fscaramuzza

Well luckily not really, their work is more foundational-y, the idea is not new, but they made it more efficient and practical. On our side, we work more on the integration of these protocol in production systems, but we are pretty much on the frontier and is not very common to see such big companies experimenting with this technology. Microsoft already tried something [1] but I'm not sure if it ever went further than a research prototype.

[1] https://arxiv.org/pdf/2402.02675

HTTPS makes sure you’re communicating on a secure channel with the API endpoint. ZKPs guarantees that the computation happened correctly behind the endpoint.

The work from Anthropic isn't a new idea, but they made the process more efficient, as inference verification is one of the biggest bottlenecks in the AI lifecycle. Beyond that, there's a lot more that can be proven, like showing a model was trained on a specific dataset with certain properties.

From an auditing perspective, this used to be handled either by showing the data (which is pretty rare), checking samples (which has been shown to be easily hackable), or similar workarounds. For inference, an auditor might want to interact with a model on demand and have absolute certainty that the model they're talking to is actually the one declared, and not a substitute.

My PhD is on auditability of AI systems, and we are exactly working on the engineering of ZKP in this context. I can confirm through our industry partner (AI Assurance and auditing), that black box auditing is indeed a huge problem for them, and snarks have shown their potential in this context. It's great to see Anthropic confirming our early results.

Google Street View is one of the most incredible engineering work produced by humans, it's just incredible.

What scares me about this new AI mode thingy is that every answer sounds like a systematic literature review, but only for the results. For example, if I look for users feedback about a specific product, it says "People think that..., but also that...; It's important to notice that some people ..." where with 'people' it means just a random comment on a random website just because it thought it was a good contribution to the results. Sounds like it's giving a ground truth from "multiple" data, when instead it's just aggregating almost random stuff. In the context of a systematic review, the feature that I would love the most is augmenting my initial query, so that I can just get more results that I could find interesting. I am 100% sure they thought about this, but ignored it for the most profitable option.

My bad, the network I was connected to didn't like the certificate for some reason. Cool project!

Am I the only one that can't access the website due to certificate problems with the .dev?

Well, Italian *nonnas rarely used more than one pan for cooking, and it was very common to just put pasta without the sauce in the plate, and a generous spoonful of sauce on top. This is what you used to find in restaurants, too.

The cooking water in the large pan is a rather new thing. Or maybe it's something just from my region :)

I like the Aluminum Finger Pieces from the first URL.

or even the "export to ascii" of draw.io? I would be happy to hear what the advantages could be.

There are plenty of these here :)

https://artsandculture.google.com/search/asset?project=art-c...

New Gemini models are quite good, Gemini 2.5 Pro is 1st in the user-benchmarks [1]. They also have Gemma, very good model that can run locally [2]. Benchmarks are not oracles of truth, but I feel like Google is not a kid who arrived late at the party anymore.

[1] https://lmarena.ai/

[2] https://blog.google/technology/developers/gemma-3/

waiting for the italian version Past

> Running overtime is the one unforgivable error a lecturer can make. This is so true. It has happened too often that a professor has decided to explain the most difficult topic in 5 minutes overtime. Before lunch. What a mistake.

GH just fixed it, but there's a snapshot from few hours ago: https://web.archive.org/web/20240608060046/https://github.co...

This kind of project reminded me of the homework we had to do in uni for the "formal methods for real-time systems" course. We had to use UPPAAL [1] to verify the stability of a transportation belt with some machines doing things with some delays and other stuff. As far as I remember UPPAAL provides also some python libraries, that can be used to make automated tests. It would be interesting to see how Hay Day scenarios could be implemented in UPPAAL and test different approaches. Anyway, very nice project, the article is really well written as well.

[1] https://uppaal.org/

OMG, the French "pissette". Last semester I was there for an internship, took me a full week to get used to it. Then there was the "râteau", used only by one of my colleagues (the strongest) when he was clearly loosing (very rare).

Thank you very much, now I want to play Universal Paperclips so bad.

https://www.decisionproblem.com/paperclips/