Michael Anderson

Im worried about that maybe happening and it being bad, I'm not worried about it "working" to keep them in power. 1/6 wasn't ever going to do that either!

Yes we've been down this road before and don't need to reiterate it

Knowing yourself is good here, but it's not an impossible task and we all should try to do better at it.

I feel like you should hold yourself to a higher standard TBH

She retiring though

Oh no

Also FWIW in the senate Dems have majority of the Senators not up this cycle

Question is whether the thugs code as legitimate or street goons.

All of this

Yes

The principle here is good but the suggestion in the last paragraphs above is still dumb. If you are going to find that the 2nd amendment imposes a restriction on legislation that needs to be grounded in some principles that dictate the bounds of where/when the legislatures can act.

How does that work? So long as we know the vote totals its up to the Senate itself to decide the winner and Dems have a majority of the members not up for reelection.

“orders the military to seize ballot boxes” is even more load-bearing; there are more than 116,000 polling locations in the united states, more ballot drop offs, and two states are mail-first, that’s an enormous logistical effort

which ballot boxes? which races? which troops? deployed from where, under what authority, and by whom?

i think we are likely to see some chance of Stop the Steal version of the Yoon Coup in a country without martial law or a centralized election administration, that is also run by whoever left in the Admin

One answer to this is that winning the elections would be preferable to pressing this button even if it works. BUT they are not smart enough to game it like this.

When I say the "same gains" what I mean is that it will not be as easy to produce the same level or rate of gains as we saw previously but its possible the jury is still out on that.

The fact that we have made gains doesnt tell you how costly it was to do that and whether and to what extent that is scalable or more importantly able to continue things on a similar growth trajectory.

Doesn't the nature of post training tend to make it less susceptible to the kinds of gains we saw from simple data scaling? Eg further advancement will require deploying new paradigms for one or both of post training or model architectures themselves.

We are trying to PREDICT what kind of future gains are reasonable to expect or not. If the amount of effort needed to produce marginal gains becomes too great than either it stops being mechanistically possible to generate improvements or more likely the cost of doing so is prohibitive.

Interesting thanks!

Early on I thought the output return for additional training inputs was an exponential improvement in performance

The point is that output usefulness alone does not tell you the nature of the trend line. Eg it matters how much "effort" is being expended to create that useful product when we are trying to forecast FUTRUE improvements

The thing to take away from this is that dems need to be reaching deep into possible races and not trying to micro target what the inexact data says will be the most "important" seats (or states)

This had been my sense of the general state of things too (especially post GPT 5), but I am not keyed in on the granular minutia

What I would pay more attention to is not hitting a hard wall but diminished returns for input effort, though that's hard to gauge externally given the proprietary nature of the training process.

Extrapolation is ok and justified, but also we should have humility that the pattern will not hold! Especially when we do not fully understand the mechanism by which the pattern has operated thus far.

Many such cases of the 5th circuit doing shit work of late

Depends on how they do the roll up to some extent

I recommend reading the actual opinion because what it makes clear is that the background immigration laws and their court interpretations are utter crap.