ResearchAtPlay

Have you ever tried to achieve consistently deterministic output from an LLM? I have, and it's not easy.

That means output differs between machines and architectures. Running inference on CPU vs GPU also affects output. Even running the same prompt twice in a row on the same machine can lead to different outputs because a prompt that was partially stored in the kv cache will result in different output than an uncached prompt.

LLM output is very much not deterministic!

The technology already dictates that you must write the first draft yourself because AI can only respond to your initial prompt. That initial prompt IS your first draft.

The title's statement (and the argument presented in the rather short article) appears misinformed about the strengths of AI and where those strengths can support an effective writing process that leads to concise, thorough, and pleasant communication with your audience.

Large language models excel at sketching initial outlines, adding new perspectives, highlighting thematic connections, and shaping format, spelling, and grammar. But their tendency to converge towards an average makes those models terrible at selecting precise terminology, avoiding verbosity, and conveying passion through careful iteration.

My advice: Let AI throw truckloads of slop onto your daunting blank page! Read it to spawn your thoughts, then re-write your message from scratch! That's right - do not attempt to salvage the slop. Borrow the inspiration but avoid copying the words. Then re-read and re-fine. And then again. And again.

And much like the principles of good design have taught us, your writing task is complete when no word can be removed and no term could be replaced to more accurately express your intent.

(Disclaimer: This comment was entirely written without AI.)

Several paid providers for X.509 certificates exist but document signing certificates cost around 80 € per year [0]. And if I want duplicate X.509 certificates for my redundant Yubikeys then the cost doubles.

Other providers require an initial deposit and then charge per signature [1], which leads to intransparent pricing. In the interest of open commerce, I strongly believe that securely signing an electronic document should cost the same as my manual signature, i.e. nothing.

A partial solution already exists because I can use my electronic ID card with the AusweisApp to prove my identity when interacting with German authorities. This feature is generally useful because I live outside of the EU, but I especially appreciate that I can have my OpenPGP key signed by Governikus (a government provider) to prove the key belongs to my name [2].

Technically, I should be able to use my certified PGP key to sign documents, but in practice most non techies don't know how to validate my signature. For the average user opening my signed PDF in Adobe Reader, I would need an X.509 certificate from a trusted Certificate Authority for users to see the green check mark.

[0] https://shop.certum.eu/documentsigning-certifcates.html

[1] https://www.entrust.com/products/electronic-digital-signing

[2] https://pgp.governikus.de/wizard/requirements

The blog post and paper [1] describe a promising approach to solving related problems at previously impossible scale and quality: I am currently exploring methods to better represent seasonal land cover changes that would improve wind power generation forecasting and this paper provides a great starting point.

I hope DINOv3 can inspire more work like this - and I would encourage any curious mind to play with that model! I was amazed by its capability to distinguish between fine object details. For example, in a photo of a bicycle, the patch embeddings cleanly separated the background from the individual spokes of the wheel.

[1] https://arxiv.org/abs/2603.06382

Would you mind sharing how you trained the model to produce the vectors? Are you using a vision transformer under the hood with contrastive training against price, product category, etc.?

EDIT: I see that the training script is included in the repo and you are using a CNN. Inspiring work!

I see the ColiVara-Eval repo in your link. If I understand correctly, ColQwen2 is the current leader followed closely by ColPali when applying those models for RAG with documents.

But how do those models compare to each other and to the llama3.2-vision embeddings when applied to, for example, sentiment analysis for photos? Do benchmarks like that exist?

For choosing a model, the article mentions the AWS Titan multimodal model, but you’d have to pay for API access to create the embeddings. Alternatively, self-hosting the CLIP model [0] to create embeddings would avoid API costs.

Follow-up question: Would the embeddings from the llama3.2-vision models be of higher quality (contain more information) than the original CLIP model?

The llama vision models use CLIP under the hood, but they add a projection head to align with the text model and the CLIP weights are mutated during alignment training, so I assume the llama vision embeddings would be of higher quality, but I don’t know for sure. Does anybody know?

(I would love to test this quality myself but Ollama does not yet support creating image embeddings from the llama vision models - a feature request with several upvotes has been opened [1].)

[0] https://github.com/openai/CLIP

[1] https://github.com/ollama/ollama/issues/5304

Bureaucracy is tool to manage large, complex, and heterogeneous systems. Ideally, efficient and effective bureaucracy goes unnoticed. Why can I plug my laptop into the power outlet anywhere in Miami or in Vancouver and it just works? Why can I drive on the right side of the road from Toronto to San Diego and be reasonably sure that everyone else will drive on the right side as well?

Because humans have self-organized into a multitude of governments, standards organizations, and corporations that all align to produce the same shape of power plug and teach compatible rules-of-the-road across vast geographical distances and unrelated communities. Without bureaucracy, we humans would not be capable of building a global society.

Pointing to broken, ineffective, and inefficient processes to scapegoat “the bureaucrat” reveals an ignorance of the underlying mechanisms that make human society function.

IBDNS fills a gap in the universe of DNS test tools by offering the possibility of deviating intentionally and on demand from the DNS specifications, and thus simulating incorrect behaviour of authoritative name servers.

Regarding "is it paying for stuff that would have happened anyway, or is it somehow net removal?": One of the requirements for generating carbon credits is additionality, i.e. a project should only receive carbon credits if it were not viable without the revenue from those credits. But as you point out, determining additionality is rather difficult and often fuzzy.

I find this site useful to get an overview of the development of voluntary carbon markets and their recent rapid expansion. Voluntary carbon markets are fractured into several market places and registries, so getting an aggregate overview over all markets was somewhat difficult prior to discovering this site. Thank you for the submission!

For example, the authors suggest rearranging machinery to limit the number of workers exposed to their noise, but I would have liked to read at least a cursory analysis of how much such measures can reduce noise, and whether administrative or engineering solutions can sufficiently meet safe noise thresholds. If such analysis wasn't conducted within the scope of this work, then literature review could inform such discussion.

Edit: added lit. review recommendation

Large loads (compressors, pumps) usually consume a share of the gas they transport because gas is so much lower cost than electricity. I wasn't aware that any facilities have a large enough electrical load to become part of the voluntary load shedding program.

There are many different residency programs, some of which take longer than others, but the Express Entry program is probably the default option for most HN readers.

[1] https://www.canada.ca/en/immigration-refugees-citizenship/se... economic immigration > skilled worker (federal) > I haven't applied yet

An offending example might be a slide showing complicated equations that remain unexplained by the presenter. I've observed many presentations where such slides are introduced as "this is the equation used to derive value X from earlier, but I am just going to move on to the next slide..."

edit: format

Rooftop mounted, optimally oriented PV systems at Seattle latitude have an annual capacity factor of ~ 14%. That means a 1 kW system will average 140 W output over a year. That system has a module area of about 4-7 m^2, equivalent to the area available on a sedan.

On a car that has suboptimal module orientation and solar exposure, "power density" (RE: Smil) is really low.

In contrast, taking your module efficiency from 20% to 21% increases electricity generation by 5% and thus reduces costs per kWh by 5%.

In my opinion the biggest issue is terrible documentation on ipfs.io (as noted by the report). The "Getting Started" page fails to explain the basic structure of ipfs and uses jargon that isn't defined. This lack of documentation prevents even tech savvy users from adopting ipfs.

The rationale for creating this model is the lack of up-to-date electric grid location data, especially in developing countries. While I find the approach fascinating and well documented, I think the error in this model’s results is much too high for further use. I would not want to rely on any electricity system planning that utilizes these model results, because the model may omit entire branches of the grid.

Can anyone think of any useful applications of these model results?

Maybe you can partner with the folks over at https://www.electricitymap.org/ to improve the emissions data?

No, they have not! I am not sure by what metric one can compare FIFA and the United Nations, but I don't think you you are familiar with the breadth of the UN's work. Selected achievements from their results page [1]:

-Every year we mobilize about $7 billion in humanitarian aid to help people affected by emergencies.

-Every year we assist over 34 million refugees and others fleeing war, famine and persecution.

-We vaccinate 40% of the world’s children, saving 2 million lives a year.

-In 2011 the UN will provide food to around 90 million people in 73 countries.

-In the past few years, the UN has expanded legal international rights to indigenous peoples, people with disabilities, migrants and their families.

FIFA is merely better at marketing...

[1]http://www.un.org/en/strengtheningtheun/results.shtml

In my opinion, Canadian banks are way overdue to switch to 2FA.