Hacker News new | past | comments | ask | show | jobs | submit | desdenova's comments login

What makes models non-deterministic isn't the training algorithm, but the initial weights being random.

Training is reproducible only if, besides the pipeline and data, you also start from the same random weights.


That would fall under "Feed the same data in and you'll get the same weights out." Lots of deterministic algorithms use a random seed.


So is there no “introduce randomness” at some step afterwards? If not, I would guess these models would be getting stuck in a local maxima


> If not, I would guess these models would be getting stuck in a local maxima

It sounds like you're referring to something like simulated annealing. Using that as an example, the fundamental requirement is to introduce arbitrary, uncorrelated steps -- there's no requirement that the steps be random, and the only potential advantage of using a random source is that it provides independence (lack of correlation) inherently; but in exchange, it makes testing and reproduction much harder. Basically every use of simulated annealing or similar I've run into uses pseudorandom numbers for this reason.


The models that run the search overview slop are probably some shitty 1B models running as cheaply as possible.


In the Ollama API, you use the "tools" parameter to describe the available tools to the model, then use the "tool_calls" from the response to call the functions and send the results back to the model using "role": "tool".

Most of this is handled very easily by the ollama-python library, so you can integrate tool calling very simply in any script.

That said, this specific model was unable to call the functions and use the results in my "hello world" tests, so it seems it expects a few very specialized tools to be provided, which are defined by that platform they're advertising.

Right now the best tool calling model I've used is still qwen3, it works very reliably, and I can give it any ability I want and it'll use it when expected, even in /no_think mode.


I did a very simple tool calling test and it was simply unable to call the tool and use the result.

Maybe it's specialized to use just a few very specific tools? Is there some documentation on how to actually set it up without requiring some weird external platform?


That 10% would be over the other taxes.

But Trump also managed to devalue the dollar so much, we won't even feel the difference.


Despite the original naming being initials, there are many other countries who are members and partners of the BRICS:

https://en.m.wikipedia.org/wiki/Member_states_of_BRICS


Drivers are precisely the reason we only have Windows, Macos and Linux as viable options.

Hardware manufacturers write Windows drivers, Linux community write drivers for basically all consumer hardware, and apple develops both the hardware and the OS with their own drivers.


>Drivers are precisely the reason we only have Windows, Macos and Linux as viable options

That is one big issue and another one is software.

Writing drivers and porting software means both time and money.

How ever if a new OS would bring lots of benefits to both users and companies, it might tip the scale and make the time and money investment worthwhile.

Of course, by a new OS I don't mean just another platform that ebables us to run software and use hardware, as existing OSes do that just fine.

By a new operating system, I mean one that enables us to use new computing paradigm, enable new types of software to software and software to hardware interactions and would make a big disruption to the market. Something with the same kind of impact as AI or the introduction of smartphones.


Everything is so fluid nowadays. Even if use-case was found, someone would find a way to run Linux on it.


Yeah but finding a way to run it is having a hobby of debugging software and work arounds on top of your job.


So now every trend that spawns a subtrend is the new real web 3.0


I don't think so, I write about why I think this one warrants it though in the article. Do you disagree with any specific points?


Why does the site have CSS then?


Its a fairly ragtag operation, they do do it, just manually and I guess nobody has gotten to it yet

https://github.com/css-naked-day/css-naked-day.github.io/com...


“Do as I say not as I do” and other hypocritical expressions come to mind.


There's plenty of OSD licenses that don't fit the FSD, but a free software license is necessarily open source, so the opposite can't happen.


I’d really be interested in an example of such a license. Where is the difference in the two definitions?


Not OP, but I linked to some details here:

https://news.ycombinator.com/item?id=43581484

The NASA Open Source Agreement is the one I found.


Thank you!


No there cannot be an OSD license that violates the FSD


GNU has a helpful chart where they clearly show that there is a sliver of "nonfree open source" licenses that are available [0].

> The term “open source” software is used by some people to mean more or less the same category as free software. It is not exactly the same class of software: they accept some licenses that we consider too restrictive, and there are free software licenses they have not accepted. However, the differences in extension of the category are small: we know of only a few cases of source code that is open source but not free.

I was able to find one example, the NASA Open Source Agreement, which is accepted by the OSI [1] but rejected by the FSF [2]:

> The NASA Open Source Agreement, version 1.3, is not a free software license because it includes a provision requiring changes to be your “original creation”. Free software development depends on combining code from third parties, and the NASA license doesn't permit this.

[0] https://www.gnu.org/philosophy/categories.html

[1] https://opensource.org/license/nasa1-3-php

[2] https://www.gnu.org/licenses/license-list.html#NASA


Guarantee that all "nonfree open source" is different readings. Take the NASA case. If youu read it as strictly as Stallman does then it violates the OSD also. The people at OSI at the time it was submitted read it more like a lawyer and decided it was compliant. Possibly today's OSI would disagree. Possibly tomorrow's FSF would agree. It's not a difference between free software and open source but a difference between how two sets of humans interpreted the text of the license.


Which point of the OSD would be violated by the NASA clause if read the same way that the FSF reads it?


OSD 3

And possibly 9


Eh, OSD 3 just says that derived works must be possible, it doesn't say that you must be able to incorporate third party source code into the derived work. Meanwhile the FSF's definition explicitly calls out this freedom as an essential component of Freedom 1:

> One important way to modify a program is by merging in available free subroutines and modules. If the program's license says that you cannot merge in a suitably licensed existing module—for instance, if it requires you to be the copyright holder of any code you add—then the license is too restrictive to qualify as free.

https://www.gnu.org/philosophy/free-sw.en.html#four-freedoms


Consider applying for YC's Fall 2025 batch! Applications are open till Aug 4

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

Search: