C.W.K.
Stream
Lesson 05 of 05 · published

The Future Is Messy — MIRAS and Coexistence

~13 min · future, miras, coexistence

Level 0Observer
0 XP0/50 lessons0/14 achievements
0/100 XP to next level100 XP to go0% complete

The field is converging on coexistence

If you've read this far, the picture should be clear: no architecture is winning a clean fight. Pure Transformers dominate frontier reasoning and few-shot ICL. Hybrids dominate production long-context serving. Pure SSMs win in extreme memory-constrained or streaming workloads. Hyena wins genomics. RWKV wins on-device. Linear attention wins long-context summarization where Kimi Linear-class quality is acceptable.

The 2026 landscape is not Transformer-vs-alternative. It's a portfolio of architectures, each best in its niche. The mature engineering question is no longer "which architecture should we use?" — it's "how do we route different workload classes to the right architecture?"

MIRAS — the theoretical unification

MIRAS framework (Google, December 2025) provides a theoretical unification of all the architectures in this quest as variants of a single underlying primitive: associative memory. Transformers, SSMs, RWKV, RetNet, Hyena, linear attention — all are MIRAS framework members with different choices of memory operator, key-value structure, and update dynamics.

The MIRAS view is more than academic. If all sequence models are associative memory variants, then "hybridizing" them is just choosing different MIRAS configurations at different layers. The distinction between "attention" and "recurrence" may eventually dissolve into a single design space parameterized by associative-memory choices. We're early in that conceptual collapse, but the framework is starting to influence how research teams describe their architectures.

What to watch in 2026–2027

  • Kimi Linear scaling. Does the Moonshot result replicate at frontier scale? If yes, full attention may be displaced even in mainstream Transformer-shape models.
  • Hardware-specific SSM/hybrid chips. Google Ironwood and AWS Trainium3 are starting to optimize for parallel scan and SSM workloads. The hardware-software co-design loop is closing.
  • Transformer distillation into hybrids. Cartesia's Llamba-8B (12× throughput from distilling Llama 3.1 8B into Mamba) hints that the fastest way to a hybrid frontier model may be to start with a Transformer and convert.
  • Multi-scale operators. StripedHyena 2's SE/MR/LI pattern could absorb back into general-purpose hybrids.

The myth to discard: "Transformers are obsolete." The reality: Transformers dominate and will for years. Alternatives solve specific bottlenecks where quadratic scaling becomes impractical. The field is about expanding the toolkit, not replacing the hammer.

External links

Exercise

Read the MIRAS abstract (or any high-level summary of the framework if the paper is too dense). Then think about your own workloads: which one would benefit from a memory-heavy variant, which from a fast-decay variant, which from a hybrid. The point is to start using the MIRAS-style language when you reason about architecture choice — "this workload needs more associative-memory capacity than recency-decay" is a more useful framing than "should I use Mamba".

Progress

Progress is local-only — sign in to sync across devices.
Spotted a bug or have feedback on this page?Report an Issue

Comments 0

🔔 Reply notifications (sign in)
Sign inPlease sign in to comment.

No comments yet — be the first.