Abhishek S.
Shipping in public. Listening in private.

Abhishek

I lead women’s Indo-Western & Premium at Max Fashion. I also wrote the AI that runs the buying floor.

Rare profile. Category operator who ships production code.

Senior Buying Leader · Max Fashion Women’s Indo-Western & Premium · 530+ India stores NIFT ’12 · Twelve years on the floor

abhishek@bengaluru ~ %
>role: senior buying lead
>dept: women’s indo-western + premium
>floor: 530+ stores india

The Arrow of Time in AI Architectures

A transformer can read a sentence as a square matrix; Mamba has to carry the sentence forward one state at a time. That difference is not philosophy dressed as engineering. It decides whether a model treats "A before B" as a learned pattern or as a constraint built into the machine.

The old physics problem is sharper: most microscopic equations work both ways, while coffee never unstirs itself. AI has a smaller version of the same problem. Some architectures know order; fewer have a built-in past.

The case

Self-attention, as formalized in Attention Is All You Need by Vaswani et al. in 2017, compares tokens pairwise. A length N context gives an N x N attention pattern. Causal masking blocks future tokens during generation, and positional encodings tell the model where each token sits, but the core operation is still a geometric comparison over positions.

Mamba, introduced by Albert Gu and Tri Dao in 2023, starts from selective state spaces. Its state update has a direction:

h_t = A(x_t) h_{t-1} + B(x_t) x_t
y_t = C(x_t) h_t

h_t is the current state. h_{t-1} is yesterday's residue. Once the update runs, the model has compressed part of the past into a state that the next token inherits. Gu and Dao report linear sequence scaling and a 5x inference-throughput claim over transformers in their 2023 Mamba paper, with Mamba-3B matching or beating transformer baselines in some language settings.

Titans goes one step further. Behrouz, Zhong, and Mirrokni's 2024 paper stores history in a neural long-term memory updated at test time. The paper reports needle-in-haystack tests beyond 2 million tokens. That is not just a longer window; it is a model changing a memory module while it reads.

Where the arrow bites

Architecture Time treatment Cost shape Failure mode to watch
Transformer Positions compared together O(N^2) attention Long contexts get expensive
Mamba Forward recurrent state linear sequence pass Reversal and symmetry tasks can hurt
Titans Test-time memory updates paper reports 2M+ token tests Memory policy can store noise

The sharpest warning came after Mamba's first wave of adoption. Chen, Lin, Wang, and Xu's 2025 paper Achilles' Heel of Mamba used synthetic tasks to show an asymmetry bias from Mamba's nonlinear convolution. The reported weak spot is not all SSM machinery. It is the token-fusing convolution before the SSM, which makes reversed matching and symmetric pattern recognition harder.

This is the trade: a built-in arrow helps when data itself has a direction, as in speech, sensor traces, markets, and biological sequences. It hurts when the task asks for invariance under reversal. A palindrome does not want a past; it wants a mirror.

What's contested

No one has isolated "temporal asymmetry" as a single variable across matched models. A fair test would need equal data, equal parameter count, equal compute budget, and tasks that separate event order from mere retrieval. As of 2026-07-28, the public evidence mixes architecture, training recipe, benchmark design, and scale.

The physics analogy also has a ceiling. Landauer's 1961 result says erasing one bit has a thermodynamic cost of k_B T ln(2), but a neural memory update is not automatically the same thing as physical erasure. The metaphor becomes useful when it asks what information cannot be recovered after an update. It becomes sloppy when it treats every gradient step as a tiny heat engine.

Why this has to do with other realms

The bridge to concept arrow of time is direct: memory is the visible scar left by irreversibility. The bridge to concept information theory is tighter: a system that forgets must choose what to compress, and compression always hides a bet about the future.

There is also a biology bridge. The concept titans hippocampus model analogy works because hippocampal memory is not a database lookup. It changes the organ that performs the lookup. AI long-term memory starts to look less like a context window and more like a nervous system when reading changes the reader.

An open question

If a model can remember only by becoming slightly different, should benchmark design measure answer accuracy alone, or should it measure the cost of becoming that kind of model?

Key Sources

Further Reading

See Also

Abhishek's take

What grabs me here is that "memory" stops being a storage metaphor and becomes a design choice with a direction. I trust causal masks less than state changes, because masks tell a model what it cannot see while memory tells it what it has already become. The hard next page is not "AI consciousness"; it is concept machine memory: when does a trace become evidence rather than residue?

Tags: #transformer #mamba #titans #arrow-of-time #temporal-asymmetry #causal-reasoning #irreversibility