AI Developers: AI “Not Comparable to Humans or even Animals”

Written by Brian Shilhavy on October 22, 2025. Posted in Current News

AI code cannot learn or reason like humans, or even animals. So says Andrej Karpathy, one of the founding members of OpenAI along with Sam Altman.

Both made a lot of waves in the AI Tech world this past weekend, when he stated that the idea of AI replacing humans as being “right around the corner” was not true, and that industry leaders were over-hyping it for financial gain.

Is Andrej Karpathy Right About Overhyped AI?

by The Information

Excerpts:

Andrej Karpathy, one of the founding members of OpenAI, on Friday threw cold water on the idea that artificial general intelligence is around the corner. He also cast doubt on various assumptions about AI made by the industry’s biggest boosters, such as Anthropic’s Dario Amodei and OpenAI’s Sam Altman.

The highly regarded Karpathy called reinforcement learning—arguably the most important area of research right now—“terrible,” said AI-powered coding agents aren’t as exciting as many people think, and said AI cannot reason about anything it hasn’t already been trained on.

His comments, from a podcast interview with Dwarkesh Patel, struck a chord with some of the AI researchers we talk to, including those who have also worked at OpenAI and Anthropic. They also echoed comments we heard from researchers at the International Conference on Machine Learning earlier this year.

A lot of Karpathy’s criticisms of his own field seem to boil down to a single point: As much as we like to anthropomorphize large language models, they’re not comparable to humans or even animals in the way they learn.

For instance, zebras are up and walking around just a few minutes after they’re born, suggesting that they’re born with some level of innate intelligence, while LLMs have to go through immense trial and error to learn any new skill, Karpathy points out.

This leads to a lot of issues, many of which center around the idea that LLMs struggle to do things that involve subjects or information outside of the data they were trained on. In AI lingo, it means models don’t “generalize.”

An example of how LLMs learn differently from humans appears in reinforcement learning, the model training technique du jour which rewards a model for accomplishing certain goals and penalizes it for other behaviors.

Karpathy points out that RL oversimplifies the learning process: just because a model ends up at the right final answer doesn’t mean that every step it did up to that point was correct, and vice versa. (Karpathy likened RL to “sucking supervision through a straw.”)

In contrast, humans are able to reflect on their thought processes more deeply, not just the final outcome. For instance, a founder who wants to build a successful company can reflect on the company-building process and think about what went right or wrong to bring those insights to their next company, regardless of whether the first company was ultimately successful.

The AI version of this would be to have a model get feedback on every step of its thought process, an approach that some have tried but can get very expensive and tedious.

It’s fitting that this weekend, shortly after the Karpathy interview was released, OpenAI landed in hot water for making claims that GPT-5 had solved new math problems that it had actually just looked up the answers to.

Maybe we’re further off from automating AI research than we thought!

All of this goes to show that it’s good to be skeptical once in a while, and to listen to experts who don’t have an incentive to hype up AI progress to land billions of dollars from VCs and other investors. (Full article – subscription needed.)

AI “Agents” are NOT Replacing Humans

In another article published today by Aaron Holmes of The Information, they looked at the hype about “AI Agents” that existed at the beginning of this year (2025), and where we actually are today in terms of the promises made. He referred also to comments made by the co-founder of OpenAI, Andrej Karpathy.

A Reality Check on Agents

Around the start of this year, executives at artificial intelligence providers like OpenAI and Salesforce said 2025 would be “the year of agents,” making it seem like fully autonomous AI would soon take over human jobs.

Since then, Salesforce, Microsoft, SAP and many others pitched roughly seven types of agent to automate white collar tasks in realms like coding, HR, finance, and sales.

But the promise of truly autonomous agents still hasn’t been fulfilled, and tech leaders are warning companies to lower their expectations.

Andrej Karpathy, one of the founding members of OpenAI and a leading voice in AI research, believes it’ll take at least a decade until AI can meaningfully automate entire jobs,

“like an employee or intern that you would hire to work with you.”

In a podcast interview, he gave a brutal assessment of AI’s current capabilities, in part because large language models lack cognition, and argued that most jobs still can’t be automated.

“Overall, the models are not there,” Karpathy told podcaster Dwarkesh Patel.

“I feel like the industry is making too big of a jump and is trying to pretend like this is amazing, and it’s not. It’s slop. They’re not coming to terms with it, and maybe they’re trying to fundraise or something like that.”

Karpathy’s remarks resonated with business leaders on the front lines of using AI agents, which have repeatedly run into speed bumps over the past year when trying to get AI to handle complex workplace tasks without errors.

“It‘s the fashion of the day to call these tools ‘agents,’ but agent doesn’t really mean autonomous,” Donzé said.

“There‘s the marketing BS that says everything will be autonomous, and then there’s some level of truth behind it, but they can only be autonomous in limited use cases.”

Others say they’re using AI agents to replace some employees, but note that the agents need to be supervised more than those employees typically would.

Sahil Lavingia, CEO of the e-commerce marketplace app Gumroad, said he’s shrunk the size of his engineering team to roughly a dozen people, down from over 40 two years ago, thanks to AI coding agents, but said remaining staff spend a lot of their time double checking the agents’ work.

“I think it’s interesting that I can replace some $400,000-a-year jobs [with agents], but I can’t completely replace them. I just have someone else I pay $400,000 to manage agents to become four times as productive and then fire three people,” Lavingia said.

“AI-generated code needs to be [checked] as the product experience can’t be validated by the AI’s yet, and that’s a different task from coding.”

Last week, OpenAI researcher Sebastien Bubeck took to X to share an impressive accomplishment by OpenAI’s GPT-5 model: the AI helped researchers find the solution to several difficult conjectures in mathematics, known as Erdos problems, which were first made decades ago by the 20th century mathematician Paul Erdos.

Kevin Weil, vice president of OpenAI for Science, touted the same accomplishment, saying GPT-5 “found solutions” to ten Erdos problems that had “been open for decades.”

Some took the announcements as proof that OpenAI’s models were getting better at complex math. But Oxford researcher Thomas Bloom, who maintains the online database of Erdos problems that Bubeck was referring to, deflated some of that enthusiasm on Friday.

That’s because, as Bloom pointed out on X, GPT-5 didn’t actually solve the math problems in question; it merely surfaced online publications by other mathematicians who had already solved those problems.

The problems were listed as “open” on Bloom’s site because Bloom himself wasn’t aware that others had solved them, he said in a response to Weil.

Weil and Bubeck deleted their posts and said in subsequent posts that they meant to clarify that GPT-5 was proving useful at turning up relevant research, not solving math problems.

Some of OpenAI’s rivals were quick to pounce, such as Google DeepMind CEO Demis Hassabis, who called Weil’s original post “embarrassing.” (Full article. Subscription needed.)

While this news shocked the AI world this week, it is all pretty basic stuff I have been writing about and saying for about 2.5 years now.

AI can NOT replace humans, but the new Large Language Models (LLM) AI that started with OpenAI and Microsoft at the end of 2022, do have some value as a tool to assist humans and make their work easier and quicker.

Once the AI bubble bursts, the real-world applications and usefulness of this AI will remain, but will our economy survive the correction in the market that is coming?

source  healthimpactnews.com

Please Donate Below To Support Our Ongoing Work To Defend the Scientific Method

PRINCIPIA SCIENTIFIC INTERNATIONAL, legally registered in the UK as a company incorporated for charitable purposes. Head Office: 27 Old Gloucester Street, London WC1N 3AX.