinating, almost contradictory tension pulling the industry in two opposite directions at once. On one end, companies like AMD are betting on etched ASICs, turning software directly into rigid silicon to squeeze out every drop of hardware efficiency. On the other end, software developers are discovering that you can unlock frontier level reasoning just by building smarter, dynamic multi agent harnesses around existing open models.

William Palmer

It leaves us with a critical question for the next era of computing. Is the future of AI going to be won in the cleanrooms of semiconductor foundries, or in the open source codebases where agent harnesses are written? My guess? It is both. We are heading toward a world where etched silicon handles the low level execution, while hyper flexible software harnesses do the actual thinking. Alright, that is the breakdown for today. Talk to you soon.