Essentially if you misalign a model in one area, say opinions on left wing people, it can start exhibiting misaligned behavior in other areas, like calling itself MechaHitler.
> The A.I. company’s advisers are pushing its chief executive, Sam Altman, to move slowly after SpaceX’s stock has been volatile and as the start-up grapples with financial challenges.
Surely if your company isn't just blowing smoke then you have nothing to worry about. Or is this an admission that the insane valuation for these companies is currently just bullshit?
> if your company isn't just blowing smoke then you have nothing to worry about
Not really. Plenty of solid companies have to wring their hands around IPO timing based on market conditions. Sometimes, this is due to valuation multiples. Sometimes it's due to fads, e.g. investors preferring capital-structure efficiency versus low leverage.
I mean, my comment wasn't necessarily meant to be some insightful analysis. But I do find it weird that OpenAI has seemingly gone from racing Anthropic to "maybe in 6 months" in the span of a week.
> OpenAI has seemingly gone from racing Anthropic to "maybe in 6 months" in the span of a week
When was the last time someone seriously asked if OpenAI was going to go public before Anthropic? For me, it's been at least months, maybe closer to a year. The corporate-governance complexity drove half of that, momentum the other half, and messaging from both companies having been consistent with that timeline for months sealed the deal.
GoldSrc (HL1 engine) is very much not open source (or even source available). There's at least one open source remake (which is possibly illegal due to using the SDK) but no official release.
Couldn't you build some internal knowledge that would stay and you could teach a model this way. A very fast local memory of some sort. You could also specialize model this way so it is very skilled in your domain. The more you use it, the smarter it gets. I guess the problem is for the model to decide whether the information stored in memory is sufficient or not.
You could, but it's driving in the wrong direction to try to build that knowledge into the model weights because you'll always run into a capacity limit sooner with a small model than with a larger one. The thing the model is specialised for is linguistic understanding and the reasoning process itself, and you max that out at the expense of domain-specific knowledge. If you take "as few weights as possible" as a given, I think the interesting question is how small you can make the model with externalised memory. The openclaw and hermes people are all over this sort of memory problem: using the local filesystem or a local database of some sort is exactly a "very fast local memory" where the more you use it, the more knowledge it gathers. Whether that translates to it being "smarter" is a deeper question than it looks.
Yeah, it was funny seeing a bunch of people going like "Anthropic is fighting for privacy" meanwhile I'm like "Uhh, what about the other 8 billion people?"
As a thought experiment - such shocks (govt pressure to use models for bad purposes and govt excluding access to non-Americans) coming early in the ‘ai revolution’ will wake up the rest of the world sooner that they have to get their act together to stay competitive without relying on USA. Just like with nato.
It can, along with a bunch of other GPOs in an admx template.
But how many companies are running Workspace + Windows with on-prem AD? I suspect that number is shrinking pretty rapidly. You can do it with InTune as well, but it starts to get real messy if your users aren't on Windows or you have non-windows endpoints.
If you're a mac shop, on google workspace, and using something like JamF (or even Intune+EntraID), you are stuck deploying .plist files to each endpoint, you don't get compliance reporting back, and you lose a ton of visibility.
These are all things that don't matter to each individual user, but are hugely important to IT/security in the company, and Firefox unfortunately just doesn't have any centralized management platform for it.
Sure. But there's generally no standardized function ensuring they're actually only using that specifically configured browser when logging in. What happens when they try to log in from some other device? What happens when they manage to load a browser on to that machine?
This feature supposedly ensures (or at least pushes users to) only the approved browsers running approved configurations are allowed to log in to the company's instances of Workspace.
What if a company decides that their preferred browser is Firefox. Can you use this feature to only enable logins from Firefox? Or is it only for Chrome?
They can't use this feature to enforce only Firefox. Firefox doesn't have support for this feature, they really don't offer anything like Chrome Enterprise. Its just as much a feature of Chrome Enterprise that Workspace leverages rather than only a feature of Workspace that leverages Chrome.
You can still use Firefox with Workspace though, but if you want the management features of Chrome Enterprise you need to use Chrome Enterprise. Firefox itself just doesn't even begin to offer the same kind of endpoint verification.
With Firefox today, how would a web app have any serious clue the client was running approved versions of Firefox configured in approved ways on approved hardware with approved OS configurations? It wouldn't, and I take it Firefox wouldn't bother implementing that kind of technology. Which is fine, but if the customer wants to be able to ensure a certain kind of policy compliance that's just not possible when using Firefox. And that's just as much if not more of the ball being in Firefox's court as it is Google Workspace's. There's nothing for Workspace to even interface with at all from the Firefox side to ensure policy compliance.
Its like asking "can I print on this printer with this app?" when the app itself doesn't even have a concept of printing things. The basic underlying feature set just doesn't even exist, before we're even talking about some form of platform compatibility.
https://arxiv.org/abs/2502.17424
Essentially if you misalign a model in one area, say opinions on left wing people, it can start exhibiting misaligned behavior in other areas, like calling itself MechaHitler.
reply