Many things that look exponential originally turn out to actually be sigmoidal.

I consider the start of this wave of AI to be approximately the 2017 Google transformer paper and yet transformers didn't really have enough datapoints to look exponential until GPT 3 in 2022.

The following is purely speculation for fun and sparking light-hearted conversation:

My gut feeling is that this generation of models transitioned out of the part of the sigmoid that looks roughly exponential after the introduction of reasoning models.

My prediction is that tranformer-based models will start to enter the phase that asymptotes to flatline in 1-2 years.

I leave open the possibility for a different form of model to emerge that is exponential but I don't believe transformers to be right now.

From the comments on the link:

> I think in general you give smart people powerful tools and good things always happen. That’s why I’m so hesitant to ban a TOOL when the problem to me is inadequate people.

This resonates strongly with me. I've seen AI weilded mindlessly, effectively offloading the burden of critical thought from the submitter to the reviewers. This inevitably annoys the reviewers. I've also seen mindless changes where AI wasn't used at all that are equally annoying.

In contrast, I've seen AI weilded by contributors who still put in the effort of thinking critically about the change, and these changes are usually no more onerous to review than a fully-human generated change.

The differtiator between a good submission and an annoying one is typically the behaviors of the human, not whether or not AI was used. The way forward is to define the behaviors we want to encourage and avoid, and figure out the incentive structures to push contributors to using the proper behaviors.

I didn't know this existed. Thanks for sharing! Very useful api

> Thorium MSRs don't make sense for the Americas, Europe or Australia. We have plenty of uranium.

That covers the input side of th equation. Thorium can help transform the outputs of our existing reactors into waste with orders of magnitude better in terms of dangerous lifespan

Thorium is and will always be a less desirable fuel source - except if you don't have access to uranium or are trying to make your MSRs work (which to date have signs of progress but no proof of commercial viability). MSR also inherently unstable due to salt.

I'm glad people are finding more research and hopefully this will unlock other tech but this has limited impact on the current trajectory of commercial nuclear and the designs currently in the labs.

Though the commentary in here does remind me how much hype has infused the nuclear space - good thing on the whole as long as an eventual AI shakeout doesn't knee cap all the good work being done.

The real challenge is that implementations aren't static.

Just because today's implementation has 4 9s that doesn't mean tomorrow's will...

Each AWS service may choose different pipeline ordering based on the risks specific to their architecture.

In general:

You don't deploy to the largest region first because of the large blast radius.

You may not want to deploy to the largest region last because then if there's an issue that only shows up at that scale you may need to roll every single region back (divergent code across regions is generally avoided as much as possible).

A middle ground is to deploy to the largest region second or third.

Former AWS employee here. There's a number of reasons but it mostly boils down to:

It's both the oldest and largest (most ec2 hosts, most objects in s3, etc) AWS region, and due to those things it's the region most likely to encounter an edge case in prod.

I think Wiim can do this

The real issue is the administration of the hospital sees every minute the doctors spend taking better records instead of seeing another patient as a loss of ability to bill someone's insurance for that time.

I'm sure there's many doctors who would like to take better notes if they were allowed the time to do so.

Maybe the case for better records reducing costs to insurance by assisting in prevention / early intervention is a path forward?

Naming may provide useful hints about some utility of a tool but naming does not bound the utility of a tool.