Hacker News: unjuno

New comment by unjuno in "Show HN: Rotation via Double Reflection"

unjuno — Sat, 22 Aug 2026 18:53:45 +0000

thank you

New comment by unjuno in "Baking a Model: A Metaphor for LLM Training"

unjuno — Tue, 18 Aug 2026 18:41:45 +0000

LLM pretraining might be viewed as learning useful state representations from sequences, rather than having humans explicitly design the states in advance.

In many conventional software systems, we define states and transitions ourselves: idle, sending, completed, error, and so on. LLM pretraining works differently: the model learns from large numbers of sequences to predict what is likely to come next.

I don’t mean that an LLM literally constructs discrete internal states. Rather, it learns latent representations of context that are useful for prediction. In that sense, these representations might be thought of as a kind of learned state.

This seems closely related to abstraction. Abstraction is not simply throwing information away, but abstracting away differences that are irrelevant to prediction while preserving those that matter.

Two sequences may look very different on the surface, yet rely on similar underlying features to predict what comes next. Even a rare feature should be preserved if it changes the prediction.

From this perspective, pretraining can be seen as a form of useful compression: learning representations that preserve what matters for prediction while abstracting away what does not.

How far can abstraction, compression, and prediction be understood as different aspects of the same process?

New comment by unjuno in "Ask HN: Does AI research need "world models" more than bigger LLMs?"

unjuno — Mon, 03 Aug 2026 23:59:59 +0000

By “world model,” I mean a model continuously updated from experimental or sensor data, not just an LLM trained on text.

Ask HN: Does AI research need "world models" more than bigger LLMs?

unjuno — Mon, 03 Aug 2026 23:59:28 +0000

It seems that generating hypotheses is no longer the main problem. The harder problem may be continuously collecting real-world data, building accurate world models, and using them to guide experiments. Is this where automated research is heading?

Comments URL: https://news.ycombinator.com/item?id=49162887

Points: 3

# Comments: 2

New comment by unjuno in "Commodification of Intelligence: Good, Bad, and Ugly Circular AI Deals"

unjuno — Wed, 29 Jul 2026 22:28:36 +0000

The problem with circular transactions is that they can make companies look healthier than they really are, distort investment decisions, and cause widespread damage when the cycle breaks.

New comment by unjuno in "crustc: entirety of `rustc`, translated to C"

unjuno — Fri, 03 Jul 2026 04:17:56 +0000

The most interesting part of this to me is not “Rust to C” by itself, but the fact that it widens the pool of people who can help debug portability problems.

There are relatively few people who understand Rust’s compiler internals, LLVM backends, and obscure target support deeply. But there are many engineers who understand C compilers, ABIs, linkers, makefiles, cross-compilation, old operating systems, and weird platform-specific compiler behavior.

If Rust can be lowered into target-specific C, then some problems stop being exclusively “Rust compiler problems” and also become C toolchain problems. That means more people can inspect the generated C, build failures, linker errors, ABI mismatches, and compiler-specific behavior.

C is obviously not a magic portability layer. ABI details, integer widths, alignment, TLS, aliasing, and undefined behavior still matter. But as an ecosystem boundary, C gives many more engineers a way to participate in debugging and porting work.

I think that social/maintenance aspect may be more important than the language translation itself.