Free Republic
Browse · Search
General/Chat
Topics · Post Article

Skip to comments.

US researchers create $50 AI model to compete with OpenAI’s o1
CNBC TV 18 ^ | 02/10/2025 | Pihu Yadav

Posted on 02/10/2025 9:08:08 AM PST by SeekAndFind

Researchers in artificial intelligence (AI), from Stanford and the University of Washington, have trained a "cutting-edge" reasoning AI model for under $50 in cloud compute credits, according to a research paper published recently.

The model, named s1, purportedly rivals industry-leading models like OpenAI's o1 and DeepSeek's R1 in tests of math and coding skills. The s1 model, along with the data and code used for training, is now available on GitHub.

The team behind s1 started with an off-the-shelf base model and fine-tuned it through distillation, a process that extracts reasoning abilities from another AI model by training on its answers. The s1 model is distilled from Google’s Gemini 2.0 Flash Thinking Experimental, a technique also used by Berkeley researchers to create a similar model for around $450 last month.

This breakthrough raises concerns about the commoditisation of AI models. If small teams can replicate expensive models with minimal investment, it challenges the notion of proprietary advantage in the AI industry. OpenAI, for instance, has accused DeepSeek of improperly harvesting data from its API for distillation purposes. OpenAI is currently fighting copyright cases in India where publishers have accused it of training its models on proprietary data without permission.

The s1 paper suggests that reasoning models can be distilled using a relatively small dataset through supervised fine-tuning (SFT), a more cost-effective method compared to large-scale reinforcement learning, which DeepSeek used to train its own model, R1. SFT allows AI models to mimic specific behaviours in a dataset, achieving high reasoning performance with lower costs.

The researchers behind s1 curated a dataset of just 1,000 questions and answers, paired with reasoning processes from Gemini 2.0 Flash Thinking Experimental.

Training s1 took less than 30 minutes using 16 Nvidia H100 GPUs, and the total cost was less than $50, with Niklas Muennighoff, a Stanford researcher involved in the project, stating that the necessary compute power could be rented for about $20.

In addition, the researchers used a clever technique to improve the model’s accuracy: instructing s1 to "wait" during its reasoning. This way, it was able to extend its thinking time and produce slightly more accurate answers.

While major AI companies like Meta, Google, and Microsoft are set to invest billions in AI infrastructure, the s1 model demonstrates how small-scale innovation is pushing the boundaries of AI capabilities.

However, experts argue that while distillation methods can replicate existing models, they won’t necessarily lead to breakthrough advancements in AI performance.


TOPICS: Computers/Internet; Society
KEYWORDS: ai; openai; sail; stanford; stanfordai
Navigation: use the links below to view more comments.
first 1-2021-23 next last

1 posted on 02/10/2025 9:08:08 AM PST by SeekAndFind
[ Post Reply | Private Reply | View Replies]

To: All
What does she look like?

2 posted on 02/10/2025 9:18:07 AM PST by BipolarBob (My pet termites name is Clint. Clint Eats Wood.)
[ Post Reply | Private Reply | To 1 | View Replies]

To: SeekAndFind

All that concern about cheaper models related to China’s AI 2 weeks ago and this news has had zero impact on the markets today....Nvidia is up 3%


3 posted on 02/10/2025 9:26:03 AM PST by reed13k
[ Post Reply | Private Reply | To 1 | View Replies]

To: reed13k

RE: Nvidia is up 3%

From the article — the model STILL needed dozens of Nvidia H100 chips. Regardless of what you do, the need for fast GPUs still exists.


4 posted on 02/10/2025 9:27:44 AM PST by SeekAndFind
[ Post Reply | Private Reply | To 3 | View Replies]

To: SeekAndFind

Wow, all of that makes my head spin!

The most important question is “Can s1, o1, R1 Gemini 2.0 Flash Thinking Experimental help Musk expose how the Deep State Swamp is stealing us blind?”

The coolest commercial on the Super Bowl yesterday was the ChatGPT one, by a country mile. It was SO innovative it had me reeling that nobody had thought of that clever technique before.


5 posted on 02/10/2025 9:30:11 AM PST by ProtectOurFreedom (They were the FA-est of times, they were the FO-est of times.)
[ Post Reply | Private Reply | To 1 | View Replies]

To: sauropod

Review


6 posted on 02/10/2025 11:01:26 AM PST by sauropod (Make sure Satan has to climb over a lot of Scripture to get to you. John MacArthur Ne supra crepidam)
[ Post Reply | Private Reply | To 1 | View Replies]

To: ProtectOurFreedom

Your sarcasm has been duly noted.

https://www.youtube.com/watch?v=kIhb5pEo_j0

This was sucky.


7 posted on 02/10/2025 11:10:36 AM PST by Responsibility2nd (Nobody elected Elon Musk? Well nobody elected the Deep State either.)
[ Post Reply | Private Reply | To 5 | View Replies]

To: usconservative; Mr. K; FreedomPoster; nathanbedford; DocRock; Nateman; Boardwalk; ...
Holy crap. I could easily run a local (unrestricted) AI in my home office.

NIIIICCCCEEEEE............

Ping me to be added to the ᎪᎡᎢᏆᎱᏆᏟᏆᎪᏞ ᏆᏁᎢᎬᏞᏞᏆᏀᎬᏁᏟᎬ ᏢᏆᏁᏀ ᏞᏆᏚᎢ


8 posted on 02/11/2025 5:10:47 AM PST by Lazamataz (The BEST birthday present I ever got WAS DONALD TRUMP WINNING IN 2024!!!)
[ Post Reply | Private Reply | To 1 | View Replies]

To: Lazamataz

Unrestricted AI?

Sounds like my ex-wife.

A lot more than $50 though...


9 posted on 02/11/2025 5:14:03 AM PST by HombreSecreto (The life of a repo man is always intense)
[ Post Reply | Private Reply | To 8 | View Replies]

To: HombreSecreto; Ciaphas Cain

If I can successfully stand up an AI that I personally own, perhaps I can create images with political figures, which is presently something forbidden in Dall-E and the like.


10 posted on 02/11/2025 5:16:20 AM PST by Lazamataz (The BEST birthday present I ever got WAS DONALD TRUMP WINNING IN 2024!!!)
[ Post Reply | Private Reply | To 9 | View Replies]

To: HombreSecreto
Sounds like my ex-wife.Angry Invective?
11 posted on 02/11/2025 5:20:12 AM PST by Sirius Lee ("Never argue with a fool, onlookers may not be able to tell the difference.")
[ Post Reply | Private Reply | To 9 | View Replies]

To: SeekAndFind

“From the article — the model STILL needed chips. Regardless of what you do, the need for fast GPUs still exists.”

This AI project bought (rented) the “dozens of Nvidia H100 chips” computer time. Cloud computer time. Used this time to train their AI model.
How they did AI on the über-cheap.


12 posted on 02/11/2025 5:34:04 AM PST by dennisw (DËMÔNràts - Truth is hate to people who hate truth.)
[ Post Reply | Private Reply | To 4 | View Replies]

To: Lazamataz

For Me,a ‘Snarky’ AI to
Respond to silly people online.
.


13 posted on 02/11/2025 6:25:28 AM PST by Big Red Badger (ALL Things Will be Revealed !)
[ Post Reply | Private Reply | To 10 | View Replies]

To: Lazamataz

I want one. I also want a 10kw ish nuclear reactor to power my home. Please ping me when those become available. Surplus would be okay too. I am not a snob.


14 posted on 02/11/2025 6:39:37 AM PST by Colorado Doug (Now I know how the Indians felt to be sold out for a few beads and trinkets)
[ Post Reply | Private Reply | To 8 | View Replies]

To: SeekAndFind

“The team behind s1 started with an off-the-shelf base model and fine-tuned it through distillation, a process that extracts reasoning abilities from another AI model by training on its answers.”

I imagine that some time in the near future, they’ll make this practice illegal. Otherwise, big corporations won’t be able to generate big bucks from their AI research.


15 posted on 02/11/2025 7:01:05 AM PST by rightwingcrazy (;-,)
[ Post Reply | Private Reply | To 1 | View Replies]

To: Lazamataz
The s1 paper suggests that reasoning models can be distilled using a relatively small dataset through supervised fine-tuning (SFT), a more cost-effective method compared to large-scale reinforcement learning, which DeepSeek used to train its own model, R1. SFT allows AI models to mimic specific behaviours in a dataset, achieving high reasoning performance with lower costs.

Dude, it's a very small dataset focused on ONE set of facts, designed for specific queries.

If the commie chinese taught us anything with DeepSeek its this:

LLM's aren't the way. Focused, well trained, smaller, distributed, FOCUSED AI engines are far faster, will scale out better, be more efficient, and deliver better, more consistent results.

The above statement comes with a large number of assumptions, the biggest being this: The smaller, more focused and trained AI becomes, the more dependent on UNBIASED information they are to be properly trained and provide reliable results.

This is where DeepSeek FAILED: heavily biased algorithms with government approved data to train them. (Knew THIS point specifically as soon as I started reading up on DeepSeek and how the Chinese did it.)

BTW, I've told you this before: It's relatively easy to run a small AI engine at home with a proper graphics card (I believe I recommended one to you when you were building your PC). The instructions on running a small Docker packaged AI engine are out there. I do it on my Ubuntu Linux server @ home.

16 posted on 02/11/2025 7:03:01 AM PST by usconservative (When The Ballot Box No Longer Counts, The Ammunition Box Does. (What's In Your Ammo Box?))
[ Post Reply | Private Reply | To 8 | View Replies]

To: Big Red Badger

Noland Baugh
.
Elon Musk .
.
NUERALINK
.
Capt. Pike-— Trekkie stuff


17 posted on 02/11/2025 7:19:25 AM PST by Big Red Badger (ALL Things Will be Revealed !)
[ Post Reply | Private Reply | To 13 | View Replies]

To: Big Red Badger

Noland ARBAUGH.
Quadriplegic in Yuma AZ
Gets implant
Fascinating story.


18 posted on 02/11/2025 8:15:24 AM PST by Big Red Badger (ALL Things Will be Revealed !)
[ Post Reply | Private Reply | To 17 | View Replies]

To: HombreSecreto

Was her name Stella.😏


19 posted on 02/11/2025 1:52:11 PM PST by BiteYourSelf ( Earth first, we'll strip mine the other planets later.)
[ Post Reply | Private Reply | To 9 | View Replies]

To: BiteYourSelf

Lol, no. But maybe?


20 posted on 02/11/2025 2:01:23 PM PST by HombreSecreto (The life of a repo man is always intense)
[ Post Reply | Private Reply | To 19 | View Replies]


Navigation: use the links below to view more comments.
first 1-2021-23 next last

Disclaimer: Opinions posted on Free Republic are those of the individual posters and do not necessarily represent the opinion of Free Republic or its management. All materials posted herein are protected by copyright law and the exemption for fair use of copyrighted works.

Free Republic
Browse · Search
General/Chat
Topics · Post Article

FreeRepublic, LLC, PO BOX 9771, FRESNO, CA 93794
FreeRepublic.com is powered by software copyright 2000-2008 John Robinson