32 Comments
User's avatar
Westley Deitchler's avatar

When we apply this technology to crypto and abolish all governments and their preying on us, the sky is the limit.

Expand full comment
Gale Pooley's avatar

We need government to protect our rights. A reminder from Mr. Jefferson;

We hold these truths to be self-evident, that all men are created equal, that they are endowed by their Creator with certain unalienable Rights, that among these are Life, Liberty and the pursuit of Happiness.—That to secure these rights, Governments are instituted among Men, deriving their just powers from the consent of the governed,—That whenever any Form of Government becomes destructive of these ends, it is the Right of the People to alter or to abolish it, and to institute new Government, laying its foundation on such principles and organizing its powers in such form, as to them shall seem most likely to effect their Safety and Happiness.

Expand full comment
Mullet Snyder, the Lying Poet's avatar

wherever you read 'Liberty', insert 'Property Rights'.

I'm a big fan, don't get me wrong; I hold patents, I've claimed copyright and I even have a house in Florida.

Just want to make sure we're defining our terms.

Expand full comment
Halftrolling's avatar

You need guys (or robots) with guns to enforce rights. We just traditionally give the government that job.

Expand full comment
Westley Deitchler's avatar

I followed Jefferson's advice and abolished 'our' government over me when it became destructive of my rights and replaced it with a 'government' more ab;e to effect my safety and happiness - my own mind. I live by the universal natural law that has governed the innocent for 11 million years but which no criminal has ever lived by. And Washington, Adams, Jefferson, Madison, et al have all been criminals preying on the innocent. I accepted the responsibility of securing my own rights. Not one person in any government on earth has ever known what rights are and was willing to protect them, even their own. They all want power which is the ability to harm the innocent on a massive scale and GET AWAY WITH IT.

Expand full comment
Same Same's avatar

You know it isn't real money right?

Expand full comment
Andy's avatar

What’s “real money”?

Expand full comment
Gale Pooley's avatar

Time

Expand full comment
Andy's avatar

Your answer is money!

Expand full comment
Same Same's avatar

For starters not invented to make drug dealing easier.

Expand full comment
Westley Deitchler's avatar

Or, rather, there are no limits.

Expand full comment
Max More's avatar

Now, if we can just do the same for space flight! Even two orders of magnitude would make a massive difference in launch costs.

Expand full comment
Gale Pooley's avatar

Elon has done a 10x in the last ten years. Claims to do 10x in the next ten. Vision and action is energizing. Elon is an energizing engineering entrepreneur. Got to be more Elons on our planet. How do we find them?

Expand full comment
Oregonian's avatar

The pathways for ideas and innovation are clogged. If you didn’t go through the right gates from age 16-22, there are almost zero additional on-ramps. The additional public channels are clogged and it is very difficult to get through. Imagine you, as a lay person or one without the education to be in the right environment , have a brilliant idea. It is very difficult to get in front of the right people. The intake and sorting methods are not calibrated to incipient creativity.

Expand full comment
Sufeitzy's avatar

These statements seem to be made purely for attention, “AI can be built for $1 wow” - being made here are in a vacuum. No proof is given for comparable results among models, it’s useless to seek benchmarks.

I built Markov-chain LLM models operating on a laptop precisely like modern GPT’s in 1993 for $0.02, trained in Project Gutenberg texts. However they’re not comparable to gpt-4o by any stretch of the imagination, and neither is the Chinese model, it’s atrocious.

Why not just say it was cost-free? Borrowed time on Google Tensor chips at Google labs to build gpt-variants? I did. The quality was very limited naturally.

There are numerous postings on Medium on how to build a gpt model. The math, the data training sources, they’re all around. You can imitate a GPT model with pencil and paper.

Only OpenAI has achieved the stunning quality they have in all the models I test:

Expand full comment
Bob's avatar

The cheaper models are more limited and more specialized. I would argue that they are likely to be more useful.

Expand full comment
Gale Pooley's avatar

Thanks for your informed thoughts. Could you comment on Grok 3? Thank you.

Expand full comment
Codebra's avatar

TinyZero is not remotely comparable to DeepSeek R1. It uses the DeepSeek training technique to train an extremely small model in order to demonstrate the effectiveness of the technique. This article is pure BS and the author should retract or clarify. This is misinformation. There is no $30 AI. It’s a small proof of concept to demonstrate how R1 training worked. A desktop model of a Boeing 777 cannot fly you to Paris.

Expand full comment
Gale Pooley's avatar

Point acknowledged. Proof of concept was the purpose. The obsession with AGI may be the wrong direction. Billions of specialized bots makes more sense. Only activate when needed. Like a market of providers instead of one central planner. Why use a Boeing 777 when most of the time you just need to drive to Walmart?

Expand full comment
Anthony Hicks's avatar

AI is the next OPEC and Military Industrial Complex’s way to funnel government money to elites. It’s all hype, search engines on steroids. It’s Artificial Programming nothing can think or has a conscience so there’s no intelligence.

Expand full comment
Dirichlet-to-Neumann's avatar

The right lesson of Deepseek is not "all those American investments in computing power are wasted." The right lesson is "use Deepseek technique but with a 100millions training run."

Expand full comment
Michael M's avatar

This is a replication of R1-zero, an experimental step in the development of R1, not R1. And it’s a replication that *builds on the weights of a pretraining run that deepseek released*. I don’t know how you ended up with the interpretation that a single RL experiment in the many-step training workflow of R1 occupied the entirety of the multi-million dollar cost of training it, but if you assign even $30 of value to your reputation as a writer then you really need to correct this or take it down.

Expand full comment
Codebra's avatar

This article is almost certainly highly misleading. We all need to do a bit of research, because the claims here are extremely suspect.

Expand full comment
Bryce E. 'Esquire' Rasmussen's avatar

Cheap Ai means it will soon be everywhere and the main working model for everything.

Yuck.

Expand full comment
Edmund Bannockburn's avatar

Could I, a normal person who knows some vague things about tech (I read Scott Alexander and built my own computer, but don't work in tech professionally), build an AI model in my basement? Anyone have a how-to guide on accomplishing this?

Expand full comment
Live Life Not Behind Glass's avatar

Now you can, if you have enough money for the hardware, and the dedication. Pretty unsettling though to watch the training iteration of how it thinks go from english, to a mishmash of languages, to mishmash with symbols, to “wtf is that demon language?!!?!?!” though. You dont see that as an end user.

Expand full comment
Halftrolling's avatar

Well, shit. Guess I’m going to need to learn to code.

Expand full comment
Overhead At Docksat's avatar

Didn’t DeepSeek train itself on already furrowed ground essentially creating an emulation? And this model is doing the same thing?

It seems the real issue is that all the upfront money and cost over the last few years produces enough data to produce a more compact model. After that you don’t need to spend the money again.

It’s a P and not-P experiment. We’re into the levels of not-P for the time being.

Expand full comment
Mullet Snyder, the Lying Poet's avatar

Sam Altman said it best, “It’s a lot easier to build the second model when the problem has already been solved.”

Expand full comment
Jon Miltimore's avatar

You should write something on this for me.

Expand full comment
Gale Pooley's avatar

Ok what would you like?

Expand full comment
Jon Miltimore's avatar

I think there are several angles you could take. I’d be happy to pay your for your time and you could republish your piece on your Substack afterwards

Can you shoot me an email at Jon.Miltimore@AIER.org and we can discuss?

Expand full comment