fbpx

Human Progress Logo Human Progress

01 / 05
AI Just Got 90 to 99.99 Percent Cheaper

Blog Post | Computing

AI Just Got 90 to 99.99 Percent Cheaper

Stargate for $500 billion or DeepSeek for $5 million?

Gale L. Pooley —

Summary: Like the telegraph ending the Pony Express through sudden transformative innovation, DeepSeek’s AI breakthroughs are a leap forward that’s challenging industry giants like OpenAI. By optimizing efficiency, cutting hardware requirements, and adopting an expert system approach, DeepSeek is exemplifying how technological progress drives abundance.

Innovation can occur in dramatic bursts, such as when the telegraph replaced the Pony Express. This iconic mail carrier cut previous delivery times in half and reigned for 18 months as the fastest way to deliver information across the United States. The Pony Express was introduced on April 3, 1860, and delivered mail between St. Joseph, Missouri, and Sacramento, California. The 2,000-mile route took approximately 10 days, with riders traveling 75 to 100 miles each and switching horses every 10 to 15 miles.

Western Union erected the first telegraph poles on July 4, 1861. Construction took 112 days to complete the first electronic transcontinental communication system on Oct. 24, 1861. Two days later, the Pony Express was discontinued.

The telegraph reduced the time it took to deliver a message by 99.93 percent, from 10 days to 10 minutes. Sending a message got 1,439,900 percent faster.

Are we witnessing another ponies to electrons innovation between OpenAI and DeepSeek? Maybe.

Peter Diamandis has noted that DeepSeek was founded less than two years ago. With only 200 employees and $5 million, DeepSeek has developed a new artificial intelligence (AI) system. By comparison, OpenAI was founded 10 years ago, has around 4,500 employees, and has raised $6.6 billion in capital. AI tech giants like OpenAI and Anthropic have been spending $100 million or more to train their models. DeepSeek has matched their systems for 95 percent of the cost. A 95 percent drop in cost means you now get 20 for the price of one, indicating a 1,900 percent increase in abundance. DeepSeek has done this with three innovations:

1. Precision reimagined. Instead of using computational overkill (32 decimal places), they proved that 8 decimal places is enough. The result? 75 percent less memory needed. Sometimes the most powerful innovations come from questioning the most basic assumptions.

2. The speed revolution. Traditional AI reads like a first-grader: “The . . . cat . . . sat . . .” But DeepSeek’s multitoken system processes whole phrases simultaneously: 2 times faster with 90 percent accuracy. When you’re processing billions of words, this is transformative.

3. The expert system. Instead of one massive AI trying to know everything (imagine one person being a doctor, a lawyer, and an engineer), they built a system of specialists. Traditional models rely on 1.8 trillion parameters being active all the time. DeepSeek, by contrast, relies on 671 billion in total, but only 37 billion are active at once (97.9 percent fewer).

Diamandis goes on to note more staggering results from DeepSeek’s innovations:

  • training costs slashed from $100 million to $5 million;
  • GPU requirements slashed from 100,000 GPUs to 2,000 GPUs;
  • 95 percent reduction in API costs;
  • runs on gaming GPUs instead of specialized hardware; and
  • done with a team of less than 200 people, not thousands.

The DeepSeek system is open source, which means anyone can verify, build upon, and implement these innovations. You can download the new app on your iPhone.

Bonus: AI now has a counterpoint to the environmentalists who say AI uses so much electricity. DeepSeek just brought down the cost of inference by 97 percent.

Cathie Wood at ARK Investment has observed “over the last few years that AI training and inference costs have been dropping 75% and 85-90%, respectively. DeepSeek may be accelerating the pace of change, but the declines already are dramatic. Faster cost declines will add to demand, more for inference, a more competitive GPU space than training, and one of the nuances.”

Moore’s law suggests that computer transistor abundance doubles every two years. That would indicate a compound rate of around 41.4 percent a year. The cost to train an AI system to recognize images fell 99.59 percent from $1,112.64 in 2017 to $4.59 in 2021. This would indicate a compound rate of 295 percent a year. AI is growing over seven times faster than Moore’s law.

Nvidia is leading the development of these systems and their CEO Jensen Huang has claimed that AI processing performance has increased by “no less than one million in the last 10 years.” This is a compound annual rate of 298 percent. He expects this rate to continue for the next 10 years. That would mean we go from one to one trillion in twenty years. We’ll be 976 million times ahead of Moore’s law. Quite astonishing.

The Pony Express needed lots of fast horses and skinny riders. The telegraph was a whole new platform that used wire and batteries and poles instead. DeepSeek may be the Western Union to the Pony Express.

So what about Stargate—a $500 billion US AI infrastructure initiative led by OpenAI’s Sam Altman, Oracle’s Larry Ellison, and Softbank’s Masayoshi Son? They want to spend 100,000 times more than DeepSeek has spent so far. Will their product be 100,000 times more valuable?

Microsoft’s CEO Satya Nadella brought up Jevon’s paradox in regard to DeepSeek. On January 26, 2025, on X, he said: “Jevons paradox strikes again! As AI gets more efficient and accessible, we will see its use skyrocket, turning it into a commodity we just can’t get enough of.”

On January 6, Nvidia announced its new Nano line of AI development hardware starting at $259 and its new Project DIGITS as the “world’s smallest AI supercomputer capable of running 200B-parameter models” and expected to be priced at around $3,000. Between DeepSeek’s open-source software and Nvidia’s hardware, the world could experience a brilliant efflorescence of superabundance in learning.

The logos of DeepSeek, an open-source AI software, and Nvidia, a leading producer of the chips used to power AI.

We expect to see AI make dramatic advances in our ability to discover valuable new knowledge, but we’ll also come to realize that the only intelligence in artificial intelligence so far has been human intelligence. If human beings have the freedom to innovate, the potential to create resources is infinite.

Find more of Gale’s work at his Substack, Gale Winds.

UCL | Communications

UK Neuralink Patient Uses Thought to Control Computer

“A patient with motor neurone disease was able to control a computer just by using his thoughts following the UK’s first Neuralink implant surgery in a study led by UCL and UCLH clinical researchers.

The surgery is part of the GB-PRIME study evaluating the safety and functionality of Neuralink’s robotically implanted brain-computer interface (BCI), which aims to improve independence for people who are paralysed. 

The surgery, which took place at UCLH’s National Hospital for Neurology and Neurosurgery (NHNN) in October 2025, went as planned, and on the day following the procedure, the patient was able to begin using their BCI implant to move a computer cursor with their thoughts and to return home from the hospital.”

From UCL.

New York Times | Computing

Google’s Quantum Computer Makes a Big Technical Leap

“On Wednesday, Dr. Devoret and his colleagues at a Google lab near Santa Barbara, Calif., said their quantum computer had successfully run a new algorithm capable of accelerating advances in drug discovery, the design of new building materials and other fields.

Leveraging the counterintuitive powers of quantum mechanics, Google’s machine ran this algorithm 13,000 times as fast as a top supercomputer executing similar code in the realm of classical physics, according to a paper written by the Google researchers in the scientific journal Nature…

In another paper published on Wednesday on the research site arXiv, the company showed that its algorithm could help improve what is called nuclear magnetic resonance, or N.M.R., which is a technique used to understand the structure of tiny molecules and how they interact with one another.

N.M.R. is a vital part of effort to develop new medicines for fighting disease and new materials for building everything from cars to buildings. It can help understand Alzheimer’s disease or drive the creation of entirely new metals, said Ashok Ajoy, an assistant professor of chemistry at Berkeley who specializes in N.M.R. and worked with Google’s researchers on the new paper.”

From New York Times.

Nature | Science & Technology

OpenAI’s GPT-5 Hallucinates Less than Previous Models Do

“In one literature-review benchmark known as ScholarQA-CS, GPT-5 ‘performs well’ when it is allowed to access the web, says Akari Asai, an AI researcher at the Allen Institute for Artificial Intelligence, based in Seattle, Washington, who ran the tests for Nature. In producing answers to open-ended computer-science questions, for example, the model performed marginally better than human experts did, with a correctness score of 55% (based on measures such as how well its statements are supported by citations) compared with 54% for scientists, but just behind a version of institute’s own LLM-based system for literature review, OpenScholar, which achieved 57%.

However, GPT-5 suffered when the model was unable to get online, says Asai. The ability to cross-check with academic databases is a key feature of most AI-powered systems designed to help with literature reviews. Without Internet access, GPT-5 fabricated or muddled half the number of citations that one of its predecessors, GPT-4o, did. But it still got them wrong 39% of the time, she says.

On the LongFact benchmark, which tests accuracy in long-form responses to prompts, OpenAI reported that GPT-5 hallucinated 0.8% of claims in responses about people or places when it was allowed to browse the web, compared with 5.1% for OpenAI’s reasoning model o3. Performance dropped when browsing was not permitted, with GPT-5’s error rate climbing to 1.4% compared with 7.9% for o3. Both models showed worse performance than did the non-reasoning model GPT-4o, which had an error rate of 1.1% when offline.”

From Nature.

Wired | Science & Technology

OpenAI Just Released Its First Open-Weight Models Since GPT-2

“OpenAI just dropped its first open-weight models in over five years. The two language models, gpt-oss-120b and gpt-oss-20b, can run locally on consumer devices and be fine-tuned for specific purposes. For OpenAI, they represent a shift away from its recent strategy of focusing on proprietary releases, as the company moves towards a wider, and more open, group of AI models that are available for users…

What sets apart an open-weight model is the fact that its ‘weights’ are publicly available, meaning that anyone can peek at the internal parameters to get an idea of how it processes information. Rather than undercutting OpenAI’s proprietary models with a free option, cofounder Greg Brockman sees this release as ‘complementary’ to the company’s paid services, like the application programming interface currently used by many developers. ‘Open-weight models have a very different set of strengths,’ said Brockman in a briefing with reporters. Unlike ChatGPT, you can run a gpt-oss model without a connection to the internet and behind a firewall.”

From Wired.