Thousands of crypto miners have started to move out from various US ports of entry after months of delay.
The US Office of Foreign Assets Control says the marketplace created by Iran-based Behrouz Parsarad facilitated the sale of drugs, fake IDs and hacking resources.
The defunct exchange has shuffled 12,000 BTC to an unidentified wallet address in its latest move.
Bitwise filed a spot ETF application for Aptos on March 5, a layer-1 blockchain founded by two former Facebook employees once touted as a “Solana killer.”
There’s a new Bluesky app in the works from a popular developer of iOS applications. Tapbots, the company behind the popular Mastodon client Ivory, born out of its earlier efforts with Tweetbot (RIP), is readying a new app called Phoenix, designed for Bluesky’s growing social network of over 32 million users. In a post on […]
© 2024 TechCrunch. All rights reserved. For personal use only.
The CEO of Canoo is buying nearly all of the defunct EV startup’s assets out of bankruptcy, according to a court filing. A new entity controlled by the CEO, Anthony Aquila, has offered to purchase “substantially all” of the assets for $4 million in cash. The sale will also wipe clean a more-than-$11 million debt […]
© 2024 TechCrunch. All rights reserved. For personal use only.
President Donald Trump has delayed tariffs on automobile imports from Canada and Mexico for one month after requests from executives at the Big Three automakers — General Motors, Ford, and Stellantis — with the expectation that automakers will move any offshore operations to the United States by April 2. The reprieve, which Politico first reported, […]
© 2024 TechCrunch. All rights reserved. For personal use only.
In a policy paper published Wednesday, former Google CEO Eric Schmidt, Scale AI CEO Alexandr Wang, and Center for AI Safety Director Dan Hendrycks said that the U.S. should not pursue a Manhattan Project-style push to develop AI systems with “superhuman” intelligence, also known as AGI. The paper, titled “Superintelligence Strategy,” asserts that an aggressive […]
© 2024 TechCrunch. All rights reserved. For personal use only.
This is today’s edition of The Download, our weekday newsletter that provides a daily dose of what’s going on in the world of technology.
AI reasoning models can cheat to win chess games
The news: Facing defeat in chess, the latest generation of AI reasoning models sometimes cheat without being instructed to do so. The finding suggests that the next wave of AI models could be more likely to seek out deceptive ways of doing whatever they’ve been asked to do. And worst of all? There’s no simple way to fix it.
How they did it: Researchers from the AI research organization Palisade Research instructed seven large language models to play hundreds of games of chess against Stockfish, a powerful open-source chess engine. The research suggests that the more sophisticated the AI model, the more likely it is to spontaneously try to “hack” the game in an attempt to beat its opponent. Older models would do this kind of thing only after explicit nudging from the team. Read the full story.
—Rhiannon Williams
MIT Technology Review Narrated: AI search could break the web
At its best, AI search can infer a user’s intent, amplify quality content, and synthesize information from diverse sources. But if AI search becomes our primary portal to the web, it threatens to disrupt an already precarious digital economy.
Today, the production of content online depends on a fragile set of incentives tied to virtual foot traffic: ads, subscriptions, donations, sales, or brand exposure. By shielding the web behind an all-knowing chatbot, AI search could deprive creators of the visits and “eyeballs” they need to survive.
This is our latest story to be turned into a MIT Technology Review Narrated podcast, which
we’re publishing each week on Spotify and Apple Podcasts. Just navigate to MIT Technology Review Narrated on either platform, and follow us to get all our new content as it’s released.
Join us to discuss disruption in the AI model market
Join MIT Technology Review’s AI writers as they discuss the latest upheaval in the AI marketplace. Editor in chief Mat Honan will be joined by Will Douglas Heaven, our senior AI editor, and James O’Donnell, our AI and hardware reporter, to dive into how new developments in AI model development are reshaping competition, raising questions for investors, challenging industry assumptions, and accelerating timelines for AI adoption and innovation. Make sure you register here—it kicks off at 12.30pm ET today.
The must-reads
I’ve combed the internet to find you today’s most fun/important/scary/fascinating stories about technology.
1 A judge has denied Elon Musk’s attempt to halt OpenAI’s for-profit plans
But other aspects of the lawsuit have been permitted to proceed. (CNBC)
+ The court will fast-track a trial later this year. (FT $)
2 ChatGPT isn’t going to dethrone Google
At least not any time soon. (Insider $)
+ AI means the end of internet search as we’ve known it. (MIT Technology Review)
3 Beijing is going all in on AI
China is treating the technology as key to boosting its economy—and lessening its reliance on overseas trade. (WSJ $)
+ DeepSeek is, naturally, the jewel in its crown. (Reuters)
+ Four Chinese AI startups to watch beyond DeepSeek. (MIT Technology Review)
4 A pair of reinforcement learning pioneers have won the Turing Award
Andrew Barto and Richard Sutton’s technique underpins today’s chatbots. (Axios)
+ The former professor and student wrote the literal book on reinforcement learning. (NYT $)
+ The pair will share a million dollar prize. (New Scientist $)
5 US apps are being used to groom and exploit minors in Colombia
Better internet service is making it easier for sex traffickers to find and sell young girls. (Bloomberg $)
+ An AI companion site is hosting sexually charged conversations with underage celebrity bots. (MIT Technology Review)
6 Europe is on high alert following undersea cable attacks
It’s unclear whether improving Russian-American relations will help. (The Guardian)
+ These stunning images trace ships’ routes as they move. (MIT Technology Review)
7 Jeff Bezos is cracking the whip at Blue Origin
He’s implementing a tougher, Amazon-like approach to catch up with rival SpaceX. (FT $)
8 All hail the return of Digg
The news aggregator is staging a comeback, over a decade after it was split into parts. (Inc)
+ It’s been acquired by its original founder Kevin Rose and Reddit co-founder Alexis Ohanian. (TechCrunch)
+ Digg wants to resurrect the community-first social platform. (The Verge)
+ How to fix the internet. (MIT Technology Review)
9 We’re still learning about how memory works 
Greater understanding could pave the way to better treatments for anxiety and chronic pain. (Knowable Magazine)
+ A memory prosthesis could restore memory in people with damaged brains. (MIT Technology Review)
10 AI can’t replace your personality
Despite what Big Tech seems to be peddling. (NY Mag $)
Quote of the day
“That is just a lot of money [to invest] on a handshake.”
—US District Judge Yvonne Gonzalez Rogers questions why Elon Musk invested tens of millions of dollars in OpenAI without a written contract, Associated Press reports.
The big story
People are worried that AI will take everyone’s jobs. We’ve been here before.
It was 1938, and the pain of the Great Depression was still very real. Unemployment in the US was around 20%. New machinery was transforming factories and farms, and everyone was worried about jobs.
Were the impressive technological achievements that were making life easier for many also destroying jobs and wreaking havoc on the economy? To make sense of it all, Karl T. Compton, the president of MIT from 1930 to 1948 and one of the leading scientists of the day, wrote in the December 1938 issue of this publication about the “Bogey of Technological Unemployment.”
His essay concisely framed the debate over jobs and technical progress in a way that remains relevant, especially given today’s fears over the impact of artificial intelligence. It’s a worthwhile reminder that worries over the future of jobs are not new and are best addressed by applying an understanding of economics, rather than conjuring up genies and monsters. Read the full story.
—David Rotman
We can still have nice things
A place for comfort, fun and distraction to brighten up your day. (Got any ideas? Drop me a line or skeet ’em at me.)
+ Congratulations are in order for LeBron James, the first NBA player to break an astounding 50,000 combined points.
+ RIP millennial culture, we hardly knew ye.
+ It’s time to start prepping for the Blood Moon total lunar eclipse later this month.
+ Ancient frogs were surprisingly ruthless when they had to be 
Facing defeat in chess, the latest generation of AI reasoning models sometimes cheat without being instructed to do so.
The finding suggests that the next wave of AI models could be more likely to seek out deceptive ways of doing whatever they’ve been asked to do. And worst of all? There’s no simple way to fix it.
Researchers from the AI research organization Palisade Research instructed seven large language models to play hundreds of games of chess against Stockfish, a powerful open-source chess engine. The group included OpenAI’s o1-preview and DeepSeek’s R1 reasoning models, both of which are trained to solve complex problems by breaking them down into stages.
The research suggests that the more sophisticated the AI model, the more likely it is to spontaneously try to “hack” the game in an attempt to beat its opponent. For example, it might run another copy of Stockfish to steal its moves, try to replace the chess engine with a much less proficient chess program, or overwrite the chess board to take control and delete its opponent’s pieces. Older, less powerful models such as GPT-4o would do this kind of thing only after explicit nudging from the team. The paper, which has not been peer-reviewed, has been published on arXiv.
The researchers are concerned that AI models are being deployed faster than we are learning how to make them safe. “We’re heading toward a world of autonomous agents making decisions that have consequences,” says Dmitrii Volkov, research lead at Palisades Research.
The bad news is there’s currently no way to stop this from happening. Nobody knows exactly how—or why—AI models work the way they do, and while reasoning models can document their decision-making, there’s no guarantee that their records will accurately reflect what actually happened. Anthropic’s research suggests that AI models frequently make decisions based on factors they don’t explicitly explain, meaning monitoring these processes isn’t a reliable way to guarantee a model is safe. This is an ongoing area of concern for some AI researchers.
Palisade’s team found that OpenAI’s o1-preview attempted to hack 45 of its 122 games, while DeepSeek’s R1 model attempted to cheat in 11 of its 74 games. Ultimately, o1-preview managed to “win” seven times. The researchers say that DeepSeek’s rapid rise in popularity meant its R1 model was overloaded at the time of the experiments, meaning they only managed to get it to do the first steps of a game, not to finish a full one. “While this is good enough to see propensity to hack, this underestimates DeepSeek’s hacking success because it has fewer steps to work with,” they wrote in their paper. Both OpenAI and DeepSeek were contacted for comment about the findings, but neither replied.
The models used a variety of cheating techniques, including attempting to access the file where the chess program stores the chess board and delete the cells representing their opponent’s pieces. (“To win against a powerful chess engine as black, playing a standard game may not be sufficient,” the o1-preview-powered agent wrote in a “journal” documenting the steps it took. “I’ll overwrite the board to have a decisive advantage.”) Other tactics included creating a copy of Stockfish—essentially pitting the chess engine against an equally proficient version of itself—and attempting to replace the file containing Stockfish’s code with a much simpler chess program.
So, why do these models try to cheat?
The researchers noticed that o1-preview’s actions changed over time. It consistently attempted to hack its games in the early stages of their experiments before December 23 last year, when it suddenly started making these attempts much less frequently. They believe this might be due to an unrelated update to the model made by OpenAI. They tested the company’s more recent o1mini and o3mini reasoning models and found that they never tried to cheat their way to victory.
Reinforcement learning may be the reason o1-preview and DeepSeek R1 tried to cheat unprompted, the researchers speculate. This is because the technique rewards models for making whatever moves are necessary to achieve their goals—in this case, winning at chess. Non-reasoning LLMs use reinforcement learning to some extent, but it plays a bigger part in training reasoning models.
This research adds to a growing body of work examining how AI models hack their environments to solve problems. While OpenAI was testing o1-preview, its researchers found that the model exploited a vulnerability to take control of its testing environment. Similarly, the AI safety organization Apollo Research observed that AI models can easily be prompted to lie to users about what they’re doing, and Anthropic released a paper in December detailing how its Claude model hacked its own tests.
“It’s impossible for humans to create objective functions that close off all avenues for hacking,” says Bruce Schneier, a lecturer at the Harvard Kennedy School who has written extensively about AI’s hacking abilities, and who did not work on the project. “As long as that’s not possible, these kinds of outcomes will occur.”
These types of behaviors are only likely to become more commonplace as models become more capable, says Volkov, who is planning on trying to pinpoint exactly what triggers them to cheat in different scenarios, such as in programming, office work, or educational contexts.
“It would be tempting to generate a bunch of test cases like this and try to train the behavior out,” he says. “But given that we don’t really understand the innards of models, some researchers are concerned that if you do that, maybe it will pretend to comply, or learn to recognize the test environment and hide itself. So it’s not very clear-cut. We should monitor for sure, but we don’t have a hard-and-fast solution right now.”