Hacker Newsnew | past | comments | ask | show | jobs | submit | ignoramous's commentslogin

See: Fast Static Symbol Table: https://github.com/duckdb/duckdb/pull/4366

FSST is based on a fixed size (255 items) dictionary of high frequency variable length strings/substrings (learned from the corpus) encoded as one byte.


I wouldn't be so sure; Ex A: The terrifying expansion of Sweden’s state surveillance, https://edri.org/our-work/the-terrifying-expansion-of-sweden...

> ... a certain popular "pro-privacy" product beloved by many here ...

If you're talking about Proton VPN, they do support "credential-less accounts" through their official apps, I believe? At least, on Android since 2024: https://www.androidpolice.com/proton-vpn-works-without-accou...


i am hesitant to really narrow it down, but it is not proton (i am a very early proton customer)

Then why comment at all? This is the danger of speaking in riddles.

> 50% cheaper

Cache read/write decrease by 50% or similar? That's where most (95%+) of the cost is for agentic coding workloads.



Per Xiaomi, MiMo v2.6 training run cost $3.47m. A far cry from the estimated costs ($100m+) for the Big 5 (MSL, xAI, GDM, OAI, Ant). I wouldn't be surprised if salaries and R&D costs have similar drastic disparities.

For a model that matches Muse Spark 1.3 in benchmarks, MiMo v2.6 Pro is incredibly cheap, given its cache rates will remain $0.0036 per million.


That is the RL training cost only. Their announcement blog mentions this: https://mimo.xiaomi.com/mimo-v2-6#scaling-rl-fully-open-sour...

My understanding of tech salaries in China is that they are pretty decent, but not as high as in SF; closer to typical European salaries.

Mostly due to lower cost of living; Shenzhen is way cheaper than SV


I seriously doubt salaries are included. It must be just the electricity and GPU costs.

In these metrics, yes. In the reported training budgets of anthropic/openai, who knows?

I sorta got the impression that the $3.47 million only covered post-training , given that few of the graphs start at zero. Is a barely-trained model going to score 48 on DeepSWE v1.1 ?

https://mimo.xiaomi.com/rl/


The Kano Model customer satisfaction theory:

  Kano predicted that users' perceptions of satisfaction with a feature will shift from delight to expectation over time. This is either because they have got used to it or because competitors have started to include it in their offerings. In the case of the touchscreen, for instance, the newness that came with the iPhone is no longer a novelty; the touchscreen has become the norm. It's no big deal anymore, but taking it away would be a big deal!
https://ixdf.org/literature/article/the-kano-model-a-tool-to...

> got lots of clever behind the scene tricks like Google or Deepseek writeups) and benchmark scores

Xiaomi MiMo is led by Luo Fuli, a former Alibaba & DeepSeek employee. Perhaps it is due to Luo just how similar Xiaomi's tech & GTM approach is to DeepSeek's.

- How Luo Fuli Keeps an Earthy Touch as she Soars Through the AI World, https://newsen.pku.edu.cn/news_events/news/people/15385.html (https://archive.vn/I8Pmu).

- Luo Fuli, the 30-year-old ‘AI genius girl’ behind DeepSeek’s success?, https://e.vnexpress.net/news/tech/personalities/who-is-luo-f... (https://archive.vn/sb3B6).


Open tech is cool. Speeds up all progress...

I can barely read Jill Lawson's letter:

  My son, Jeffrey, was a very tiny, very sick premature baby, born Feb. 9, 1985, at a gestational age of 25-26 weeks. During the almost two months of his life, he was on a respirator, with several lung diseases, a heart problem, kidney problems, and a brain bleed.

  ... Jeffrey had holes cut on both sides of his neck, another hole cut in his right chest, an incision from his breastbone around to his backbone, his ribs pried apart, and an extra artery near his heart tied off. This was topped off with another hole cut in his left side for a chest tube. The operation lasted 12 hours. Jeffrey was awake through it all.

  The anesthesiologist paralyzed him with Pavulon, a curare drug that left him unable to move, but totally conscious. When I questioned the anesthesiologist ... she said Jeffrey was too sick to tolerate powerful anesthetics. Anyway, she said, it had never been demonstrated to her that premature babies feel pain. She seemed sincerely puzzled as to why I was concerned. It turns out that such care, or lack thereof, is possible because, as a neonatologist explained, babies, unlike adults, don't go into shock no matter how much agony they suffer.
From: https://massagefitnessmag.com/massage/pediatric-pain-history... / https://archive.vn/5hTuT

Can we unbury the lede here? Did Jeffrey survive?

"such care, or lack thereof" is heavily loaded; about the only objectively-sounding statements in the last paragraph are that Jeffrey wasn't able to tolerate anesthetics, and that "babies, unlike adults, don't go into shock", both of which add up to possibility of care where alternative was to leave the child to die.


> Can we unbury the lede here? Did Jeffrey survive?

It's right there at the start:

During the almost two months of his life [...]


From the link in your parent: "Jeffrey passed away five weeks after the surgery at Children’s Hospital National Medical Center."

Can we unbury your inability to RFTA lmao

There's Descartes dog too.

> At $0.04/M

Unless you meant step-3.7-flash, the input cache hits are $0.05 per mil for step-5-preview.

> $8-12 per day only for cached reads

Pretty decent "API" rates for ~500M+ tokens on Step Fun 5, a Kimi K3 / GLM 5.3 level model?

Their "Step Plan" is ridiculous, by comparison: ~$60 usage on $6.99/mo; ~$220 on $9.99/mo. https://platform.stepfun.ai/docs/en/step-plan/overview


Yes, I meant the Flash version.

I have used Kimi 2.5 and GLM 5.3 (& 5.3 Flash). Do not need them for what I do outside of spec hardening (basically, a lot of chatting).

I tend to know exactly what I want and most of the weaker models are enough to get me there. I have mainly been using MiMo, DeepSeek V4 Flash and MuseSpark Contributor over the last month or so.


Not generous enough? How token efficient and fast is it compared with American models?

LLM copyedits such as these aren't my idea of clear.

Guidelines | FAQ | Lists | API | Security | Legal | Apply to YC | Contact

Search: