Overall it's worse than the other frontier models, but it's decent for queries about breaking news, due to being trained on twitter data. It's also better for queries about controversial topics, and topics that the other labs have deemed to be "unsafe".

Politically, it differs quite a bit from other models.[0] It's right leaning, although it's closer neutral than other models, defining what neutral is a challenge though.

[0]: https://arxiv.org/abs/2603.23841

The study you link to doesn't take into consideration the Overton window of opinions. Perhaps there's some dimension along which you could say that one ideology lies 'opposite' to another political persuasion, but that doesn't necessarily mean that the two ideologies are equally acceptable to support in a given society.

I don't think calling defining neutral a 'challenge' does the question justice - neutral will always be context-dependent, and what may be in the center of the Overton window of one society may be unpopular or even highly illegal in a different society.

Wasn't it just, likely, a Claude proxy, then a local LLM for a while, then now-ish an OpenRouter proxy?

> due to being trained on twitter data

twitter data is 70%+ bots (probably more than that now)

Grok is of course also trained on the same giant blob of "all human writing" that the other models are trained on.

stated goal ≠ output

see: democratic people's republic of korea, the chinese communist party, american first

It's a good comment, it explains the reason for the setting.

They didn't expect to leak their source code.

It's hardly a trade secret, what value is this to a competitor?

I wish there was a way to cap the lockout time.

It makes sense for 4 digit codes, but I have a 20ish character password, I once locked myself out, and it was an incredibly frustrating experience.

My password can't be brute forced even with offline access to the hash, there is no risk of it being brute forced from keyboard input.

The inevitability is that these kinds of services just won't be offered without identifying yourself.

Claude's free tier requires a phone number just to try it.

PRISM as a Service.

I like to think of Palantir as JIRA or salesforce for killing people.

You can run it on-prem, where you can actually technologically enforce data custody.

Custody enforcement using the cloud hosted product, is mostly contractual, although they do offer some technical features, like encrypting all data using a AWS KMS key in the customer's AWS account.

Still, this relies on trusting that they won't make their own separate copies of the data.

Gemini also won't help you with C++ if you are under 18, since it would be unsafe.

https://qht.co/item?id=39632959

Is it still true? That's two years old

It's improved significantly in that time, but relative to the other frontier models, it is still the one that is the most condescending and coddling.

I don't quite understand the 5 second overlap. I assume it's so that events that occur over the chunk boundary don't get missed, but is there any examples or benchmarking to examine how useful this is?

yea, it's so events on a chunk boundary still get captured in at least one chunk. i haven't had the chance to do formal benchmarks on overlap vs. no-overlap yet. the 5s default is a pragmatic choice, long enough to catch most events that would otherwise be split, short enough to not add much cost (120 chunks/hr to ~138). also it's configurable via the --overlap flag.

> took a private key from KMS

They used KMS to sign the minting operation, but they didn't "take" the key, AWS KMS doesn't let you extract keys.

^ this is a common security misconception in crypto. "We're using an HSM, they can't steal our private key." OK genius now you still have to secure the HSM.

There's no shortcut to MPC/multisig with 3+ keyholders.

It's still significantly better, since access can be revoked, vs a leaked key where you're permanently fucked

Not much better because even a single signature can drain your whole wallet.

> you still have to secure the HSM

Obviously.

> There's no shortcut to MPC/multisig with 3+ keyholders.

The whole concept of a stablecoin seems to be based on centralised trust. Ultimately there is some org that has the fiat bank account, that mints and redeems the coins.

Nope, that is the foundation of bad stablecoin. Trustless decentralized stablecoin like DAI exist. People just largely don't do their homework and prefer scams that lure them in with promises of 'yield'

DAI and SKY are backed in large part by USDC, so they are not truly decentralized. It is possible in theory, but nobody has successfully done it so far.

It's possible in practice: that's how DAI worked originally. It's just not very competitive where the main customer -- traders -- want a lot of liquidity and razor thin spread.

DAI made some dumb decisions for market reasons recently but it was an actual stablecoin for a long time. It worked fine, they just decided to make it worse for some reason.

If their coin requires a web service to process each transaction, then an offline key isn't very useful.

You can criticize their design, but you can't have a dude burning a CD-ROM every time someone wants some coins.