impossiblefork

It could be simple training tips that are obvious, but which any particular builder might miss.

It's nice that they're passing some laws like this allowing people some chance to a get a remedy.

I have a sort of science-fiction story premise that I haven't worked out. A US president finds twenty or so monsters in the security state, people with the highest type of the security clearances and who can possibly be induced to become something like Beria, and he makes an executive order using something like the Atomic Energy Act to create an office for securing US nuclear material-- very obviously this office must be exceedingly secret and have, within its domain, very great powers. But he appoints these twenty men he finds to it, and he arranges for many of his political opponents to travel to the District of Columbia, where he uses ordinary policemen to arrest them contrary to law, but using administrative warrants so as to secure the co-operation of the policemen, in bringing his political opponents to places where they are killed to by these men working for this new office; he continues to use administrative warrants and certain other deception to get people to D.C. to be killed.

Now of course, you could imagine that this would be some sort of fascist takeover story, but instead the guy lets courts continue, with them working as they do in real life, unhindered-- no threats, no special instructions, nothing.

Instead we get to see prosecutors not bringing suit, because they can't, because they of course work for the executive branch, state prosecutors not being able to bring suit, because it's D.C., civil lawsuits dismissed because the executive invokes the state secrecy privilege-- all while the government is ordering people to D.C. and conducting mass executions.

I was inspired by the El-Masri case, and some other cases of this sort and it eventually became clear to me that it the US executive could basically conduct an actual straight-up holocaust, obviously illegally, but without the courts intervening. This is why I think things like the duty to prosecute and the right of courts to evaluate evidence freely, that we have in Sweden, and the ECHR right to an effective remedy are so important.

We've failed in this sort of thing too though. In North Macedonia the men working for North Macedonia who held El-Masri for some time have not been prosecuted even now, even though North Macedonia has sent an official apology and paid him what the the ECtHR asked. Still the CoM of the CoE should have intervened more aggressively and ensured that North Macedonia actually prosecuted at least those of the perpetrators who are among their politicians and bureaucrats.

I think we in the EU might even need to bring back some limited form of private prosecutions or other measures to deal with the worst-case situations, where states go after people, because there should not be impunity even for people working in state security.

It beats z.ai's GLM 5.3, which everyone agreed was interesting a couple of weeks ago when it was released.

It's exactly where a model trained with these resources in a modern and reasonable way should be expected to be.

I also think it demonstrates that GLM 5.3's performance is probably not primarily due to distillation but simply the performance one can get out of a a model of that size and with that architecture.

Very nice. I always expected it to beat z.ai's model by at least a little bit, since it's so similar and that it would be slightly behind Kimi. Maybe this can change with better post-training, but it's great to go from not having a model in this size class to making something competitive with z.ai, who of course have made several good large models.

A first step on a trip which will hopefully go quite far.

I actually see the point of discord as the fact that it is less public.

It means I can put half-finished story summaries there, without spoilering or alerting potential future readers to the plot ideas, that sort of thing.

A hierarchy of privacy levels is necessary in order for people to speak as they like.

This is only good for Micron if 4F^2 is actually unachievable though.

If others end up making denser cells and Micron can't they will genuinely be behind. 6/4=1.5. It is not a small improvement.

I think this is definitely the "interest rates are high, dropping research and just pushing out what we already make gives us a higher valuation" type thing which can kill companies.

If it targets combatants, it isn't terrorism.

If it targets others, it is. Some countries have ideas like that only non-state entities can commit terrorism, but that is nonsense. My own country's laws don't distinguish between state and non-state acts in its terrorism legislation, for example.

I never believed the EHT, but I can't prove I didn't and I only didn't believe it for the shallow reason that I didn't feel that there was a strong argument that it should actually be exponential, since there are so many almost-exponential things that it could equally well be.

Surely not really?

The card emphasizes training resources, the other comment emphasizes ingenuity and few people?

I actually don't think so.

I think Anthropic actually has enough revenue to function. I think they cover their compute obligations with revenue, I think they can probably cover reasonable salaries too.

I made some calculations earlier, but with 16.8B for Q3, what is the problem? 2.5B per annum should cover the necessary salaries. After that they have 64.7B. Compute obligations over the coming ten years are only 518B. 8x their their current revenue.

I think this is actually a stable business. Probably not compatible with a 2T valuation though, especially not at present interest rates.

I mean, the model architecture is basically that of GLM 5.3 with slightly more parameters and slightly more active parameters, so it should probably be able to beat GLM 5.3 enough post-training.

I seem to have used the wrong numbers. Revenue is apparently reported per quarter.

They aren't in trouble at all.

Yeah, i calculated something like that earlier.

But with a 2T valuation-- if they did raise, let's say, 300B, that's a 15% float, maybe they'd have a chance.

518 B in future compute obligations for the coming decade. 300 B, then 218 B is left. That's 13x their current revenue, not too far from 10x their current revenue.

Perhaps they could survive if they raised 300B. With that they'd only need to increase revenue a little bit, and that's probably possible, what with inflation and all.

The big problem is if there's improvements in AI accelerators that aren't going to be be bought by the companies that the people to whom Anthropic have obligations are likely to buy from. So if Google, AWS or NVIDIA end up not being the leaders in tokens/$, then Anthropic are screwed.

I'm not sure how the total picture ends up if one adds OpenAI to the picture.

I think the arguments presented here are really in favour of ensembles of simple stories rather than in favour of pattern matching.

I think I could believe that ensembles of simply stories could be powerful. After all, ensembles of decision trees are still a thing.

Yes, I think some of these competitive MMA things are illegal here in Sweden.

But when I think sports I think things like soccer, tennis, swimming, wrestling etc.

These things you bring up are very niche, and I don't understand their relevance to whether it's a good idea to use the law to make choking in pornography less visible.