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

The approach in the paper specifically addresses the case where an LLM can usually solve a task when it requires few steps, but fails for the same kind of task with more steps because it randomly gets a step in the middle wrong and then derails. It can't do anything for tasks that the LLM can't solve even when there's just a few steps.

In other words, it compensates for random error, not systematic error.



The comparison is against general models which are explicitly fine-tuned. Specifically, they pre-train their models on unlabeled in-domain images and take DINO models pre-trained on internet-scale general images, then fine-tune both of them on a small number of labeled in-domain images.

The idea is to show that unsupervised pre-training on your target data, even if you don't have a lot of it, can beat transfer learning from a larger, but less focused dataset.



I recommend writing a blog post explaining the constraints in the FPGA example in detail and pointing out the constraint conflict your method identified. That would make for a better sales pitch than merely asserting the existence of this example.

Deterministic output is a property of the inference framework, not the model. E.g.

https://docs.sglang.ai/advanced_features/deterministic_infer...

https://github.com/ggml-org/llama.cpp/pull/16016


Really appreciate this, this is exactly what I was looking for.

It cannot. But it can detect whether the terminal emulator is shifting the cursor by the correct number of cells when displaying a character.

Glyph support is typically left to the system font stack, but positioning and hence cursor movement needs to be handled by the terminal emulator.


In that case, I want a detection software that can screen-grab the terminal to check if the glyphs are correct.

This is to protect against shoulder-surfing. Having the information present on the client and accessible to the user while appearing redacted is very much the point.

A jpeg of a copyrighted image can be copyright infringement, but isn't necessarily. A trained model can be copyright infringement, but isn't necessarily. A human reciting poetry can be copyright infringement, but isn't necessarily.

The means of reproduction are immaterial; what matters is whether a specific use is permitted or not. That a reproduction of a work is found to be infringing in one context doesn't mean it is always infringing in all contexts; conversely, that a reproduction is considered fair use doesn't mean all uses of that reproduction will be considered fair.


(I'm not an egyptologist.) The Thesaurus Linguae Aegyptiae has a breakdown of 𓅓𓂋𓄋𓏏𓏝𓊹𓊵𓏏𓅓𓇾𓂋𓇦𓂋𓆑 / jm.j-rʾ-wpw.wt-ḥtp.w-nṯr-m-tꜣ-r-ḏr=f / "overseer of apportionments of the god's offering(s) in the entire land" https://tla.digital/lemma/850281 with 𓇾 / tꜣ / "land" and 𓂋𓇥𓂋 / r-ḏr / "entire".

So we have most of 𓇍𓇋𓏭𓂻𓍘𓇋𓇾𓂋𓇥𓂋𓈐𓆑 / jy.tj-tꜣ-r-ḏr-?? / "welcome land entire ??" except for the 𓈐𓆑 at the end where I have no idea whether it's phonetic ḥr=f or a determinative https://en.wiktionary.org/wiki/%F0%93%88%90 or something else.


I think the =f at the end is a possessive, because r ḏr=f is an idiomatic phrase literally meaning “to its limit”.

𓇍𓇋𓏭𓂻𓍘𓇋𓇾𓂋𓇥𓂋𓈐𓆑

jy.tj t3 r ḏr=f

come [STATIVE] land [VOCATIVE] to limit its

“Welcome, entire land”

(I’m not an Egyptologist either.)


Did you forget about the 𓈐 or does it get merged into the preceding word?

I think 𓈐 is the end of the preceding word, ḏr. It is functioning as a logogram, not a phonogram:

https://en.wikipedia.org/wiki/Determinative


What a time to be alive: my iPhone has enough Unicode coverage to display hieroglyphics… imagine thinking two decades ago that you could use disk space on a mobile device like that.

(It doesn’t have the glyph layout chops, though…)


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

Search: