Hacker News: aaronblohowiak

New comment by aaronblohowiak in ">10x More Efficient Pretraining"

aaronblohowiak — Thu, 10 Sep 2026 18:04:35 +0000

more experiments.

New comment by aaronblohowiak in "As AI eats the web, the internet’s collective memory is disappearing"

aaronblohowiak — Tue, 11 Aug 2026 10:49:57 +0000

It’s good at all kinds of stuff that is search adjacent, “find amesent parks with water feature within 100 miles of me that have rv parking nearby”, etc

New comment by aaronblohowiak in "Show HN: I left Figma to build a diffusion-based UI design tool"

aaronblohowiak — Fri, 31 Jul 2026 01:40:52 +0000

Anthropic should buy you

New comment by aaronblohowiak in "Steam Machine launches today"

aaronblohowiak — Mon, 22 Jun 2026 23:02:54 +0000

China

New comment by aaronblohowiak in "Ten years of ClickHouse in open source"

aaronblohowiak — Fri, 19 Jun 2026 16:52:50 +0000

while ClickHouse can scale down to compete with duckdb, I dont believe (but happy to be corrected) that duckdb can scale up like ClickHouse can.

most people dont have that scale problems, but when you do...

New comment by aaronblohowiak in "Surprise, pay $1000"

aaronblohowiak — Wed, 10 Jun 2026 06:17:22 +0000

I can see their logic - instead of breaking people's builds, they are being the nice guys and letting you pay them back later. BUT, any time you break convention, you've dipped into your trust budget (even if you communicate it way more clearly than was apparently done here.)

New comment by aaronblohowiak in "Claude Opus 4.8"

aaronblohowiak — Thu, 28 May 2026 16:53:03 +0000

Same price for regular and cheaper fast mode. Happy for these incremental improvements.

New comment by aaronblohowiak in "Gemini 3.5 Flash"

aaronblohowiak — Wed, 20 May 2026 17:47:12 +0000

for LLMs as implemented today?

New comment by aaronblohowiak in "Cutting inference cold starts by 40x with LP, FUSE, C/R, and CUDA-checkpoint"

aaronblohowiak — Mon, 18 May 2026 19:29:28 +0000

Put differently, 1/40 is not the same as 1x - 40x. I’d phrase as Reduced by 97.5% or 0.975x

New comment by aaronblohowiak in "Cutting inference cold starts by 40x with LP, FUSE, C/R, and CUDA-checkpoint"

aaronblohowiak — Mon, 18 May 2026 19:26:41 +0000

How can you cut latency by more than 1x? I am no intending to be snarky, it just doesn’t fit my brain how you can reduce a measure time by more than the original starting time.

New comment by aaronblohowiak in "Local AI needs to be the norm"

aaronblohowiak — Mon, 11 May 2026 06:12:44 +0000

taalas!!!

New comment by aaronblohowiak in "Teaching Claude Why"

aaronblohowiak — Sat, 09 May 2026 05:27:39 +0000

Should according to your ethos, not should according to history, sadly.

New comment by aaronblohowiak in "SpaceX says it has agreement to acquire Cursor for $60B"

aaronblohowiak — Wed, 22 Apr 2026 01:13:45 +0000

Only penalties if you withdraw from 401k. Most 401k plans have some kind of moneymarket, bond fund, or similar

New comment by aaronblohowiak in "Measuring Claude 4.7's tokenizer costs"

aaronblohowiak — Fri, 17 Apr 2026 22:59:47 +0000

taalas is amazing. id gladly spend 5-15k on something that matched that performance with opus 4.6 quality

New comment by aaronblohowiak in "Lean proved this program correct; then I found a bug"

aaronblohowiak — Tue, 14 Apr 2026 01:57:35 +0000

I disagree but the constraints required to get there are probably untenable for most practical applications, so in practice I agree.

New comment by aaronblohowiak in "Pro Max 5x quota exhausted in 1.5 hours despite moderate usage"

aaronblohowiak — Mon, 13 Apr 2026 00:16:33 +0000

how long they stay around after the cache miss is irrelevant if I am burning all the prior tokens again. also, how much context they have depends entirely on the task and your workflow. I you have a subagent implement a feature and use the compile + test loop to ensure it is implemented correctly before a supervisor agent reviews what was implemented vs asked then yes, subagents do have a lot of context.

New comment by aaronblohowiak in "Pro Max 5x quota exhausted in 1.5 hours despite moderate usage"

aaronblohowiak — Sun, 12 Apr 2026 15:36:31 +0000

So if I run a test suite or compile my rust program in a sub agent I’m going to get cache misses? Boo.

New comment by aaronblohowiak in "Build123d: A Python CAD programming library"

aaronblohowiak — Mon, 30 Mar 2026 17:48:44 +0000

Kernel here is ambiguous.. I get what you mean, but parasolid is usually the thing described as the cad kernel.

New comment by aaronblohowiak in "Apple discontinues the Mac Pro"

aaronblohowiak — Fri, 27 Mar 2026 15:42:34 +0000

External drive is workaround to apple’s pricing scheme, often purchased at same time as computer.

New comment by aaronblohowiak in "Rob Pike's Rules of Programming (1989)"

aaronblohowiak — Wed, 18 Mar 2026 18:29:57 +0000

The fundamental tension between nouns and verbs and the attempts to unify them like events have made programming a long art form to study.

It's all use-case and priority-specific, and I think the more varied your experience and more tools you have in the tool belt, the better off you can be to bring the right solution to bear. Of course, then you think you have the right solution in mind (lets say using partitions in postgres for something) but you find the ORM your service is using doesn't support it, then what is "best" becomes not only problem-specific but also tool-specific. Finally, even if you have the best solution and your existing ecosystem supports it but the rest of the engineering staff you have is unfamiliar with it, it may again no longer be "best".

this ladder of problem-fit, ecosystem-fit, staffing-fit is something I have grappled with in my career.

LLMs are only so-so at any of the above (even when including the agent as "staff".)