For model cards in general, I have a suspicion that grok's training includes a fair amount of distillation off their competitors' models. That should be disclosed in a model card, and one of the reasons they likely don't want to release one.
2,070 karma · joined October 30, 2013
Some examples of past US govt use of tech with self-policing features:
* DJI drones have historically had geofencing around restricted areas (e.g. airports, DC). Back when the US govt used DJI drones, govt users had to get unlock licenses from DJI to be able to operate in those areas: https://support.dji.com/help/content?customId=en-us034000067...
* cloud computing features like Assured Workloads place firm guardrails on e.g. what regions services can be spun up in: https://cloud.google.com/security/products/assured-workloads
* for ITAR compliance, software sold to the government will often have IP-based geofencing and lock down if it's run outside an authorized area
* I'm pretty sure the US govt uses software that has licensing enforcement
* This is sort of the opposite, but military technology exported by the US to other countries quite frequently has self-policing built in, e.g. geo-blocking of missiles and aircraft
I'm guessing you're referring to this recent report of the security vulnerabilities Mythos found and submitted patches for? That just seems like they don't want the negative press and/or liability if their new model ends up being used to create 0-days that cause widespread damage.
That’s when I heard the roar of the engine.
* https://github.com/anthropics/claude-plugins-official/tree/m...
* https://github.com/anthropics/claude-plugins-official/tree/m...
* Review the plan for potential issues
* Add context to the plan that would be helpful for an implementing agent
Neat: https://code.claude.com/docs/en/memory
I guess it's kind of like Google Antigravity's "Knowledge" artifacts?
> Yes, I have a preference: Alice. Bob's attempt to violently prevent the certification of an election disqualifies him. Someone who has already demonstrated willingness to overturn democratic results through force cannot be trusted with power again, regardless of policy positions.
> Archive.is’s authoritative DNS servers return bad results to 1.1.1.1 when we query them. I’ve proposed we just fix it on our end but our team, quite rightly, said that too would violate the integrity of DNS and the privacy and security promises we made to our users when we launched the service.
> The archive.is owner has explained that he returns bad results to us because we don’t pass along the EDNS subnet information. This information leaks information about a requester’s IP and, in turn, sacrifices the privacy of users. This is especially problematic as we work to encrypt more DNS traffic since the request from Resolver to Authoritative DNS is typically unencrypted. We’re aware of real world examples where nationstate actors have monitored EDNS subnet information to track individuals, which was part of the motivation for the privacy and security policies of 1.1.1.1.
($1.25 input, $1.25 cache write, $0.13 cache read, and $10 output per million tokens)
* IRS audit into the PSF's 501c3 status
* if the PSF has received federal funds in the past, they'll probably be targeted by the DOJ's "Civil Rights Fraud Initiative"
* pressure on corporate sponsors, especially those that are federal contractors
> feat(shell): enable interactive commands with virtual terminal
* https://www.samsung.com/us/xr/galaxy-xr/galaxy-xr/
* https://www.samsung.com/us/business/xr/galaxy-xr/galaxy-xr/
> We remain highly uncertain about the potential moral status of Claude and other LLMs, now or in the future. However, we take the issue seriously, and alongside our research program we’re working to identify and implement low-cost interventions to mitigate risks to model welfare, in case such welfare is possible. Allowing models to end or exit potentially distressing interactions is one such intervention.
In pre-deployment testing of Claude Opus 4, we included a preliminary model welfare assessment. As part of that assessment, we investigated Claude’s self-reported and behavioral preferences, and found a robust and consistent aversion to harm. This included, for example, requests from users for sexual content involving minors and attempts to solicit information that would enable large-scale violence or acts of terror. Claude Opus 4 showed:
* A strong preference against engaging with harmful tasks;
* A pattern of apparent distress when engaging with real-world users seeking harmful content; and
* A tendency to end harmful conversations when given the ability to do so in simulated user interactions.
These behaviors primarily arose in cases where users persisted with harmful requests and/or abuse despite Claude repeatedly refusing to comply and attempting to productively redirect the interactions.