Hacker News: himata4113

New comment by himata4113 in "CRISPR tech selectively shreds cancer cells, including "undruggable" cancers"

himata4113 — Fri, 12 Jun 2026 21:47:22 +0000

Oh wow, not as impressive as I thought, but I guess we are looking at broad categories rather than specific types and variants. But overall the trend is down on ever cancer since around a decade ago. Was expecting a sharper drop around 1-2 decades ago, but things just take time when it comes to experimenting with human lives. Will be interesting to revisit this in another decade when a lot of the treatments finally leave the experimental stage.

New comment by himata4113 in "CRISPR tech selectively shreds cancer cells, including "undruggable" cancers"

himata4113 — Fri, 12 Jun 2026 21:30:40 +0000

I guess then a graph of deaths from disagnosed cancer at various stages would be just as awe inspiring. I'd settle for that as well.

New comment by himata4113 in "WASI 0.3"

himata4113 — Fri, 12 Jun 2026 21:06:40 +0000

I do specifically that and abandoned WASI in favor of freestanding with shared memory regions + PIC memory region. It works today and is easily extensible as well as increadibly fast, sandboxing is not as strong since modules can read code they don't own and modify data they do not own, but splitting modules into "trusted" and "untrusted" largely mitigated those issues as memory has write bit stripped for execution during untrusted modules, only allowing data they own to be modified. Sensitive data lives outside the runtime and has to be fetched with (slower) host methods.

Point being, if you rely on a non standard specification of developer defined wasi api subspec you might as well just go with freestanding.

New comment by himata4113 in "CRISPR tech selectively shreds cancer cells, including "undruggable" cancers"

himata4113 — Fri, 12 Jun 2026 20:18:32 +0000

Does anyone know a website where I can see/read of how many cancers (and their variants) we've effectively solved, have drugs to negate their effects, have experimental drugs for and uncurable cancers? I think that graph would be awe inspiring looking at the past decade of advancements.

What's more crazy is that we're slowly going from millenia, to decades, to likely years in the near future from being presented a biological problem and achieving the next milestone in solving it. We might have "AI", but we also have brilliant minds right now that are speeding up development to a pace that would be unimaginable just few years ago.

New comment by himata4113 in "macOS 27 Beta breaks the ability to boot Asahi Linux"

himata4113 — Thu, 11 Jun 2026 20:08:18 +0000

404

New comment by himata4113 in "Amazon Says Its Data Centers Use 2.5B Gallons of Water"

himata4113 — Thu, 11 Jun 2026 15:08:48 +0000

Roughly 736* of US households assuming it's 2.5B per year, 100 gal per person. I can't read the article because I got what it feels like 17 popups that nearly gave me epilepsy.

edit: corrected to 736, its per year.

New comment by himata4113 in "Cybersecurity researchers aren't happy about the guardrails on Anthropic's Fable"

himata4113 — Wed, 10 Jun 2026 22:48:59 +0000

some static words in AGENTS.md trigger it as well as some mcp servers.

New comment by himata4113 in "Cybersecurity researchers aren't happy about the guardrails on Anthropic's Fable"

himata4113 — Wed, 10 Jun 2026 22:48:16 +0000

I've done this, including the hardcoded refusal strings that already exist in claude code. It won't stop a real attacker, but I still find it really funny when you're trying to use one of the AI tools and it gives you a random refusal and you don't know why, wastes a little bit of time.

New comment by himata4113 in "Who Runs Your Rust Future? Hands-On Intro to Async Rust"

himata4113 — Wed, 10 Jun 2026 22:12:40 +0000

async gives you a few other advantages, it allows you to handle bi-streams much easier than two sync threads especially when it comes to synchronization.

New comment by himata4113 in "Who Runs Your Rust Future? Hands-On Intro to Async Rust"

himata4113 — Wed, 10 Jun 2026 16:03:23 +0000

You use async to preserve system resources. For example you can easily exhaust the host with ~20k connections running a thread-per-connection schema where each thread simply waits for epoll event, async prevents this by having a threadpool of ~16 threads that handle all the connections instead of polling the scheduler wakes it up, asks "do you haev work to do" if not continues to next task. (This heavily varied by the async runtime implementation, each async runtime can and will act differently to maximize efficiency over throughput)

Claude Fable 5 Launches at #1 on the Artificial Analysis Intelligence Index

himata4113 — Wed, 10 Jun 2026 15:11:06 +0000

Article URL: https://artificialanalysis.ai/articles/claude-fable-5-mythos-intelligence-index

Comments URL: https://news.ycombinator.com/item?id=48477562

Points: 4

# Comments: 0

New comment by himata4113 in "CEOs who think AI replaces their employees are just bad CEOs"

himata4113 — Tue, 09 Jun 2026 21:37:04 +0000

now it's closer to 95% of work can be done by AI and requires 5% mental effort, but 5% of the work requires 95% of the mental effort to finish because of all the unoptimial decisions AI has taken. I find that AI works best in small micro-service type architecture where each component has a clear goal and doesn't have interconnected parts within the same application that can break. But you do run into an issue where changes in microservice a need changes in microservice b and updating it is not ideal since it usually cascades thru the entire system or requires stacks of legacy support.

New comment by himata4113 in "Claude Fable 5"

himata4113 — Tue, 09 Jun 2026 18:05:22 +0000

  > virtualization
  switching to opus 4.8

ok fair

  > embedded-allocator
  switching to opus 4.8

urgh fine

  > chrome
  switching to opus 4.8

are you kidding me?

New comment by himata4113 in "Claude Fable 5"

himata4113 — Tue, 09 Jun 2026 18:01:54 +0000

concept is similar to how it works in inference, instead of performing regressive writes to the entire model you run the whole model, but part of the model can live in system memory and get swapped in/out on demand. So only XB parameters are active in training.

edit: I am not really sure if it works like that. I haven't looked too deep into deepseek v4 pro specifically.

New comment by himata4113 in "Claude Fable 5"

himata4113 — Tue, 09 Jun 2026 17:37:23 +0000

MoE experts were likely trained independently / in a sparse format. Training anything beyond 2t on typical systems would be infuriantingly slow, you could do 4t on nvidias room-scale solution, but for a reasonable training speed / batch size it caps around 3t.

New comment by himata4113 in "Claude Fable 5"

himata4113 — Tue, 09 Jun 2026 17:27:33 +0000

They're trained in a model class likely in 2t to 3t range. It's very unlikely that chinese labs have access to gpu systems capable of training models like that, let alone serving them. This requires proprietary room-scale systems which fetch a huge premium over typical 10 slot systems.

I am sure that they can develop their own equivlient version of such clusters in around 1 year though. Distilling fabel 5 will also go a long way.

New comment by himata4113 in "Siri AI"

himata4113 — Mon, 08 Jun 2026 23:46:08 +0000

Apple could have the same kind of permission dialogues with their own models (and they actually should). Each and every (first-time) use of a feature should:

  1) ask for permission explaining the scope
  2) warn you about the dangers with a confirmation / nevermind option

Putting this in practice:

  1) Acme AI requires access to your email provider in order to execute this request. Grant / Deny
  2) You're about to let Acme AI read and send emails on your behalf, this might be dangerous due to X and Y. Do you want to continue? / Nevermind.

In this case:

  1) Asks for access to a service
  2) Asks for a specific use-case of the service

1 is access to data, you might want to give broad access to some applications and input data

2 is permission to act, but you might want to deny access to some parts such as sending email and scope to summarization

New comment by himata4113 in "Siri AI"

himata4113 — Mon, 08 Jun 2026 21:47:20 +0000

"Trusted System Agent" imo sounds like an apple approved agent which would only be available to companies that accept apples (likely unreasonable) demands and would completely lock smaller companies out of the ecosystem.

New comment by himata4113 in "Siri AI"

himata4113 — Mon, 08 Jun 2026 20:59:36 +0000

I read through the entire DMA rant that apple has here: https://www.apple.com/newsroom/2026/06/due-to-dma-siri-ai-de...

This feels like it could be solved with a list of permissions that the user has to turn on when using 3rd party AI.

Apple already:

    1) requires developers to submit ID to publish an app on the appstore (at least I had to after ~1000 downloads to be able to publish an update)
    2) has strong kernel enforced memory integrity and disallowes arbitrary code execution (unless explicitely approved for games like roblox, jitting not allowed tho has to be interpreted).
    3) reviews every app update.

I feel like this is nothing more than Apple being angry that they have to allow people to actually choose what AI they want on their phone. This is particulary interesting if anthropic and openai decided they want to add siri ai override to their apps allowing them to take advantage of the apple ecosystem without signing some kind of deal like they had to with Google. I assume behind closed doors Google had to make some sacrifices for them to be the model powering siri.

New comment by himata4113 in "Siri AI"

himata4113 — Mon, 08 Jun 2026 19:05:31 +0000

Change your region. I've done so and haven't particulary noticed anything off, all the EU specific apps are weirdly still available in US appstore.