Starlevel004
Starlevel004 on Hacker News
I cannot say this fills me with confidence about alignment.
I have a small token/s extension for Pi. It's regularly around ~15-20t/s for Sol 6.1.
Sol 6.1 is incredibly incredibly slow, 20t/s slow (vs the ~70-80 of Terra 6.1 and Luna 6). It wouldn't be so much slower if it was just a distillation, surely? There's something else going on there.
It's widely suspected that sol 6.1 is a slightly detuned version of astra 6
This is the stock claude UI
> WebAuthn lets a site work with Security Keys and that makes such a loss basically harmless unlike for password hashes.
Yes, because all of those teenage fanfiction readers are going to have yubikeys.
> Mining ore in Minecraft shouldn't be such a challenge for LLMs and yet apparently it is!
It seems like it wasn't really a challenge at first, but GPT has a major anxiety disorder and immediately retreats at the sign of any adversity (in this case, dying) and tries to be as safe as possible.
> GPT almost immediately started using truncated forms of speech: “First leg complete x32 z9..24 y100.”
I'm pretty sure the 6.1 models see "communicate with another agent" and immediately start spewing thinking-like nonsense into the tool. The leaked traces look pretty similar.
Looks like every other claude game and it's laggy to the point I just closed it immediately.
That terminal bench 4.0 result is... bad? Terra XHigh more expensive than Sol Low and still does worse.
It's very likely this is a next checkpoint of GPT-6 "Sol", over-optimised for the web chat and offered to Codex subscribers as a "please go away" offering.
If this was half the price, I'd be very interested, but at 1/5/0.05? Feels like another panic release because they can't get Astra 6.1 out of the door...
OpenAI have an incredibly inconvenient "one billion free users" shackle attached to their leg that they need to fix first.
> It turned out that GPT-6 Sol was a dud performancewise, so a week later OpenAI announced GPT-6.1 Sol,
I strongly believe that GPT-6 Sol was only ever meant to be the web chat model, but it was a panic public release to respond to Opus 5.5. It's clearly been over fitted for drawing diagrams in the chat.
> I don't understand why it has to be non-electromagnetically interacting
Because we have zero evidence of it electromagnetically interacting. We have evidence of other particles like this, like neutrinos.
> Why can't it just be the equivalent of non-light-emitting asteroids and planets
Because they've been ruled out: https://en.wikipedia.org/wiki/Massive_compact_halo_object
You are not proposing anything new. Your ideas have been thought of before, and discarded because they lack evidence.
Well, any other theory would have to explain all of the following better than cold dark matter does:
1. Galactic rotation curves (galaxies spin too fast)
2. Galactic gravitional lensing (galaxies bend too much light)
3. The shape of the cosmic microwave background (it only matches if you have dark matter)
4. The bullet cluster (visible matter collided, but the gravitional lensing kept going)
Non-electromagnetically interacting matter solves all of these, and the same ratio of matter:dark matter solves all of them. That would be one hell of a coincidence.
Very big object x Very small chance x Very long time = normal probabilities
> But if dark matter is also interacting gravitationally, why wouldn’t it also be tangled up?
Friction between regular matter causes it to lose energy and thus velocity. Dark matter doesn't have self-interaction outside of gravity, so it just keeps going.
Being able to directly distill their huge frontier models into Luna/Haiku-level models puts them a step ahead still.
The people who work at Claude seem to talk in claudlish, so I doubt they care.
Unlike everyone else, which at least has to pretend not to train on user data, they gave out their model for basically free if you agreed to let them train on it. Opencode shows nearly 40 trillion tokens going through muse spark 1.3 contributor this week alone. That almost definitely helped close the gap a lot.
An AVIF and a WebM + AV1 is basically identical file-size wise, but you can zoom in on an AVIF so for actual animated images it wins on that front.
I tried using a passkey to log in to something and it gave me an error and I simply didn't care about trying to fix it. So passwords it is for me.
I've noticed that code mode seems to cause progressive lobotomisation in Sol 6.1 around subagents. The more it uses it, the worse it gets at giving prompts to subagents:
URLs, datasourceproxy accesspathonlyallowed anddefault404. Do not set authbasic onprivate ports; perplanPodmannetwork trustedinfrastructure nottenant boundary. Publicroutes internal /tinyauth protected internal, proxyGETemptybody /api/auth/nginx toprivate tinyauth3000; proxy_pass_request_headersoff, CookieonlyTinyauthSession header extracted map name/value actual runtime pattern tinyauth-session-[0-9a-f]{8}, X-Original-URL constructed
This is unedited; it's merging words together and spamming keywords.
For anyone not willing to click through, by "Bahrain" they actually mean Malaysia.
I'll consider it if they let me use a subscription with a custom harness, like Codex does. Until then, no thanks.
> OpenAI models used to be the prototype for robotic text, but lately I've been finding them much more natural.
This was the biggest thing I noticed in the 6 models; their conversational prose is dramatically less grating.
Okay, now price cut 6 Sol (and rename it to Terra again).
Perhaps we should never ever use hashed class names?
I'm fairly sure it's just luck of the draw if you get put onto a quant'd model or not. I've seen luna xhigh change intelligence fairly drastically on a day to day basis.
Assembly is honestly very easy to learn, especially ARM assembly, for the basics (even if whole-program assembly is still difficult). In the worst case, you end up learning something interesting.
> On top of my on-going battle with Windows + Firefox + DNS/TLS resolution sometimes stalling for seconds at a time,
Have you tried disabling HTTP/2?