Fascinating work and inspiring application of the underlying DINOv3 image segmentation model!
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.
This article discusses the DGX Spark GB10 GPU architecture from a hardware engineering perspective. The authors explain the trade-offs between datacenter Blackwell and consumer Blackwell chips, list key hardware features that evolved between GPU generations, and highlight some of the challenges that have resulted on the software side with incomplete support from Triton, Flashinfer, and various kernel incompatibilities.
This is great! I've forwarded the site to my wife.
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!
Yup, it's a small model I trained on my Mac mini! The model itself just classifies product attributes like keywords, price, retailer, etc. The features it learns are then used as embeddings
Yes, you could implement image similarity search using embeddings: Create embeddings for the entire image set, save the embeddings in a database, and add embeddings incrementally as new images come in. To search for a similar image, create the embedding for the image that you are looking for and compute the cosine similarity between that embedding and the embeddings in your database. The closer the cosine similarity is to 1.0 the more similar the images.
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].)
Thanks for the link to the ColPali implementation - interesting! I am specifically interested in evaluation benchmarks for different image embedding models.
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?
The “equivalent” here would be Jina-Clip (architecture-wise), not necessarily performance.
The ColPali paper(1) does a good job explaining why you don’t really want to directly use vision embeddings; and how you are much better off optimizing for RAG with a ColPali like setup. Basically, it is not optimized for textual understanding, it works if you are searching for the word bird; and images of birds. But doesn’t work well to pull a document where it’s a paper about birds.
Makes sense. My main takeaway from the ColPali paper (and your comments) is that ColPali works best for document RAG, whereas vision model embeddings are best used for image similarity search or sentiment analysis. So to answer my own question: The best model to use depends on the application.
This article fundamentally misunderstands the role and purpose of bureaucracy.
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.
EDIT:
To those of you downvoting this comment, please let me elaborate.
I am tired of the trope of the lazy bureaucrat because I refuse to believe that inefficient government and corporations are inevitable.
I do believe that we must strive for efficient and effective government to improve our society because the potential benefits are immense.
Those improvements must be driven by competent and qualified leaders who understand and foster the advantages that result from collaboration, communication, and making choices that benefit society as whole.
A failure of bureaucracy is a failure of leadership!
The purpose of this tool is testing if a domain name system follows (or does not follow) the correct specifications:
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.
Voluntary carbon markets typically trade in both avoidance and removal (i.e. sequestration) credits. Since this site is called Carbon Dioxide Removal I would assume listed trades only cover sequestration, and my cursory review of the listed removal methods appears to confirm that only sequestered carbon trading is listed (but the methods section does not state this explicitly).
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.
Fantastic! This site provides an intuitive overview of carbon dioxide removal purchases sourced from six different market places and two registries. Metrics include carbon credit sales and deliveries, prices, names of suppliers and purchasers, and the method of carbon dioxide removal (e.g. direct air capture or biochar production).
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!
I'm very surprised if they're able to actually get really good quality data on what constitutes removed carbon. Most carbon offsets for example are a scam, in the sense that one of the most common 'offsets' is for example paying people to not cut down trees - that they may or may not have been planning to cut down in the first place. There's no international standard or monitoring bodies, and the registries are generally incentivized to make the people paying them (industry) look good.
The whole concept is silly, we don't offset crime for instance. I can't just pay someone in advance to not punch someone so that I can do it.
Offsetting is just a way of propping up existing unsustainable business models in the eyes of ESG investors. It's greenwashing.
I'm not saying none of this is real, but I'm sure it varies dramatically and the data is probably epically unreliable.
Here's some write-ups from Greenpeace [1], NRDC [2] to the Center for American Progress [3].
As I recall, there are a variety of economical methods to remove carbon. In particular, olivine coating of beaches promises to be both cheap and effective.
I do not see a world in which all carbon emissions are eliminated, but a net zero world seems to be close. An efficient means of offsetting individual carbon expenditures is necessary for this task.
Olivine? This is mined, which means eco-system destruction from wherever it is being mined from. It means fossil fuel inputs to mine it, process it, transport it. What's the ratio of carbon-stored to carbon-released here? Any studies on the effect of of both marine animals and the beach/dune/fringe dwellers on this?
Economical? Maybe. Actually effective? Doubt it.
People are so in love with the idea of a technical solution, because deep down they don't want to give up their SUVs, their large families, their expensive gadgets and cheap food. It ain't never gonna work that way.
If people were serious about solving the climate problem, they would have smaller families, drive smaller cars, cut-out air travel, stop buying plastic shit they don't need, stop upgrading their phone every year ...
... but I'm not seeing many people doing it. Pretty much nobody in fact.
According to Schuiling's calculations, about 7 cubic kilometres of olivine a year would offset all of humanity's emissions. For reference, the largest mine in the world is the Bingham copper mine in Utah, which is (IIRC) about 23 cubic kilometres. So, it's a lot, but it's within the reach of current technology. Schuiling estimated it would take an industry roughly the size of today's oil industry to operate, probably around a million employees. Again, it's big, but it doesn't require any new magic air-sucking machines or anything like that, it's just what we already know how to do, on a scale that we already do it. And hopefully we'll have lots of coal miners out of work soon who would know how to do it.
If it can make your day a little happier: I change car/gsm/whatever when they die and can't be repaired anymore (my PC is 13 years old for example, it's alittle too weak for rust-analyze, but well :-))). I haven't taken a plane in 20 years; use the train when possible. I buy online less than once a year. I don't have a swimming pool.
But I'm not involved in politics, so I don't spend much time on helping others to change. And my house's insulation is not good (old house).
> “And my house's insulation is not good (old house).”
Depending on where you live and how you heat your house, fixing this one thing might be more effective at reducing emissions than everything else you’re doing combined.
Isn't the net result the same? Is it better to pay (for example) $500 to remove a tonne of CO2 from the atmosphere, or to pay $100 to prevent the same amount of CO2 from being emitted in the first place?
If CO2 removal costs us more than emissions reduction technologies, isn't it a false economy?
> olivine coating of beaches promises to be both cheap and effective
The problem with all these novel get-carbon-credits-quick ideas is that they are difficult to verify. How do we know that the outputs aren't taken in by some other ecological process and spat back out as CO2? How do we know the process doesn't reach some saturation point? How do we know how well it generalizes to all coastlines?
... and if we find out 5 years from now that it's bunk, do those carbon credits get revoked?
The the reason you can’t offset violent crime in the same way is that no amount of violence is acceptable but some amount of carbon emissions is acceptable.
Also in some cases damages are used to offset wrongdoing, especially in civil litigation which is a better comparison.
I would argue this is false and that if you live in a western country, someone else is performing violence for you and your ideals, so that you can live a relatively peaceful way of life
I should have been more specific: no amount of assault is acceptable. Although the other follow up that brings up the fungibility of carbon emissions versus the fungibility of assault is better I think.
>Most carbon offsets for example are a scam, in the sense that one of the most common 'offsets' is for example paying people to not cut down trees - that they may or may not have been planning to cut down in the first place. [...]
>The whole concept is silly, we don't offset crime for instance. I can't just pay someone in advance to not punch someone so that I can do it.
Governments do something similar though: gun buyback schemes are basically the same thing. Sure, maybe that gun that the government bought would have been used to commit a crime, or it was held by a responsible gun owner who wouldn't have committed any crimes to begin with.
Not sure guns are the same thing. 200-500,000 guns are stolen each year in the US and then used to commit crimes. Really anything you can do to reduce the number of guns in the US is strictly a good thing. [1]
> The large number of guns stolen each year suggests that theft may be an important source of “crime guns.” Indeed, it appears that while gun theft is often not the proximate source of firearms for most criminals [...], it is often the ultimate source-the way guns initially enter into the illegal market.
It's not 'maybe a responsible gun owner would or wouldn't use it for crime' so much as reducing the surface area of a massive vulnerability.
Harvard sounds like a reputable source. But I didn’t even need to click it to know that would be a David Hemenway link.
That guy is an absolute propagandist. A complete fraud given a platform by Michael Bloomberg who single handedly funds the majority of gun control efforts in the USA.
If you do any research on him at all, you’ll find he was center of the Clinton Admin and 1994 Congress efforts to push gun control, including weaponizing the CDC, funding compete junk studies where data was never released and extremely unethical methodology like a phone survey of 250 people who filed police reports that mentioned someone was shot in a home to conclude that for all people and all reasons “you are x% more likely to die if you own a gun in your home” (eneter Hemenway, and data never released), and all the other lies they used to push the Clinton AWB, which they were severely punished for in the 1996 elections.
Hemenway is a liar and a hack. I would highly suggest finding sources that stay far away from him.
Without addressing anything you wrote, here's three other sources that yields the same results, including one from the ATF. Regardless of your opinions regarding the author there is plenty of data that backs up this assertion.
If you're trying to convince someone that is pro 2a of anything, the ATF is hardly a source of authority. They're pretty high up there in the list of reviled 3 letter agencies.
If you are OK with no addressing content; For your reference… NIH, ATF, and Washington Post are all anti-gun organizations.
For your ATF link, it’s all University criminology depts sourcing data, typically universities are anti-gun; and the sociology depts especially so.
It is very difficult to get unbiased results on guns. And ultra easy to see authority and confirmation biases. If you are mindful of that fact, the landscape and topic takes on a new dimension.
Why is the link that direction? Could it be that universities, etc are anti-gun because the data shows they are harmful?
Overall I find it odd that this standard is applied this way. No one accuses Universities of being anti cancer when they publish a study on how cancer harms the body. Or for example dismissing any study on the effectiveness of a vaccine because the group doing the study are pro vaccines.
Yes thanks to the power of imagination guns actually shoot candies and rainbows, thanks for setting me straight. The number of stolen guns shouldn’t be controversial lol
I thought the same thing, but none of the carbon sinks listed in the database are offset companies. They are all actually producing biologically stable physical chunks of fixed carbon.
Incidentally: We own some trees that we will not cut down, but that would not have been cut down by others.
There is a scam carbon offset site that will pay us for each year we don’t cut down an existing tree (if I remember right, it’s the biggest carbon offset market; can’t remember the name, but that’s not important, if you care, I’m guessing one of the articles in the comment I am replying to names and shames them).
Is there a tax efficient way for me to sell scam carbon offsets, then funnel the money into actual carbon capture businesses? How much would the brokers take?
(Or, better, is there a set of criteria that could be encoded into law to shut down the offset scam industry?)
What if I bought up a bunch of coal on the open market, and sold carbon offsets against it? For each offset sold I’d bury the corresponding coal again, permanently preventing it from being burned.
Seems like this could be quite profitable, especially since I can sell 2.86 tonnes of carbon offsets for each tonne of coal.
Perhaps we could go a step further, skip the intermediate steps, and just pay coal miners to leave the coal in the ground?
Taking the site at face value, the claim is removal (not offsets). However, even at face value, the amount of carbon removed is far less than what has been sold.
(I agree that offsets seem unhelpful at best, just not sure it's worth deep diving on that when this isn't obviously about offsets.)
Is there a difference between removal and offsets? I tried to research the consensus but couldn't find anything. For instance the dashboard shows Microsoft purchase 2,819,637 tons of 'removed carbon' - surely to offset their own activities no? Am I missing something obvious, or is this just re-branded carbon offsets?
I think the conceptual ideal of the offsets marketplace is that it is equivalent to removal, but in practice, I don't think it lives up to that. E.g., as a sketchy offset seller, one can sell the promise not to cut down the same tree repeatedly and that dilutes the carbon offset. Or sell credits against a forest you weren't going to harvest anyway.
Removing carbon seems fundamentally different -- if the carbon is actually removed as sold, and measured correctly.
(What Microsoft is doing isn't selling into the offsets marketplace.)
I disagree with your sort of moral argument against buying offsets / removal. I think it's fine to pressure companies into paying for externalities (either via ESG shareholder advocacy or regulation) and this is just one flavor of that.
Removal: We take a specific measure to grab carbon from "the environment" and "tow it outside the environment" (capture it in some somewhat-permanent form).
Offset: We go to this poor village and teach them about the importance of not deforesting their surroundings, generously guesstimate how much deforestation that avoids, generously guesstimate how much carbon that will avoid releasing, then sell offsets (twice, because why sell them only once).
Removal is a subset of offsetting, but generally considered a) more concrete, b) actually removing carbon vs. preventing hypothetical future emissions, and as a result, c) more verifiable.
Removal subsidizes the development of the technology, whether or not it's just zeroing emissions. So regardless of the instant impact, it's progress towards the long term commercial viability of large scale sequestration.
Is your critique of the current system or the idea in general?
Would you feel the same way if the voluntary markets became less voluntary with a true carbon tax?
The companies on that list capture carbon dioxide from the air as a service for money, and they would not exist or be scaling their operations without demand.
Absolutely! I didn’t pass a judgement on those companies, just pointing out that you need some form of demand (greenwashers) in order to have any supply (CDR companies) because it’s the classic two-sided marketplace chicken-and-egg bootstrapping problem
The purpose of the market can also be clear (removing CO2 from the air) while simultaneously the “value to humanity” is unclear (hard to measure)
This paper analyzes measurements of sound pressure levels on construction sites in Brazil to argue for better worker protection. The measurements provide valuable evidence to support better worker protection, but the manuscript would benefit from a more comprehensive discussion of the effectiveness of control measures.
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.
Does the article mention what type of natural gas facilities were shut down?
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.
They keep referring to the list in "firms" instead of facilities or something like that. It sounds like they qualify at a corporate level instead of a firm, so if they're vertically integrated then it could impact all the facilities.
The dollar amount they spent also seems high enough that it may have been more profitable to shut down than actually operate.
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
reply