inkysigma

Depending on the implementation, I could see that having rate limiting effects. There're only finitely many IDs so scaling sockpuppeting will saturate these IDs quickly but it's quite easy to spin up a new anonymous account. For example, I think the EU ID system has an upcoming way to create pseudo anonymous identifiers that can identify a user per website.

This presents the problem of governments being able to gatekeep speech which I am quite uncomfortable with but maybe there's some safeguard within the eIDAS proposal that makes this idea incorrect?

Just an FYI, but I don't know if being in the website field of GitHub really helps since there's a rel nofollow on the link.

I mean in the sense that they seem to literally be random names. I don't even think they're people associated with Palantir in anyway.

I'm assuming this is satire but I'm wondering why include names of seemingly random people? Why not leave it empty or make it signed by high level known executives.

Once that side channel was found, it was kind of inevitable it would be plugged. Even under a normal administration, that's an opsec leak.

Passkeys are an open standard? You might as well argue against SSH keys.

I'm not completely familiar with bidding procedures but don't bidding procedures usually have requirements? Why not just list a requirement of unrestricted usage? Or state, we require models to be available for AI murder drones or whatever. Anthropic then can't bid and there's no need to designate them a supply chain risk.

What are you talking about? Google employees and the corporation itself in particular overwhelmingly donated to the Harris campaign.

https://www.opensecrets.org/orgs/alphabet-inc/recipients?id=...

The corporation gave millions _after_ Trump had already won. If your criticism is that, then that does not apply to the people signing.

I think the bigger insanity here is the labeling of a supply chain risk. It prohibits DoD agencies and contractors from using Anthropic services. It'd be one thing if the DoD simply didn't use Anthropic. It's another when it actively attempts to isolate Anthropic for political reasons.

I think those are two different orders: one with a gag order and one without.

In cases without gag orders, Google has pushed back or requested users fight the subpoena.

In this instance, Google got a gag order while Meta doesn't appear to have gotten one. I'm not sure how gag orders like this can be legal. I'm sure there's like Nat Sec defenses but it sure seems dangerous to say the target cannot be notified of such requests.

I think because here there's no single correct answer that the model is allowed to be fuzzier. You still mix in real training data and maybe more physics based simulation of course but it does seem acceptable that you synthesize extremely tail evaluations since there isn't really a "better" way by definition and you can evaluate the end driving behavior after training.

You can also probably still use it for some kinds of evaluation as well since you can detect if two point clouds intersect presumably.

In much a similar way that LLMs are not perfect at translation but are widely used anyway for NMT.

Would it actually be a good idea to operate a car near an active tornado?

Maybe this is why execs love LLMs a little too much? It's probably not unconnected at the very least.

I don't think the environment being cool is a factor in current data center designs is it? Otherwise, the northern US or Alaska would be candidates. Instead, a lot of the data center boom is in states like Texas or the south.

I think some interviewer with Trump did actually ask him the question you posed and he said something to the effect of "ownership is important" for him _personally_ not necessarily for the _US_ which is the a ridiculous thing to hear from a leader of a country.

Wasn't the original ARPANET entirely owned and controlled by the US government? I think it might've been resilient against attacks from people who didn't own the network but I would also be surprised to learn the US government couldn't shut it down if it wanted.

Not to be too pedantic, but I think you mean Grok with a k. Groq with a q is a separate AI hardware company.

It's the frictionless aspect of it. It requires basically no user effort to do some serious harassment. I would say there's some spectrum of effort that impacts who is liable along with a cost/benefit analysis of some safe guards. If users were required to give paragraph long jailbreaks to achieve this and xAI had implemented ML filters, then I think there could be a more reasonable case that xAI wasn't being completely negligent here. Instead, it looks like almost no effort was put into restricting Grok from doing something ridiculous. The cost here is restricting AI image generation which isn't necessarily that much of a burden on society.

It is difficult to put similar safeguards into Photoshop and the difficulty of doing the same in Photoshop is much higher.

I wonder what the terms of this new deal are. Goldman Sachs had accepted absolutely ridiculous terms like forgoing fees because they were apparently desperate for consumer business. Chase has an existing consumer business so I can't imagine they would accept the same terms but I wonder what this would mean for the card benefits.

While there isn't a way to differentiate between scraping for training data and content accessed in response to a user request, I think you can block Googlebot-extended to block training access.

Iirc, the problem with Bayesian neural networks is that they're significantly more difficult to train. Using stuff like SVI reduces a lot of the representational ability of the distribution over weights. It's also questionable how useful the uncertainty over weights is.

I suppose in the tradition of Bayesian influence, VAEs and the like are still common though.

It's weird to me that when people talk about Tesla, the business, is doing poorly, the retort is citing the stock price as if that's somehow what the person is talking about and not going to be seen as disconnected from fundamentals by the person you're talking to. So many things were being used to justify the stock price have simply not materialized. For example, the idea that Dojo would be a lasting moat does not appear to be materially true.

As a business, they've seen declining yearly revenues (2025 is expected to be lower overall) and declining sales on a YoY basis in a sector that has seen tremendous growth. The article we're on literally is how Tesla squandered their lead over BYD.

Autopilot and FSD are neat (for some definition of neat) but have clear competitors with Waymo being far ahead in deployment. BYD has their own developing system that's not as good but it's not clear to be that FSD will be some kind of enduring advantage either given that parallel systems are also being developed.

What makes you think that the entire process isn't being made more efficient? There are entire papers dedicated to pulling out more FLOPs from GPUs so that less energy is being wasted on simply moving memory around. Of course, there's also inference side optimizations like speculative decoding and MoE. Some of these make the training process more expensive.

The other big problem is that you can always increase the scale to compensate for the energy efficiency. I do wonder if they'll eventually level this off though. If performance somehow plateaus then presumably the efficiency gains will catch up. That being said, that doesn't seem to be a thing in the near future.

I think the concern is that this might somehow enable a privacy policy they weren't aware of that permits training over the entire Drive. However, I think the primary reason for this is that these products generally would like to store data on the user's Google Drive but Google Drive doesn't have super granular permission structure to be able to set up a partitioned directory for the app alone. I actually think that might be a good thing to work on next?

It's actually a different crawler for training data: Googlebot-extended so you can exclude yourself from the training data though not the search summaries.

> Large language models are something else entirely*. They are black boxes. You cannot audit them. You cannot truly understand what they do with your data. You cannot verify their behaviour. And Mozilla wants to put them at the heart of the browser and that doesn't sit well.

Am I being overly critical here or is this kind of a silly position to have right after talking about how neural machine translation is okay? Many of Firefox's LLM features like summarization afaik are powered by local models (hell even Chrome has local model options). It's weird to say neural translation is not a black box but LLMs are somehow black boxes that we cannot hope to understand what they do with the data, especially when viewed a bit fuzzily LLMs are scaled up versions of an architecture that was originally used for neural translation. Neural translation also has unverifiable behavior in the same sense.

I could interpret some of the data talk as talking about non local models but this very much seems like a more general criticism of LLMs as a whole when talking about Firefox features. Moreover, some of the critiques like verifiability of outputs and unlimited scope still don't make sense in this context. Browser LLM features except for explicitly AI browsers like Comet have so far had some scoping to their behavior, either in very narrow scopes like translation or summarization. The broadest scope I can think of is the side panels that show up which allow you to ask about a web page with context. Even then, I do not see what is inherently problematic about such scoping since the output behavior is confined to the side panel.

The answer should be no right? I think BPP is expected to be equal to P and BQP to be not equal to P.

1p solutions still require consent since the analytics banners are also there to enable processing of personal information in the first place (on the most primitive level IP address)

Do you know an analytics service that actually does this? I've seen a bunch of "consentless" analytics solutions that seem to be violating GDPR one way or another because they use the IP address as an identifier (or as part of one).

I think this scheme still requires consent since you are processing pseudo anonymous identifiers that fall under personal information without the essential function basis. Hashing is considered insufficient under the GDPR iirc. Have you asked a lawyer about this?

Can you actually do meaningful analytics without the banner at all? You need to identify the endpoint to deduplicate web page interactions and this isn't covered under essential use afaik. I think this means you need consent though I don't know if this covered under GDPR or ePrivacy or one of the other myriad of regulations on this.