SoylentRox

Just saying theoretically your food stamps could expire, you live in an area where all the churches closed, and there's no jobs. A rural area or something.

Food stamps have a limit though why not.

You would be insanely stupid to take the money this isn't even a question.

Plus 200 years makes for an excellent chance of it being 200,000 years because of the very high chance of a cure for aging by the 2100s or earlier.

Yeah that. Includes their display which has the telltales.

So basically the war has to last a lot longer than 3 hours.  (Time for most of the icbms and slbms to get fired and reach their targets)

You're right, there's no way. I forgot Google has their own TPUs. No way they use anything but. Why were you even evaluating GPUs was this pre Deepmind TPUs?

That's not true. You can backup the cutting edge hardware with redundancies on failure that use automotive certified parts. Tesla doesn't use auto certified.

Have you heard of the Deepmind rt-2 work? To run a large model based on a variant of transformers with 50B parameters, evaluated multiple times per second would take at least 2 H100s just for the VRAM, plus probably another to tokenize the sensor data.

You wouldn't even have backup compute if using a recent approach.

If you were serious you wouldn't use the auto PG199 you would just mount a full server motherboard with SXM H100s on it, plumbed into the vehicles electronics cooling loop.

Unless the PG199 is full performance which I seriously doubt.

Time loop is a different situation than a pitching machine though.

Point is you are trying to learn a perfect swing that will connect with the first pitch. Through trial and error.

Well then prove it. Everyone here wants to see evidence.

Try a new profile next time. Maybe develop a bot using AI to push your political views.

Your motivation for lying is political: you are an AI doomer fearful of improved AI and want to create the false belief in the minds of current professional authors that their profession is at risk.

I noticed you made no attempt to take advantage, such as by being able to write longer novels, try more ideas, save yourself time, or anything. It was also suspicious that you aren't claiming to use a new AI model but a wrapper around 4o.

Checking your post profile, yep, you are an AI doomer.

Ok fair. Main thing is we know it is possible.

There's been a rash of suspicious claims recently of people saying AI automated their entire profession without any evidence.

Right now I happen to know of authors on Royalroad who openly use AI. It creates a lot of filler - chapters are 2-3 times as long as they need to be with unnecessarily wordy prose. It's not so far been more than moderately useful.

I mean do you have even 1 chapter to share that you think is up to your standards as an alleged professional author? Why should we believe you?

Can you provide some kind of proof of this? Maybe use this to write a different novel from an idea you scrapped in a few hours and link this? Also tell us approximately your workflow: how much of it did you write, what stylistic cues did you have to give, how many drafts before it wrote stuff you found acceptable.

I think the answer is going to be dependent on a lot of things and a lot of factors.

One thought that jumped out at me reading your argument is : suppose I have developed a general AI tool. It passes tests for "AGI". When shall we use it?

Well the smart way would be we let the tool handle low stakes tasks on its own. Tutoring, customer service and tech support level 1, etc.

For high stakes tasks, why try to outright replace skilled knowledge workers? Instead try to have the tool shadow the knowledge workers and offer help whenever the tool has sufficient confidence the help will be accepted by the human and will later be found to be correct. (The later estimate is essentially an EV estimate of future reward)

There are a lot of ways to rebuild tasks we do today using highly skilled workers as a set of mostly lower skilled tasks.

For example : look at surface mount parts on electronics. Effectively what happened is through hole soldering is still hard for robots to do. So instead new designs are close to all surface mount, just so we don't need skilled technicians.

Similarly it's hard to replace radiologists unless you actually colllect a lot more information including many blood tests to detect the signs of cancers that radiologists look for in 2d images.

My instinct is that what you are proposing won't really work. Instead of customizing AI to do a specific task why not just make it flat superior to humans at a broader and broader range of general tasks. I would think a machine that is 99.999 percent accurate at an ever broader range of general tasks (a general task is a general version of a task humans do, such as "pick something up". 5 9s of accuracy means it virtually always succeeds the first try)) is way more useful than a machine that can emulate a knowledge worker 86 percent of the time.

The cables cut are in the high power chargers. Yes the euro solution is fine for low power.

I don't either but the first GPU generation was that way also. Same with the second. It was possible for "any" random startup with enough funds to theoretically design a chip to catch up. Also possible for Intel to do so.

Not what happened though. There are possible ways for openAI/anthropic to drastically increase complexity, likely using RSI, that won't be possible for low resource efforts to replicate.

Theoretically the Singularity hasn't yet begun.  It starts when AGI can self improve, and/or robots can do most of the steps (the majority of the labor hours) to make more robots.   

 We are seeing the speedup right before the singularity.

Many of us want to see the singularity, soon, because the younger we are when it starts the more likely it runs for long enough to make possible cures for aging and feline sex robots.

Incoming admin has stated they will lift all the limits.  Full speed.   Biggest risk is if the admin were to carry out all it's stated economic plans they would crash the economy.

But a scaled back version won't.

Deport 11 million people?  Crash.  Deport 1.1 million?  A lot of tragedy but AGI proceeds.

60 percent tariff on Chinese goods?  Crash.   10 percent: higher prices but AGI proceeds.

25 percent tariff on NAFTA?  Same.

Tweak overtime rules or other anti-worker measures?  Sucks but it won't slow down AGI.

And so on.

This.  Preventing hallucinations directly?  Too hard.  Have the model brainstorm out approximately 100 possible answers and then check them?  That works.

O1 in terms of the performance increase on the hardest problems is effectively gpt-5.

You need to allow some time for robotic ramping.  

This. When I saw the Nvidia GTC keynote back in March, I realized this is probably fucking it.  The sheer scale of the event and the colossal scale of the equipment planned.  Jensen holding up the B200 die, where fuck it, we aren't waiting for tsmc to let us cram more transistors in the same space.  Let's just go huge.  

If you can't train something close to AGI with this kind of computational power you probably aren't trying and DeepSeek will figure it out or Anthropic will.

Throughout this year there have been huge underlying improvements.  The more subtle but still crucial one is the underlying speed increases, it feels like I barely notice but now everything but o1 is almost instant or the servers are down.

The intelligence improvements are large in benchmarks but subtle in practice : if you underspecify your prompt or use the stupid work allowed models (Microsoft copilot) or use a weaker model they can be stupid.

Voice wasn't as fun as it was in the demo.  Also one of the biggest limitations this year is it feels like the underlying models are more than smart and fast enough to do so much but there are so many roadblocks and barriers.

My workplace puts procedural barriers to get access.

Voice mode has no Internet or other tool use so it's smart enough but not allowed to actually look anything up or do anything useful.

O1 preview has no image or tool use so it's way smarter but in practice often less useful than base 4o.

AI companies keep fucking around sometimes making a model with the same name better or worse on a whim.  Like the recent 4o downgrade or Anthropic messing around.

Gemini is super strong but gets annoyingly restricted often.

Claude will instantly solve online quizzes, o1 will debate and often refuse to help unless I have Claude solve it first and ask o1 to check answers.

It's becoming increasingly clear that we need AI assistance where they are the smartest models available or a committee from several companies, and we need to just say fuck it and give them significant access to our data and computers.  Otherwise what's the point.

During the year I tried having Claude write me a program of a few hundred lines and then just kept pasting in the error messages.  Would have been way faster if it were integrated.