Making AI do less work is a strategy the entire industry should learn

I noticed an interesting detail: the enhanced PSSR for PS5 Pro that Sony presented at SIGGRAPH 2026 has a core idea of making AI do less work.

This is exactly opposite to the mainstream trend of recent years. Large models are competing on parameter count, super-resolution tech is competing on network scale, everyone is adding more to AI, but Sony is doing subtraction. The redesign approach Daniel Craig talked about, as I understand it, comes down to one sentence: previously, PSSR had AI process every single pixel; now it's split into two paths—critical parts go to AI, while tasks with clear rules are handled by dedicated modules. AI only handles the parts it's truly good at, and surprisingly, image quality gets better, plus each frame saves about 100 microseconds.

100 microseconds sounds tiny, but in game rendering, this is fixed overhead saved on every single frame. Last week I wrote a post about how to check the hype in AI company valuations, where I mentioned a criterion: truly valuable AI implementation means spending compute power where it counts most. This change in PSSR is the extreme version of that logic.

Let's talk about the technical side first. The new PSSR changed the underlying architecture, extracting predictable parts like lighting and shadows from the neural network and handing them over to the traditional ray-tracing pipeline. The AI network is now only responsible for ambiguous areas without clear rules, such as texture detail inference and temporal stability in dynamic scenes. After splitting things up, the AI model size decreased, inference frequency dropped, but the output results became more solid.

One detail stuck with me: the old PSSR often suffered from flickering in fast-moving scenes, like those fine highlight edges drifting around. Simply put, neural networks make probabilistic judgments on pixels—they're accurate when the screen is static but tend to guess wildly when things move. By moving motion-related tasks to deterministic algorithms, this flickering issue was essentially bypassed. It's not that the model was fixed; it's just no longer responsible for that part.

This kind of thinking is common in traditional software engineering but rare in AI products. Most current AI features follow the approach of letting one model do everything it can, even if there are cheaper and more reliable traditional algorithms available for certain parts.

However, we need to be clear here: PSSR dares to make these changes because its application scenario is extremely narrow. Super-resolution for game graphics is a highly constrained problem—developers know what the input is, what the output should be, and how much hardware capability exists. In this closed environment, doing subtraction makes the overall system more reliable.

In general-purpose AI scenarios, it's different. You might say let LLMs do less work and offload some tasks to traditional rule engines, which sounds reasonable, but in an open domain, it's hard to determine which parts AI excels at and which can safely be handed to deterministic algorithms. Last week I looked at an AI-native startup product, and they did exactly this: user inputs first pass through rule matching filters, and only uncertain cases are sent to the large model. After running for a month, they saved about 60% on API call costs, and performance was more stable. But their scenario was customer support ticket classification, where fields and intents are limited, so rule matching could handle it.

This precisely shows that the value of PSSR isn't just in the technical solution; it reminds us to ask a question first when designing AI systems: Which tasks really need AI?

Back to the gaming scenario, the PS5 Pro's hardware specs are fixed, and compute power is limited. Instead of having AI spend massive computation on parts solvable by traditional methods, it's better to concentrate resources on aspects that truly enhance visual perception. Sony made PSSR compatible with over 50 games and enabled it by default, proving this path works on actual hardware. I saw players test comparisons across twenty-plus games, and the most obvious change was in details like hair and fabric—previously, AI calculations often blurred them together, but now the edges are much cleaner.

Some might argue that iterating the old PSSR a few more times to train a stronger model could also solve the flickering issue. But that trades more data and longer training time for quality improvements, whereas Sony chose the cheaper route: reducing AI's burden at the system design level.

This decision feels familiar to me. Recently, while analyzing AI company valuations, I found a pattern: the AI products burning cash the hardest are often those letting models do all the work; meanwhile, products achieving better results per unit cost generally implemented similar system layering. In the end, both might achieve similar image quality, but saving 100 microseconds per frame accumulates into a dual advantage in experience and cost.

I noticed that the PS5 Pro generation hasn't reached the sales heat Sony expected, but PSSR's engineering philosophy of subtraction might become the most reference-worthy tech case study during this period. For people building AI products, instead of thinking about how to make models stronger, consider first how to make models do less. This is a direction worth trying out.